Scientist Stories: Jonathan Weissman, Organizational Principles of Biological Systems

Um, so first I just want to, um, say I'm one of the 3 of the 4 organizers who is stuck at home and can't be there, but I've been watching online and it's been a fantastic meeting so far. So I'm really pleased with that. And I also want to say thank you to John for being literally the man on the ground running the meeting and all the hard work that you're doing on behalf of all the organizers. So, with that said, I want to introduce our next keynote speaker, Jonathan Weissman, who to this genomics crowd likely needs no introduction, but I'm going to do it anyway, in part because I'm such a huge fan. So, after almost 25 years at UCSF, Jonathan has recently moved east to my neck of the woods, which is great, and is now the Landon T. Clay Professor of Biology at MIT. A core member at the Whitehead Institute, and still a Howard Hughes Medical Institute investigator. So it's a homecoming of sorts for Jonathan because he got his undergraduate degree summa cum laude here at Harvard, his PhD at MIT, and then went on to do a postdoc at Yale. So in his lab, Jonathan investigates one of the main biological questions is how proteins fold into their correct shape and how misfolding impacts disease as well as normal physiology, while at the same time building innovative tools for exploring biological systems from— you guys know this, right— ribosome profiling, lineage tracing, to developing CRISPR-Cas gene editing to regulate gene expression in humans. So I think this is one of the most impressive aspects of Jonathan's research program is that he's tackling both fundamental questions in biology and also at the same time developing tools. And to be a leader in both of those arenas is incredibly impressive. So he's won a number of awards, again, too many to mention. Not only being a member of the National Academy, but winning the NAS Award for Scientific Discovery a few years ago, to most recently winning the Ira Herzkovitz Award from the Genetic Society of America. He is also mentoring the next generation of scientists, some of of whom are attending or presenting at this meeting. So with no further ado, Jonathan.

Thank you so much, Hopi, for the kind introduction.

Can you hear me all?

Yeah, great. And it's wonderful to be back here. So this is very much a homecoming for me. I've been to Cold Spring Harbor many, many times. Some of you— most of you would not know this. My first time here was, I think, when I was about 6 years old, coming with my dad to meetings. So I won't even mention how many years it's been. I took a little gap between high school and college and came back in graduate school, and I've been here multiple times a year, but this is the It's the first time back for me since 2019, since the CRISPR meeting, and that's a great chance for me to do a plug that we're going to be having—

I'm not sure which one—

the annual CRISPR meeting in August. So anyone who wants to know how to edit these genomes, come there. So I'm going to tell you today about some of our work, both developing new tools for understanding the information encoded in genomes and how it's expressed in space and time, and a few vignettes that illustrate how we can use these tools to uncover biological principles in a hopefully systematic way. So it's also a great year to have the biology of genomes. The human genome is coming of age. It was 21 years ago that the first draft of the human genome was published, and just last month we got the complete sequence of a human genome, the telomere-to-telomere nucleotide sequence. Although if you actually look in the paper, it was missing the Y chromosome, so that was either a little bit of poetic license or it was published— publication day, a wry commentary on the humanity of the male half of our species. But the Y chromosome will fall. And so in some sense, this is the end of an era. We really know to the nucleotide, at least for one genome, and we certainly will for many more, what the nucleotide sequence of these genomes are. By contrast, we're at the very infancy of understanding the information encoded in these genomes and how they're expressed. But a critical tool in deciphering this has been the CRISPR-Cas technology, and what's What's so important about this is that now, even in complex genomes like yours and mine, it allows us to turn up, turn down, knock out, and alter in increasingly sophisticated ways any gene or any sets of genes.

And these, in particular, have been combined with pooled screening technologies where you do things like make a lentiviral—

what's now become sort of the standard approach of making a lentiviral library with a different guide targeting a different element in the genome.

And then transfect these so each cell gets typically only a single modification.

You can then put a selective pressure, so for example, look for resistance to chemotherapeutic agent, and at the end of the experiment, count the presence of the guides in the untreated and the treated and try to understand from that, understand how altering a gene impacts the selective pressure. And then, in sort of the next level of sophistication is you build a reporter, so intricately build a reporter that, for example, turns on a fluorescent protein in response to some biology you're interested in. So if you're interested, as we were, in understanding how the ER acts as a protein folding chamber, we hooked up GFP to the unfolded protein response in ER-specific stress response and then identified all the genes that when then when compromised, lead to dysregulation of folding in the ER as evidenced by induction of this response. So these many, many thousands of these screens have been done and an enormous amount of biology has come from them. But they really have two challenges. First, it takes a lot of prep time to build these reporter and really the art is in building the reporter or doing the selection so that it pinpoints the biology, the predefined biology that you're interested in. And then the second is, and really in a way probably more profound, is that at the end of the day you get these lists of genes. So you get most genes have no fitness effect and then some will either induce your reporter or have a fitness effect. But of course cells exist in a high-dimensional manifold and you could push a cell in two very different directions and when you do this dimensionality reduction through the lens of your fitness or your reporter, you've lost that information. There's no way to reverse it. Understand this process. This is not a controlled UMAP.

It's completely—

it's not designed to not lose information. And so what you end up with is a long list of genes, and now increasingly those genes are true. So unlike some of the early days when you would do RNAi-based screens and half of the time, or maybe more, they'd be due to off-target, you'd really believe that perturbing this gene gives you this phenotype. You just have no idea why. And then it's sort of a PhD thesis to figure out each gene. This motivated us and several other labs in what was a wonderful collaboration with Aviv Regev to develop a pooled screen that then took advantage of single-cell RNA sequencing to read out the transcriptional response in high-dimensional phenotype. In response to doing an alteration. And the basic idea was to, as with the pooled screens before, was to put a guide in each cell, but now you do a single-cell— make a single-cell RNA-seq prep.

And the trick was that—

so all the RNAs from individual cells were marked by a barcode, but in addition to getting the transcriptional response in each cell, you also were able to identify the perturbation, the guide RNA. And in this way, you could do a pooled screen that would map out where the perturbation was and where it was pushing you in this high-dimensional manifold. And we and others have now applied this to a very broad range of biological problems, including a very nice paper from Marco Hain, who's in my lab, is now in Vienna, studying, for example, how the host and the CMV genes alter the viral propagation. But in our studies and pretty much everyone else's, the general strategy was to, because of the scale that you could do Perturb-seq, was to do a screen ahead of time. So for example, do a screen for genes that affected CMV propagation and take the couple hundred best hits and then use the perturb-seq as sort of a high-resolution lens to look at those— how the list of those genes worked. The problems with that is you then again had to, ahead of time, predefine the biological question that you were getting at, so it was going to be a long slog to really get broadly at all the different types of biological problems.

And it—

you didn't know what you didn't know, so it was not sort of a great discovery tool an optimal discovery tool. And we reasoned that because what's so important when you do these perturb-seqs is to compare one gene to another one, that the cost of doing them grew linearly. If you do 200, it costs twice as much as 100, but the information grows as n squared because you go from 10,000 combinations to 40,000 combinations.

So it's really made more sense to do this at to scale. And Joseph Replogle and Britt Adamson—

Britt was a postdoc in the lab, is now an assistant professor at Princeton, and Joseph was an MD-PhD student who's recently gone back to the clinic—

built a lot of foundational work sort of over the beginning of the pandemic to allow us to scale this up. And this allowed Joseph and Ruben Saunders, another graduate student in the lab, to do to do genome-wide Perturb-seqs. So we did this in every expressed gene in the BCR-ABL leukemia cell line, K562, and then all the central genes in 4 other cell lines. At the sort of most basic level, what you get from this type of screen is you take all the cells that have given perturbations, so all the cells that are, say, knocking down Hsp70, and average them and then get the transcriptional response for each of these perturbations, for each of these genes that you turned off or on or mutated.

So for example, if you knock down Hsp70, you see predominantly that you'll get a heat shock response, or if you knock down a mitochondrial gene, you get predominantly an integrated stress response.

You can then in a—

do basically an unsupervised clustering on the gene, the expression responses, and from this derive all the different transcriptional profiles. And these, in a principled, unbiased way, undirected way, allows you to sort of rediscover the integrated stress response and the heat shock response and the unfolded protein response, as well as a number of responses for— many responses for which we don't have a name.

The other way to look at this, and I think even more informative, is to take for each genetic perturbation, so if I'm knocking down Hsp70 or Hsp40, take the transcriptional response as a high-resolution phenotype.

You might not understand anything why one gene went up and one gene went down when you perturbed that— perturbed the expression of Hsp70 and Hsp40, but if Hsp70 and Hsp40 cause a very similar transcriptional response, it's very likely that their loss is doing the same thing to the cell, and therefore that those two genes are acting in a similar pathway. In this way, it lets you group genes by their function, again, in an unbiased way that doesn't require you to any way presuppose you know what all the cell is doing.

And these give you really very rich information, and I'm going to I'll just spend a little time blowing up this— unpacking this a little bit.

So here we do the correlations of transcriptional responses of one perturbation to another.

Of course, on the diagonal it's all 1 because that's just correlation with itself. But the data, when you then cluster it— and we do this in a number of ways, but in some ways just a simple hierarchical clustering of heat maps is a nice way to display display it. So you now—

you can blow up a little region of this and you can see there's really a great deal of substructure, and even blow up a subregion of this and you see that there's again these sort of clusters of similarly acting genes.

Now when you put your names on them, it's really very satisfying.

So here you get the COP9 signalosome, this is methylation machinery, elongator complex, stargardt complex, and integrator complex.

And basically what we're able to do is recapitulate thousands of known relationships and predict, and then through work that Reuben Saunders led, validate new genes for functions in many different diverse processes like ribosome biogenesis, transcription, and mitochondrial respiration.

The other thing this lets you do is dissect large protein complexes. So in a really elegant example, the integrator, which is a multiprotein complex that's very extensively studied, So a metazone-specific transcription of vertebrate-specific transcription response element, see how that different proteins within the integrator cluster together but away from each other, and then in this way assign, mapping this on the structure, assign different functions to different submodules within this complex, and as well as identify a novel component that had missed characterization. So, in a sense, this was what we hoped for going in, that, you know, that we would be able to get expression networks, we'd be able to get gene functions. In some ways, it's similar to how one might use DepMap to— in coessentiality to cluster genes, or protein complex data to cluster genes. And that was very satisfying and I think, you know, a highly valuable endeavor. But what I'd like to spend the rest of this part of the talk about is something that I think is actually a much more profound use of this, which is that you can then go in, because you have this very rich sequencing information that's much more than just expressions that we've heard about, how you can look at introns and exons and variant, and I'll show you other types of information that you can get.

That you can then computationally—

so instead of building a reporter with a GFP and putting it in a cell, you can computationally build a reporter. And now in silico you can screen for that phenotype and get out the behavior, the genetic basis of it, and understand even something about the mechanism. And I'd like to— I'll tell two stories about this, one about aneuploidy and chromosome instability, and the other one about the logic by which a mitochondria communicates communicate with the nucleus. So I'll start with the chromosome instability, and I have to say this was a really a favorite of mine for a sad reason, which is my colleague and friend Angelica Amon really pioneered the study of aneuploidy, understanding that— basing on some fact that a hallmark of cancer is that their cells are aneuploid, they can't reverse this, so she was very interested in what were the cellular responses to aneuploidy.

And, uh, Joseph's brother John was a graduate student with Angelica who sadly, in October 2020, passed away from—

at a young age from cancer.

So it's nice to be able to, in a small way, carry on the baton on this.

So we got into this by being interested in the single-cell nature of this. So everything I've shown you has been ensemble averages. We take the 200 or so gene cells which have knocked down for each gene and do an average so we get a better signal. But we know that in addition to this ensemble averages that there can be cell-to-cell variability and that taking an average can obscure a lot of information. So I'll do one slide.

I'm not sure it's the most relevant, and I actually borrowed it from a former student, Sylvia Ruskin, who's now assistant professor at Harvard Med School, and, you know, I will say I'll do it for no other reason than it will be the one slide you'll remember from the whole talk.

So, in 1985, the average starting salary of a geography major from UNC was $130,000.

Think UNC is a good school.

It's 1985, $130,000 is a pretty good salary, so what's going on?

You see there were 9 8 students who made sort of just north of $10,000 and one that was $1 million, and that was Michael Jordan.

So ensemble averages can obscure the truth of what's going on.

So we— and I should say we is really Tom Norman, who has been our collaborator, is now assistant professor at— and so Tom and Britt Adamson developed developed the Perturb-seq method when they were in the lab, and Tom is now an assistant professor at Memorial Sloan-Kettering and did all this analysis as an independent investigator. So Tom developed basically an algorithm that took the single-cell nature of this and looked not what the overall impact of the—

average impact of the perturbation was, but what the variance was.

And this sort of shows you that you can have some genes where there's very little variant and other genes that the mean variance would be the same, but do that because they have a very different cell-to-cell variability.

You might think, you know, if you think about this for a moment, you think, well, that's probably because some guides are lousy and in some cells they knock down the gene and some cells they don't, and that causes the variability.

But because we're using CRISPR interference, which knocks down the expression of the gene, we can ask, are the cells that are not responding, not responding because the gene wasn't knocked down, or were the— and these, the gene was knocked down, or is the level of knockdown in these non-responding and responding cells the same and there's something more interesting biologically going on? And the answer is, for some cases, there's a problem with the reagent and we were able to exclude those.

But for the large majority, that this noise, this variability, represented something about the biology.

And for some of these,—

we're still trying to figure out why they cause a variable response, but for a subset of these, what we noticed is that they were genes that were important for chromosome segregation. And we hypothesized that the noise was because the underlying defect was that chromosomes were missegregating, but which chromosome in which cell missegregated was random. So there would be a different transcriptional response depending on whether you lost chromosome 6 or gained chromosome 16.

And fortunately, you might say, all right, that's great, but where do you go with this? Fortunately, we knew from work that had been done in the field, particularly this paper from 2014, that you could use single-cell RNA-seq data to study aneuploidy. The basic idea is that if you had a cell that's aneuploid and the genes are in a random order, you'd see some genes go up and some genes go down..

But if they had lost chromosome 2 and you order the genes by the chromosome, what you'll see is for other chromosomes there'll sort of be random things going up and down, but for chromosome 2, across the chromosome, you'll see a depletion of the genes by roughly a factor of 2 because they have only one copy of the gene instead of two.

So in this way, in a single-cell basis, we can judge the aneuploid state. And what we saw is in a non-targeting— if we don't target any gene, each one of these cells and the expression pattern, it's just sort of random noise. But when we knocked down TTK, one of those very noisy genes, this is a kinase involved in spindle pole checkpoint, now we see that some of the cells lost chromosome 1, others gained chromosome 1, other cells lost chromosome 2, et cetera.

So this really did sort of validate our hypothesis that the source of the noise was that although there was one single underlying molecular defect, the tendency to gain or lose chromosomes to become aneuploid, that which chromosome you gained or lost was relatively random, and the transcriptional response to gaining or losing different chromosomes was different. So when you look at the level of a single-cell RNA-seq, you see this high variability. This also, because—

and this is what I mean by we can now create this synthetic phenotype based on the propensity to gain and lose chromosomes. Something that would be very hard to design a single reporter for because you have all these different chromosomes and how would you know which ones you gained and lost? But in silico, this is very straightforward to do, and then we can now do a genome-wide screen for the propensity to gain or lose chromosomes. And what we found was that in K562, a highly aneuploid cancer cell line, and RPE1, a non-transformed immortalized nearly euploid cell, that there's very good correspondence between the genes, the SIN score, which measures basically the aneuploidy propensity. There's very good correspondence between genes that were important for modulating aneuploidy in RPE1 and K562, and then when we put these— the function of these, it's sort of a who's who in chromosomal chromosome segregation, things involved in spindle and centrioles and cell cycle regulation, centromeres and DNA repair, and then a few others. But importantly, only a subset of these showing that these were really the critical modulators of the ability to robustly segregate the chromosomes. So one reason that we're excited about this for the biology of aneuploidy is that now we can really catch cells in the act. Of losing or gaining individual chromosomes. So we can understand, you know, how— what is the immediate response, general generic response to being aneuploid, or chromosome-specific response? Are there global karyotype-dependent cellular stresses? Are there karyotypes that are favored in their formation or selected immediately after doing this?

All right, so I'm going to go on now to the second The second vignette, second and last vignette from this part of the talk, and this is our insights into the logic by which mitochondria signal stress and that the nucleus in the cell responds to this. So as you all know, mitochondria are endosymbionts.

They have a bacterial origin, but slowly over time they've lost the vast majority of their genes, so it's only 99 plus of the proteins that are localized to mitochondria, those genes have migrated into the nucleus and they're expressed in the nucleus. So they then are turned on by sort of normal transcription, translated into cytoplasm, and then delivered as a protein into mitochondria.

But there's still a handful, in the case of humans, 13 genes that are encoded by the mitochondrial genome translated by ribosomes inside of the mitochondria. And why mitochondria would lose 99% of the genes and not lose the last 13, which would have allowed them to give up having to make mitochondrial ribosomes and all this extra machinery, is really, I think, a major open question in cell biology.

So we've perturbed all the nuclear-encoded mitochondrial genes because we perturbed all the expressed genes, and what we now asked was as we turned up and turned down different functions in mitochondria, this must transmit a signal to the nucleus that's sensed, induced transcription, and responded to. So we wanted to understand how these different defects in mitochondria are sensed by nucleus and how the nucleus responds through gene expression.

And what we got was at first really rather disappointing. Unlike this, you know, intricate structure where the elongator and the COF9 signalosome gave clearly distinct transcriptional responses, basically to first approximation, any perturbation, whether it's to cardiolipins or to ribosomal proteins or to translocation into mitochondria, all of these very different response— different biochemical functions of mitochondria, when you knock them down, gave a single uniform transcriptional response.. That response was this stress pathway that I've talked about, the integrated stress response. So basically, to first approximation, all the nucleus is doing is saying, the mitochondria are not working, I'm gonna turn to ISR, which is a massive stress response that turns down translation and sort of—

but there doesn't seem to be any nuanced response. But as I said, in addition to the 1,000 or so— sorry, nucleon-coded mitochondrial genes, there's 13 protein-coding genes that are encoded in the human mitochondrial DNA.

And when you do the single-cell RNA-seq, you pick up these 13 mitochondrial genes, and although they're only 13, they're well expressed and we can measure them very well.

So we can now ask, what if we clustered the 1,000 or so nuclear-encoded mitochondrial genes, not on the nuclear response but on the mitochondrial response, so using only these 13 genes to cluster them. And now, as you see, there's much more structure. When you go into this, the structure makes a lot of sense.

So it very nicely separates, you know, the large subunit, the small subunit, import and quality control, ATP synthase, and the various different OXPHOS complexes. Axis.

And again, when you blow this up, you get this sort of intricate structure that recapitulates a lot of known biology and allowed us to identify a new component involved in biogenesis, the ATP synthase, which we and independently another group, John Walker's group, actually validated.

John Walker beat us by a couple months. We got the result results, and literally the next day we see John Walker's paper in PNAS. I have to say, being beat by John Walker, who of course won the Nobel Prize for his characterization of ATP synthase, I don't feel too badly. So this was a bit of a surprise, that you'd get these different expressions of these 13 mitochondrial genes.

And the reason why is the transcription of these genes is really kind of dull in that it's actually only 2 transcripts, one that encodes one of the 13, another that— and then 12 of these open reading frames are transcribed on a single transcript. So there's no way for differential transcription to alter the relative abundance of these different genes in a nuanced way in response to stress. However, there's elaborate processing, polyadenylation, and degradation of these different transcripts.

And so what this is really saying is that it's all of these differential expressions really has to be post-transcriptional.

And finally, I'll end this part with a little bit of speculation or synthesis. What we're seeing is there's a logic to the response to perturbations in mitochondria is that whether it's loss of ATP synthesis large subunit, etc. There's a uniform ISR activation nucleus, but there's then a stress-specific response of the mitochondrial genes. And this is interesting because it actually allows mitochondria to do something they wouldn't otherwise be able to do. So a typical cell may have 1,000 different mitochondria, and if one is stressed in one way and another stressed another way, they have to speak to the same nucleus. Nucleus. So this mitochondria would have no way of telling the nucleus to turn on some genes for me, this one say some genes for another one. But now basically what you have the potential of doing is having a generic nuclear response that sort of try to repair things, shut down translation, hope to survive, and then a mitochondrial, mitochondrial autonomous response by the internal transcription and translation of the mitochondrial genes that is specific to the stress that that mitochondria is feeling. So just a bit of speculation of how you can— I hope to open up a new piece of biology. So again, I hope to sort of give you a sense that you can really, in an unbiased principled way, start getting very rich biology.

And I have to say, we did not think ahead of time that we would be studying aneuploidy or mitochondrial stress response. Process.

I would like to have said we thought we were clever enough to have done this, but really we just let the data lead us this way, and we think— we hope that these will be, for us and many other people, a great resource, and that— and more generally, that these types of information-rich readouts— and you heard a beautiful story from Britt Adamson, who worked— did this repair-seq work to study DNA response, that these types of systematic, information-rich phenotypes will allow one to explore very broad biology and even do this in a principled, unsupervised way. And so with the last few minutes of the talk, I step back a little bit and talk about another, but actually quite related, area of the lab. So really what we're doing here, we and of course many others, are using single cell technology to allow us to have each cell be a test tube. And then the trick is to make each— and so now you can do a million experiments at once at single cell level.

And so then the trick is to sort of make each experiment as interesting as possible, and then to be able to read out in the cell about what the experiment was and what the response was.

So we did this by perturbing different genes in each cell and then looking at the response. But more broadly, we— and I should say many, several other labs, Jay Shindori in particular has been a pioneer in this— thought that you could go beyond this, that you could actually record in a live animal the history of the cell in its DNA, in a DNA scratchpad. And then have this recorded information be transcribed.

So then when you do your single-cell experiment, you'll be reading out the response that where a cell is, its present state, and the history that was recorded. In some sense, this is like a molecular version of the flight recorder on a plane where you are recording what was the pilot doing 10 minutes before a crash and what was fuel like and what was the altitude., and in that way have a much better understanding of the series of events that led to an accident.

So as I said, a number of labs have come up, I think, with similar approaches.

Many of these—

but basically you need something to write into DNA and you need something to read it out. A favorite of ours and several others is to use CRISPR-Cas9 to mark the DNA by insertions and deletions. In a controlled way. A key feature of the technology that we've been pioneering is that it has very high recording potential, so we could—

and we're going to be—

I'm going to talk about using this for phylogeny, for family trees, and we could in principle uniquely mark every cell in even a complex animal. The other is that we have a continuous in vivo recording over months, so as cells grow and divide, they continue to accrue these indels in these scratchpads, and this then allows you to reconstruct very rich phylogenetic trees. And then, finally, they're transcribed so they can be read out— you can read out the cell lineage in the transcriptome. And just give you an example of one of these phylogenies, this is a tumor that's been growing in a mouse for almost 2 months.— those 5,000 cells, each of these is a cell, and then they're ordered based on their phylogenetic relationship. And again, as you blow it up and then blow it up further, you see that there are really multiple different indels that were accruing over time that support each of these clades and subclades and sub-subclades. So you really can sort of see who is your sibling, who is your uncle, who is your 15th cousin once removed. We had a beautiful introduction to the power of this type of phylogenetic analysis from Pardis, who's used this to study both the evolution and the geographic spread of viruses. So you study geographic spread by looking at where variants are appearing and how they're spreading, but also, as illustrated by this phylogenetic tree of SARS-CoV-2, you see branches that expand, like the Delta branch and then later the Omicron. And the fact that every new place it goes, it expands more rapidly, tells you that the, the ability of this virus to spread more rapidly between individuals has changed, unless you've linked this change to the variant itself. So before you do any experiments— and so, Pardis, I'm a believer— before you do any experiments, if you've collected a an appropriate phylogenetic tree, you can be very confident that the Delta and then later the Omicron variant has changed the behavior of the virus to make it more readily spread. And basically, we reasoned that we could use the same type of logic to study tumors as they grow and metastasize. In a paper we published last year, we used this by a transplant to study how a tumor metastasized, how basically basically moves from one place in the body to another in the same way that you study how a virus is transmitted from one place to another. And then in more recent work, we've used this to study how you go from a KRAS P53 mutation to oncogenic mutations that we can introduce in an adult animal thanks to beautiful technology from Tyler Jacks' lab that allows you in individual cells in an adult lung to induce these mutations, and then we knew from their work that these then go over a period of months from hyperplasia to adenoma to adenocarcinoma, et cetera, in a way that very nicely recapitulated what the human tumors— human lung tumors with KRAS P53 mutations have. But now, in much more detail, we can follow the evolution as you go from different transcriptional to high plasticity states, to an outgrowth, sort of your Omicron and Delta branches coming out that let you have aggressive tumors, and to how these can then be the origin of the metastases. And we can overlay other oncogenic loss of tumor suppressor genes like LKB1 and APC and rerun the clock of evolution multiple times and see how changing the genetics of the tumor alters its evolutionary history. And in this way, we think we can reconstruct key distant and rare events in the tumor in increasing depth. So, just to finish, we hope we can now apply this to study where resistance comes from, to understand using spatial approaches the relationship with the stroma, and then to record other molecular events. Attack from the immune system or hypoxia. So with that, I'll finish and thank the people who did the work. I've had tremendous good fortune of having wonderful people work in the lab and wonderful collaborators. These are very multidisciplinary approaches and really takes multiple different people with different expertise to come together. So Joseph Replogle with Britt Adamson led the development of the expansion of the Perturb-seq approach and much of the data analysis of the large-scale Perturb-seq with Tom. And then his partner in crime in all of this is Reuben Saunders, who led much of the follow-up on this and is continuing this work. Michelle Shan, when she was a postdoctoral fellow— she's now assistant professor at Princeton with Alex Meissner's lab, and Zach Smith and Alex Meissner in Professor's lab developed the first lineage tracing to study mouse development. Jeff Quinn, in a wonderful collaboration with Trevor Bevona and Nir Yosef, did the metastasis work. And Dian and Matt Jones, in a terrific collaboration with Nir and Tyler, did the KRAS-P53 work.

So I'll end and take any questions. Okay, so thank you so much, Jonathan.

That was absolutely fantastic, and we already have questions, so in the middle. So you mentioned having done this perturbation across several cell types for the Perturb-seq of all genes. How similar were the responses by the different cell types, and do you think there's eventually a way that we're going to be able to translate information from the cell types you've done into other cell types as well?

The correlation—

so if two genes have a similar perturbation indicating they're doing similar function, that is very well— that's very well preserved. Not perfectly. Sometimes genes are doing a different thing in different cell types or different importance, like the moon laying function, for example. But in general, the rule is those are very However, the underlying transcriptional response that's driving that similarity can vary from some cell types to different cell types. And I should say, we've really— we've sort of analyzed the K562 and the RPE1 data. We're really sort of just digging into the other datasets to really do more of this comparative analysis.

Hi, Jonathan. That was a beautiful talk. I have a question. You might have alluded to this in the aneuploidy part of your talk. Your talk, that you should be able to infer the fitness cost of the different chromosome duplications, at least in the context of the knockdown. And what about— is there any evidence that chromosomes are gained and lost at different rates per chromosome, and can you—

do you think you can disentangle that? So I think we should very well be able to disentangle that. So we can either— there are drugs that you can use that would promote aneuploidy, or we can do a pulse knockdown because it's CRISPRi, it's reversible. So we'd like to sort of give a quick impulse that promotes gain or loss, and then within a cell division or two, look at the abundance of gain and loss of different chromosomes, the immediate transcriptional response, and then as you pass with time, how does that change? So that gives you the forward rate— I mean, you have to do the right time course, but in principle, you should be able to see the forward rate at which chromosomes are gained and lost, and then the selective pressure on the cells that have gained and lost those chromosomes.

[Speaker:JOSHUA_ELKINGTON] So I might ask a question. So when you have a CRISPR screen and you're looking at a stress response, the cells are very stressed, how do you disentangle the cell signaling to one another from the actual perturbation? And particularly in the mitochondria, where it could I can imagine these are very stressed and unhappy cells. So is some of the correlation, because at the nuclear genome level at least, related to the fact that there's just a lot of signaling going on and they're talking to one another and that's why there's less signal than you might have expected?

[Speaker:JOSEPH EISENBERG] So we are using— those were predominantly in K562, but we also saw similar things in RPE1s. But K562 are suspension cells, so they're sort— they're very rapidly becoming uncorrelated with their neighbors. And remember, the vast majority of cells don't have perturbations in the mitochondria, so the vast majority— and those cells are not responding. But this is predominantly about cell autonomous responses, and we sort of designed K562 as sort of our yeast because they're in suspension, and so their environment is averaged very rapidly. Really what I think is that this— these should now be done in a spatial— by spatial readouts. So now you can start to see if I perturb gene cell A, what is its neighbors doing, and get nearest neighbors effects. And so it's huge. So I think these datasets are perfectly capable of being collected.

Fully analyzing them is going to be a challenge.

Absolutely. Other questions in the room? So in the functional dissection of the mitochondrial biology, you mentioned that like taking the correlation across those like 13 genes is going to be, you know, informative to reveal the fine structure. Do you think in other contexts of biological phenomena, like not only computing the correlation across all the genes but focusing on subset of genes that captures the types of biology you are interested in? Going to be like a generally useful strategy to, you know, reveal the substructure?

[Speaker:CHRISTOPHER_MILLER] Yeah, so for sure, and beyond just the transcriptional responses. So, for example, in work I didn't talk about, one of the things we did was use this as a splicing. So we heard from Leo about that you can look at different variants, so we can now look at how perturbations change the abundance of different variants or globally change different splicing, or we can look at reactivation of retrotransposons. So basically, anything you can pull out of the sequencing data, you can now develop either in a hypothesis-oriented way, like we did with the aneuploidy or with the mitochondria, or even in a purely data-driven way, you can get phenotypes that are just by unsupervised clustering and then and you don't even know what the phenotype is, you just know it comes out as unsupervised learning approach, and then you can understand the genetics of that. And I should say, I've been really focusing on transcription readouts in part because that's what we can do well, and it's a nice rich readout. I alluded to the repair-seq that Britt Adamson has done, which was for DNA repair, looking at the variants of repair, was a much better readout for getting at that biology. And then on top of this, there's been beautiful work from Xiaowei Zhang and work I helped a bit out with Ron Vale and most prominently Paul Blaney doing pooled image-based screens. And that, I think, is going to be very useful revealing things that are sort of largely invisible from—

or hard to pull out from transcriptional responses. At the back there. I had a question about the scratchpad stuff you mentioned at the very end. So the analogy with the virus is interesting, because presumably in the virus, the rate at which these things are coming in is pretty constant. You have this kind of constant molecular clock introduction of these edits. How true is that in the scratchpad example? Presumably there's some kind of conditional dependence of the rate at which they occur based on what's already happened.

So I mean, we're engineering them into, you know, we hope neutral parts of the DNA. And what could vary is the expression of the Cas9. And that could make it go— that could make indels go faster or slower. And this gets to maybe— I'm not sure I'm entirely answering your question, but one of the things we'd really like to do is to have an absolute absolute sense of time, to actually sort of build a counter that would count every time a cell divided, or sort of a clock. And the idea is you would then know when something happened the way you know when the fire happened back when we had analog clocks, because you'd watch the melted minute hand, and if it melted at 2:10, that's when the fire happened. So we'd love to build a sense of a clock or time in these. You can also— there's some very nice work, including really elegant set of papers from Jay Shendure's lab that are on Bio Archive. He calls it a ticker tape approach, which lets you look at the order. So you don't know exactly when something happens, but you can say that A happened before B, and B happened before C.

I'm curious, based on what you've seen so far about tracking molecular events in cancer, from that data, have you— has something emerged on how you might be able to change treatments to mitigate resistance, for example, ahead of time?

Combination treatments, what have you seen from that? Yeah, so in the metastasis paper, we, in a sense, this was, we thought this was like our simplest system, but it was a nice proof of principle. We worked with the A549, a long passage cancer lung cell, and what we found was that there was heterogeneity, not in how rapidly they grew, but in the propensity for these different subpopulations to metastasize. We get transcriptional signature of that. We could show that this identified new genes that were driving tumor invasion in an in vitro model. So that's a way we can get at sort of the molecular nature. Going from that to how do we target that, of course, is its own challenge, but it's a start. I would not do that in A549s, to be honest. I'd much rather do this in the autochthonous model, Tyler Jackson's model, where you We are introducing in otherwise genetically wild-type mouse the KRAS and p53 mutations. So we are able to also see signatures that drive the tumor aggressiveness as a sort of equivalent of the Omicron and Delta variant. And when you take those signatures, you can see that in human patients who have high signatures of aggressive— that drive these expansions, they have a poorer prognosis.

Metastasis.

We're very interested, though, of going beyond that to try to understand sort of the signatures of metastasis, what the cells were doing, from which the metastasis came, what state they were in, and/or looking for the origin of resistance. So that's a long-winded way of saying I certainly hope so.

I don't have a great answer at the moment from the data we have. One final question. So going back to the neoploidy, so you showed like those those knockouts that are associated with high variance seem clear to be enriched for that type of explanation. But I was just curious, is that really the universal, you know, answer to those high variance responses, or do you have— you know, how frequently do you have, like, genes that, you know, you knockout, show that huge variation, and it doesn't seem to be explained by that?

[Speaker:JOSEPH EISENBERG] No, it's only a subset. So it's one that we could get a nice story on and, you know, understand, but then there were others that were clearly not working. Through that. And actually, even in the first— Britt and Tom's first Perturb-seq, they saw that when you knock down BIP, the HSP70 in the ER, that you were getting variable transcriptional responses. And, you know, we hypothesized that maybe whether there's three branches of the UPR, maybe if stochastically if one goes up first, it may take over. But we really have not disentangled what the origin of this variability is.

Excellent.

So please join me in thanking Jonathan again for a fantastic talk.