Skip to main content
Loading Explore guides…
Genetic Interaction Networks Mapping · Acedra
Home
Explore
Biology
Genetic Interaction Networks Mapping
Sign in to save
Chat
If YouTube asks you to sign in, open the video on YouTube.
Watch on YouTube
Reload player
Transcript
85 segments
Copy
Hide video
0:00
[BRENDA ANDREWS] Experience of expertise, which one doesn't normally get at these workshops and meetings. In any case, what I'd like to do today is tell you about our ongoing work to map genetic interaction networks. And the real goal here, like most of the work that we've been describing at this meeting, is to contribute to the mapping of the genotype, the phenotype relationship. And we want to use information we get from mapping networks and
0:24
model systems to help discover potentially principles that would allow us to more cogently map genetic networks in more complex organisms, such as human cell lines, for example. So I'll start by giving you an overview of our yeast genetic interaction network project, because it provides the essential foundation for unpublished work, where we started to explore a scaffold genetic interaction network in a human cell line, and hopefully I'll have time at the end, since we've been discussing
0:51
a lot of these things, and actually your talk brought up some unsupervised clustering methods, where we're trying to integrate our automated genetics platform with single cell imaging, so that instead of using a fitness-based readout, we can get multiparametric phenotypic profiles, and that'll allow, because it's single cell imaging, we can explore not only gene function, but penetrance of genetic mutations and pleiotropy, which Pascal was talking about in his talk. So we'll see how far I get.
1:17
I wouldn't-- I don't wanna dig into your time for being outside on this wonderful day. Anyway, I don't have to tell this audience about the importance of developing systematic methods to explore the genotype, the phenotype relationship, and really our interest here is exploring the role of genetic interactions in this relationship. And the big groundwork that we've been working on for well over 15 years now is to map a referenced genetic internet--
1:41
action network in the budding yeast using reagents that are available for doing systematic genetics. And this hasn't been discussed yet, but although we've had a couple of yeast talks the reason that we were able to start contemplating doing this in yeast is because of the availability of reagents for doing forward genetic screens, which includes an array of 5,000 mutant strains, Each deleted for a nonessential gene, so single mutants there, and then mutant arrays that allow us to query the
2:08
function of essential genes. And so our lab and others have made a variety of these collections. The one that we use in these projects contain hypomorphic or temperature-sensitive alleles of essential genes. So even this result though, that you can delete any one of 5,000 genes in yeast and you don't kill the cell, for example, illustrates that there is genetic redundancy here, and in order to begin to address that, we reason that we should be able to develop,
2:32
we should develop some method for automating yeast genetics so that we can use these arrays and make double mutant arrays, so then we can explore the difference in phenotype between the single mutants and the double mutants, and start to build genetic interaction networks. So, many years ago, we developed a method called synthetic genetic array or SGA analysis, which does automate yeast genetics, and it allows us to introduce any marked allele of a gene of interest into these
2:57
yeast mutant arrays, or any other mutant array that you have available. And so in the case of the project I'm describing today, we're using SGA to make double mutant arrays, which we then are gonna assess for unusual phenotypes, and specifically, we're assessing for genetic interactions, and this occurs when an allele of one gene combines with the allele of another gene to et-- generate an unexpected double mutant phenotype. And so we can use any, any phenotype in theory to
3:23
measure a genetic interaction, including sub-cellular morphology phenotypes, which I hope I'll get to later. But by far the easiest and most scalable phenotype to assay in yeast is cell growth. And we can do this very easily in an array-based format by measuring the colony size of single mutants simply by taking pictures of arrays and looking at them, and measuring the number of pixels associated with the colony, and then we can measure the colony size
3:48
of the double mutant, and then ask if that deviates from expected based on the single mutant phenotypes. Super easy, easy to scale and quantitative. So this is what we've been focusing on a lot. And getting back to your point, Therese, about non-additive effects, what we're testing here is a multiplicative model, where we expect that the double mutant will show both of the phenotypes associated with the single mutant. And so we expect, for example, if a single mutant reduces
4:13
fitness by 30% or so, and another single mutant by 30% 50%, if we combine both mutations in the same cell, we expect to see the multiplicative effect of these two single mutants. So we expect 35% the growth of wild type. And anything that deviates from this, we would score as a genetic interaction, as a so-called negative class, where the double mutant is growing less well than expected based on the single mutant phenotypes. The most dramatic example of that is synthetic lethality,
4:41
which we've heard people mention several times in the context of cancer, for example. But we can also score positive genetic interactions, also a very interesting class of genetic interactions, particularly from a human disease gene genetics perspective. So we can score positive interactions, including these dramatic interactions where the strain is growing better than the worst-growing single mutant.
5:06
So anyway, about six years ago now, we published a complete reference genetic interaction network for budding yeast where we tested most of the 6,000 genes in yeast for possible pairwise genetic interactions. Uh, this project ha-- has been led throughout by Michael Costanza, with senior research associate in our group, and Michael identified nearly a million interactions 550,000 negative interactions and 350,000 positive interactions, and this network spans about
5:33
90% of, of all yeast genes, so it's reasonably complete. Amongst the negative genetic interactions, there's about 10,000 synthetic lethal interactions, so these are double mutant combinations where the double mutant is dead. And again, this sort of defines genetic background-specific gene essentiality, which is something that's interesting to explore, for example, in other strains of yeast that we heard about from Andreas, where we might be able to ask
5:59
do we see differences in the central gene sets in different backgrou-- backgrounds, and can we define the genetic effects that underlie that dramatic phenotype? It's also an important consideration for people interesting in designing synthetic genomes. It's important to know what kinds of combinations you you can't make. What about genetic network properties? I motivated this by saying we wanted to find features of the network that we might be useful for exploring networks
6:24
in human cells, for example. And to discover these kinds of features, the way we display our data typically is to generate genetic interaction profiles for each of the query genes. So genetic interactions are rare, so most of the genes don't interact with your query, but you'll see some positive and negative interactions associated with each gene mutation, and that's its genetic interaction profile. And then we simply group or cluster those genetic
6:47
interaction profiles together and look for high-correlation modules in our data. So we're looking for genes that when mutated share genetic interactions in common. And we reason that this would likely identify genes who are performing related functions. That was the sort of major hypothesis going in. And we can display these data in a number of ways, but we just simply use a network display where the nodes are the genes that we've taken through the SGA process,
7:15
and they're connected by edges that reflect the degree to which they share genetic interactions in common. So things that are real close together on the network share a lot of genetic interactions in common, and you can dial that up and down depending on how stringent you wanna be. So this is a picture of the global all-by-all similarity network for the budding yeast, looking at double mutants in a fitness-based readout. And I think without annotation, you can clearly see there are
7:40
clusters of correlated genetic interaction profiles on the network, and then we can functionally annotate it using various means. And for example here, we're functionally annotating the work using gene ontology, bioprocess terms, and we can see about 500 terms that are enriched on the network, and they tend to map to 17 unique network regions or clusters on the diagram. And you can't see this, but these correspond to the major bioprocesses within the eukaryotic cell,
8:09
such as ribosome biogenesis, vesicle trafficking, cytokinesis, peroxisome function, respire data vox-- data phosphorylation, et cetera. So we're clearly able to, using this kind of genetic interaction profile approach, identify bioprocesses within the yeast cell, and we reason that the position and connectivity on these networks would likely help us to predict, for example, the function of uncharacterized genes. We can also use other functional annotation standards to explore the network.
8:36
For example, here, we're looking at a sub-part of the network, only now we're using a protein complex standard, and we're demanding that these genetic interaction profiles are more similar between the genes. And in this case, we clearly start to see major protein complexes that we know were involved in, for example, transcription and DNA damage and repair. So there's a very high correlation between known protein complexes with correlated genetic interaction profiles between these mutants.
9:02
And then you can demand less correlation and use, let's say, protein localization standard to see if there's any enrichment on the network, and indeed, we see that. So for example, we can annotate with a standard that was based on GFP fusion protein localization in yeast, and we see regions of the network that correspond to the major subcellular compartments, such as nucleus, nuclear periphery, et cetera. So to me this was kind of remarkable at the time,
9:28
and what Michael uh, dubbed this was a functional wiring diagram of the cell or a hierarchical model of cell function, where we can make these networks and then functionally annotate them to re-- reveal different layers or levels of functional meat-- uh, information, like cell compartments, bioprocess, and pathways. Not pathways, protein complexes and sub-modules. Anyway, so this is sort of the background. And so what we wanted to do was also look within these,
9:56
this information to see if we can see any properties of the network that we could use to start to map the human cell. And this is one that I like to show because it's turns out to be a motif that we see recurring in the human network that we're working on. And that is that when we look at genes that when that code components of known and well-conserved protein complexes, the 19S proteasome here, and we're looking at both the positive and negative genetic interactions that this components
10:22
of the proteasome have, we see very coherent interactions with other conserved protein complexes, such as the DNA replication factor C, the COMPASS complex, APC trap, et cetera. And we sort of say coherent because they tend to be all-negative or all-positive genetic interactions that we see between components of the proteasome and other complexes, and we see this kind of pattern duplicated when we look at other protein complexes throughout the network.
10:47
So that was kind of a striking motif, which we thought that we could try to use to pick query genes that we might use to map a scaffold network in human cells. So that's one thing I'll come back to in a bit. And of course, we can start to look at networks in human cells because we can now do genome-scale genome editing screens, where we can make loss-of-function mutations simply by using libraries of guide RNAs to target all the genes, for example, in the human genome.
11:12
And we've been using the we've been doing this for the last few years in collaboration with our colleague in the Donnelly Center, Jason Moffat, who has set up a pipeline for doing these CRISPR-Cas-based screens. And we have to choose a reference cell line, and we could argue about this forever, but the cell line that we've been using is the HAP1 cell line that was introduced to you earlier this morning.
11:33
And so the challenge is massive, though. Of course, I've just told you about yeast, a project that took us 15 years, where we did an all-by-all map of the 6,000 genes double genes and their double mutants. In humans, the scope of the problem is much larger, and so we, again, this is what's motivating us to kind of try to go back and forth from yeast to human cells to take a logical approach to approaching this problem, maximize the amount of information we're gonna get
11:58
from each screen that we do. It's time-consuming and expensive. So again, this was to apply the rules learned from yeast to efficiently map interactions for key human genes. And so the most important thing predictor of whether or not you're a network hub, a genetic network hub, that is you're gonna have lots of genetic interactions, is whether or not the single mutant has a fitness defect. So essential genes are clear hubs on the genetic interaction network.
12:24
They're a little bit hard, harder to study as queries. In yeast, we have reagents. In human cells, not so much as queries. And also, non-essential genes that have a fitness defect. So if a single mutant already has a fitness defect to start with, it will give you a lot more genetic interactions than one that doesn't, at least using a fitness-based readout. So the first thing we wanted to do was discover these non-essential fitness genes in the HAP1 cell model
12:46
so that we could then make queries and start doing genetic interaction screening. And so to make that catalog Katy Chan uh, o-- s-- supervised a technical team in the lab that simply did about 40 different genome-wide screens in the HAP1 parental cell line, looking and then just following the dropout of guides as we outgrow these cells. And so we, of course, see an, a large number we see a few thousand, I'll show you this in a second, genes that we d--
13:11
define as essential because they drop out very, quickly from the population, and then a number of different genes that are nonessential sick mutants, and we can quantify that. And then of course, some things that don't change at all, so we wouldn't choose those to do our screens. So having done this many, many times and controlling for guide quality and all this stuff, we see about 1,568 essential genes in the human haploid cell line. This is co-- uh, comparable to the you know, the number of
13:37
essential genes people see in a bunch of cancer cell lines, for example, in the Cancer Dep Map. And there's a very large overlap between the core essential gene sets that people see in all of these different cell lines, and a couple thousand of nonessential genes that have a significant fitness defect. So, we did indeed narrow it down, but still that's a couple of thousand, and so we've applied some other criteria to come up with a couple hundred mutants that we
14:00
might like to make to begin mapping our scaffold network. And so we did the nonessential fitness genes. Obviously, they have to be expressed. We started to choose functionally diverse genes so that they could cover a lot of bioprocesses as we're starting to build out the scaffold, and then we also threw in a few conserved genes of unknown function so that we could start to explore these unmapped functional areas of the human genome, so to speak.
14:23
So, we kind of came up with a list of a couple of hundred genes that we thought we would begin to do screens with to start to build out our map. And so the general protocol is quite straightforward. We just used CRISPR-Cas-based guide RNA disruption to make mutant query cell lines in these genes that we've chosen, and then we simply repeat the screen that I just described in these different mutant backgrounds. And there's always wild type controls. We keep doing them over and over again.
14:51
So we'll have the ability, just like in yeast, to compare the single mutant fitness from our wild type screens, dropout screens, to the double mutant fitness that we see in the different query cell lines.
15:03
That's the assay, and of course, it's a next-gen sequencing readout of the bar codes and their abundance in the population. So as always figuring out the batch effects and scoring these data are of-- is often a real bottleneck at the beginning of these projects, and Max Billman, who was a postdoctoral fellow in our collaborator Chad Myers' lab at the University of Minnesota has done a, a great job of looking through our data and developing a score
15:28
that we feel quite comfortable in terms of calling genetic interactions from these data. He accounts there's a lot of data normalization steps, including identifying and removing frequent flyers, batch correction, which is a huge problem always, and of course, a lot of guide RNAQC has gone on throughout the whole process. Pardon? Can you tell us about the frequent flyer? The frequent flyers? Mitochondrial genes. They're still things that appear in all the screens. Oh okay. Good. Yeah. Yeah.
15:55
So, and so we've done a couple hundred of these genome-wide screens with 108 unique query mutant cell lines, and so far this project has given us three million double mutants and 80,000 genetic interactions. And so, it's nowhere near the yeast network yet, but we're s-- we can still start to ask, do we see any structures that remind us of what we see in the yeast network? So I'll show you that data. And this has been done by Henry Ward, a graduate student in Chad's lab.
16:20
I should have mentioned at the beginning, all of our computational work on genetic networks is an ongoing, very productive collaboration with Chad Myers and his team. And when Henry did this work --And so all he did was make profile, genetic profile similarity matrix using the exact same approach we used in yeast. And here's the matrix that he's developed so far. There's about 4,200 genes on this network so far, and they're mostly on the array side, right,
16:45
because the query strain side is quite sparse at this point, but we're still getting genes onto the network because they interact with the query genes. And again, I hope you can appreciate that we are starting to see clusters of correlated genetic interaction profiles on the network, which is exactly what we were seeing in yeast. And then we can go ahead and try to functionally annotate the network as we did in yeast. And this is a bioprocess annotation that Michael,
17:09
Michael did where he used just a-- an unbiased annotation algorithm to identify 44 network clusters that were enriched for Gene Ontology bioprocess terms. We were talking about this yesterday that because of the, the Gene Ontology's n-- not as well developed in human cells as in yeast, Michael then did some manual annotation because some of the Gene Ontology bioprocess terms were basically identifying the same thing. So he merged various domains and ended up with 27 domains on this
17:39
network that correspond again, just like we saw in yeast, to major bioprocesses in the cell. So we think that this kind of approach is going to be useful for mapping out the genetic interaction network in human cells as well. Well, what about those network structures that I was hinting about? Do we have enough data, or can we start to see these kind of coherent structures in the human cell data? And this is totally fun because this gets to what we were talking about that you can start to look in
18:07
more detail at these kinds of things, and then really--- you get clues where you, that prompt you to do mechanistic experiments. So you can actually start with this very high-level data, and quite quickly come up with specific hypotheses that if you're interested in exploring, you can. So I, I don't wanna belabor this since I'm, I, I only have about five minutes left or so. Um, but I'll just try to tell, So this is the, a network motif surrounding the PTAR1 gene in, in human cells, which encodes
18:37
a conserved protein parental transferase alpha subunit, and it normally puts geranylgeranyl residues o-- onto its target proteins. And it shows these sort of coherent patterns of negative or positive genetic interactions with a variety of protein complexes. And one of the most dramatic motifs is sort of a interactions with things involved in vesicle trafficking and intracellular transport, things like that, like a whole bunch of SNAREs the CORVE complex et cetera. So there's a--
19:06
And the COG complex, for example. Now, that's cool because it suggests these network motifs exist, so we looked back in our yeast data. So this is an example of the human data motivating us to look back at yeast. And the yeast, a version of PTAR1, is called ECM9. Ptar's not essential. Ecm9 is essential in yeast, so this was done with a TS allele, and we see the same thing. We see genetic interactions with SNAREs and other genes and, or protein complexes involved in intracellular
19:33
vesicle transport, et cetera. What was kind of interesting too is we saw this And we thought, "Well can we get any clues about what these genetic interactions are, mean for the biology underlying these geranyltransferases?" And because ECM9 is an essential gene, we have another project we were looking for deletion mutants that can bypass the function of essential genes in yeast. And we found an unknown open reading frame that completely bypasses the requirement of ECM9 for viability in yeast.
20:01
And then we go and we look at this particular gene, it doesn't matter what it's called, YNL320E and W or something, and it encodes a palmylyltransferase, and we started to look at YKT6 the ECM9 in yeast, and we can see that it's palmitylated at very low levels in wild type cells, but when you get rid of that suppressor, you see the palmitylation go way up. So we hypothesize that normally this parental transferase, geranylgeranyltransferase, is hitting its targets and
20:30
activating things and secretion go on normally. When that's perturbed these, this sort of other mechanism, the palmitoyltransferase, can take over. And remarkably, when you look now back at the human network, we see these positive genetic interactions with two genes that encode depalmitoylases in human cells. I don't know, I really like this example because it shows how leveraging both networks and going back and forth between these models can really prompt you to explore
20:58
kind of deep mechanistic questions about this kind of post-translational modification. Okay, so I don't really have a lot of time left, but I'm, I am gonna be around all day tomorrow. I'm just gonna throw you a teaser of what I kind of wanted to talk about, but I hope that I've shown you that these, I think it's useful to map reference genetic interaction networks, and we can argue about, you know, tissue specificity and whatnot, but everything we know from yeast
21:24
suggests that having a reference network is going to be useful, right? So that we can try to understand or focus our efforts on looking at what might be different in other situations. We've also, as I mentioned at the beginning used a single-cell image analysis, automated image analysis, to explore changes in the subcellular morphology of various compartments in the yeast mutant collections. And the reason we, this is an important thing to do is there are a subset of curated genes that
21:50
don't yield great GI profiles using a fitness-based readout, and of course that's a population-level measurement. We're using any information about single-cell variability, which is a pretty important thing when trying to understand the genotype to phenotype relationship. So that's why we've sort of done this, and we've undertaken a marker project where we have adapted the SGA platform to introduce fluorescent markers of these various subcellular compartments into the yeast mutant collections.
22:15
Then we image them and use automated image analysis to identify morphological outliers or phenotypes associated with it. So far most of our work has been on single mutants, but we want to go ahead and start testing selectively double mutants to see if we get new information about genetic interactions by using these different readouts. So we did-- uh, we published a paper, I'll just put this up here, a couple years ago looking at endocytosis markers, and in this case because it
22:43
was a very specific bioprocess, we were able to build a training set by looking in the literature and basically looking at the images, and we could build a training set of what phenotypes we thought we were seeing in our outliers. So we do outlier detection, we find phenotypes, and we make classifiers that identify all these phenotypes. So that worked for this project. We were able to identify about 17 different morphological phenotypes. And related to things that Pascal was talking about
23:09
earlier morphological pleiotropy is quite prevalent. Although some mutants only affect one marker, most mutants affect more than one marker. And also incomplete penetrance is the rule rather than the exception, and actually you can make penetrance profiles that provide extra phenotypic information than just whether or not something has a phenotype. So we've explored possible factors underlying incomplete penetrance. So now we have this big data set, though, and it's just
23:36
not practical if we wanna ask what are the possible morphological phenotypes associated with genetic perturbation in this cell. These supervised approaches simply don't scale well. There are some compartments for which we have not much information in the literature, and to make these training sets would be, first of all, biased and just almost impossible to do, although we have some training data. So we wanted unsupervised phenotypic detection, which
23:59
may allow identification of unexpected phenotypes, and also it's very clear that subtle biological significant effects with variable penetrance might not be easy to detect with some sort of-- without an unsupervised approach. So I'm gonna stop because actually it's a very much an ongoing project. I'll tell you what we have tried. The beginning is the same. We make our mutant arrays. We've tested various outlier detection and feature extraction methods.
24:25
We find that using cell profiler features and segmentation is the best way to do things for this type of project so far, although that could be played with. We've tried deep learning features too. We've made a couple of classifiers that get rid of garbage cells that are not segmented properly, and because it's cell morphology with changes with the cell cycle, we also use cell cycle classifiers, and then we do outlier detection. Like I told you, we tested various
24:50
types of features and various types of methods, and we've settled on a Gaussian mixture model method for doing our outlier detection. So we get all the outliers from all of these different strains, and then now we want to classify them or cluster them in an unsupervised way to see how many phenotypes are in there. So these are some penetrance profiles, and I think right away you can see that these are all the different mutants that we've taken through the screen.
25:12
These are some of the different markers, and you can see that there are some mutants that seem to just affect all of the markers. And actually, these tend to be essential genes, some of which are just dy-- you know. So there's various things we have to remove in order to reveal the underlying biology. So, we've tried NMF to try to cluster these data, and other unsupervised clustering approaches. I don't have time to go into it. It's really hard for us to know what's the best way to
25:39
do this in order to detect useful information, biologically relevant information. So that's a real bottleneck for us right now. We're trying various things, removing the cell cycle, removing the essential genes, blah, blah, blah, blah, blah, to see if we start to make more biological sense of these data. I'd be delighted to talk with anybody about ideas of how to analyze these data, and also to collaborate with people who might be interested in analyzing these data.
26:02
Again, I've been talking about work that's mainly been done by Neil Sahin, a graduate student in the lab, and Mojca Matijas-Lusage, a former postdoctoral fellow in the lab. Carles Pons in the Alloy Lab has been involved in analyzing these marker project datas, so super fun. And Quade Morris was a, a collaborator on all of the automated image analysis. And there's a large number of people that worked on the genetic networks, including the yeast networks and the human networks.
26:26
And I tried to mention key people when I was showing the slides. Thank you. (applause)
26:36
[AUDIENCE MEMBER] Really fascinating. Have you considered using parallelized CRISPR screens, rather than one guide RNA Yeah, [LECTURER] We considered it. It rapidly becomes a problem with the library size and the scale at which you have to do it. We've done it on a more focused level, for example, to test some of the interactions we see in other cell lines. Then you can do it, but to do it on this large scale, it rapidly becomes just too many cells. Just becomes a ridiculously large library.
27:07
Yes. [AUDIENCE MEMBER] So for the unsupervised cluster, so how do you assess whether something is biologically relevant? [LECTURER] Indeed. [AUDIENCE MEMBER] We just compare it to what we know, we could well use supervised method. [LECTURER] Is that-- Well, we, you could use a supervised method, but then you'd have to have some sort of training data. So we'd like to avoid that. We could do it for some of the markers, for example, like we did for endocytosis.
27:31
Right. So what we've been doing is looking at these various clusters, and looking at the cells, looking at any enrichment for bioprocess or some sort of functional information that's in there. And we see it, right? But that doesn't really give you enough information. Like I don't, I'm not comfortable looking at that cluster and saying, "There are 463 different morphological mu-- phenotypes that are possibly mutant." And I'd like to be able to say that, but I, I'm not sure how to get there.
27:58
[AUDIENCE MEMBER] Yeah, but still it's using
28:01
known biology to assess-- --the [LECTURER] Meaning [AUDIENCE MEMBER] Of [LECTURER] Clusters. Exactly. [AUDIENCE MEMBER] Exactly. Yeah. Can we spot, see in all these genetic models you talk about non-additivity, that's the default, non-additivity means there's some genetic interaction. [MODERATOR] Often [AUDIENCE MEMBER] It's hard to know what is the right indication of the. For example, for growth, you take the growth rate as within the exponential So [LECTURER] Just measuring colony size.
28:31
[AUDIENCE MEMBER] Colony size. So you're saying, but I think you have 0.7, 0.5. Isn't that 0.7 of the growth rate? [LECTURER] Yeah, that would be, if it grows 70, like so you take images of the plates at set times after you plate them. And then you can measure the colony size at a set time. Right? So you're measuring the colony size at the double mutants, single mutants and the double mutant [AUDIENCE MEMBER] So-- --and the wild type. You're taking a log of the colony size? [LECTURER] Yep.
29:00
Okay. It's actually remarkably quantitative. We've done some comparisons with liquid growth assays. [AUDIENCE MEMBER] So okay, let's use colony size as an example. If you take all your N million pairs, and say, you know, what is your null?
29:19
Is a typical interaction matching your null model well or [LECTURER] Additive? Yes. No, not for additive. For multiplicative. [AUDIENCE MEMBER] Multiplicative. Sorry. [LECTURER] Yeah. Yeah. The multiplicative model explains the data. Yes. [AUDIENCE MEMBER] Exactly? [LECTURER] Yes. Yeah.
29:36
[AUDIENCE MEMBER] Sorry, with the data generated so far, have you had any early glimpses into how cell type specific the data might be? Either like porting your systems to another cell line, or some of the other data that's out there with like combinatorial CRISPR screens?
29:51
[LECTURER] Well, a lot of the data that are out there are kind of bioprocess specific. And so we can do some comparisons there. We've tried to compare to the DepMap to see if we can get any clues about background effects. So yeah, we're doing that.
30:07
I don't have any overall conclusion for you. Okay. What was it, question? Yeah sorry. Well, I was [AUDIENCE MEMBER] Just wondering if there was, if you had any like early indications as to how cell type-- Oh [LECTURER] I see. I think probably 'cause of the types of query genes we're choosing, they're probably not gonna be that cell type specific And as I just said, we have taken a subset of genes that we've got genetic interaction profiles off fairly early on,
30:29
and tested them in other cell lines, and they seem to reoccur in other cell lines. But yeah, hopefully we can look at our data maybe and perhaps start to predict what sort of modules you might have to use in another cell background to, just to explore it. Yeah. For sure. But I'm gonna update you on it. Yeah. So I assume that all the screens are based on genes, not transcripts. So they are not looking at, you know, what's alternative splicing or something like that to supplement. Yeah. Yeah.
31:05
[AUDIENCE MEMBER] And I think you mentioned readout is through sequencing basically the barcode. Through the [LECTURER] Barcodes. Abundance. Barcode abundance. [AUDIENCE MEMBER] Abundance. So it's like, is that equivalent to the growth?
31:19
[LECTURER] We're suggesting it's a measurement of the relative growth rates in the population. So again, it's a population growth assay instead of an RNA-based assay. But yes, we're inferring a growth phenotype from the guide abundance. [AUDIENCE MEMBER] So maybe the phenotype is most relevant to the diabetes.
31:43
[LECTURER] Yeah, possibly. Yeah possibly. Thank you. Not in detail. Yeah, yeah, not yet. Yeah. We should--We will when we get more data.
32:04
We have done some of that. Um, Trajikas' lab has also done some nice work looking at DNA damage networks, et cetera. And we do-- So, what we try to do is sample a space. We can't do that experiment over and over again in plain conditions, so we took what we knew from study, Making the big network and design a much smaller array that gives us most of the information. And that way, we could test condition specific single mutants effects and other effects.
32:30
And most of the time, you see the interaction in the referenced condition, and it's exacerbated or suppressed in some other conditions. We do see a set of things that we didn't detect in the reference, but by and large, most of it is there in the reference network. So again, we find that
32:48
encouraging in terms of thinking about reference networks.
32:54
[MODERATOR] Trey quick comment. A couple years ago, he developed this classifier for cancer cells using transfer tools and all kinds of networks that was oriented at the gene ontology trigger-- Right yeah. Right. And I think one thing that emerged while he was doing those experiments that he needed to propose new Gene Ontology categories and terms. And I was wondering if you see evidence for those in your gene ontology annotation of the clusters that you see in the human [LECTURER] Samples?
33:23
[MODERATOR] We haven't looked at that, but we should. Yeah, me too. Okay. It's time. [LECTURER] It's time for break time. [MODERATOR] It's time to move on. Thank you very much. (applause)
33:35
Now a short wait outside, and then we go and
33:38
relax for the hiatus.
Here to help you learn
Sign in to keep a private conversation with your course tutor.
Ask me anything about the material…
Sign in to chat