Founder Stories: Brendan Frey, Deep Genomics & Changing the Course of Genomic Medicine
In this talk Brendan Frey (Deep Genomics) outlines how the company, spun out of the University of Toronto, aims to bridge the genotype‑phenotype gap using large‑scale machine learning. He begins with a rapid biology primer (DNA → RNA → protein) and explains why genomic data is now amenable to ML: massive, noisy, but increasingly affordable sequencing datasets. Frey contrasts early rule‑based and stochastic‑equation approaches with modern deep learning, arguing that ML can model the complex, combinatorial effects of mutations. He illustrates the approach with spinal muscular atrophy, showing how a single nucleotide change can be detected and how therapies (gene editing, RNA‑based) can modulate the effect. He then describes Deep Genomics’ team composition (ML, data generation, biology, diagnostics, therapeutics) and its collaborative culture, including interns and external partners. The core technical pipeline is detailed: high‑throughput experiments generate billions of data points across tissues; models are trained (generative and discriminative) to predict DNA‑RNA, RNA‑protein, protein‑DNA binding, chromatin structure, and therapeutic response. Performance on retrospective disease‑mutation classification is presented, highlighting high accuracy but also the variability of human pathologist labels. Frey discusses broader challenges for the ML community: the need for “smart data” (minimal, well‑structured datasets), building trust through transparency and explainability, and the limits of current models for design‑oriented tasks. He concludes with a vision of super‑human AI that can solve genomic problems beyond human capability, calling for advances in generative models and trustworthy AI.
Chapters
circumstances. Because obviously we're not all genetically identical, and even in the case of identical twins, there will be modifications that are different between the twins in terms of somatic mutations. So, that's what's needed. So that's a machine learning problem. It's perfect for our community. Machine learning is needed to close this genotype-phenotype gap. I'm going to talk about deep genomics today, but the University of Toronto is well-positioned in this area. Down the street, the Center for Cellular and Biomolecular Research has some of the researchers producing the best leading-edge high-throughput data in the world. Okay? So, today's talk is not just about deep genomics. It's about the Toronto environment. So, we have excellent machine learning researchers and some of the best experimental biologists in the world. Generating data that's amenable to machine learning. All right. So, the goal here at Deep Genomics, which is spun out of the University of Toronto 2 years ago, is to close this genotype-phenotype gap using machine learning. And it basically comes down to detecting and curing disease. So, Alexandra talked about detecting disease using image classification techniques. We're also interested in curing disease. By building machine learning models of the genome itself, the genetics. So just a very quick example. There's a recent announcement in Science just last week. Spinal muscular atrophy, which is the leading cause of infant mortality, so 1 in 10,000 births. Usually those kids suffocate within 2 years. They die of suffocation within 2 years of birth. So it's a dreadful disorder. But good news. November 7th, it was determined that the clinical trial, it was a Phase 3 clinical trial, I'm not sure how far through the clinical trial they were, but they terminated the clinical trial because they realized they had to give the drug to the control cases. So that's a big success, okay. And it's a genome-based therapy, all right, derived using the genome by looking at the genome. And it was in development for 13 years. So, the goal of deep genomics is to use machine learning to reduce that pipeline to a 3-year period. Okay? So, just a very quick review of biology. This is sort of a 1-minute review of something that really takes a couple of years to understand well. But if you look at DNA, we sort of think of DNA as encoding the instructions for life. Okay? And replication is when you have kids or when your cells divide. But these other processes that, of course, lead to what's going on within the cell. Okay, so what's going on in your cells? It's not just the DNA. There's RNA molecules. There's protein. And essentially what happens is pieces of DNA are transcribed into RNA sequences. Those RNA sequences are transcribed and translated into protein. And each of these can be described as a sequence of letters. Okay? So, in the case of DNA, it's a sequence of A, C, Gs, and Ts. And RNA, basically the same letters. T is replaced with U. And then protein is also a sequence of amino acids. All right? And so, I think a big issue for machine learning researchers is, "Can I easily understand the data and get to work on it quickly?" Okay. And I guess one point I want to make here is for genomics, the answer now is yes. We can encode the data easily. We can give you datasets where you can assume that different data points are IID. We can give you large datasets. And a lot of this has changed just in the past 6, 7 years, as I said, because of people at the— like people at the CCVR. All right. So, these are— if you think about this process, if you can understand how mutation in the DNA leads to a change in the RNA, which leads to a change in the protein, which goes on to lead to a change, for example, in intracellular signaling, then that could lead to a disease. So, by understanding these processes, or more importantly, by accurately modeling them, Okay? We can hope to connect genetics to disease, to phenotype. So, about 6 or 7 years ago, people were trying to do this by writing computer programs using the good old technique of if-then statements and control logic. Okay? Another approach that was used is to write down stochastic differential equations. So you can look at balances of different concentrations of molecules. And write down stochastic differential equations. And those approaches are very brittle. The equations are usually wrong because there's unknowns. And there's a lot of noise in the biology. Obviously, the logical approach to writing programs with if-then statements doesn't really make sense for this kind of system. And so machine learning, of course, is the solution. And one reason is that biology is complex and noisy. And another reason is that we can't precisely measure all of these molecules and what's going on in the cell. So what we have is datasets that allow us to peer inside of the cell and look at these quantities, but the measurements are noisy. Okay. So we have incomplete data. We have noisy data. And we have a lot of data now. So it's a perfect fit to machine learning. So let's go back to spinal muscular atrophy. I just want to show you this example. It really is quite simple to understand. So most people in the room here have this piece of DNA, this sequence in your DNA. If we sequenced your genome, we could map this to your genome and we'd find it. Children born with spinal muscular atrophy have a mutation here. So that's all it is, is the C— whoops— the C is mutated to a T. It's a little bit more complicated than this, but this is what it boils down to. All right, and that's what leads to this disease. This disease. So that's it. So detecting the problem is about finding letters in that sequence that lead to changes in phenotype. Okay. So think about that as the input to the neural net. You're trying to find changes in those letters that lead to a change in the output of the neural net. How about treating it? So treating it, the therapy I mentioned earlier essentially interferes with this part of that sequence down here. There's different kinds of therapies. There's gene editing therapies like CRISPR-Cas9, where you have these molecules that will go in and actually modify the DNA. Or, you can have a therapeutic treatment that involves a protein binding or some other molecule binding, like an RNA molecule binding. But, essentially what these therapies do, these genetic therapies, is they modulate the effect of that sequence. They alter it. Okay, so what's interesting here is that the therapy does not target this particular mutation. Those therapies didn't work out. The therapy that works operates down here. Okay, so that's another exciting thing from the point of view of machine learning, is that simple correlation is not enough to solve the problem. We do need inductive systems that are capable of understanding how the inputs cooperate or compete, how there's some sort of combinatorial logic or combinatorial combination of inputs that lead to the output. Okay? So this mutation causes a problem. This modification fixes it. So that's also exciting from a machine learning perspective. All right. So I don't have a lot of time, but I'll just give you a glimpse into what's going on at deep genomics to address this, and how we're using machine learning, and kind of the structure of deep genomics, how the workflow operates. So, just briefly who we are. So, we've got people from all over the world. By the way, blue here is land and white is water. I should have downloaded a different image from the web. So, people from all over the world, from different universities in the US, Canada, and England. And here's a recent group picture. We have— we had 3 interns last summer, Victoria, Michael, and Omar. And various people here do the work on machine learning. Others work on data generation. And others work on molecular diagnostics, trying to understand how our modifications that are predicted by our systems are going to affect phenotype. Here's a few pictures of people being happy at Deep Genomics. Eating cakes. Looking at cakes. The opening ceremony was, I guess, 2 years ago. One year ago. We had a couple guest speakers give talks in our facility. So, this is located over at Mars. And here's a reception. You can see, I think Rich is there, and Jeff is there, and I think that's Ilya, the back of Ilya's head. All right. And as I mentioned, our summer interns. Okay. So, I'm going to now just give you a couple examples of how our technology works. So, this machine learning technology for understanding the genome.. And then, also show you one example of how we actually train a model. Give you a little bit of detail. So, basically, we detect things and we fix things. All right? This is an example where we did a retrospective analysis of mutations across a whole wide suite of diseases. So, spinal muscular atrophy would have been in there. Would have been one of those diseases. And each of these little orange people is equivalent to 30 patients with a serious disease. And the green people is equivalent to 30 healthy cases. So those are controls. And so then we just took these people, looked at their mutations, ran them through our machine learning system, which flagged them as either pathogenic or non-pathogenic. And the ones that were classified as disease or pathogenic, you can see we got most of them correct. And in the cases where our machine learning system said these are okay, no problem here, we also got most of them correct. So there's the false— we've got one false case there and three over there. So, it actually works quite well. Now, one thing we've realized is it's really hard to say precisely how well, because there's a lot of bias. So, if you look at how do we actually obtain these labels for mutations in patients with a serious disease, well, those are obtained by pathologists. They looked at the data and then analyzed it. And pathologists, it's a little bit hard to figure out exactly what the reasoning was. In some cases, it's clear-cut, like that spinal muscular atrophy mutation. It's been known for a long time to be the mutation. But in many other cases, it's not clear-cut. And so, a big issue that we face here, actually, is the fact that the— is that the human pathologists don't really have a lot of data on how well they're doing. Okay? There's been two recent studies in the last two years where— excuse me— Mutations were sent to two different diagnostic labs, and then they assessed the mutations. And the level of agreement between the two labs— or let's say the level of disagreement— ranged from 15% to 50%. Okay? So that means you're going to get two different answers, depending on who you send your mutation to. All right? One clinic versus another. So what's interesting, and I think this will evolve over the next few years, is a more data-driven approach to justifying how well systems work or validating systems, both human and AI-based systems. So, let's look at fixing it. So, this is— by the way, that was work done by Danielle and Amit. And let's look at— so now let's look at fixing it. And this is work done by Mike Leung and Andrew DeLong. So, what you're seeing here is different therapies. This is a therapy that that has to do with polyadenylation, one of these processes that occurs in the cell. And what you're seeing in different columns here are different possible therapies, different drugs if you like. And what you're seeing vertically, this pattern here, is just a readout that tells you whether that therapy is doing what you want it to do. Okay, so a bar up here says it is, and a bar down there basically means it isn't, very crudely speaking. So now I'll overlay on top of that the prediction from our machine learning system. And then I'll toggle back and forth so you can compare the two. Yeah. So, really good agreement. Okay. So, how do we do this? So, this is the machine learning part. We've got many different models in development, but I'll just talk about one to give you an idea for the workflow. So, in this case, we're looking at predicting how proteins interact with DNA. Which is an important process that leads from DNA, of course, to proteins and to health or disease. So, this diagram here is showing a protein interacting with DNA. And this is basically the workflow. So, there's different kinds of high-throughput experiments we use to gather big datasets. And I'll briefly talk about the sizes of the datasets we have in a moment. And then, just the usual machine learning techniques, both generative models and discriminative models, trained usually using GPUs so we can push through a lot of data quickly.. And then that leads to the model, a machine learning model that can be then used for different needs— diagnostics, therapeutics, and also just understanding. So, it's just the scientific use as well. All right. So, what about the data? And I think from a machine learning perspective, the first thing you want is a dataset. That's when you believe you can actually contribute. You can download a dataset and look at it, peer at the data. And one of the biggest enablers for us is the cost of genome sequencing. So, as I mentioned, it's dropped drastically, and it'll be— so right now it's about $1,000. It'll be less than $20 in a couple years. At the same time, our ability to train models has gone up, thanks to NVIDIA. So, we have different kinds of models in development here, different kinds of training sets, of course, associated with those models. DNA to RNA models, RNA to protein, protein DNA binding, so interaction models. Models that describe how the chromatin is structured, the chromosomes are structured, models of proteins, and therapeutic models, and others. So, here's just one of the kinds of models we train. We have 9,661 datasets, 73 trillion data points in total, covering 54 tissues. So, when Ofer was in my lab, we were always struggling to get small datasets. You know, we'd sort of spend 2 years putting together a dataset of 10,000 training cases. This is the situation now. Thousands of datasets. And here's a histogram showing you the number of datasets versus the number of data points in each dataset. So, lots of datasets with 75 million data points. Some with as many as 200 million. All right, so we have a lot of data that allows us to sort of fill in the gap between genomics and phenotype. And we have computing power that allows us to train these models. And that's basically what we need to solve the problem. So, here's the organization at Deep Genomics. We have the biology team that figures out what we need to do in terms of these biological components in the system. A data generation team, which produces these giant datasets. The machine learning team, which then trains the models. And the stakeholders for the machine learning team include the diagnostics team and the therapeutics team. So, these are the ones detecting it, and these are the ones fixing it. And they interact with our clients in terms of producing value externally. So I'm going to finish up now by talking about a few interesting challenges, I think, for the machine learning community that emerge from genomic medicine. These, of course, are related to other challenges in other fields, but in particular, in the context of genomic medicine, this is a big deal. So, one thing that has always bothered me is big data. People say, "We've got lots of data, lots and lots of data." And I'd like to think about it as the needle in the haystack. If you're trying to find a needle in a haystack, it doesn't help to throw a lot more hay on the stack, okay? So, what you really want to do is remove hay until you have a small haystack and the needle is in there. And so, that's the analogy here. What we want is smart data. All right. You don't want lots of data. You want the smallest amount of data possible that will help you solve the problem or that will enable you to solve the problem. And I think that's a big challenge for us, especially when machine learning people talk to other communities. Don't just give us everything. Give us the smallest amount of data we need so we can solve the problem. And we'd like it to be preferably in sort of a matrix format. Okay. Yes, that's right. Okay. And that's really important. And so, you know, that was the question I had for you at the end of your talk was, Alexander, you talked about images. That's well-organized data. I'm sure the machine learning people like that. Ah, images. I can just take my ImageNet neural net and apply it to radiology images. In the case of the health system, I think from a the perspective of a machine learning researcher or a student, they're going to wonder, "Okay, that sounds really complicated. Can you just give me a matrix?" And so, I think what's needed really is going to be people that are the glue between machine learning researchers and practitioners and the other participants in the process. So, at Deep Genomics, we have that. It's an internal team. It's our data generation team. The other aspect is trust, okay? 40 years ago, if you'd said to someone, "We're going to get the computer to land your airplane," they probably wouldn't have gotten on the airplane. And nowadays, computers land our airplanes all the time. So what happened, right? Now, landing an airplane is fairly straightforward. You can actually look at different steps. You have to do something with the ailerons. You have to make sure the wheels are deployed, things like that. And so there's this notion of transparency, right? You can— You can show experts, look at that, the wheels were deployed, right? So transparency is very important for building trust. So when it comes to machine learning in medicine, and whether it's image analysis or other aspects of medicine, I think what we need is machine learning techniques that are capable of being transparent. Or another way to talk about it is providing explanations to the human experts. I like transparency more because I think it doesn't necessarily need to be an explanation. It just needs to be a piece of data that the human expert can look at so that they understand why the system is doing what it did. And so, when I talk about the genotype-phenotype gap, we've got DNA here, we've got disease over there. The systems of deep genomics can say, "Well, this mutation is going to cause this disease, and the reason is that the protein didn't bind to the DNA correctly because of that mutation." Okay, so that's the level of transparency that we can provide. And I think that's true more generally in healthcare, especially because it's so complex. So this is deep art. So this is what happens if you train a neural net to classify images. And like we heard about ImageNet earlier. And then what you do is you take an image, a Van Gogh image, I guess, and then you say, well, let's change the image so as to maximize the label. At the output of the neural net that says this is a, I think maybe a dog. Okay. So you get this really, this psychedelic image which you could print that out and put it on your wall for sure. So can we do anything useful with it? So Ian Goodfellow tweeted recently that, you know, he thought we could do design. We could use neural nets for design. You'd literally adjust the inputs to the neural net to get the desired output and you'd design what you like at the input. So this is something we've been thinking about at Deep Genomics for a while, which is if we have good neural nets that can tell us what the phenotype is going to be, we can take someone's DNA and we can say, "How should you change the DNA to get the desired phenotype?" That is, correction of the disease. And what this basically shows you right here is that at this point in time, there's no way in hell you'd trust that system. Okay? So, but it's an interesting question for machine learning researchers, right? How can we build models so that we can? Alter the input to the model to achieve the desired output and do so in a way that works. So the problem here is that this image is completely outside of the space of images that are natural. And so the machine learning technique has never seen an image like this, okay? So that leads to ideas like generative models, generative models which can account for the data and not just perform a classification task. Anyway, so an interesting open question. And then, the last thing I'll mention is just to connect the different applications of AI with genomic medicine. So, you know, there have been some recent fairly spectacular achievements. Google's AlphaGo system achieving, well, beyond human-level performance on Go. So, that's kind of like, you can think of that as human-level AI, matching human performance.. But in health, and at Deep Genomics, what we're trying to do actually is build systems that can do things humans can't. Okay, so the situation, the way I see it, is we'll need more advances in AI that allow us to move up this curve to the point where we can crack the kinds of problems we're working on at Deep Genomics. So, superhuman AI. And the reason it's superhuman, well, obviously, people aren't good at understanding DNA or genetics. No one person ever will be, and no group of people ever will be. We're naturally good at vision. We invented language, so hopefully we're pretty good at language translation, things like that. But genomics we're not good at. We weren't built to be good at it. So we need superhuman-level AI to solve the problem. Thanks.
More from Joshua
Scientist Stories: Hans Waterham, The value of phenotypic assays in peroxisomal disorders
Scientist Stories: Joel Zlotogora, Genetic carrier screening programs for reproductive purposes