Skip to main content

Transforming the understanding
and treatment of mental illnesses.

Workshop: Genes to Biology: Integrative Systematic Approaches for Revealing Biological Functions of Psychiatric Risk Genes and Alleles

Watch on YouTube.

Transcript

Transcript

Geetha Senthil: Good morning everyone, I welcome you all. My name is Geeta Senthil, I'm the program officer at NIMH. And this workshop is going to be a four- and half-hour workshop. For any technical problems, please email Lora Bingaman and Nicole North. We provided emails on the first slide, but we will project that again. This is an overview of the program. We will be four panels. And combined with each panel has a set of talks and followed by discussions except the Future Directions panel where it will be entirely discussions. So please, this is a reminder. At 3:00 pm, we will be switching from Zoom Webinar to Zoom Meeting. There's going to be a large audience so please be mindful of time. At 2:50 pm, we will break. At that time, you will have to switch over to Zoom Meeting link. So the Zoom Meeting link was sent to all the attendees. People who are who are attendee role will be joining in the listen and view only mode, so you won't be-- but however, you can ask questions during discussions, general discussions, so Q&A talks. And as audience, you have an option to vote-up some questions for moderators. If a large number of the audience has the same set of questions, you can just vote on the prior question, and then that will be prioritized for the moderators to address those set of questions. And so, I will now switch-- turn it over to Dr. Josh Gordon to provide opening remarks. Thank you.

Joshua Gordon: Welcome everyone, thank you for joining us today. I want to thank Geetha and Lora and other members of the Genomics Team at an NIMH for organizing this workshop. And, I want to thank Steven McCarroll and Anne Bang for chairing it and for bringing everyone on board and setting the agenda. So why are we here today? In case it's not obvious, the goals are to discuss potential systematic approaches that might be needed to gain insight into how psychiatric risk genes lead to the disorders. We're in a space, now as you all know and as we'll review throughout the day, where we have increasing numbers of both rare and common human genetic variants that raise risk for psychiatric disorders. Some with larger effect sizes, some with smaller effect sizes. And we recognize that we need a path, or I should probably say multiple paths to interpret these genetic findings and understand the biological pathways that result from them that raise risk. But it's not clear to me at least what those pathways might be. It's also not clear to me whether we have the right scalable technologies to engage in systematic interrogation of these gene variants, and/or the pathway to developing those scalable technologies. And why now? Two reasons one, of course is, we have lots of variants now. So, if we had a systematic way to interrogate them we should probably start doing it. The second reason why now is that we are really trying to think hard about how we at NIMH can devise encourage support such systematic efforts.

Joshua Gordon: As many of you know, the approach to understanding risk variants predominant in the field right now are one offs; let's make a knockout mouse, or let's make a gene variant mouse, let's study 30 cell lines with this mutation, or let's understand the impact of this mutation in an “X” system. The rest of neuroscience is moving towards more systematic technologies, large scale recording, thorough descriptions of cell types throughout the brain. And borrowing from genomics frankly, using high throughput approaches to approach questions in parallel is something that we recognize is going to be really important. So, the question is what approaches can we identify now that are promising, that need further development, or that might be ready to go right now that we could use to systematically rather than one off approach the biological consequences of these genes? We don't know from NIMH’s perspective, we're not convinced that we're ready to scale up, but we are convinced that we need to think about what might be ready to scale up or what we need to know about these various approaches in order to identify those that are ready to scale up. But once we are ready to scale up, we're committed to make those investments and to encourage the kind of-- the kinds of research that will examine this landscape in a systematic coordinated fashion. So that's the motivation for the workshop today. It won't be the last workshop on this issue. It's probably not the first. But I think it's important that we assess the state of the field right now so that we can try to figure out ways forward. And with that, I'll turn it I guess over to the chairs, to Steve and Anne, and thank you again.

Steven McCarroll: Thank you, Josh and Geetha. So human genetics struggled for many years to find genes in alleles that shaped the biology underlying major psychiatric illnesses. As you'll hear from Mark Daly in a few minutes, that hard work is finally bearing a real harvest in terms of quickly expanding sets of genes and alleles that human genetics tells us contribute to the biology. And the question that presents itself with the fiercest urgency is evolving from how do we find genes in alleles that matter, to the question of how do we benefit, and in particular, how does our biological understanding benefit from having found these genes in alleles. It's a deeply challenging question because if human genetics is clear on one thing, it's that all of these are highly polygenic illnesses in which phenotypes are shaped by variation in many different genes. None resembles say Huntington's Disease with a single locus that can be the sort of focus of a field or Alzheimer's Disease with a common allele with very large effect. So, I'd point out that even for those illnesses, the actual pathophysiological mechanisms are still unclear after decades of work. But the problem that we increasingly need to understand is how did genetic influences percolate through complex biological systems. What are the key biological processes upon which they converge? And the exciting thing is that many kinds of innovations in molecular biology and neuroscience and physics and computation are making it newly possible to measure a far wider range of molecular and cellular phenotypes and data rich ways and to relate them to genetic perturbations. So today we're going to get to hear about that and discuss it all together. The speakers today are really ambassadors for their fields and areas rather than just for their own research. So, they they've been charged with delineating how capability is evolving in their fields and what questions its newly possible or imminently possible to ask and answer, so we're expecting an exciting day. Anne, do you want to say anything also by way of welcoming or introduction?

Anne Bang: No, no. That was that was great, Steve. Thanks.

Steven McCarroll: Okay, great. Well, I'm excited to introduce someone who's made an enormous number of good things happen in psychiatric genetics and in human genetics more generally. Mark Daly.

[silence]

Mark Daly: It is a real pleasure to be here. I have been charged with kicking things off in a sense reviewing the current progress and state of the art [inaudible] and gene discovery and mental illness. I'm going to focus today on using schizophrenia as the example because there's been great risk profile less common and rare variation in schizophrenia. But it's really simply a model, and nearly every other mental illness is on the same trajectory though in various places not quite as far down the road necessarily as schizophrenia. So, the motivation, of course, of what brings us all around the table is unquestionable, psychiatric illness cause, the among the greatest causes of disability lost years of function in life of any disease area that we focus on. And yet, unlike cardiovascular disease, unlike cancer have really been next to no novel therapeutics introduced in decades for psychiatric illness as there has in those other examples. And that fundamentally, we conclude arises from the fact that we really don't understand the root causes of disease. Yet, we have understood, for quite some time, that, in fact, the root causes of disease are residing, to some extent, in the genome because we have known back to work of guardsmen for decades and decades that despite our challenges in articulating the causal biology of psychiatric diseases, they are essentially among the most heritable of common human phenotypes that we study.

Mark Daly: And so this at simultaneously confounds us because we all - if you're of my age or older - lived through decades in which the search for the genes underlying mental illness of our kind was largely a fruitless and frustrating activity compared to many other such activities. And among those activities were sort of the decades of the 1990s and in the early part of the last decade in which enormous progress was made in the identification of genes and mutations from Mendelian disease. But next to no progress in psychiatry or really any complex human trait [inaudible]. So when we really started to take on this question fundamentally was after the sequencing of the human genome, after we began to develop comprehensive tools for studying the genome in larger and larger sets of families and individuals, we were faced with an interesting debate that it was clear that our failure to identify genes via traditional approaches such as linkage analysis in families, didn't tell us much. But what it did tell us was that psychiatric illness was not caused by one or a small number of genes in a Mendelian fashion. And so, we were left with multiple hypotheses that really, for all the effort we put in, that wasn't really limiting the search space very, very aggressively. But we could see two possibilities going forward. One was that we have common variation in trying to understand the role of common variation with respect to a risk for psychiatric illness. And the second was, of course, the study of rare high impact variation with the possibility being that rather than clustering in a small number of genes, it might be found in many genes scattered across the genome.

Mark Daly: The first of these was much more technically approachable in the era of 2005 to 2010, with the advent of sequencing of the human genome followed by variation pattern compendium such as Hap Map and the 1000 Genomes Project and the advent of genome-wide genotyping activities that came along with technological advances that enabled two things. And it really first allowed the identification of a widespread role for submicroscopic copy number variation. And I'll circle back to this a little bit later. I'm not going to spend a lot of time on this. But the first definitive genetic lesions that were associated to psychiatric were found by this--

Hae Kyung Im: Sorry Mark for interrupting you.

Mark Daly: Yes.

Hae Kyung Im: People are complaining about the audio, I think people are saying it's too low, the volume, sorry.

Mark Daly: Too low the volume, okay. That's interesting. I don't-- let me pause and see if I can improve that. I don't quite understand why it would be too, but.

[inaudible].

Mark Daly: Okay, that's a bit too echoey. How does that sound? Is that the reasonable volume?

Geetha Senthil: You're muted, Mark. We can't hear you.

Mark Daly: I'm muted, okay, that's very interesting.

Stephan Sanders: I don't think you are muted.

Mark Daly: I'm not muted.

Joshua Gordon: No, I think maybe you're using a different microphone and is ideal, but you sound very distant.

Mark Daly: Maybe I'll just increase the volume. How's that?

Steven McCarroll: That's much better.

Daniel Geschwind: That works.

Mark Daly: Okay, great. Let's continue. Okay. So with respect to the new technologies that came online in the midpoint of the last decade prior to the one we just exited, I guess, not only was the recognition of copy number variation for the first time possible at scale, but also the detail the value of the role of common genetic variation as had been long postulated could finally be explored for real. And these technologies ushered in the possibility of two quite distinct mechanisms by which we might perform unbiased scans of the genome for discovery. And really, the first new approaches to gene discovery since family-based linkage analysis had been introduced more than half a century earlier. So with that in mind, the question arose-- well, there was some debate as to whether there would be any role for common genetic variation with respect to psychiatric illness to common disease and just such a-- such things would play a role. Of course, dates back considerably to population genetics which has identified for quite a long time that the vast majority of genetic differences between any two members of the human population reside in common variants that arose a long time ago when the human population was, for a very long time, small and resided entirely in Africa. As a result, an obvious hypothesis had existed for a long time that most genetic variation in common traits to human populations would also be explained by common variation for the simple reason of that's where most of the differences are.

Mark Daly: And this hypothesis actually more than half a century ago was clearly laid out by Irv Gottesman and James Shields who described that the most parsimonious way of explaining a common trait with a high heritability across all cultures backgrounds, nationalities, types of relative pairs you might expose - explore would be through a constitutional predisposition or liability, essentially, background genetic variation as you would refer to within a Mendelian setting that would ultimately explain most of quantitative traits in general, and even psychiatric disease specifically. And the era of the GWAS over the last 15 years has borne out this prediction which turns out to be remarkably prescient even in the case of schizophrenia. However, it wasn't an easy road to those discoveries because, of course, the first generation of these genome wide studies came online in psychiatry were not at all successful and this caused some hand-wringing, and also made us recognize and do some hard digging into the fact that genome wide association studies seemingly worked out of the box a little bit more easily in autoimmune diseases, in cardiometabolic diseases, and in late onset diseases. And that brought us back to sort of an obvious feature, which was in fact included in Gottesman's model of the fact that unlike those other disease areas, there is a dramatically reduced fitness, a very strong evolutionary pressure against schizophrenia and autism specifically, but really all psychiatric illness, if you look at large Danish, Swedish, Finnish registry studies that have been published over the past few decades. What the ramifications of this are is that no genetic variant that exerts even a relatively modest effect on risk is allowed to persist in the population at any frequency.

Mark Daly: So a variant arises if you imagine it has to have a very, very mild effect on any trait like this in order for it to have any chance of rising in frequency to become a common polymorphism or even a low frequency variant. As it turns out, the math around this suggests that the only possible things that could work are actually common variant studies where variants become so frequent that you can detect very weak effects or de novo mutations studies and those predictions have largely been borne out with what's been discovered in recent years. So conceptually, the polygenic model would be fine even with natural selection, but required a much larger scale to gain access to the individual elements, the genetic associations themselves. This is where the work of the PGC, led by Pat Sullivan and so many others; my late friend Pam Sklar, Nick O'Donovan were instrumental in driving forward a worldwide collaboration to tackle the problem head on. Rather than giving up and saying, "Well, it's a hard problem. Who's ever going to have 10,000 samples of any mental illness?" They pressed on and pushed the community over the past 15 years to really embrace a collaboration in a real sense in which data is shared, everyone shares in the ultimate discoveries that come from that. Or as I like to say, each of us had our own study. We were discovering 0 genes. So, anything we discover together, even if we partition the credit out, is greater than zero. So with that in mind, the landmark paper after several papers over the course of 5 years previous to this built up to this point was published in 2014 in which we went from a very small number of uncertain genetic findings to now more than 100 distinct regions of the genome associated to schizophrenia.

Mark Daly: There are a few noteworthy places on this plot that I flagged here, and I'll come back to some of them later. The vast majority of these remain not completely annotated, which is one of the main challenges that will be [inaudible]. This work has continued and led by Stephan and James Walters and Nick O'Donovan over in Cardiff. A new posting to the bio-- to the med archive just a few weeks ago has articulated an updated version of this PGC effort. And so, this is now the definitive reference for common genetic variation association in schizophrenia. As you can see, we're up beyond 70,000 cases of schizophrenia. The ancestry has been diversified to a small extent. There's a very large East Asian component to the study, but we are still sadly and badly lagging behind in the introduction of African ancestry and Central and South American ancestry individuals into these studies. The number of discoveries and many other things have all gone up dramatically, and in fact the case sample size has gone up by 80%, but the number of definitive discoveries has gone up by 150%, so that's a pretty good exchange. We've finally started to get some insights, but these are such still early and not generally encouraging, but not specifically actionable findings, but you can point to certain enrichment, gene sets of many different flavors. Clearly, they are relevant and correct statistical results, but they don't give us the hard experimental handles that we might need to move forward on. A few examples that do get us a little further by scrutinizing these results and attempting to take the latest algorithms for refining the mapping of these genetic signals down to in some cases single variants or a very small number of variants and utilizing this fine mapping activity to then as a jumping off point for experimental follow up of specific SNPs confirmed to be definitively associated to disease.

Mark Daly: There are actually quite a number of these-- well, not spend too much time on this, but through various prioritization techniques, we can provide some insights that allow us to sort of generally say that, okay, we have a very good sense of the genes that are being prioritized either by this fine mapping pointing to a nonsynonymous variant or to a highly probable variant in a UTR of the gene, or that all of the genetic association exists in a small locust confined to a single gene. And we have a number of examples that these are being followed up and pursued. The best evidence I won't describe for the validity of this prioritized list of genes they won't all be 100% accurate, but they show a very strong excess in the second study that I'm going to describe, the SCHEMA study with respect to exome sequencing. And so that's of particular importance. I've highlighted a few in bold, but I'll come back to in a few minutes describing further work on those. As has been from the case—the specific-- the general enrichment of different categories has been great. It doesn't give you a real biological building block to anchor on. What does are a few examples. The most noteworthy of it was Steve McCarroll in his lab, leading a careful dissection of the strongest of the peaks on that Manhattan plot, as it were, triangulating it down to a copy number variant of affecting the copy number of C4A, where additional copies of C4A, obviously, create more expression of the gene and create a higher risk for schizophrenia.

Mark Daly: And this dovetails nicely with a lot of historical work with respect to synaptic pruning and the role of compliment for and extensive work continues. Steven and his colleagues, Beth Stevens, Mike Carroll, and many others pursuing this potential building block that we can put into a more firm foundation. A second of the stamped number of really firm building blocks that we can anchor on that's of a high interest is an extremely  pleiotropic variant in a gene with the informative name of SLC39A8. This gene is not only has this amino acid substitution is not only associated to schizophrenia, but also associated to a variety of other diseases brain and non-brain. This gene is a ubiquitously expressed transporter of many metals, has a higher affinity for manganese, and the genetic functional allele here is specifically associated with a defect in manganese transport but not transport of any of the other. One thing that's of particular interest with respect to this is that the same allele increases risk to schizophrenia protects from Parkinson's and is clearly defective in manganese transport, and therefore manganese availability throughout the body. And that's interesting because manganese toxicity, so the opposite of what this allele would be doing, creates a Parkinson's like phenotype with tremor and bradykinesia. And my colleague Ron Xavier has not only created a mouse, studied it, demonstrated that it mimics the same manganese and glycosylation deficiencies but also has demonstrated that in wild type mice, the same deficits in the gut and with respect to manganese availability and glycosylation are mimicked by creating a restricted manganese diet in wild type mice. And this is quite interesting because it highlights that sometimes the genetic clues might lead us to environmental factors and potentially public health interventions and not down a normal path of, "We found the gene, let's mimic it and make a drug or something."

Mark Daly: But the reality of the situation is that despite hundreds of genome wide association signals, the vast majority of them have proven very, very difficult to pin down to single variants or any function whatsoever that we can be truly confident in. And that's because the majority of these with the exception of only a dozen or so generally implicate wide genomic regions that none of the variants implied in the region changing a protein sequence that would give us a very clear biological clue that we could take in model. And so, the challenge then is that with the vast majority pointing to some sort of regulatory difference, we have an additional challenge is that the context in which that regulatory-- regulatory tissue-- regulatory change exerts its influence. The cells, stimuli require, developmental time points, the surrounding milieu of cells that might be influencing it are all unknowns. And so, it's extra seemingly difficult to study these in comparison to what you can do more directly when you have a broken gene and you know how to model that. And so that's where I think we need to think about new tools but also think carefully about how much of this activity can be systematized versus how much does need to involve some hands to hand combat or bespoke activities such as Steve and Beth Stevens described for the C4 variation.

Mark Daly: So common genetic variation is proving quite productive, and actually now is explaining a significant amount of the heritability of disease. And this is great. However, it has its challenges with the follow up and really allowing us to anchor biological inquiry. And so, since it's the case that the vast majority of these associations are not so easily implicating a specific gene or variant, the question is where do we turn next? And obviously, with the advent of next generation sequencing, we have a possibility to look at much, much rarer variation and even de novo variation. And we know this has played a role because we have seen documented over decades, in fact, the role of microscopic and submicroscopic deletions and duplications in the genome as contributing in a major way to psychiatric disease risk as well. And so, they don't contribute so much to the heritability, but they may provide biological insights. The challenges of following those up are quite different because now you're looking at regions with many genes so they also are not providing super specific biological clues but it does open up the possibility that with the advent of next generation sequencing, pinpointing individual genes with point mutations might now be possible. This has been a tremendous effort of TJ Singh on behalf of the entire community and as many of you know, relatively small and then slightly larger sequencing studies began to come online five to seven, years ago. And this sequencing activity continued to expand quite significantly investments from the Stanley Center, investments from the NIMH, investments from many fronts building, again, the collaborative possibility that like the PGC and following directly in that model, we might bring together many studies and that's what TJ has been focused on.

Mark Daly: His effort over the last two years has been enormous in terms of its technical opportunity-- the technical challenges in bringing together diverse sequencing [inaudible] captures and so forth. We won't talk too much about that, we'll just thank TJ for tremendous effort over the last few years. This paper, like the latest PGC paper is also now a medRxiv, and the two papers have a considerable amount of crosstalk. So, thinking about the analytical approach to this data, we're now not thinking about one variant at a time. We are aggregating variants by their annotation category across a gene. This gives us then the opportunity to sort of build up even from ultra-rare and de novo mutations evidence for genes involved in schizophrenia. TJ took a look at the latest strategies for weighting missense mutations and created weights on loss of function in missense mutation according to their global risk and then brought those forward into gene base tests. A single test for each gene which identified now in the end 10 clearly genome-wide significant genes and another 22 that have a high probability of ultimately being true because of the false discovery rate analysis saying they are more than 95% above this threshold are likely to be true. So, with this, TJ was then able to now in this effort begin to fill in or start to complete the picture of the genetic architecture of schizophrenia. And so we have ultra-rare point mutations in red, the CNVs that I described earlier in green, and the common variation down in blue as you can see the C4 in the SLC39A8 standing out a little bit from their cohort but still far down in terms of the individual risk conferred by those variants.

Mark Daly: What do we learn from this? Well, we see a number of instances where there's a collision of results. For example, the genome wide association study and the exome sequencing study point independently through different mechanisms to GRIN2A, which is, of course, a wonderful candidate gene. Everyone in this audience I'm sure understands the construction of NMDA receptors. And one thing that's particularly interesting about GRIN2A is that it has its primary peak of expression after birth, whereas, it's essentially in the construct of NMDA receptors swaps places or swaps importance with GRIN2B. And if we look at the sequencing results from the Autism Sequencing Consortium, we see a very strong signal for risk of autism from loss of function mutations in GRIN2B, whereas, the later arriving GRIN2A is clearly a very strong and dual hit in the schizophrenia studies. Overall, that's not the only example where genes previously defined in the deeper sequencing studies of autism and DD/ID have demonstrated some overlap, but there's actually quite a bit more genes in the schizophrenia study that do not overlap the DD/ID and autism than there are that do. It's a pretty comparable number. So clearly, there is novel information coming from the focus on the sequencing of schizophrenia independent of what we've seen and previously documented in DD/ID and autism. Among those are some interesting--

Miri Gitik: Sorry.

Mark Daly: Yes.

Miri Gitik: We're over time. The buzzer already gone three times.

Mark Daly: Oh, okay.

Mark Daly: Alright.

Mark Daly: Great. I can't hear a buzzer, but that's good. There are plenty of other examples of interesting functional variants and sets of functional variants, allelic series that provide now some concrete handles into the biology of disease. And I think the last message then is that we need to be paying attention to those clues and putting the common variants and the rare variance in this same spectrum that I described a few slides back with respect to creating a holistic picture of genetic variation and creating a more comprehensive picture of where the genetics is clearly pointing to biology. And as the last point, you'll notice there's still a considerable hole in this distribution. That's the outcome of natural selection, which there are, of course, variants of intermediate effect that are low frequency in the population. This is the most productive part of most gene discovery efforts in autoimmune, disease cardio metabolic disease. Unfortunately, the strong force of natural selection kicks those variants down to such a low frequency that only variants with a very, very weak effect still can even reach unseasoned in the population. So those variants will eventually be there. It will just require even larger studies to define them. So with that, I would be very happy to take some questions and then thank you for your attention.

Steven McCarroll: Thank you so much, Mark for kind of setting the genetic table for the whole afternoon of biological discussion. So, one question I actually would love to have you comment on first, and then we'll open it up. And I should say about that the discussion and this take place throughout the day that the-- we only have a certain amount of time, so I just want to really encourage everyone to be pithy, to remember that brevity is the soul of wit and to have contributions to the discussion half a minute maybe a minute in extreme cases. But Mark, I'd like to start with you on this first question. So, beyond that kind of hand-to-hand combat with individual genes and loci, which seems when successful might be helpful for any illness which psychiatric illnesses do you see as having a sufficiently number is strongly implicated genes for meaningful large scale biological follow-up? Or that even the right way to think about it?

Mark Daly: That's a great question Steve. I think-- no, no. I think it's a perfectly reasonable question. I think because of the emphasis of the studies and the sort of power advantage in sequencing studies to having families from which de novo mutations can be derived, which is generally not the case in schizophrenia. The results from autism have been leading the way with respect to implication of genes with these ultra-rare burdens of truncating mutations, missense mutations, and so forth. Those, there's now a very large solid list of genes that are ready for meaningful biological follow up. And as those are coding variants, those lend themselves to more straightforward screening mechanisms, potentially. I think the other psychiatric illnesses are generally speaking, there's a lot more similarity in the genetic architecture of all mental illnesses across the spectrum. Schizophrenia is a little bit further ahead because of sample size, how much has how much effort has been applied to these studies to develop them. And also, perhaps phenotypically, slightly less confused or mixed with any other diseases. However, I think many of the others are now following in lockstep behind. Bipolar will not be very far behind at all in terms of its sequencing results. And I think even ADHD has showed some very strong promise with respect to these studies.

Steven McCarroll: Looks like someone just asked what about PTSD?

Mark Daly: PTSD, it certainly has-- there's no evidence that it has a distinct genetic architecture under the hood as it is a consequence of a strong or severe environmental exposure. It is that much more challenging to make the fundamental gene discoveries. And so, I think that adds a layer of complexity to the studies. But the genetic contribution under the hood doesn't necessarily appear any different, it's just going to be a little bit harder to get to.

Steven McCarroll: I should make sure people are aware of the Q&A feature within Zoom. I think questions are starting to come up in that and there's also people-- there's a sort of lively people can also help respond to one another's questions. Someone's just asked about the genetics of resilience. Protective alleles, for example.

Mark Daly: Sorry, I don't see the question but-- oh, okay. So, with respect to protective alleles, those have been-- we've had a great thought about how we might approach finding them. As it turns out, it's fundamentally more difficult to find protective allele than risk alleles because we ascertain on the extreme phenotype we're interested in studying. We don't have a mechanism to sort of ascertain the opposite of that. One thought to pursuing that going forward is we can identify individuals with high polygenic risks so their high background natural risk who don't develop the consequent mental illness. And enriching our future studies for that cohort might strongly enrich for the ability to discover protective rare variants, which might lead to some new biology.

Nicole Soranzo: My question--

Steve McCarroll: Oh, go ahead.

Nicole Soranzo: Can I ask you a question? Sorry, I can't type in the Q&A. So, it's Nicole here. I'm just wondering, so the point you made about some potential public health interaction could ameliorate some of these symptoms. I thought it was a very interest one, and I guess the question is, do you see evidence that particular the low end of the spectrum there are different types of molecular mechanisms so there is an arrangement for certain types of putative pathogenic effects through certain mechanisms compared to what you learned from the common disease side of the spectrum. So, put it another way, so if you have different types of pathologies-- well, I think you understood my question.

Mark Daly: Yeah, no, I think there's a little bit of a-- there's a bit of a hint of that in a few of the genes which overlap between schizophrenia and more "severe" neurodevelopmental disorders where you can see in the schizophrenia cohort that carry, for example, SETD1A mutations a sort of lower cognitive function, lower or worse life outcomes than the remainder of the schizophrenia population. I think these are pretty rare in terms of their contribution. And in fact, most of the genetic variance we're finding even at the ultra-rare end of the spectrum are more of the flavor of GRIN2A that clearly is a general and more specific contributor to schizophrenia, or are simply in genes that are clearly not documented as contributors to earlier neurodevelopmental phenotypes.

Steven McCarroll: Rick Huganir, you had your hand up?

Richard Huganir: Yeah. Hi, Mark. It's Rick.

Mark Daly: Hey, Rick, how are you?

Richard Huganir: I'm good. So GRIN2A is a de novo mutation in epilepsy. So, it's interesting that that's a loss of function and schizophrenia genetic suggests there also a loss of function. So, I just--

Mark Daly: Absolutely.

Richard Huganir: --wanted you to comment on that and could we use the de novo mutations to address the biology rapidly?

Mark Daly: Yeah, I think there's a couple of ways in which that might be possible. One that we're pursuing, not just in psychiatric and brain disorders, but across the spectrum is that identifying from among a very large set of these associations that come in in any one disease, what subsets overlap specifically with another disease. So, there may be a subset of associations that overlap with epilepsy and a subset that overlap with autism, and obviously, a large subset that overlaps with bipolar. Partitioning this way may provide us sort of ways of lenses through which to look at functional data, ultimately, that could be very, very productive. And while we have a number of examples where different mutation types control risk to different phenotypes in the same gene, that's not unusual from the history of Mendelian genetics, the GRIN2A role in both epilepsy and schizophrenia, quite a high degree of confidence in both of those observations is quite interesting because the mutations are not in any way, obviously different.

Stephan Sanders: Thanks for these great responses to this question. I think this plays naturally onto the next question we've posed, which is should we be studying risk alleles and haplotypes or risk genes? I think this idea that there's a gradation between complete loss of function versus maybe a small amount of loss of function possibly leads into different phenotypes for the responses to this, our panelists, if you would like to respond or comment, please use the Raise Hand function. And as Steve said, we can try and keep our responses quick. For those of us who are attendees, you can use the questions and answers function, the Q&A, which you get to on the More in zoom. And I'll do my best to try and look it back. And I do see there's one question there already from Sema about C4 alleles? So, Steve, can I pass that question about risk alleles and  haplotypes  vs. risk genes to you?

Steven McCarroll: Sure, sure. I mean, I think one way to think about it is that the human allele is just a tool for finding the gene. And in thinking about the kinds of biological experiments that come next, one's not limited to the human allele in designing experiments. One can certainly design a biological probation and experiment that represents a stronger gain or loss of gene function than the human allele represents. And so, it's not strictly necessary to do experiments on the allele that are ascertained in the human genetic studies. And especially when it comes to the discoveries from rare loss of function variants, which most of the rare variant discoveries are, they may be more or less interchangeable with loss of function variants that are just generated. It doesn't have to be the same loss of function there. There's a small number of cases in which evidence is coming from missense alleles, and probably in those cases it actually functional experiments might be helpful for attaching-- for really knowing which missense alleles matter in the  allelic series. But then for common variants, as Mark said, it'll in many cases, be unclear, which is the important gene at the locus. And so there, you at least that-- of those lights clear, and again, you don't have-- then you don't have to stay limited to the allele. But where it is unclear it may be necessary to start by thinking about the haplotype and what's the haplotype doing.

Stephan Sanders: [inaudible] Where does the specificity of phenotype come from? Mark, is that something you feel you could, you can take a stab at? You've mentioned already this sort of GRIN2A vs.  the example in trios, but from a 20,000-foot view, where does that supposed to come from?

Mark Daly: That's a great question. I think that's really one of the mysteries. And from the outset of discovery, even with the high penetrant CNVs 10, 12, 15 years ago, we were struck by the remarkable pleiotropy of effects that nearly every one of the major submicroscopic deletions that were discovered were not at all specific to one form of mental illness, but confer risk across a variety of cognitive, developmental, and psychiatric phenotypes. And whether this comes from the polygenic background, which whatever we think is the major heritable contributor and that the same sort of major allele, for example, in GRIN2A exposes some people who have a natural susceptibility to go to schizophrenia versus epilepsy, or whether environmental factors come in or whether there's simply elements of this that we haven't yet conceived of with respect to the biology and pure chance. It remains a mystery, but a very, very compelling one. I think we will have now the opportunity in cases like GRIN2A to actually begin to look at that interaction with other genes, with a pleiotropic background and with other lifestyle and prenatal factors to see where might we get some insight into that question.

Stephan Sanders: Thanks Mark. Dan, I see you have your hand up, can I pass to you?

Daniel Geschwind: Yes, yes. Hi. Just a quick kind of note and then kind of question for discussion. I think that probably with regard to that phenotype beyond the diagnosis is going to be important to understand what these pleiotropic loci actually mean. And one notion might be that one could study the smaller of-- that large effect size variants are like a crowbar in brain development to the brain. They're very large and big effect. And hence, maybe pleiotropy is more or less expected with a smaller SNPs. It becomes an interesting issue as to what phenotype are they actually contributing to in brains are kind of quantitative phenotype studies beyond the diagnosis seem pretty relevant. I'm just wondering-- yep, yep, that point of comments if you guys have.

Stephan Sanders: I think there's a key tool which needs to be developed following on from that. And that's the ability to dial down gene function in a gradiated manner. So at the moment we really get it knocking out genes completely, starting to get to the stage where a 50% knockout is possible so that you get constituent models, but the ability to dial it down and say 5% increments, I think that would be an extremely useful tool, but also a very, very challenging one to make.

Steven McCarroll: So, let's make sure we address this question about gene prioritization. So how do you go about taking this human genetics data and prioritizing genes and alleles for functional follow-ups? I know this is something you've thought about quite a bit.

Stephan Sanders: Thanks Steve. So, I think there's two key dimensions in thinking about this. The first one is focusing on the P-value piece, which as a geneticist, you would not be a surprise to hear. I think genes need to pass genome-wide significance, really strong follow-up. But within that, thinking about the rank of them within the order, there's a reason that genes get be the lowest P-values or the sort of full the highest on that curve, which Mark showed. And that's going to be driven by population frequency and effect size, both of which are great reasons to go after it. I think in reviewing these in grants, having that experiment that's put in there is very important, but we also need to be conscious of some genes, which, or at least I'm just, can't get through that. For example, low gene regions or something like Fragile X, where we just remove those cases before, they get into the cohort. And I think the second thing is making the arguments by function. Is there something which makes this gene more interesting vantage rank on the list? Is it specific to a certain cell type? Is there a therapeutic target? But we need to hit the drive gene-wide significance because otherwise, it's a very unique gene function, its risks being false positive. So, I think trying to dispel those two sides of the biological neuroscience view versus the statistical genomics view together is really the key to laying out the arguments.

Dan Geschwind

Stephan Sanders: Dan, I see you have your hand up, is that from before? Ah, okay. Other thoughts about gene prioritization?

Mark, is this something you would be happy to talk to?

Mark Daly: With respect to gene prioritization?

Stephan Sanders: Exactly. And locus prioritization.

You use spelled it out quite well, Stephan. I think we should pay attention pretty much to the statistical significance for starters, but it's still quite a considerable challenge, and I'm honestly, there's a huge amount of effort was put into this in the PGC paper. That's just now on medRxiv. And it's still not-- it doesn't give you a sense that we're really particularly accurate at this at all. And I think it's hard to prove that you're a significantly outperforming proximity to the gene, really in terms of many of these prioritization methods. So, I think we will need to begin to, as Steve puts it, seeing if the functional perturbation introduced by the haplotype is something that begins to give us clues. But of course, in the non-coding scenario, what model, what cell type perturbations, etc. to look at, even when you get a readout, you can't have ultimate contact confidence that you've got to the right answer because you sort of got to an answer, but it's hard to prove that thread all the way through to the causation of disease.

Steven McCarroll: There's a question in the chat about why-- maybe you can think of this as a strawman, but the functional follow-up should it-- should we begin functional follow-up with the genes implicated by rare variants together with the fine-  from the common variant association studies? Does that sound like a good sort of draft one, or would you revise or add to that in some way?

Mark Daly: That sounds like a logical starting place for me. I mean, in particular, we get a high degree of certainty from the multiple mutations clearly pinpointing a gene. And so that's clearly the best place to start, though, there are an equivalent number of the several hundred from the GWA that could outweigh a, single variant or single causal variant as you did with C4 or the manganese transporter. We know what the variant is. We know what gene is perturbing And so those are great. In an emerging set in which noncoding variation in a region is adjacent to a gene that has significant results from the exome study, and this gives you a particularly intriguing opportunity to pursue in a real series of that gene, where you now have hard genetic evidence that whatever this non-coding variant is doing, it's likely path to relevance is through the adjacent gene, but has the rare variant signal. And so, I think there are definitely opportunities that we can take now, not all of the results from the GWAS, but if we're selective about it, I think we can do a pretty good job.

Steven McCarroll: I see Rick and Kevin both have their hands up. We're going to call on both of you guys and then probably wrap up.

Richard Huganir: Kevin, do you want to go first?

Kevin Eggan: Yeah, sure. I was just going to propose biochemical interactions as another means of grouping and prioritizing genes of interest for study. I think there's good precedence for this already with the old type calcium channel. And I think that focus on these sort of locations in the genetics where the biology concentrates in a particular location can also be a key biological lead. And I thought that that would be useful to inject in the conversation.

Richard Huganir: Yeah, that was my comment too, basically looking at the pathways, several of these rare variants are uncommon pathways, common locations, in fact, in some cases interact with each other. So, I think that's a real key to the biology.

Daniel Geschwind: Yeah. I want a second or third, that's exactly what I was going to say. I think from a neurobiological standpoint, that begins to make sense.

Mark Daly: So, I think that's unquestionably correct. And I would inject that what the field needs to do is to develop a rigor around these sorts of analyses such that if we predefine pairs of genes that we consider to have evidence or predefined the biological pathways specified in a certain way, we can then apply in an analytically rigorous way, the same sort of discovery techniques just as you would like. My reading of the literature is that that's not generally what's done, but people take sort of the most optimistic opportunity, things that they recognize in the list of genes and say, "Isn't it great that such and such is important?" And there's absolutely no way to know consider that when we're looking at list of a hundred genes in autism, 300 genome-wide association hits in schizophrenia. But I think that is something that with the NIMH could encourage the development of real rigor around this area, because I think you're absolutely correct, Rick, Kevin, Dan, that this is a key and essential path to interpretation of these genetic results. It's just not done very well right now.

Daniel Geschwind: So, let me put you on that a little bit, Mark. It's interesting because-- so Kevin, you mentioned biochemical interaction, so you get protein interactions, and those are kind of-- can be kind of well-defined, of course, they're just the tip of the iceberg, but one way to look at things. And so, you can basically ask, "Do risk genes fall into certain protein-protein interaction networks, more than by chance using kind of standard permutation approaches that are they're used across a wide array of genetic and genomic studies?" And I'm wondering what your thought about that is.

Mark Daly: Yeah. I mean, we've put a lot of thinking actually more than a decade now ago into using the protein/protein data in this way. And I think it can be employed very, very successfully. The challenge is when you take that step to sort of the definition of networks, everyone has a different approach to defining networks. And so, the same data will have countless people asking related but not identical questions of it. So, there's some challenges in that, but I think we can do these things in a rigorous way and the protein-protein data, receptor-ligand pairs.. As long as we predefine the space that we're going to apply ourselves to, I totally agree, Dan, can be regeneration.

Steven McCarroll: So, this discussion is so lively and has so much moment that I hesitate to cut it short and I've already forgone my five-minute summary and wrap-up, which I think was unnecessary because the discussion itself was actually really structured and clear. But I can't forget that time for our afternoon speakers. So first I just, I want to really thank you, Mark, for preparing this. This was a terrific talk and everyone for the role in the discussion after, and then I want to pass the baton to Anne for the next step and Kevin for the next session.

Mark Daly: Thanks, Steve.

Anne Bang: All right. So maybe I can start by introducing the session here. So in this next session, the speakers who are all leaders in their field, I think taking very creative approaches to a difficult problem will address the question of how do we leverage the amazing progress we just heard about and the identification of genetic variants associated with neuropsychiatric and neurodevelopmental disorders. So, the focus of this section is really how do we methodically screen at scale? These putative risk genes are not one at a time, but many for biological function and understanding the impact of their disruption with an eye in the long-term on identifying potential drug targets or entry points for development of therapeutics. Of course, the challenge of doing this at scale is that the phenotypes are relatively complex. So, with that, I'll pass it over to Kevin. Who's our first speaker. Kevin.

Kevin Eggan: Thanks, Anne. Hopefully, I can do this expeditiously. Okay. Is everybody seeing the presentation or the presenter tools right now? Looks good? Okay, great. So, I want to thank Steven and Anne for inviting me to be a part of the workshop today [crosstalk]. That's easy to fix. Here we go. How's that? Great. So, thanks again for the opportunity to be here today and tell you about some of the approaches that we've been taking to attempt to solve these challenges. I think Marcus always did an exceptional job of really laying out what an incredible job the genetics teams around the world have been doing, connecting variants to human phenotypes and psychiatric illness. And we're all cognizant that laying in between these alleles and the effects on individuals must be many different layers of biology. There are immediate effects of these variants on transcription which then cascade down through levels of behaviors of proteins and cells and tissues, circuitry to ultimately affect the health of people that we'd like to provide solutions for in the longer term. The question is what are these effects? And I think the main topic of the discussion today is how do we devise systems for addressing them, understanding them, and then knowing whether or not we can ultimately successfully intervene. And really, I guess you could say the tribe of people in this area that I've been asked to represent are broadly those that have been trying to deploy different types of human cell models for solving these types of problems.

Kevin Eggan: And I think the attraction of doing this as is obvious because of the nature of the genetic problem that we confront in psychiatric illness, actually being able to study models that have human genotypes both to look at these interesting biological consequences of common variance, but also to ask how variants come together to affect a particular form of neurobiology, I think is important in conditions with genic architecture like psychiatric illnesses. Now, this is an area that's been progressing rapidly. Thousands of cell lines have been made by different consortiums around the world. Many from individuals that harbor psychiatric illness and methods are advancing, although there's still a lot of work to do to make many different brain cell types of interest for studies. And of course, also these cell lines provide an opportunity for intervention that we don't necessarily have. Of course, it's an inpatient themselves. I think overall, there are many beneficial trade-offs between scale and more brain-like complexity that can consider in models like this, where you could screen individual cell types like cultured neurons, but you can also do more complex studies in organoids that will culture it for many weeks months or even years. The problem with this is that as one begins to scale these sorts of experiments, to really try to embrace the mandate that Josh mentioned at the beginning, to kind of transcend this study of one mutation or genotype at a time that it becomes laborious, it becomes expensive, and it becomes challenging to balance the desire to measure many things with the rigor of doing great experiment, and that variance in self constituency, the state of maturation and differentiating cultures, or just natural forms of experimental variation begin to creep in and dampen the signal in experiments like this.

Kevin Eggan: And so, at the Cerner Center, in collaboration with Steve and others, we've been trying to pioneer new approaches to overcome these types of challenges. And I think combining new single-cell methods and sequencing approaches with just the simple idea of beginning to move to experiments, where we grow the cells of many, many people together in a single environment has begun to provide really remarkable returns on investments. And I think that the benefits from doing this are really appreciated almost every single level of the experiment. And I'm happy to answer many questions about that. But really the goal is, instead of using hundreds of dishes to do a single experiment where we ask a question about a particular variant, we're now doing hundreds of experiments with hundreds of people on a regular basis. One of the methods that I think is easiest to think about is just to take a cell village like this, and to subject it to single-cell sequencing and in an experiment like this, instead of reading out how the genotype of a single person affects transcriptional output, we can literally ask about the relative effects of gender variation amongst dozens, or even hundreds of people in a single experiment. And by using intrinsic genotypes and common genic variation as an intrinsic barcode, this experiment can be carried out very rapidly. Just a quick vignette and an example of the sorts of things that can be successful in this, this kind of drop relation genetics or single-cell sequencing and expression, quantitative trait identification allowed us to get some insight into one of the common haplotypes implicated in schizophrenia. So in the upper left-hand corner here, you can see on the X-axis or region of chromosome nine, where SNPss that are associated with schizophrenia reside, and you can see this group of SNPss that all share about the same P-value for significance of enrichment and schizophrenia.

Kevin Eggan: But there are a variety of different genes that reside in this particular locus. And it's hard to a priori to guess which one might be important since they all encode several different subunits of acetylcholine receptor. But when we carried out an experiment like this, we could readily see that a subset of those SNPss also were highly correlated with expression of just one of those subunits of the receptor in the area. And in this case, Charney five and its expression were closely coupled with the SNPss that resided within that region of the genome. And I think this is a great example of how we learned, certainly, an interesting hypothesis from this one about a series of psychiatric risk variants, but in that one experiment, we also identified almost a thousand expression QTL in other genes that exist in excitatory neurons. And so this one single experiment taught us a lot about different variants and how they behave in that particular cell. Just as the second thing I'll highlight is that we can begin to go beyond this and not just look at how the variants affect transcription in these villages, but to take it beyond that, to ask about expression of proteins or even different functional consequences of variants in these contexts. And we can do this by using a simple method, sequencing method, called census sequencing that Steve and Jim in his group developed where we can basically take a village, and by using genome sequencing, readily identify the constituency of cells within that village. Just take a census and ask how common each person's cells are within that pool. And by really subjecting that pool to any kind of segregating selection or sorting method, we can begin to ask who cells within that pool have enrichment for a particular type of phenotypic trait of interest?

Kevin Eggan: And so you could think about this as anything, and it's kind of like the proverbial lining up of individuals on the football field by height and by selecting people cells that have either a higher or lower phenotype segregating them into these two different populations for further analysis. And then we can use census sequence again to basically define those pools and use that as a quantitative trait that relates to that selective force. And we can then correlate that to genotype. And I'll show you just one quick example of this. And just to show you that this method works well, here's another example where we had done dropulation  genetics. This time on MPC is from 48 individuals it's identified at about 1,500 expression QTS within those NPCs. And one of those happened to be in a gene called IFITM3 and three that had previously been implicated in viral susceptibility. And when we went on and looked at another pool of MPCs made from 37 donors, we could perform see census sequencing where we infected that pool with secret virus to ask whether or not that expression QTL was also a functional QTL for viral susceptibility. And when we did that, we found that those same SNPs, which govern either IFITM3 expression were also associated with variation in the susceptibility to infection of the virus itself in that pool. So I think that there's still, of course, a long way to go in developing methods like this, but this represents a kind of simple playbook where you can take advantage of pools of cells to not only discover how common variants express transcription of genes near where they reside, but also to then immediately go forward to ask whether or not those also govern important functional responses to the biology of interest. So thanks a lot for letting me  participate today.

Anne Bang: Right. Thank you, Kevin. I think we'll move on to the next speaker and then hold all the questions until the end. That was great. So next I'd like to introduce Jay Shendure. Jay is in the department of genome sciences at University of Washington, and he's going to talk about massively parallel methods for functionally characterizing genes and variants. Welcome, Jay. Thank you.

Jay Shendure: Can hear me now?

Anne Bang: Yeah.

Jay Shendure: Great. Okay. I'll try it. Yeah. Okay. How are you doing everyone? So thanks for the invitation and the opportunity for a quick talk here. So I just have a handful of slides. I thought what I would do is kind of just give an overview of some of the technologies that our group and collaborators have developed that I think are relevant to the general topic of this sub-session. So kind of at a high level, I think so our group is broadly interested in technologies for genomics, and these are kind of some of the questions that I think some of the methods that I'll briefly mention on the upcoming slides trying to address. So I don't think these are necessarily brand new questions. Some are quite standing. So how do they say which benign from pathogenic sequence variation? How do I identify what genes any given digital regulatory element is regulating, and then how do we interpret a signal from GWAS studies? So broadly-speaking, quite a few of the methods that we have developed and others as well. So some of the methods that Kevin was just talking about as well, I think are trying to leverage the barcoding power of sequencing. So effectively enable a large number of experiments to be carried out in the context of a single volume or space or whatever it is, right. So analogous to the next-generation sequencing. So I'm just going to run through a few of these, I think they're all relevant to this question of functionalizing genetic hits, and just kind of the state of the art, at least from our perspective in each so massively parallel reporter assays, I don't think these need too much of an interruption there now on the order of 10 or 12 years old.

Jay Shendure: The latest round, at least from us, has been to take regulatory elements that have been implicated in either rare or common human diseases across a broad range of disease areas and subject those to saturation mutagenesis. So here we're doing that effectively through PCR mutagenesis, and then kind of a straightforward NPRA, but altogether in a single set of experiments generating on the order of 30,000 empirical measurements of the functional, I should say, the approximate molecular consequences of SNVs in this kind of a model system. And so I think in general, one of the reasons that I like these kinds of experiments, where one is densely sampling variation over a locus, whether that's a protein-coding region or a regulatory element, is that you really get this distribution of effect sizes. And so you're able to contextualize the consequences of any given variant by putting it alongside all possible changes in that same element. And then you can also often, at least for example, the SORT1enhancers shown here, you can see additional structure such as motifs and such that connect you more directly to function. So another example here also from the last couple of years is saturation genome editing. So the essays that I mentioned in the last slide were episomal, and that's kind of historically been the way these sorts of large-scale functional assets have been done. So here, this is a method that we came up with a few years ago for basically doing multiplex introduction of different repair templates to a locus of interest. So you're cutting with the guide in a single location, and then you're appearing with many different templates.

Jay Shendure: And so most recently, we've applied this to BRCA1 exome in the context of that [inaudible] is essential. And we're able to generate these sorts of genome editing based measurements for on the order of several thousand different point mutations of these 13, exome actually nearly saturating. And a couple of things that are nice about what came out of this is that we get-- one, we get this kind of bi-modal distribution of effects, where we're looking at depletion of the variants. Again, this is essential BRCA1is the central in this outline that we're using. And I'm not showing it here. We also see that, two, the extent that clinical data is available or clinical adjudications are available for individual variants within the set, we see very strong agreements sort of indicating that these sorts of functional measurements can potentially be used to inform clinical decision-making obviously a different scenario than the one we're talking about here, but I think more generally points to the organismal phenotype, having some relation to these proximate molecular phenotypes that we're observing in a dish. Third one here, CRISPR QTL. So this was a framework that we designed with the goal of trying to have a scalable approach for connecting distal regulatory elements to the genes that they regulate. And in a nutshell, you can think about this analogous to an eQTL study, where the individuals are cells, and it's a nice genetic background, but we're introducing starts as a neurogenic background, but then we're programming in a large number of CRISPR perturbations kind of randomly to each cell, and then we can pack in. And then we're doing effectively an association study between the presence versus absence of individual guides in subsets of cells and the expression of genes that are located nearby. The CRISPRi target of that guide, right?

Jay Shendure: So we're leveraging a single experiment to do many tests, right, of individual perturbations of enhancers, analogous to how human genetics leverages the population of combinatorially assorted humans to do many tests, right. And down in the bottom left, it's kind of a QQ plot showing associations over background. And we were able to use them to come up to test a large number of hypotheses and pair this will regulators to their likely target genes. And probably, the thing that I enjoy most about this work is on the bottom right here, where there's a lot of speculation about how far enhancers are from the genes that they regulate, and we're able to kind of empirically ground that and quantify how often these sorts of one mega base regulatory relationships are actually out there, which turns out, at least in the context of this outline we're looking at, is not very often. Okay. Last example here, so this is moving from actual functional provisions to measurements or readout. So we've developed, over the last couple of years, a series of technology-- or a set of technologies that are complementary to Drop-seq, and those kinds of methods pioneered by Steve and colleagues that relies on this concept of combinatorial   indexing, where you're pulling and splitting populations of cells or nuclei to sets of wells and you're performing in situ molecular barcoding along the way. And the advantages of this framework are, I think two, one is that through several rounds of barcoding, if you can get past three or more rounds, you get access to exponentially, larger numbers of cells. So in kind of current experiments, we're able to profile several million cells for a quite reasonable cost and work done by a single trainee.

Jay Shendure: And then we're also just-- there's relatively straightforward molecular biology able to access a broad swath of different kinds of things that you might want to profile. So not just gene expression, but also chromosome accessibility, Hi-C methylation etc., right? And I'm actually going to stop there to stay on time. I guess the only point I was going to make here is that we can leverage this kind of single cell, a taxi data to identify cell types that are relevant to neuropsychiatric disorders. And we're now working to try and couple these sorts of single-cell readouts with the sorts of large-scale or multiplex perturbations that I mentioned on the previous slides. And I will stop there.

Anne Bang: Right. Thank you, Jay. That was great. And so in the next talk here, we're going to move on to screening and actual whole organisms. And so Randy Peterson, from University of Utah is going to present zebrafish behaviors, scalable high content meetups of genetic and chemical perturbations, and brain function. You ready Randy?

Randall Peterson: Yeah. Can you hear me?

Anne Bang: Yep.

Randall Peterson: Great. Hi everyone. I'm really grateful to be able to participate in this workshop and tell you a little bit about zebrafish behaviors and how they can be used as scalable high content tools for assessing brain function. Really appreciate the funding of various agencies, particularly NIMH and other NIH institutes who have supported this work over the years. So as a couple of people have already alluded to, when you're considering the use of a scalable technology for a particular biological application, one of the main considerations is how to balance trade-offs between scalability and phenotypic richness. So on the one hand, we have, for example, cultured cells, which, as we've just seen in the last couple of talks, have just very powerful scalability and can enable really high throughput experimentation, but they do have some limitations on the phenotypic richness that they exhibit. On the other end of the spectrum, we have, for example, rodents, which have a lot more complexity and richness of phenotypes that are harder to scale to the same degree. And some of us who have been working with zebrafish were attracted initially because of this ability to get some of the best of both worlds. You have an intact organism with a lot of phenotypic richness, but because of their size and also their fecundity , you can perform experiments on tens or even hundreds of thousands of animals at a time. So it gives you some of the best of both worlds. Of course, also, another advantage of working with animals is that they have behaviors which have traditionally been among the most powerful ways of assessing brain function. And so in my lab, we've been interested in trying to scale some fundamental zebrafish behaviors to a degree where they could be useful for assessing large numbers of pharmacological or genetic perturbations.

Randall Peterson: I'll just give you a couple of examples. This first one here was developed by Yijie Geng, who's been interested in sociality. And so he 3D printed some arrays where he can put zebrafish in an arena where they can choose to either spend time around a transparent window where they can see another zebrafish or near an identical window through which they can see nothing. And most zebrafish are social, and they choose to spend their time hanging out around another stimulus fish, but these kinds of assays can readily identify fish who are ambivalent to a social stimulus. In another example, Gabriel Bosse has developed an automated assay for conditioning fish to self-administer opioids to themselves. And after five training sessions, you can see that the fish spend all of their time hanging out around a platform where they can self-administer opioids to themselves. This assay is relatively low throughput can only screen hundreds to thousands of conditions as opposed to this sociality assay which can be scalable to the thousands or tens of thousands of conditions. Andrew Rennekamp developed an essay where he came and assess the effect of drugs or genetic mutations on threat responses. In this assay, he uses 96 well plates to assess the way that fish respond to a threatening stimulus, which in this, case is a strobe light. And it's highly scalable, attend to the fourth to ten to the fifth conditions at a time, and the endpoint can either be unidimensional or it can be multidimensional using AI to assess and develop complex signatures for behaviors.

Randall Peterson: This last example here is developed by David Kokel, he is actually using 96 well plates filled with 10 fish per well, and then subjecting these plates to a full battery of different behavioral assays, where he presents the fish with various light vibration or acoustic stimuli and then quantifies the way that they respond to those different stimuli across this whole battery of assays. This also is scalable to ten to the fourth, to ten to the fifth conditions and is really highly dimensional. So what he does is basically put the fish in the plates, add different compounds and then record videos of them responding to this battery of challenges that they're presented with, and then measure up to 14,000 different behavioral features in each of those wells. And what this does is to generate a high resolution, very rich mathematical descriptor of the impact of that mutation or that pharmacological perturbation on the behavior of the fish. And here's an example of shown below, where in the red line, you can see over time - this is over about 20 minutes - wild type fish, or control fish, responding to a whole battery of different stimuli that are shown down along the bottom. The green line contrasts a fish that had been exposed to this chemical compound, VU compound, which is an inhibitor of a KCC2. And you can see that it has a very distinctive impact on the way the fish respond to these different stimuli.

Randall Peterson: And so that green line becomes essentially a barcode or a behavioral barcode or signature that mathematically describes the impact of that compound, on the brain of the fish. And what's powerful about that is that we've been able to go through and generate a whole library of tens of thousands of different chemical compounds and the signatures that those compounds produce. And we can then take a new compound of interest such as this fast-acting antidepressant, run it through the assay to generate a behavioral signature, and then essentially blast this against the library of different behavioral signatures to identify new compounds that share the same behavioral effects as the fast-acting antidepressant. And in this way, you can find relationships between different compounds and genes through the same battery of assays. What we found is that the compounds we discover with these assays are well conserved in mammalian models. For example, here's a compound Finazine that was discovered in a zebrafish assay as having antipsychotic-like features. And we find that in a mouse model of PCP-induced a hyper-locomtion commonly used schizophrenia model in mouse, that the compound shows at a antipsychotic-like activity there as well. A different screen identified Finasteride as a potent suppressor of opioid self-administration in zebrafish. And we found that it had the same effect on rats without affecting other behaviors. And in the third example, ICRF-193, a ablate sociality in zebrafish when they're exposed at an early age, and we found the same thing to be true in mice reproduced in several of the key features of autism.

Randall Peterson: I'll just end with a couple of limitations here, although I think many of these assays are scalable to the tens or hundreds of thousands. Some of them are still not scaled sufficiently. They can only be done at the scales of hundreds or thousands of compounds. We'd like to increase that. And then I think another major limitation in the zebrafish field is that the scalability of genetic perturbations is still currently limited to usually tens or maybe a few dozens of genes at a time. And I think one of the big needs and something that I expect will be coming in the next year or two will be methodologies that allow us to do much larger genomic perturbation in zebrafish on the order of thousands of gene perturbations at a time. With that, I'll thank all the people who've contributed, and hand it off to the next speaker.

Thank you, Randy. So our final talk addresses a really crucial component. So Alex Battle is going to speak about integrative computational approaches for understanding genetics of psychiatric disease, how do you bring together all of this data that's being generated in these high throughput screens? Alex

Alexis Battle: Hi. Let me share my screen. Okay. Is everyone able to see the slides? Yes. Okay. So I'm Alexis Battle, I'm faculty at Johns Hopkins in biomedical engineering. And I have my acknowledgments on the first slide. So I do want to go ahead and say, I'm going to try to represent what's happening sort of across the field, beyond my own work, but I will go ahead and thank a lot of members that have contributed to some of the work that I am showing here, including Ben Strober and Karl Tayeb, and Marios Arvanitis , and also the work of the entire GTEx Consortium and my close collaborators there, who I've worked with us on as well. So let's see, I'm not advancing. Okay. So as you've seen in the past talks, including Mark's and the others in my session, now that we have so many different genetic low side that appear to be associated with psychiatric disease and other diverse common diseases and rare diseases, we're also collecting really diverse data types now for large population studies in addition to individual patients or families across multiple different layers of resolution. Now, the focus of in my lab is largely on looking at sort of multi-functional molecular data. So this will include things like RNA sequencing, ATAC-seq, etc., that are giving us a picture into what's happening in the cell beyond just the genetic variants themselves. But of course, we have data that starting to look at things like imaging lab tests, other into phenotype data that is really giving us a much bigger picture of what might be happening in individuals harboring some of these variants. So I'm kind of taking it for granted that talented individuals, such as those you've heard in session so far are indeed collecting this data. And what I'm going to focus on are some of the opportunities and challenges that we can then use computational approaches to address some of these questions. So I'm really thinking about computational integration of all these different data types that can hopefully offer us a fast and scalable way to provide preliminary characterization of what these diseases lociloci might be doing, including down to sort of mechanistic predictions of individual loci characterization of the disease architecture as a whole or interpretation of individual rare variants, if we're looking at personal genomics.

Alexis Battle: So the tools that people use really vary quite widely all the way from very simple, straightforward statistical hypothesis testing to probabilistic modeling, which is kind of the focus of my lab's methods and approaches to deep learning and other forms of machine learning and things like this. So as we think about what kinds of tools are appropriate, I will circle back to this at the very end of the talk, but the considerations that you can keep in mind include things like speed, really, how scalable is it? Flexibility to incorporate all these different data types that we might be interested in, and finally, of course, interpretability because if we really want to use those to generate better understanding of disease, loci, or of disease biology, we need to understand what these models are telling us, of course. Now, I do want to be very clear that my view of this, even though I am a computational PI is that these tools are used to generate hypothesis and are certainly not a replacement either for collecting data upstream of all these different types we'd like to integrate or making a final determination. They're not a replacement for experimental approaches that will validate the predictions we're making about what these loci might be doing. So this is kind of in the middle, what can we do to leverage large scale public data or individual targeted experiments to provide diverse data types that we can use to sort of characterize these loci before we go and run maybe a potentially much more extensive or challenging validation experiment. So a big focus in my lab is on integrating gene expression data to characterize common disease loci. And here our goals - it's already been brought up today - that it can sometimes be challenging even to identify the gene that's affected by a risk variant. So one of the goals of integrating expression data is to help identify target genes that might be affected by those loci. And also to understand beyond that, what are the cell types, the environmental context, the temporal stages, where these loci are actually active.

Alexis Battle: So I've been involved in the GTEx project, essentially since it started. You may have seen recent papers that marked the endpoint of that analysis. So in the GTEx project, they were able to collect gene expression data for nearly 1,000 individuals across multiple tissues of the human body. So the most recent analysis includes about 50 different tissues and these multiple different brain subregions that could be relevant to the neuropsychiatric diseases we're discussing today; of course, liver, blood, lung, etc. So we have multiple different tissues across about 1,000 individuals, of course, representing genetic diversity. So we're able to look at association between all these common genetic variants and expression across multiple different tissues, giving us expression QTLs, splicing QTLs, and other things, and all of these diverse tissues. And from that, we can go and take - as many of you have probably done - a disease locus of interest and look it up in a resource like GTEx or the many other QTL studies that are now available, and say, "Okay, is my disease locus a clear QTL affecting a given target gene?" And you can, of course, check and see what tissues that EQTL is actually active in as well. You may see that it is specific to the brain in some cases, or even specific to a very particular brain sub-region, which could give you some insight into what it might be doing. Now, of course, this is not as simple as it sounds, because there is so much common genetic variation that does affect gene expression. Nearly, any locus you lookup will appear to have some association with a nearby gene in some tissue, but that doesn't necessarily mean that they actually share a causal variant. And so what I'm showing here in the bottom left is a figure from the co-loc paper. So co-localization seeks to take two different association studies - in our case, a GWAS and an eQTL study - and look at the association signal you see across a particular region and try to assess whether or not they actually share the causal variant, or whether they simply have different causal signals arising from distinct variants that maybe an LD or nearby each other.

Alexis Battle: So co-loc is one of the sort of key computational approaches, and there's been lots of others that have been developed in the field that are related that try to tell us if you have QTL data, can you identify what is-- can you identify whether a GWAS locus actually shares a causal signal with a particular change in gene expression or splicing or other molecular data that you may have available. In our lab, have recently developed a method called CAFEH, which actually seeks to take association studies from not just two different studies, but actually aggregate the signal that you see, for example, across all the tissues in GTEx potentially multiple different GWAS at the same time and find the factors that represent different underlying causal signals across these different studies and also handle the fact that in many cases, you may have more than one causal variant, or a allelic heterogeneity, even for the same phenotype in different tissues or in different studies and actually sort of separate that out into its causal components. So one thing I do want to highlight here is that I do think that it's really important to consider the different contexts, the different time points, the different environmental stresses that may actually be important to the activity of the disease locus. It may not actually affect just sort of steady-state adult expression levels of a gene in the brain, for example, it may only be active during development [inaudible], etc. So we really have to consider what data do we need to collect in order to identify these actual shared causal signals and to identify the correct genes and the correct cell types. And that's something that my lab is particularly excited about. We also can use gene expression and other functional data to help characterize rare genetic variation. So we've heard a little bit about rare genetic today, but most of that focus is usually on protein coding variants, and loss of function variants.

Alexis Battle: So what do we do with the tens of thousands of rare genetic variants that all individuals are harboring that may actually be impacting our phenotypes and may actually explain some of the rare disease cases that we do see in autism and others? So non-coding variants and regulatory variants that are rare are extremely difficult to characterize. And the approach that we take again is an integrative approach where we think that if we also collect in the same individual where we have whole genome sequencing, collect RNA sequencing or other functional molecular data, we can combine these computationally to try to see, "Okay, you have this individual that has a potential rare regulatory variants, and yes, we also see an extreme change in gene expression of a nearby gene compared to the population or an extreme change in patterns of splicing." So we and others have really been focusing on understanding how often rare regulatory variants kind of correspond to major changes that are evident at a cellular level. If you collect molecular data and developing computational approaches that are appropriate for actually integrating these diverse signals to try to more accurately or accurately at all prioritize some of these rare non-coding variants that may explain phenotype in cases where you don't see a clear coding variant, that explains the phenotype you observed. So finally, I wouldn't say that, of course, even though it's not my lab's focus, lots of people are working on doing association studies and creating models that integrate lots of different data types. And that ranges from looking at protein levels all the way up to looking at imaging and the higher level of phenotypic data that you can integrate into some of these models. And that includes things like functional MRI, MRI, structural MRI, histological slides, collecting more detailed environmental data, wearables, trying to look at what's happening, not just to the proximal gene, but to downstream genes in the cellular network. And I've listed here, and I can share the slide, some of the considerations that you need to take into account, which includes that these data have very different distributional properties. They often have different missing this where subsets of subjects have one data type, but not the other.

Alexis Battle: So as you create models, you do have to-- or computational pipelines, you do have to really consider these. And some methods, for example, like deep learning are very good at automatic feature construction. You don't have to have as much expert sort of specification of the model. They are a bit more challenging to potentially interpret in some cases. My lab tends to focus on probabilistic modeling which is very flexible and powerful and can be highly interpretable, but in many cases, does require more significant expert specification of the distributions and the structure of the model that you want to use. That gives you a very quick high-level overview of what I think is exciting and our potential application of computational pipelines to characterize these variants. And I'll try to stick to time. So with that all, I'll pass it on. Thank you.

Anne Bang: Okay. Thank you to all the speakers. I think, Kevin and I are going to moderate a discussion session here. Talks were great. So maybe we can start with the first question, which is, to what technologies both in vitro and in vivo to the panel are ready to be scaled for unbiased biological assessment of human risk variance and/or genes? Throw that to you, Kevin.

Kevin Eggan: Yeah. I mean, I think this is one that I hope many panelists will weigh in on. I think that it's been quite exciting to see the way that single-cell transcriptional profiling, chromatin accessibility, and other all of the kinds of bolt-ons that came along with RNA sequencing are becoming rapidly deployed in a single cell setting. And I think that a lot of the molecular biology was done so well in the development of those initial technology is it's that that's really making that possible. I think that also the more kind of high-throughput approaches that people are taking with biochemistry around systematic tagging of proteins to allow kind of fair, equitable comparison of protein/protein interactions with many, many different proteins is becoming practical. These are the things that really kind of are rising to the occasion today and seem ripe for deployment. And I think as was really nicely said by many people in this session can really-- I believe provide a lot of texture with respect to thinking about what the variants that are being identified mean, what the haplotypes that they reside in point to  and which genes are being affected, and how they're being procured by that common variation. So I think that it's a time for taking advantage of these remarkable advances in molecular biology for moving forward. And I think we should see who else is really diving in with their hands up.

Anne Bang: Let's see, Jay.

Jay Shendure: I don't know how to put my hand up electronically. But I think--

Anne Bang: Here you go. Raise hand, yeah.

Jay Shendure: So I think most of what's been talked about, I think, feels very ready to scale, right. I think there's no reason why we can't and shouldn't have continuous developmental atlases of humans and model organisms with single-cell resolution methods like MPRAs and CRISPR QTL are eminently scalable, right? I do think there's kind of another generation I think coming with these prime editing methods that allow quite a bit more precision that I think maybe have a little more maturing to do, but I do think will be a very powerful addition to the arsenal. But I think that the challenge is less the technologies in my head and more kind of coming up with good in vitro models that are compatible with the technologies, with their limitations. And that the community also feels good about being reasonable models for particular diseases.

Anne Bang: Helen Willsey.

Helen Willsey: Yes. Helen Lindsay from UCF. Randy, I really liked your graph talking about the scalability of different models in relationship to phenotypic complexity. And one of the models that I think is emerging and is now ready for kind of large-scale genetic interrogation of these disorders is xenopus tropicalis diploid frogs. And so people may be familiar with xenopus laevis, which has been a cornerstone of developmental neurobiology for decades. But of course, as a duplicated genome, but now as xenopus tropicalis is online with all the same tools, all the same advantages of an aquatic organism like you highlighted. But in this case with xenopus, you're actually able to make half mutant brains. So have one half of the brain carry a mutation in a risk gene of interest and compare it to the other half of the animal to understand what that gene does during development, during behavior, other things like that. And so I think I'm in terms of thinking about maybe new or technologies or models in vivo that could be used in a more hypothesis naive way to think about what these genes are doing, I think it's important to know xenopus tropicalis, I think, is ready.

Kevin Eggan: So let's perhaps move on to this next question about what do we think is next? What's coming next? And I think Helen from you, this information about xenopus tropicalis is quite interesting and promising, but I'd love to hear from someone like Alexis or others, what's down the pipe, what should we be watching for that's coming along? And that maybe seems like it's in the distance, but we'll be here before we know it.

Nevan Krogan: , Kevin can I chime in here?

Kevin Eggan: Yeah, of course. Evan, please.

Nevan Krogan: This is Nevan Krogan from UCF. It's kind of an extension of what your comments were, Kevin, and maybe what I'm going to describe as somewhere in between the first and the second question, but obviously looking at protein interactions, protein complexes through mass spectrometry-based approaches is going to be crucial. We've heard a little bit about that already today, but one of the next frontiers is I would say is structural biology and using cryo-EM, and once we've identified these key complexes, there's been these amazing breakthroughs in cryo-EM, where it's almost medium throughput analysis. That's coming up here where you can go right from tagging and purification from cells going right to cryo-EM, and you can do that plus and minus mutations that are associated with these different neuropsychiatric disorder. So for me, getting to the biophysical to the structural level is very unique. And then tying this in with all this other great data that's been discussed already, I just think that's probably not going to be discussed a lot today, but the structural biology kind of medium to high throughput approaches that are coming online, I think are going to be really exciting in this particular realm.

Kevin Eggan: So maybe it just-- since we're hovering over that topic for a second, can you give us a sense of how you think the throughput of that and how the cost of that will evolve from where is it now and what will it look like in three years, five years?

Nevan Krogan: Well, I mean, in the past obviously structural biology approaches have been leveraging reconstitution of complexes and that takes time. But the affinity tags that are now being CRISPR-ed in the genome for purification mass spec approaches, you can use the same-- similar if not the same tags after you've purified, you can take an aliquot and analyze it by mass spec. You can then do cross-linking mass spec to get information about the typology of the complexes, but then you can take an aliquot and go straight onto these cryo-EM grids. And this has worked being spearheaded by people like Dave [inaudible] at UCFS. And I mean, in my mind, I think we can go a lot quicker than what the structural biologists say, but are we going to be getting a structure a week here? I mean, it seems like we're getting potentially close to that and they all don't work, but if you're going to screen 50 complexes and kind of them work, that's really exciting, right? And then you go deeper there. That leads to more hypothesis-driven research. You can then go back to genetics and go back and forth this way. So I would say probably over the next couple of years, we're going to see a lot more structures. Cryo-EM structures out a resolution that's as good, if not better than crystallography. So to me again, combining that structural information with all these other high throughputs, genetic and proteomic approaches is going to be very exciting.

Kevin Eggan: Does anyone else want to look into the crystal ball of the near future and make some predictions about things that are going to be exciting or that would be useful to tackle these problems? I don't see a lot of hands up.

Alexis Battle: Yeah. So I'll say again that, I think that understanding not just sort of static effects of these genetic loci, but temporarily variable effects is going to require both new experimental methods that actually allow us to look including single-cell technologies that allow us to look across multiple time points in development and things like that, but also computational approaches that actually model temporal processes appropriately at very, very large dimensional data that we're talking about in genomics. So I think temporal dynamics is going to be critical. I think single-cell technology is of course going to be critical. And we're just kind of at the point I think the computational approaches are particularly exciting right now because we finally have sort of the scale and sample size and diversity of data where these computational approaches can really contribute, I think, to the neuropsychiatric disorders. So I'm excited to see what happens.

Daniel Geschwind: It’s Dan. I realized my-- am I in line to talk? Is that okay?

Anne Bang: Yeah. Go ahead.

Daniel Geschwind: I was just going to emphasize what Alexis said and just to add one additional layer on to that. It's not just temporal it's that the nervous system is a kind of a very dynamic system. Not just, I guess it turns out to be in temporal, but in terms of having quite different states at different times as well, even when it is kind of fully developed, let's say, from a developmental standpoint, so kind of exploring how genetic variation might act differently in different states of the nervous system in terms of a homeostatic or under certain kinds of stressors etc., in learning environments, I think. Yeah. Does that make sense?

Anne Bang: Yeah, I think that's a really good point sort of approaches to reveal vulnerabilities in your cell system. Diane Dickel has your hand raised.

Diane Dickel: Hi. I was going to-- this is less of a scientific thing that I'm excited about, but what I would really like to see more in the future is more coordination, I think, between different institutes within a NIMH to drive some of this. So kind of building upon a little bit what Jay said about finding better-- we have scalable tools, a lot of which have been developed by funding through NHGRI, such as ENCODE, we don't always have really great disease applicable systems and so really partnering with groups that are primarily funded by NHGRI, who are building a lot of these kinds of scalable tools with people who have a more specialized understanding of the disease and the underlying biology, I think would be pretty powerful for kind of scaling up some of these methods in a way that would be more beneficial.

Anne Bang: One more question then maybe we can move on to the next question. So Hae Kyung Im?

Hae Kyung Im: Yeah. Hi. Do you hear me?

Anne Bang: Yes.

Hae Kyung Im: So I'm so excited about all these new technologies and the data that are coming out. And what I would like to say that we really need to kind of invest in making this available to analysts because I feel like there's kind of so much cool data, but it's so hard to actually get access to them in some cases. So investing in open data sharing, but also reproducible pipelines so that we don't have to redo everything people have already done, right. So kind of facilitate things for there are a lot of analysts that could take advantage of this, do the integrative analysis that Alexis was talking about, but make it easier because that's what I do. I feel like I spend so much time just trying to get access to data. And it's a bit of a waste.

Anne Bang: Alex, do you feel the same way? Actually accessing the available data is difficult.

Alexis Battle: This is certainly still a challenge both in terms of data being released quickly after it's collected so that it can be taken by advantage by computational individuals, but just the barriers that exist in sharing data between institutions and sharing it from NIH to investigators is of course still a challenge. And I know that there's huge efforts like Anvil and others that are trying to address this, but certainly, it remains a waste of time for many of us at the moment.

Anne Bang: No, we're work in progress. So maybe we can move on to the next question. So what are the advantages and disadvantages of these technologies, both in vitro and in vivo for large-scale screening, and what will be missed by current technologies? Any comments from the panel here?

Kevin Eggan: Well, others are familiar in their answers, and maybe I could answer one of the questions I see within the Q&A, which directly relates to this topic, which is given the current state of human cell models, what can we learn about processes that are relevant to psychiatric genetics? And I think that this also relates to another question - which I hope we can have a discussion on - and that is that given that especially disorders like schizophrenia may have their fundamental pathology at the level of circuit dysfunction, how do we think about deploying these single-cell technologies in animal models towards understanding circuit-level dysfunction as it is relevant in patients? And I think these are two kind of broad themes that we could delve into in a number of different contexts. And first, I would just say that I think that making functional cell types has come a long way in human biology, and actually the functionalities that cells can now achieve in certain culture conditions in co-culture with CLIA or with other cell types or excitatory neurons cultured with interneurons, or even organoids, I think would surprise people with respect to the state of appearance of the synapses and their apparent morphological maturation, as well as the genes that many of those cells come to express. And I think the proof is in the pudding in really showing that these cell types are really valid tools for leading to hypotheses about what the implicated haplotypes mean biologically, and that will take the types of approaches that have been supported by the convergence neuroscience mechanism to go just beyond variant to gene, to some kind of biology that's I think of broad and validated importance.

Kevin Eggan: And so we'll need more examples like that which has been pioneered around C4 by Steve and Beth and other's groups to be able to show that they're valid over time. But I think that the biology has improved dramatically from early days. And with respect to, I think, circuit biology, almost all models systems suffer from the problem of lack of human circuitry, which is a direct analog to that, which is dysfunctional in patients. And I think it'd be great to hear from those like Randy who were working in zebrafish about how they think about those kinds of problems.

Randall Peterson: Yeah. I think that those are great points Kevin, and certainly getting to the circuitry is the ultimate goal. I'm not convinced that we've done it very well yet in the zebrafish, but I think having an intact organism is, of course, first step towards trying to do that.

Kevin Eggan: So I see Nicole has her hand up.

Nicole Soranzo: Yeah, hi, So I just wanted to check on what you said, Kevin, and also linking back to the talk that Alex gave me we are, even in terms of human genetic studies, we know there's extensive pleiotropy for some of these phenotypes. And so the ability to really dissect to regulatory and other type of signals, a single-cell solution, meaning I've been able to associate with as many cell populations as possible will be incredibly valuable in terms of dissecting this pleiotropy from causality and really make it the next step of being able to rapidly interpret genetic data, and obviously, with respect to that, also the ability to go into different populations with different LD patterns is going to be incredibly valuable. So invest into the next generation of studies with single-cell resolution different populations would be I think fundamentally important.

Anne Bang: I see Neville.

Neville Sanjana: Hi, everybody. This has been a really nice set of talks to attend. And just to, I guess, echo what Nicole just said and what Kevin also said, I'm very excited by these human cell models. I think over the last like 5, 10 years, there's just been this explosion in these organoid models. And I think now we can actually cope with having all these different cell types, because we can combine things like the CRISPR style approaches that we've seen here, where your opinion per term, all these cells with single-cell sequencing, where you can actually de-mix the cells and figure out cell types. But I think this seems to be technology that probably has a long runway ahead of it. And so thinking of the next, whatever, 3 to 10 years, I think right now we're starting to see the emergence of people trying to integrate other very important cell types that we know are important for things like schizophrenia, like microglia, astrocytes better integrate them into these models that sometimes naturally don't produce them these cortical organoid models. And so I think having-- right now, it seems like there's just like kind of a diversity of different protocols and approaches, but I think more invested effort that could lead to some standardization really careful comparisons between this. I saw in the Q&A there's some questions about how does this compare to neonatal brains, like very careful comparison, so we can understand which of these protocols to make these cortical organoids is the right, the appropriate one to use, to model different diseases, whether it's schizophrenia or looking at autism, things like this. And so that's certainly what I hope to take advantage of and the next upcoming years as those protocols develop.

Anne Bang: So, Kevin, you referred to the improvements in synaptically mature human cells. Do you think organoid formats are necessary for that, which are naturally lower throughput?

Kevin Eggan: I don't think that they are in all circumstances at all, actually. I mean, it's interesting. I think we learned a lot by studying rodent synopsis and rodent electrophysiology in 2D for a long time. And we've also learned a lot by studying sys cultures of rodents, and we've come to kind of learn what we can and cannot learn in each of these two different contexts and how going in vivo to optical forms of electrophysiology, as I hope we'll hear about from Adam Cohen later can teach us even more. And I think as long as we kind of stay in our lanes with respect to what we could hope to learn from different models, and I think we're honest with ourselves about that, then a lot can be gained. And so I don't think, strictly speaking, it will be necessary for all studies to go to the level of a three-year-old organoid, but it probably will be necessary for instances, mouse modelers have known for many years to include things like astrocytes and culture. And I think that the hopes from kind of early stem cell biologists, who may be weren't schooled in the art of neurobiology, just hoping that synopsis would just kind of appear on their own without other supportive cells from the brain, I think might have I think unnecessarily negatively colored the view of many people in the neuroscience community about what can be achieved with these human models.

Anne Bang: I would agree. Okay, we should move on to the next question here, and we're sort of running out of time for the session, but what is the minimum number of variance genes that could be assayed to identify the most relevant biological pathways and disease-relevant mechanisms? Anyone on the panel here. I guess in the interest of time, we can move on to the next question. So what questions can and cannot be systematically addressed in cells in vitro? What are the most relevant factors to consider, so when do you really need a whole organism? Comments? Imagineto look at an intact circuit.

Kevin Eggan: Yeah. I think we're definitely hoping for input from a variety of people, maybe a while we're waiting for others to raise their hand. Ellen, I see that you have a point that you'd like to make.

Ellen Hoffman: Yeah. So I wanted to come back to this question about circuitry and in particular, with the zebrafish model, someone asked about whether there's conservation of circuitry with mammalian circuits. And I think it's a really valid question. I think what Randall said is true that we haven't really gotten to the point where we can completely translate from early or lower model systems to mammalian circuits. But at the same time, I think if we think there is convergence in molecular mechanisms across genes, and I think that's what we were talking about in the first half. And I think as Randall showed, there's convergence at sort of the pharmacological level from fish to mammals, then maybe we can use these, these so-called lower model systems, fish, frog, or even just looking at the human cell culture to try to identify sort of basic biological mechanisms, that link genes. And those are very likely to represent conserved pathways. So maybe we can use these other systems, even if the complete circuitry is not recapitulated in a fish to a mammal. We can rely on these scalable systems to try to gain new insights into some more basic mechanisms and what might link risk genes. Maybe we can use them to identify convergent basic biological pathways. And Kevin, you made this point as well, that maybe we can look at that sort of basic biological or biochemical pathways, linking risk genes, and these would be more likely to be conserved.

Kevin Eggan: So I'm mindful of the time. So I suggest that perhaps, Anne, if you agree that we could hear first from Gavin Rumbaugh, let Rick have a chance to comment, and then we might pass the baton to Guo-li Ming to begin to kind of summarize this portion of the day. Does that sound great?

Anne Bang: Yep. Sounds good.

Kevin Eggan: Okay. Over to you Gavin.

Gavin Rumbaugh: Yeah, thanks a lot. So I just had a small comment about us. I mean, obviously, cellular models are amazing for understanding cellular functions of genes and risk factors, but there are limiting, right. And so a lot of the drawbacks are limiting. And of course, one of the limitations is that if you want to understand behaviorally relevant circuits, you have to move on to more complex organisms. And so I think one thing that hasn't been touched on is that one thing that a lot of us learned from biology of studying like risk factors in-depth genes in depth is that they have tremendous spatial-temporal regulation across the brain, right. And that spatial and temporal organ regulation can be actually across different splice variants of the same gene. And so that combination of regulation is going to regulate specific subsets of to then drive phenotypes, right? And so moving on to complex organisms, we can start to think about, well, maybe this gene works in this way only at this time in this circuit. And that's how the brain is left with this mark that then drives phenotypes, right? So where I think the scalability of reduced systems is critically important to move forward, we also have to be doing more in-depth and more sophisticated biological experiments in the models to be able to understand the gene complexity.

Kevin Eggan: Yeah. And I think I would even double down on that by saying the more we could know about how haplotypes change in their biology and how risk variants change the behavior of genes encoded nearby over time and in different geographies and the brain, the more we'll be able to say when and where the variants are contributing to dysfunction. And I think that that's a very important area we, we should continue to try to stay unbiased in our view about that and open-minded about when pathology arises, some could be early and some could be quite late. And so I think there's still a lot to learn about that. Rick, over to you.

Richard Huganir: Yeah. So I just want to say these scalable technology is really mind-blowing to me, but what we can do now that we couldn't do 10 years ago is totally amazing. One of the key things, I know one people that think about who are developing the scalpel technologies, if you're really interested in psychiatric diseases is really the assay. And so how do we know what the assay in these especially non-organismal assays. And so I think it's really a comment that really I hope people are thinking of that while they're developing the techniques that-- and I think this was a sort of a transition to the next session where we'll be talking more about assays and but for the people who are developing the techniques really to think hard how to use these techniques, and then how are you going to get, have an assay to read out something that's relevant for schizophrenia.

Kevin Eggan: Yeah, Rick, I think that's such an important point that you raised. I think trying to marry these ensemble sequencing approaches with things like 2D in situ sequencing, which could be combined with functional electrophysiological readouts either at a low time scale through calcium imaging at first, or through something like opto patch. It seems really essential to take and create a situation where you can cross over from these important molecular readouts into something which at least initially in culture would provide some aspect of the functionality neurobiologically that we could hang our hats on, even if we could get from those biologies to the type of synaptic assay that your lab has specialized in for so many years in a static readout around synapse number, I think would be a huge would be a huge advance. And we didn't probably spend enough time, I think, today talking about things like barcoded rabies viruses that maybe could be used for measuring in an ensemble culture, other contexts in CRISPR screens to ask about synaptic transmission of the virus, to the extent which we think that would be irrelevant phenotype or not. So I think you're right. We heard a lot of things, but there's still a piece that needs to be well articulated. So I think duly noted. So Guo-li, over to you.

Guo-li Ming: Okay. So I think that this is a very informative session and people talk about different models and how can we take advantage of different models and the different technologies, especially newly developed single-cell omics to address the problem, how risk genes are rare variants contributing to psychiatric disorders. So we know that they're different, especially we have now developed the vitro human cell models, as well as some of the vivo organisms, which can be used in a scalable fashion to address some of the disk questions at different levels, for example, at the transcriptome level at and then at the behavioral level and the something that is lacking, I guess, from all the systems is at the circuitry level. And we have some nice discussions. How can we take advantage of different methodologies, and how to use these different assays to address the best question that we could ask, especially how the risk genes or variants a copy number alterations contribute to different potentially disease-relevant phenotypes. And one of the things that we start to realize, a lot of the things and the risk variance effect so functions and the circuitry infections actually are dynamic. And it's also temporally different that then cost for more essays that address this question, like when and where this happens and how does that contribute to the cellular phenotype or molecular phenotype or circuitry phenotype, or even behavioral phenotype.

Guo-li Ming: So I guess that's the general discussion and some things we talk about what can be done, maybe in large scale high throughput using different assays in vitro. The human cells has the potential technology has the potential to go to large scale and also the zebrafish technology, as well as mentioned, the xenopus tropicalis has the potential to go for as a in vivo organism for large scale assets. I guess, Kevin, do I miss anything? Maybe you can add some more on this.

Kevin Eggan: I mean, I think we're, of course, highly time-limited, and I think the main goal here has been to try to alight from as many mountain tops as possible over the short period of time. And I think there are many things that we could go into more deeply. But I think just as a closing thought because I really want to preserve this 10-minute break for everybody in the day. I would ask the panelists to please look in the Q&A box and to read as many of the questions as possible. And many of these are directed to-- I would say particular groups of us that have expertise in that area. And I would suggest that any panelists feel empowered to answer the questions that have been raised and to try to provide an answer. And I think that we can have a much richer kind of sidebar conversation that can go on through the Q&A box, and hopefully to explore many of these items that we've just run out of time to talk about today. I heard everyone to do that. And then Steve, if you think it's appropriate, then that could feel like a nice note to end on and we could go into the break.

Steven McCarroll: It sounds perfect.

Anne Bang: Yeah. Thank you, Kevin.

Steven McCarroll: We will reconvene at 1:30. Enjoy the shortest lunch break ever.

[silence]

Steven McCarroll: I just checked in on remote third grade, which appears to be going better than it did in the spring.

Alexander Arguello: Steve, I've been seeing a lot of bloopers online about Zoom elementary school. Some of them are quite funny.

We have elementary key remotely over in the next room and it's actually worked.

Steven McCarroll: I have to admit, it is really fun to eavesdrop on your children's education.

Totally. Totally.

Steven McCarroll: In fact, I actually, my son's second-grade teacher that he had in the spring, she became so good at leading class interactions on Zoom that I would eavesdrop just to try to get better at it by myself. And in fact, one day I missed my own lab meeting because I was busy eavesdropping on her and I forgot that the lab meeting time had come and passed.

Kevin Eggan: My son taught me that Zoom has a whiteboard that I didn't know about.

Oh yeah. My fourth grader is amazing at Zoom. Basically, if my husband doesn't know something I'm like, "Ask Garrett," because he knows all the short cut items. He's figured out stuff on my Mac that on kids own old Mac that I never knew you could do unbelievable.

[silence]

Sergiu Pasca: The youngest participant on unregistered, I guess.

How cute.

Stephan Sanders: [inaudible] say hi.

Stephan Sanders: Hello?

How old are you?

Stephan Sanders: Three.

Wow. Are having fun helping dad and fixing all the problems?

Stephan Sanders: Are you having fun?

Stephan Sanders: Yep.'

Stephan Sanders: [inaudible] button sounds, don't you?

Stephan Sanders: Yep.

Stephan Sanders: There you go.

Stephan Sanders: [inaudible].

Stephan Sanders: They do.

[inaudible], where's your little one?

Sergiu Pasca: Don't ask. The two-year-old is just an entire adventure there in the spirit. Second grade is easy versus a daycare. Yeah, still we're doing what we can.

Guo-li Ming: Is the daycare operating?

Sergiu Pasca: It's partly operating. At Stanford, things are very complicated. We're only really not very open, I must say, so yeah.

Guo-li Ming: Are you able to go to the lab?

Sergiu Pasca: No, we cannot. I am not actually. If I go, somebody else cannot go. So yeah.

Guo-li Ming: Yeah, me either.

Sergiu Pasca: Very, very complicated.

Steven McCarroll: : So let's see. Who's chairing the session? Anne, do you want to introduce Guo-li or actually,  Guo-li you can just go ahead and start, actually.

Guo-li Ming: I think someone needs to authorize me for-- let me see. Maybe I can do that.

Gavin Rumbaugh: I think Sergiudoing the session, I believe, is that right?

Sergiu Pasca: Yes.

Gavin Rumbaugh: Yeah. So I can introduce her, if you're ready to start.

Guo-li Ming: Yep. Can you see my screen?

Gavin Rumbaugh: Yeah. Hi. So we're going to start the session, the next session this afternoon, which is focusing on mechanisms and assay's on how they relate to neuropsychiatric disorders and phenotypes related neuropsychiatric disorders. The first talk is Guo-li Ming from the University of Pennsylvania and functional characterization of neurons derived from patient-specific iPSCs.

Guo-li Ming: Okay. Thank you so much for having me here. So I'm going to tell you a little bit about our effort trying to develop different functional essays of human neurons derived from patient-specific IPS. One of the things that I guess that has mentioned already that Rick Huganir there has developed to look at the statistics of the synapse formation by immunostaining of the synopsis of the neurons. In this case, we use human neurons to arrive directly derived from patient iPSCs in this case. The first example I'm going to show is that from patients harbors this one or a mutation, and what we've shown is that in the patient, the neurons with this specific mutation shows much less reduced synapse formation during development in 2D cultures of excitatory, glutamatergic neurons differentiated from this patient’s iPSCsIPSCScompared to could family controls and other functional assay. Of course, you can do is directly do electrophysiology on these neurons. I have to say, this is a low throughput because you have to patch individual neurons. So it's labor-intensive, but nevertheless, you can get as a put forward, as you can get a lot more information in terms of phenotype. So there's a phenotype richness by performing this whole-cell recording experiments in human neurons, derived from patient iPSCs. And you can do this, of course, you can try to connect the gene information to the phenotypes, so genotype- phenotype correlations.

Guo-li Ming: And over the years we have [inaudible] iPSCs with different mutations that have been mentioned in the first session, for example, the copy number variations 16P and others. And also another technology is combined the CRISPR technology genome editing with the iPSCs by reducing the genetic background variability. For example, recently, we have generated introduced mutation as agentic iPSCs lines of mutation at SETD1A. This is mentioned in Mark's talk one of the high-risk genes for schizophrenia identified recently. Of course, we can also generate iPSC lines from sporadic patients. At the old days, you can say these patients as sporadic, but I guess nowadays, you can also calculate the risk scores on this on those patients. And we have collected several IPC lines with the first episode psychosis and the age and sex matched  controls. And we can do similar asset functional assets in terms of synaptic development. And this is basically using the cultures that specifically differentiate IPSCs into excited or in neurons. Of course, we can make the system more complicated because we also have the ability to differentiate this IPSCs into GABAergic interneurons with high efficiency and to make a more complex circuitry. What you can do is to differentiate these excitatory and inhibitory neurons separately and mixed them.

Guo-li Ming: And the better you can actually mix and match this different type of cells, mix and a matched fashion. For example, you can have excitatory neurons from the control, well, have the inhibitory neurons from the patients with mutations. And with that, you can actually isolate the differential contribution from the excitatory neurons and inhibitory neurons. And to do this in a more-- in a higher throughput fashion, we use the MEA system, which is the multielectrode array system. Of course, this is an extrasolar recording, essentially what you're recording our spike information from the neurons, and just showing you some examples while we do this type of mix and match experiments, you can see that, well, on the controls conditions the excitatory neurons and inhibitory  neurons mix, the cultures generate very synchronized spike activities. And this is much more reduced if you use the excitatory control, but disease, the inhibitory neurons. And you can see that while you have the controlling inhibitory neurons, you can still have some activities. But what do you use is that the synchronization. And if you have neurons with mutations for both excitory and inhibitory, not only you lost a lot of the activities, you also lose the synchrony indicating different the contribution of excitatory neurons  or neurons and inhibitory neurons.

Guo-li Ming: And finally, another way to look at the functionality that in a high throughput manner is to do the calcium imaging, which has also been mentioned earlier in earliest session. So this is something that can be done, for example, by the highest throughput imaging system, we can perform large-scale calcium imaging at the single-cell level, and the calcium responses can be calculated at the single-cell level. This is also amenable for high throughput drug screening. Just to show you some examples, for example, KCl elicits activity of all neurons. However, if you look at the glutamate response in control and neurons with mutations, they show differential activities in response to glutamate. And we can task the responsiveness high throughput, for example, by applying different stimulants, al neutrophin, potassium, and MTA glutamate model, which are either excitatory agonist, glutamate receptor agonist or GABA receptor agonist. So you can look at a differential responses calcium responses in different types of neurons. So in general, there are different types of functional assays from static to  temporal regulation at a different age. You can do either whole cell recordings, or you can do extracellular multielectrode recordings, or you can do high throughput calcium imaging, each with its advantage and the disadvantages in terms of their scalability and the throughput. But these are all the systems, I think in combination, we can get a lot of the information from this human cell model systems. And with that, I'd to just conclude and thank everyone for paying attention.

Gavin Rumbaugh: Thank you, [inaudible]. That was a wonderful talk. And I think in the interest of time, we're going to have the discussion at the end like the previous session. So we'll move on to Bennett’s talk next, which is measuring neural network activities in brain organoid models of human disease.

Bennett Novitch: Okay. Let's get my little timer set up here so I can keep tracking myself. So I'd to first thank the organizers for allowing me to participate in this. So the question that my lab is interested in and also kind of touching on some of the things that came up in the last sexual discussion and also [inaudible] talk. The thing that we're particularly interested in is trying to get to this question of neural circuits. So over hundreds of years of research, we come to realize that the brain-- understanding the brain in total is really important to understanding disease processes. But I think increasingly, we are coming to understand that really, to understand gene impact on this, we really need to be able to focus on the circuits. And so the real question and challenge that we have is how do we actually make functional circuits that can be done at scale, and with the depth of complexity that we would use to model disease. Another consideration is that we can do this in using animal systems like mice, and we've been doing this of course now for decades, and we've made a lot of progress in that regard. However, a mouse brain is not the same as the human brain, I think as we've gotten to know more about the neuroanatomy of human versus mice. We know that there are some specializations, and also the genes that impact disease tend to be selective to human brain structures and cell types in some cases. So where we're coming in is trying to-- coming up with ways in which we can use the power of stem cell technologies to be able to bridge this gap and to have something that we can access, something where we can generate tissue and then see whether or not we can form functional connections.

Bennett Novitch: Now we're not yet ready to kind of jump from stem cell to the adult brain, but where we're making a lot of progress is actually taking the stem cells and turning them into a fetal brain like-structures, which are called organoids, which we're going to talk about in a second. I just want to also just introduce the circuit that we're studying in most cases, which is the cerebral cortex, and it has both excitatory and inhibitory neurons as Guo-li was just describing, and really the magic comes when you bring these two pieces together. So this is just some examples of the brain organoid technologies. My lab and many other labs have developed these technologies where we can do this now with some degree of reproducibility. And if you zoom in at the bulges that are coming off of one of these organoids, you can see-- oops, and I don't know why the timing is going off here. What I'm finding-- what you can see is that you get layered organization and a variety of cell types that are forming. If you do a comparison to the human brain, you can actually see that side by side that many of the same developmental markers and positions of the cells are approximated in the organoid as they would with the human brain. We also find form astrocytes in these organisms-- sorry, in these organoids - and I'm having a little problem with my clicking - so we can also assemble remake ganglionic eminence, organoids which have progenitor in differentiated interneuron markers. And the real magic comes when we bring these two pieces together. When we start to get the intermixing of the excitatory and inhibitory neuron subtypes, and we know that there's a variety of different types of excitatory and inhibitory interneurons that are forming. And we are also able to see the establishment of both excitatory and inhibitory synapses, thus sort of we believe are recapitulating some of the network architecture.

Bennett Novitch: The real proof, of course, is, are these actually functional? And so we've been going about this in two different ways. So one is applying some of the same sorts of technologies that are currently-- that are frequently used in rodent systems using GCaMP imaging and two-photon microscopy, where we can use viruses to deliver GCaMP to the organoids. And this allows us to measure the calcium transients with individual cells, rather than showing you the usual peak graphs, what I'm showing you here is a heat map that just indicates the peaks of activities across this with each horizontal line, being an individual neuron. And in doing this, we're able to make some measurements about the overall net network activity, such as frequency of spikes, the clustering in terms of the order in which these things fire, and so on. And we can measure the degree of synchronization. Just to show you that these are really very much dependent upon the inputs of things the inhibitory neurons, we've done experiments, proof of concept of adding in drugs that block GABA receptors. And what you can see is the same organoid treated with these drugs. It exhibits a much different level of network behavior punctuated by these massive hyper synchronies, which resembled seizure-like events. The other approach that we use is to measure local field potentials extracellularly kind of what Guo-li was talking about that rather than using MEA, which we find doesn't give us very much complexity in terms of the types of potentials that we were able to measure. We use microelectrode. So it's very much not the high throughput approach, but with this approach, we're able to probe into the depth of the organoids where we know that there's rich cell mixtures and what we're able to measure are oscillatory rhythms. And what's I think important to recognize about the oscillatory rhythms is we know that the human brain - and actually any brain - complex brain has a variety of brainwave  rhythms. And the frequency rhythm of these rhythms actually is an indication of both the network complexity, but also the state of activation.

Bennett Novitch: And we also know that many diseases manifest in changes at the level of brainwave wave activity that you can measure with techniques such as EEG. And so it's quite satisfying to us, as we found, is that when we do these types of recordings, what we actually can find is the existence of multi-frequency brain rhythms within these organoids. This is showing here at just a spectrogram plot, which basically just shows the power of across frequency, and then something called the spectrum density plot, where we just put the power against the frequency here. And that these peaks of activities are just indications for the prominence of a given rhythm within the organoid. So where are we now? So where we are now is, we're trying to now apply this to look at how different disorders and in particular, individual gene mutations impact the network activities. And so this is just showing an example of-- I don't know why my keyboard is giving me trouble. Oops. But what we're able to find here is that this is modeling for Rett syndrome, which is a mutation in the MECP2 gene. And what I'm showing you here are two different patients that have different mutations. And then just looking at first, the calcium imaging plot. And one of the things I wanted to just point out is that we can see that there's changes in the network level of activity. In this one patient, we see these bouts of hyper synchronies. And this patient, we see that there's this an unusual repetitive firing behavior of individual neurons. In both cases, what we find-- and this is compared to the healthy control, and this one's an isogenic control for this patient in particular. But we also can see that at the oscillatory rhythm activities, they actually look kind of similar. The patients look kind of similar to one another, and very distinct from the controls. And we're punctuated here is a very shallow, but very fast rhythm, which actually manifests is what we term high high-frequency oscillations, which are kind of not represented here in these spectral density plots, because they're very transient - oops - and but you can see that there's a sort of an ablation of most of these lower frequency rhythms within the samples.

Bennett Novitch: I don't understand why this is happening. And so I just wanted to just sort of-- just bring this out to say that we also have been looking at some other disorders. This is a mutation in the CHD2 gene, which is associated with autism, intellectual disability, and also is frequently associated with epilepsy. And we can see that this has a spectrum of behavior, which is different than the controls, but yet also different from the Rett syndrome patients. And then this is an example of a very severe form of epilepsy that's associated with an SCN8A, sodium channel mutation. And this one actually looks quite a lot when the organoids, which we added by QQ, and we can see this really massive level of hyper synchrony. Actually, the scale here of the spiking behavior here it's an order of magnitude stronger than all these. And here we can actually see the high-frequency oscillation aspect of it, which is kind of the hallmark of a seizure really coming up here in the spatial spectral density plot. So to sort of just wrap up, where are we going with this, and what can we do with this? And so one of the first things we need to do is to really try to use these approaches to define the network profiles or signatures of different diseases, and to also to really focus on the reproducibility within a given disease, and also the differences between individual patients. We're also because this is brain tissue, we can actually use this as an approach to be able to identify genes and molecular pathways of interest and important start to actually conduct mechanistic studies to ask questions about how a given mutation or change in it in a pathway is changing the network architecture. And lastly, we're also able to use the system much as what we was talking about as an opportunity to screen for neuromodulatory drugs. So with that, I will just end, and just sort of just acknowledge the people in my group who have done all this organoid work. I really want to put up, we put up some [inaudible] and Ron Weil who spearheaded this work. We've been really working with a great team of collaborators at UCLA and elsewhere. Thank also my funding sources. Okay. I'll pass it off.

Thank you very much, and excellent talk. Next, I think on the agenda is Ryohei Yasuda across the street neighbor at Max Planck Florida Institute, and his talk is titled understanding intracellular signaling in neurons. Take it away, Ryohei. We can't hear you. It seems to be muted for us

Ryohei Yasuda: Do you hear me?

Gavin Rumbaugh: Yes. Great. Thank you.

Ryohei Yasuda: Great. Okay. So I'm going to talk about intracellular signaling in neurons, so we have been working on a understanding signaling transduction because this is one of the important activity in the neuron which basically represents slower signaling in neurons and as you perhaps know, neural circuits in the brain acts on different timescales and electrophysiological activity represents a fast neuronal activity across a millisecond timescale to seconds. And which of course activates neural circuits and are important for behavioral regulation. But at the same time, the source signaling regulated by intracellular signaling is important for folks operating in memory for behavior plasticity. So we would like to understand how this intracellular signaling controls behavior through the neural activity or regulation of the neural activity, so timescale.  And eventually this becomes important because a lot of diseases. mental diseases. are targeting to this intracellular signaling and eventually to changes in behavior. Okay, good. So one of the consequence of this r intracellular signaling is for example structural change in synapses like dendritic spine, structure plasticity, and so on and so forth. Now intracellular signaling or crossing a synapse is pretty complicated because we have like hundreds of signaling proteins expressed in  each synapse, and they interact with each other extensively. So basically, it is a very complicated signaling system packed into an extremely tiny volume of synapses.

Ryohei Yasuda: So we use  several techniques to solve this difficult problem of understanding intracellular signaling transduction. So one is to image activity of a signal transduction using a FRET/FLIM sensor expressing a neuron in slices and imaging dendritic spines that are stimulating the spine. So as you can see in this diagram here, basically this allows us to access the timescale of similar transduction from milliseconds to hours. So you can see actually the very short calcium signaling events actually activates a cascade of events or different enzymes. Like you can see PCK, Calcineurin, CA2 to define a protein to keep these proteins or neurotropic activity receptors and sunspots. And also it activates gene transcription. And another important aspect is that because neuron is extremely polarized and compartmentalized cells, we have we have a spatial regulation with a a signaling transduction. So we see that how signal transduction propagates from one spine to two different compartments of the neuron And eventually some signals actually regulates gene transcription. So this technique it's not high throughput but basically as you can see we have made out about 10 to 20 proteins. We and other groups made it about 10 to 20 proteins and this gave us a lot of insights into how signal transduction occurs in synapses and how it how it regulates synaptic plasticity.

Ryohei Yasuda: About another aspect we are very excited is the recent advance in vivo genome editing. So we and other groups have developed CRISPR-mediated genome editing in vivo to basically insert tags, for example here, I'll show you several proteins inserted with the tags so that we can determine the protein localization of endogenous proteins. We can also insert single point mutation. But what [inaudible] specification to that subset of a specific subset of neurons. And this allows us to do-- and this is what we have to be high throughput. It's not really high throughput, but basically  we can finish this experiment in few days on average. And basically just this could provide us with a lot of different information for localization of the proteins and also hopefully in the future we can make other dynamics  of endogenous p proteins which may be able to make other interactions using optical systems or biochemical work. And also we can assess of course the function of endogenous proteins by point mutation to a subset of neurons or something like drug photo sensitive domain to precisely address the function of the protein. Well, of course, we have some problems, but it's good enough to, for example, map the localization of proteins right now and scale up to hundreds of proteins probably. So signal transduction is pretty complicated. We here mapped the activity of different proteins measured so far. And so you can see here we have this synapse, nucleus and so forth.

Ryohei Yasuda: Signal basically propagates from different signaling circuits and different domains over time. And it is important to probably have to collaborate with computational biology to understand how this complicated network works to regulate the synaptic plasticity and behavior process. So that's pretty much it, thank you.

Gavin Rumbaugh: Thank you, Ryohei . Excellent. Great summary of your work. Thanks for that. And to move on, the next speaker in this session is Adam Cohen from Harvard University, and his title is potential high throughput technologies for majoring neuronal synaptic properties. Adam.

Adam Cohen: Thank you. Okay, so yeah, thank you to the organizers, for the invitation to join you all today. I'll tell you about some tools that were first developed in my lab for measuring the electrophysiological properties of neurons with  high throughput. These technologies have really been refined in the context of a company, which I'm one of the co-founders of called Q-State Biosciences. Basically, the challenge is to measure the changes in membrane potential of neurons with as high throughput as possible. And we're interested in this because this is sort of the core of where neural activity starts, and it's a signature-- it's the function of neurons, and it's a signature of - that integrates many different biochemical and molecular input. The challenge is that it's very hard to see the neural activity on its own. And so over the last few years, we've developed some protein-based voltage indicators. These are derived from some microbial rhodopsin proteins found in the Dead Sea. These proteins are excited by red light and they emit in the near-infrared part of the spectrum. And the fluorescence is very sensitive to the membrane voltage. You can co-express these voltage indicators with optogenetic actuators, so that you can stimulate neurons with light of one color and record the fluorescence with light of a different color. And there, here is a cartoon showing that a process, and here's some data. These are from neurons in culture where the blue represents flashes of light delivered to the neurons.

Adam Cohen: The red is the fluorescents and the black is a simultaneous high-speed patch clamp recording. And you can see the optical and electrical signals respond. So this gives us an optical interface that gives patch clamp-like data, but with the throughput of a microscopy technique. Here's an example recording where we're looking at some neurons in culture and stimulating them all in parallel. This movie is taking at 500 frames a second and slowed down about tenfold so we can see what's going on. And now you can see when we stimulate the cells with the blue light, which is everywhere, you can see the patterns of spikes from the sounds, and you can really see the details of the spikes and the subfascial voltages and basically everything going on in these cultures. So we've taken this platform and-- or this technology and turned it into a platform for very high throughput recordings from many thousands or tens of thousands of cells over the course of a day. Here's an example. These are from primary DRG neurons. And we've also done this with human iPS derived neurons, where we're looking at the excitability of sensory neurons under different modulators of different that mimic different painful conditions. These sensitize the neurons in different ways. And so we stimulate the cells with this pattern of blue light, and these are optical recordings of their firing. We find all the spikes, and here is data from an afternoon where the data are arranged and [inaudible]or graphed. So each row is one cell, and each point is one spike from one cell. And you can see over the course of the afternoon, we recorded from more than 15,000 neurons. Doing this with manual patch, clamp recording at one cell per hour would be something like 10 person years of work.

Adam Cohen: And so you can really do much higher throughput measurements here. Another important aspect of synaptic-- of neural function is to look at synaptic transmission. And so we've also developed some assays to study synaptic function. Using some genetic tricks, we express the channelrhodopsin in some of the cells. And the voltage indicator in other cells. By flashing light onto the culture, the channelrhodopsin expressing cells spike, but we don't see that activity. We only see the postsynaptic responses in the downstream cells. By using different genetic tricks to target the channel option either to excitatory or inhibitory neurons, we can pull out specifically the excitatory or inhibitory components of synaptic transmission. And this lets us now measure it in a very high throughput way the distributions of synaptic strength in a culture. Here's just one example of an application of this, where we were looking at a specific mutation in this case associated with epilepsy. And we saw a clear difference in the excitatory postsynaptic potentials associated with this mutation. And now over the years, we've looked at probably several hundred different human iPS-based disease lines and compare them to controls. And you can get very rich information through these high throughput phenotyping assays. And of course, you can also apply them to a genetically modified cell or to pharmacology to explore possible rescue experiments. Thank you.

Gavin Rumbaugh: Adam, fantastic. Thanks a lot, man. Great job staying on time. So I think now I think we have the final talk of this session, and so Sergiu Pasca, a co-moderator is also going to give a talk from Stanford University, and his title is Brain Organoids and Next Generation Self-Organizing Models to Study Development and Disease.

Sergiu Pasca: Thank you so much. It's really a great pleasure to be part of this very stimulating meeting today. And thank you again, Steven and Anne and Geetha and Lora for putting this together. So what I thought I'll do today is just very briefly highlight some of the approaches that we've been developing over the last few years and maybe highlight a few applications that I thought will be relevant for the discussions today. I think that one of the major challenges that we're facing - and I think this was brought up again a couple of times a day - is the fact that to a large extent, human brain development is inaccessible. And so by inaccessible, I mean, the many of the cellular processes that are taking place in the human brain have not really been studied at the molecular and cellular level. So that primarily applies for instance, for late stages of corticogenesis, if we were to think about the human cerebral cortex, so for instance, the generation of glial cells, astrocytes and oligodendrocytes most of the migratory processes, but maybe more importantly, a lot of long-range connectivity has been very difficult to study with human cells. So what we've been trying to do over the last years is really try to develop in isolation, of course, not in total assays, that would allow us to capture some of the cellular processes with human cells. So to do this, we've been developing a very straightforward method for deriving three-dimensional cultures that resemble specific brain regions, so brains region specific organoids primarily by leveraging instructing single small molecules and growth factors. And there are a number of advantages to these cultures versus even monolayer and vice versa. And I think that would be part of the conversations that we will have today as well. But one of the advantages is that the sculptures can actually be maintained for very long periods of time. So for instance, I'm showing you here an example of 140, 20 weeks old corticogenesis that shows the generation of both deep and superficial layer neurons, which are formed after 20 to 25 weeks of gestation in the human cerebral cortex.

Sergiu Pasca: Now, this long-term culture has also allowed the generation of astrocytes. And so for instance, you can see here, one of these human astrocytes at about 350 days or so in culture, which again, shows how one can recapitulate some of the morphological features of this [inaudible]… some of the transitions that they're undergoing, going from fetal to postnatal stages. Now, the protocols can also be tweaked very often to include other cell types. And I think that was other conversations. So for instance, one can also derive oligodendrocytes and some of these cultures that in some instances are capable of myelination. And something that we've been doing over the last few years quite systematically is to try to assess what is the stage of maturation of this culture. So there are a number of papers that we've been publishing on this, including some work that is unpublished, but one of the most interesting things that we've discovered both at the chromatin transcriptional level and functional level is this transition that happens at a very similar pace to human brain development. And somewhere around, we now know around 250 to 280 days, there's a slow transition into postnatal stages. And just to give you an example of one of these transitions in work that is unpublished with the Geschwind lab we discover, for instance, that the NR2B and NR2A  subunits are actually switching at around 280 days, not just in their gene expression, but in their protein expression and even functionally when you patch cells up to 500 days or so in culture. And in order to develop a disease model, we'd be spending a lot of time in assessing reproducibility and scalability of this work, both across a lot of lines, but also across a lot of experiments and making sure that the patterning is reliable. And we've been differentiating now over 120 lines for 150 days or so each and in a larger effort with the Geschwind lab over the last 5 or 6 years, we've been trying, for instance, to assess as part of a U01 NIMH grant, the potential convergence, and divergence of transcriptional pathways.

Sergiu Pasca: And we think that some of this scalability makes actually this long-term cultures-- enabled by this long-term culture as possible to start asking some of these questions. And I want to give you a very short example of how we've been using this large-scale differentiations to gain insights into these diseases. And this is a paper that is about to come out, but it's a model of one of this common copy number variants 22q11 deletion, which is obviously a very high risk for disease. And in this study, we've actually differentiated over 40 lines for 125 days each in large-scale differentiation. And we've since that differentiation are actually there are no defecting corticogenesis across patients and controls versus systematically, but when you actually started looking with a very simple calcium assay, you can actually identify a very robust decrease in the calcium amplitude. And then this was actually done in over 5,000 to 6,000 cells that we've imaged multiple options and goals. It turns out that this defect is actually caused by L-type calcium channels, but not by L-type calcium channels themselves, but actually by a change in the resting membrane potential that we assess by patching over 500 cells in this cultures and show that indeed what this defect is doing is actually changing the fraction of available calcium channels in patients versus controls, although the L-type calcium channels, and this is the CACNA1C, which I think was part of some of the conversations today it's actually in its function.

We also want on to identify the gene behind this, and it turns out that it's DGCR8in the region and essentially DGCR8 can recapitulate all the phenotypes types and can rescue all the phenotypes if it's put back. But just want to give you an example of how large-scale differentiations and physiology, in this case, with brute force physiology can be used to gain some insights into disease. Now that was about generating one brain region. So in a couple of minutes, I want to apply some of the approaches that I think are relevant, maybe not as scalable, but I think are relevant to some of these conversation of capturing the aspects of human brain development would previously be inaccessible. So the first one is this approach that we introduced many years ago now in which you generates this brain region separately. And then you put them together essentially at the bottom of an Epi tube to generate what we called an assembloid. And there would be many examples already, both from our group and for others of assembloids that have been developed, the classic one is dorsal versus ventral forebrain that allows you, for instance, to look at the migration of interneurons, which again happens in the second and third trimester and in humans actually up to the second year of life quite uniquely to the extent that we know, and use this to capture an interesting phenotype in a specific form of genetic form of autism called by again, a mutation in a calcium channel. And this is really in studying a very simple preparation, very  simple cells of interaction, but in work that we haven't published, but I wanted to highlight here, we've been trying to build more complicated circuits by assembling multiple brain regions. And so, for instance, this is the most of say advanced assembly that we've been building in which we build the corticospinal muscle circuit from three parts. And eventually, first generating a spinal cord human brain organoid that recapitulates a lot of the cell diversity in the spinal cord, including the generation of motor neurons, but also the presence of V1, V0 and also for cell types, also glial cells, including astrocytes and astrocytes. And of course the cell can be patched.

Sergiu Pasca: Now, these cells can actually be put together with the cortex and you see a long-range connection coming from the cortex into the spinal cord. And then you can use rabies virus to actually demonstrate that the cells that are projecting are not random, but they're primarily deep layer neurons, deep layer, five neurons, which are critical spinal neurons. Despite the lack of fuse, there is some specificity in this connectivity. Now to show that the spinal cord is functional, we'd been attaching it to a dissected mouse limb bud, where you can actually see that the spinal cord is immediately able to trigger contractions in the mice limb bud, and this is actually immediately blocked by the use of Curare. So in the last type of experience that we've been doing is trying to put all three together, so deriving them separately from stem cells and then putting together cortex, spinal cord, and 3D human muscle that is derived from either iPS cells or myoblasts. Then you get this large structure that we call a corticospinal muscle assembloid. It's rather large. It's about seven to eight millimeters in size. Although, we know we can keep them for about four months or so, it has beautiful ultrastructure, both in the spinal cord, as well as the muscle. But what I want to show is a series of assays that we've been developing one in particular, where you can actually infect with an opposite the cortex and then each image, for instance, calcium and contraction at the same time in the muscle part. And first, as you can see in this example that we're imaging just the muscle part of this large structure while stimulating optogenetically the cortex. And in general, and quite reliably, we can trigger a contract and calcium rise reliably every time we do this. And this has been done not just with opticogenetics, but also with glutamate uncaging and with various other controls experiment. And I thought this type of assembly would be interesting for thinking about some of the disease models that moving forward one would want to develop. And with this, I want to thank you and look forward to the discussions and other questions.

Gavin Rumbaugh: Thank you, Sergiu. Do you have the slide with the questions supplied in your presentation?

Sergiu Pasca: I did have it cause I thought that [crosstalk].

Geetha Senthil: We can project that shortly. Lora is going to.

Gavin Rumbaugh: Okay, great. So I think that that that's an outstanding session. And so I think based on what the panelists-- there they are, thank you. Based on what the panelists discussed today, and so the goal of this discussion is to kind of intersect with some of the questions that have been posed during the meeting planning. And so by reading the questions that were posed by NIMH and listening to the panel, it seems there's a kind of a crossroads of a question, which is what are the major-- what are the measures that are possible with current techniques versus what are the best measures that are most informed formative about neuropsychiatric disorders? And so, I want to pose that question. Basically, the first question is, how do we identify the most informative functional assays for scalable assays, right? And then understand how they depend on the experimental system. But I think the idea is to understand how the assays may inform basic biology versus how they may inform mechanism of neuropsychiatric disorder. So feel free to any of the panelists to chime in.

Bennett Novitch: Well, actually, I'll give you kind of a throw out, one thought. One thing that I think we can all say from the systems that we were describing is that they're really best suited at looking - such as the stem cell-derived ones or the organoids - at looking at the intrinsic network activities in terms of function and dysfunction. One of the things that I think that Sergiu is working towards is trying to build in some sort of a more of a stimulus-response type of pathway which presumably it could also be done through using opticogenetic approaches as well as direct physical stimulation things. And so I think that's one distinction from what we're able to do and what we'd want to do.

Sergiu Pasca: I wanted to engage Guo-li because she highlighted this very interesting approach at deriving this GABAergic and glutamatergic separately, but of course, she's also very well known for pioneering brain organized approaches. So I was wondering if you could maybe highlight some of the advantages therebetween-- for instance, in terms of scalability or what you lose and what you gain by using a 2D versus a 3D approach. Because I think one of the major questions is the fifth of the assay to the question, right, and trying to understand disease.

Donald Arnold: Yes. So one of the advantages, I think, by combined inhibitory and excitatory neurons is that you can put them in the correct ratio, that's basically one advantage, right? And we know in organoids, while we can have both excitatory and inhibitory, especially in assembloids mode, however, the ratio, I believe, is still difficult to control. Right? So whether that resembles what's exactly happening in vivo and in human brains, I think that the mixture can have a better control on that aspect. But I think in terms of diversity of different cell types that we now know with this other single-cell analysis, there are hundreds of different types of excitatory neurons and hundreds of different types of inhibitor neurons. Whether that can be recapitulated better using the differentiation or in the organoids, I think that's still subject to the question and the better characterization I see in the future. Do you have anything to add? Have you looked into [inaudible]?

Sergiu Pasca: I mean, one of the questions I totally agree, right. It's much more difficult obviously to control some of these proportions in three dimensional cultures. I think one of the questions that I think still remains to be answered is whether there are any emergent properties of the cells and I mean, this in a very simple, biological way that arise by more physiological interactions. So for instance, is the function of an interneuron the same, whether you just make them together or does it have to migrate and receive some cues during that migratory, right? The same thing with, I guess, conductivity. And I think that is still not really fully understood, but we do know for instance, from work from Gordon Fishell and others, that interneurons do undergo the stepwise developmental progression. And so I guess that in that context, and also in the context that Ben was also mentioning for some of this activity arising, the question is to what extent, I guess, cell proportions or the right cell types are important for this activity to arise or for capturing specific features of the cells. I don't know if there are any other thoughts on this, because this is, I guess, an important aspect of fitting the assay, the functional assay to the question.

Guo-li Ming: Yeah. I thought this approach in using this physiology recordings in the organoids, could be very interesting especially those-- as to synchronize the activity at a different frequency, I think that's very important, but that remains to be seen, I guess, in the 2D culture systems.

Gavin Rumbaugh: Yeah. Yeah. So, yeah. Okay. I was just going to try to direct this more because I think the underlying theme here is that, so organoids are maybe more biologically informative, but they're not as scalable. And so 2D cultures perhaps could be more scalable. But so the question is even in 2D cultures, if you can mix different types of neurons together to form rudimentary circuits, to what extent does that serve as a model for high throughput screening and what would be the meaningful measures even there, right, that could perhaps be scaled to learn more about risk factors?

Guo-li Ming: Yeah, I agree. That's definitely something we need to do more, to look at more lines and more with different genetic variations. It's definitely, I think, there could be some conversion pathways, for example, the intracellular signaling underlying the pathology, even with different mutations, different variants, but that something I think we need to work on in the future to really dissect out what could be the potential convergence, right? Which presumably this is the most targetable druggable pathway, when you think about clinical applications.

Sergiu Pasca: And I think maybe the question of scalability is not necessarily of scaling and maybe it would be good to point this out. It's not necessarily in scaling up the cultures, which 3D cultures may be easily scalable, but it's I think mostly an issue of scalability of the assay itself, right. I'm sure that for instance, Adam's very elegant approach could be implemented in 3D cultures, I guess probably would require some more developments to really make it truly scalable in a 3D culture and capture the same amount of information in the most reliable way. I don't know maybe Adam, you can comment whether such an approach would be easily implemented in 3D.

Adam Cohen: Yeah. So of course, things are harder from an optical perspective when you're in three dimensions, because things are out of focus and you have light scatter and so on. In my lab, we've been now transitioning to doing a lot of measurements in vivo, in live mice, and have developed optics for trying to peer through some of the haze of the brain. And I think similar approaches could be applied to 3D culture, but there's definitely a hit in throughput just associated with the fact that you have some complex geometry and you have to find the cells in the three-dimensional space. I mean, those are solvable problems. You can try to automate that, but another trade off, which I think is harder to get around is the sort of more fundamental one between the complexity of your sample and your ability to measure with higher throughput. If you only have one cell type or two cell types in your culture, and it's a relatively homogeneous culture, then you can get sufficient statistics with a modest number of cells. If you have an organoid which has complex networks in it, that's a fundamentally complicated object in order to understand what's going on, you would have to look at a lot of cells and maybe many different parts of the organoid and that I think would be a challenge for throughput.

Sergiu Pasca: Yeah, those are great. Great, that's great point. So I wonder whether we should move to the next point here about synaptic function in some of this models and what is really possible to measure today. I don't know whether Gavin, do you want to [crosstalk]--

Gavin Rumbaugh: Yeah, so I feel the first two questions are kind of related, right. And so this is just what struck me as I watched this as someone who follows the fields of human neurons with obvious interest because of its clear importance, especially in understanding developmental biology of the brain. But so I'm stuck with this idea of, you know, what measures are possible, but what are the best majors, right. So I think we have to ask ourselves as a field, our synapse - is basic setups measurements is that the right measure, right? And if so, what are the evidence that a change in synapse that we can measure? And Rick alluded to this earlier what assay is that, right? What type of reliability does that measure have and what does it actually tell us about neuropsychiatric disorders? So it'd be nice if some other people maybe feel free to chime in on these types of questions.

Donald Arnold: So I just wanted to mention something about these recombinant probes, which are so valuable and scalable. Adam has mentioned that, and so has Ryohei, I just wanted to mention that prompts that were developed in my lab, they called finger proteins and they label endogenous synaptic proteins. So you can express these either using plasmids are AAVs and they give you a real-time readout of the distribution of excitatory and inhibitory synapses. So the very basic-- the most basic measure of synapses is where are they, How big are they, and what is their distribution? And this allows you to do that.

Ryohei Yasuda: Yeah. So that's why I think my bias is that protein distribution or localization probably tell us about the function of the synapses and the excitatory neurons. So these are probably relatively  scalable and [inaudible] of synaptic function.

Adam Cohen: I'd to speak about a little bit about Gavin's question about what are the right things to measure the best things to measure. And maybe this will reveal my bias, but I think that there's a lot of merit to trying to pick things that are sort of stable and insensitive to the specific details of the neuromodulatory environment and the activity, because we know even with the most realistic organoids you could post possibly hope to make, we're still a very long way from a brain, and the emergent dynamics, things the network activity can often be so sensitive to these hard to control factors that even if you see an effect there, and you might see a robust phenotype, you're back to the problem of having something, which might be very hard to interpret. And whereas, if you have a rich system an organoid or a more simplified one picking measures monosynaptic transmission, or intrinsic excitability, or where are the proteins and what's their molecular state are things that are probably more likely both to be robust and informative about the underlying biology.

Guo-li Ming: Yeah, I agree.

Gavin Rumbaugh: I think that's an excellent point. And so back to Don's comment in Ryohei’s comments, so if you have of the way to label stably, the location of the GABAergic and glutamatergic system and two colors, well, I think one thing the field definitely agrees on is a microcircuit is fundamental to cortical function, right? And so having an imbalance of those anatomical proxies of synapses of two different types of synapses, I think it would be very meaningful.

Donald Arnold: Yeah, I think it would be particularly useful for - as I said - for the organoid and for the iPSC cultures, because, yeah, for exactly the reasons you said.

Ryohei Yasuda: I also want to add actually probably ion channels because that's pretty much the functional and there are tons of different channels on the neuron and probably they are they are different in disease states.

Bennett Novitch: Yeah. I'd to also just like to agree with what Don is saying, having these sort of better tools to measure synaptic densities, because we've tried to do it in the organoid using, just antibody saving for pre and post synaptic markers. And when you're looking at a three-dimensional object it's, you know, it’s quite a mess, and we need to have that sort of isolated cell resolution to be able to sort of see what you can-- what we're used to seeing in a 2D setting. So one of the things we observed in our Rett syndrome models, where I was showing you the hyperexcitability is that there was a change in the ratios of excitatory to inhibitory synapses. And so I'd to be able to quantify that more. And so I think that it's these sort of tools, and I think I agree with that, looking other ion channels, etc., are going to be really useful.

Gavin Rumbaugh: And Dan's got a question and I'll call him in one second, but just as a way to kind of daisy chain, this into other questions that were posed is you need computational frameworks and scalability. And so small defined structures with high contrast are very easily picked it up by machine learning tools [inaudible] amorphous structures, right, which are complex and have less contrast. So I think there's a consensus building there for that. I think Dan Geschwind had his hand up? Dan?

Daniel Geschwind: Yeah. Hi. yeah, I was just thinking about something that usually when we're doing kind of functional studies, whether it's in an animal or in vitro, we're usually studying, we prefer to study large effect size mutations, de novo mutations, or things with large effects. But as we're coming to grips with the polygenicity of these neuropsychiatric disorders, we're kind of challenged to begin to think about small effect sizes and subtle effects of even, even groups of genetic mutations. Polymorphisms may have subtle effects on phenotypes. And I'm wondering what you guys think about this and whether that's feasible, it changes the notion from kind of an…to identify small effects in vitro, one is going to need to measure much larger numbers and have kind of much stabler systems perhaps than we're used to in biology. And I just wanted to throw that out there and get some reaction to that.

Sergiu Pasca: Yeah, that's a great point, if somebody wants to comment.

Guo-li Ming: Yeah. I think that's something where actually we are trying to do is to put the, for example, the SETD1A mutation on top of the generic background with high scores or with lower scores of the polygenetic score. And that way we can assess different ways. And one of the things I think in terms of the sensitivity and still seeing the physiology is one of the most sensitive ways to look into so we can see some differences when we do recordings from those cells.

Gavin Rumbaugh: Yeah. I think Rick, you can hear has his hand up. Rick, you have a question?

Richard Huganir: Yeah. So I think one of the things looking, at synapses in the organoids or in vivo is really powerful way to do this to analyze synaptic content and especially if you have two colors at least to look at inhibitory excitatory. But what we're running up against is the complexity of it. So we've made knock-in tagged AMPA receptors in vivo. So every excitatory synapse in the brain is labeled with a super [inaudible]. And so we can actually resolve the synapses in vivo and looking down through the layers of the cortex we're imaging several hundred thousand pumps at a time, and trying to look at dynamics with, with behavior and other in other disease states, etc. But we can't. It's very difficult to quantify very difficult to monitor over time, just because of the data is so rich. So we're trying to develop in collaboration with people at Hopkins, you know, sort of artificial intelligence or machine learning approaches to interpret this data. So I think also this is a huge need in the field. And I think we need to collaborate with people who can do that sort of thing.

Sergiu Pasca: I think also I think Helen made a very good point. I don't, Helen, you put a question here from panelists, but maybe you can bring up this issue that you--?

Yeah, sure. Thanks Sergiu. I mean, I think before we start thinking about particular assays and what are the best assays, I think a critical gap in our understanding right now is what are the relevant cell types, brain regions developmental epochs that we need to be looking at for each one of these disorders? I mean, for many of them, we still don't know the relevant cell type. And so I think that's what we need to get at first, before we start thinking about synapses or neuroprogenitor biology or everything in between. I obviously think that xenopus is a great way to start it because you can look at all brain regions through all of the development and figure out for multiple risk genes in parallel where you see phenotypes, because I think we should remember that each one of these genes is highly pleiotropic. And what we know about a gene is just the first thing that we know about a gene. And so I think having an ontologies is great, but I think we can all agree that they're limited. And so I think having strategies to identify the relevant cell types is going to be critical.

Gavin Rumbaugh: So, Helen, I completely agree with you. And so while this was, except for Ryohei and Adam, I mean, there was a very iPSC organoid heavy, but of course we can't forget the power and necessity of whole organisms, right? I mean, we alluded to that in the last session. I think I might've said something in that even for someone brought up Huntington's disease, right? I've heard other people talk about this. The gene's been known for three decades, and it's not agreed upon what the cellular manifestation of that is. And so now we're talking about all these different-- even the rare variants, right? There's a handful of them, two handfuls of them. And those of us who study them in depth, find out that, you know, while they're not actually, Sergiu, you I talked about this, what we thought were synapse, we thought were "synapse" genes are synaptic genes, but there are other genes they regulate differentiation clocks and things. And so in that-- and so I think that why we need to have these human cell models. We have to communicate and we have to remember that whole organisms are important for understanding relevant circuits and relevant brain areas.

Helen Willsey: I couldn't agree more, but I think there could be a wonderful strategy going forward of frog and fish people working hand in hand with in vitro cell culture models, to be able to leverage the high throughput and total brain advantages of these in vivo models to pinpoint time points, brain regions, cell types that are relevant, and then translate that into more human derived cell culture models, to be able to leverage both of them for their strengths and deal with the weaknesses, both take what you learn, cell culture, bring it back into the frog, make sure in vivo, you still see those same effects when you have all of the cell types there, right? Once you have microglia and astrocytes and everybody having their own choreography. So I think being able to flip back and forth, I think could be a really productive strategy going forward to kind of deal with the limitations and strengths of all the systems.

Gavin Rumbaugh: I agree. I think we have lots of hands up now. This is great. So slow rolling, but it's gone. So I think Steve you had a question?

Steven McCarroll: Well, I just wanted to amplify Helen's point, which I really agree with. We really don't know what the key cell populations are. To put another way, the key cellular contexts are in most of these disorders. And so the best biological systems in the world won't help you if you don't know what you're trying to model and what the right experiment to do is. And, I think it's just-- it's important emphasis that that's still an unsolved problem. And we can't even take for granted that it's a cortical cell type. I mean, clearly the cortex is impaired in schizophrenia, but that doesn't mean that that's where the genetics of the illness is primarily acting. And that's just still a big unknown. We could make the best cortical cell models in the world, and they might still might not be the right models. And so that's very much, I think, still unfinished business at the interface of human genetics and single cell biology. And we're fortunate to have Ed Lein on this panel. And I just, I just wanted to invite Ed to comment on any of these kinds of questions also.

Ed Lein: Thanks, Steve. Well, I mean, I'd like to very much amplify what Helen was saying and others. I mean, I think it's very difficult to model a system that you don't understand very well. I think unfortunately, there are some very effective methodologies and some big coordinated efforts now to try to start to understand that better. And I think the more that we understand the basics of how the system is organized and its cellular components and the genes that are responsible for the properties of those cells, the more you'll be able to move from this sort of genetic axis to being able to predict where that locus of action likely is. Also, those technology has also worked well across various species and can be aligned across species to understand what is actually conserved in different model organisms like mouse. So I'm very optimistic about the--

Yeah, I can't right now [crosstalk]. So I'll come down in a second. Thanks.

Ed Lein: I think that was not part of the conversation [laughter]. But anyway, I think it's really useful to merge these efforts and I think taking advantage of the efforts, for example, the cell census work to really be a key component of trying to make this movement from the genetics to effective model organisms to start to understand the mechanism is really important.

Gavin Rumbaugh: I think we have a couple more hands up to see if these points are further amplified. I think Ella your hands have been up for a while.

Ela Knapik: Yeah. So I'd like to add points. I very much liked this discussion to start with. I would like to add the point about extracellular matrix because extracellular matrix is something so unique that we can only study in vivo. And even if we generate it in the dish, it's not really representative of the native environment in which the architecture of the brain sets place. And many of the neuropsychiatric disorders and early developmental deficits that have been identified, they actually affect either the matrix proteins, the cargo, or the secretory machinery that plays that matrix in. So then the synapses conform and everything starts jamming the usual way. So other model organisms and I study zebrafish and human cell culture, we have learned that working hand in hand learning cell biology in the cell and sort of a larger architecture in the model organism it's very productive. And, the guiding light to all of it is actually human phenome because as it has been already mentioned most genes are pleiotropic, they have many functions and we only know a tip of the iceberg for most of them. So looking at the human phenome in TWAS type studies can really give us a list where we should be looking and what impact that has on biology and which part of this biology actually pertains to neurological and neuropsychiatric disorder. So I think this is an area of integration that we should be looking into. Thank you.

Gavin Rumbaugh: Great. I think we have only a minute or two left. We still have three hands up. So should we stop now or should we let these other three hands go? I'm talking to the organizers here.

Alexander Arguello: We can take the last questions with hands up.

Gavin Rumbaugh: Okay. So I'll just start from the bottom here. Trey?

Trey Ideker: Yeah. And that was me talking to my son earlier. So just to keep it quick, because I realized we're out of time. This is a point about the ontologies discussion from a few moments ago, which I totally agree with that the ontologies are limited. And I think I would argue that they're limited in a variety of reasons, but they're not fundamentally flawed in terms of ontologies, you know, sets of sets of genes being one of the right approaches here to bring in prior knowledge. I think the challenge is that prior knowledge is often formulated by limited teams, curating literature over a long period of time. I understand most human and mouse studies haven't even been curated by the GO curators and it's not their fault, it just doesn't scale. But I think a route forward is to formulate ontologies directly from data. These are just, you know, a GO term is a structure or function, a group of genes in coding a common structure or function in one or more cell types. And so ways of formulating these directly through some of the systems approaches we've been talking about today, I think provides a pretty powerful way to go. And is it likely that there's not one GO to rule them all as we have now, right? We have one tool for the unification of biology. That was the GO paper name originally back in 2001? No, I would argue that we're going to have custom GOs for different diseases and custom GOs for different dynamics and so on and so forth. And the only way to get there is to tie those directly to data. And essentially- and if not, take literature out of the equation, stop relying so much on curation of literature by limited teams.

Gavin Rumbaugh: That's a great point. And Michael?

Michael Hawrylycz: Yeah, just a quick point here that perhaps-- this also could play into the next section a little bit, but just a very introductory comment. It concerns, the last two points of the thing is that whereas we might be able to model and have infrastructure supporting, the understanding of these mechanisms that we're going after, as we've seen from the talks, the biology here is so pathologically complicated that I would argue that we've have far from the right computational frameworks and the entire model is inadequate to really understand and to facilitate the kind of discovery that we would like to do here. The concept of the publication model with data, having to look and the unavailability of data sharing mechanisms. We really could improve that dramatically, I think, to make better progress. But perhaps we can talk about that in the next session too.

Gavin Rumbaugh: I think that's an excellent segway, I think, into the second part of the meeting. Sergiu, is there anything else?

Sergiu Pasca: I think we should probably move into the summary part of the session. So that's Adam and--

Sergiu Pasca: Unless there are any other pressing issues that we should bring up if the organizers think there's anything else that we should discuss.

Alexander Arguello: We can sum up.

Gavin Rumbaugh: Great Randall and Adam.

Randall Peterson: Great. Adam, do you want to go, or do you want me to go first?

Adam Cohen: You can go first?

Randall Peterson: Alright. Well, I think the best way to summarize this session is that it was fantastic, incredible work being presented and amazing to me how fast things are evolving. I think there were at least sort of four themes and associated questions that jumped out to me at least. One was how successfully and rapidly we're moving beyond morphological descriptions to functional assessment of the cellular phenotypes. And I think the associated question is now that we have electrophysiological measurements functionally calcium imaging, other kinds of in-vivo reporters, is that enough? And are there specific disease states that those functional measures are going to be best at detecting and describing? In other words, maybe neurodevelopmental diseases might be more easily described with those functional readouts than something depression or paranoia, and what additional functional readouts do we need to be developing? Another big theme seemed to be, how do we combine cell types? What are the right cells to type? Several speakers talked about the importance of combining both excitatory and inhibitory cell types but is that enough? And how far do we need to go to really have a relevant model? Another theme was talking about sort of spatial relationships and how important they are, the relationships between different cells and do we have the methods that we need to detect and track those spatial relationships? And then finally, I think there was a lot of discussion around a maturation stage and whether we are currently able to look at the right maturation stages of these different cellular models. So I'm sure Adam has a lot of other interesting ideas, but those were four themes that jumped out to me.

Adam Cohen: Thanks Randy. And I'd to echo Randy's comment with appropriate modesty that I thought the speakers were excellent and that it was a very interesting panel, so thank you all. The sense I get from listening to the presentations and hearing the discussion is that the technical challenges of how to do various high throughput measures of neural function or structure are really not the main challenge. We have good microscopes. We have good electrical devices. People are amazing at cooking up different [inaudible] reporters. If there isn't the one you want, somebody can invent one, with relatively straightforward approaches, and that the challenges really are more around what to measure rather than how to measure it. And there's this discussion of, well, we don't even know which part of the brain we're supposed to be looking at. I think it gets at a critical challenge for the future. And so I think thinking-- the notion that there's going to be some high throughput gizmo, where you're going to put in a mutation and get out a phenotype. I think it's just overly simplistic and that we really have to treat the biology with appropriate deference and appreciate that each mutation is probably different and there are different regions of the brain. And we've been talking about, for instance, excitatory versus inhibitory neurons, but there are dozens of different types of inhibitory neurons and every particular synapse between each pair of neuron types in each brain region has its own attributes. And I mean, this might not be an overly uplifting conclusion. I mean, I think we can't ignore all of that richness and complexity, and we shouldn't be seduced by the fact that the sequencing can sort of span across all of these different areas of biology and everything comes out in a common language as soon as you get to proteins and cells I think the diversity becomes much more important and important to respect.

Gavin Rumbaugh: I think those are fantastic summaries. And since this is an NIMH-sponsored meeting, I'll just end it with one statement that is, I think-- and really echoing Adam's summary is that, I don't think it would be advantageous to abandon the wonderful support that NIMH has given to biology, right? So, I mean, I still think that there is a continued major investment in biology because I think we have so much more to learn because of all of the wonderful advances in genetics.

Geetha Senthil: Thank you, all. Thank you to all panelists. So everyone, including attendees we have to switch to Zoom meeting. So you should have gotten an email from Lora with the Zoom meeting information. So please use that link. This webinar will be disconnected. We'll be ending this webinar, so all will be kicked out. So please try to join Zoom meeting. Attendees, please stay on the meeting but you will not be able to join breakouts. You can stay in the meeting, main meeting room for 30 minutes. Our roundtable discussants and moderators of the breakout groups that report out at 3:30 pm to the main meeting room. So please use the Zoom meeting. Thank you again, everybody. And if you have any problems in connecting to that Zoom meeting, please reach out to Lora and Nikki, their emails are provided here.

Randall Peterson: Geetha, could you put the link right into the chat?

Geetha Senthil: Yeah, that's a good idea. Okay, so--

Sergiu Pasca: And the meeting starts in five minutes, right? Or six-minute?

Geetha Senthil: Yeah, I'll ask Lora to do that because she has to find the right one. Lora, can you please post that for all attendees and the panelist in the chat box for Q&A wherever is possible? Okay. I think she will do it. We'll see you on Zoom meeting. Thank you all.

[silence]