Founder Stories: Brendan Frey, Co-Founder and CEO of Deep Genomics

Right, it is now my very distinct pleasure to introduce our first speaker, Brendan Frey, who is many things. He's the founder, CEO, and chief engineer at Deep Genomics. He is the co-founder of the Vector Institute, for which we are eternally grateful, and he is professor of engineering medicine at the University of Toronto, where he also earned his PhD. Brendan is an internationally acclaimed entrepreneur, engineer, and scientist. He has made fundamental contributions in the fields of deep learning, genomic medicine, and information technology. He has co-authored over 200 papers in these areas, including over a dozen that have appeared in Nature, Science, and Cell. His work with Geoffrey Hinton on the wake-sleep algorithm helped launch the field of deep learning. His work on deep learning and on factor graphs forms the foundation of thousands of products used worldwide. Welcome, Brendan.

Thank you, Melissa. It's a pleasure to be here. And hello, everyone. I was speaking with Farnoosh from Cyclica just a few minutes ago about the panel that'll be happening a bit later. And that's what this— what today is all about is going from university, from learning about machine learning and deep learning and AI and other areas having to do with computer science into potentially industry and having a career in industry. And it can seem like a big jump, but I, I think one of the goals today is to show you how easily you can make that transition and all the wonderful opportunities that are available, I guess, this afternoon in the area of health, but, but of course more generally in other areas too. So I'm going to fire up my presentation, and the goal here is for me to talk to you about a few of the breakthroughs in artificial intelligence of deep genomics that have enabled us to discover drugs that could not be discovered without the AI systems that we've created. All right, so let's get going here. So here's a quick snapshot of our current team. We've got almost 100 people at this point in the company. And one of the exciting things about deep genomics, and I'm sure you'll see this in the other kinds of applications that you hear about today, is that it's a multidisciplinary team. Obviously AI and machine learning, but people from computational biology, software engineering, experimental biology, We work closely with our chemists and our chemistry team, and also deep genomics drug developers, and of course business development people who, who make the connection between the molecules that we're discovering and all the opportunities to, to help people with those molecules, you know, to put them into patients essentially. We have two headquarters. We have a Toronto headquarters, which you're all familiar with, in the Mars Building and the surrounding area. We also have an office in Boston just off of Kendall Square. And just a quick snapshot of our team, a couple people. Amit Deshwar, who's our head of predictive systems, which is really our AI machine learning group at Deep Genomics. And our predictive systems team, they do everything from create new datasets, and I'll talk a little bit about some of those datasets in a moment, to of course building machine learning and AI systems based on those datasets, to then ensuring that those systems that are built in conjunction with our software engineers enable our drug developers, our biologists and our chemists and our people who conduct animal studies to actually make the molecules that can save people's lives. I'll just point out one or two other people. Ferd Massary, he's our Chief Medical Officer based in Boston, and he has over 30 years of drug development experience at places like Pfizer and Shire, which develops rare drugs, and a variety of other companies. Also on our strategic advisory board, we have Tal Zaks, who is the chief medical officer of Moderna. And he's really the person who is behind the COVID-19 vaccines that, of course, a lot of us have been injected with recently. So really great group of people, all the team members and all the advisors as well. And as for you, we're hiring. So if you're interested in joining Deep Genomics, visit our Lever page. You can get to it from the Deep Genomics page. And check out the positions that we're looking to, to fill. So now a little bit about the company. So I mentioned the COVID-19 vaccine, and really in the last few years we've all become quite familiar with the concept of an RNA therapy. We've really seen a revolution because of this, this change. What's cool about RNA therapies, and this is not true for small molecules or peptides or biologics, but for RNA therapies, the medicine is literally a digital sequence. So the sequence of Cs, As, and Ts, and Gs that I'm showing there, that is, that is one of the compounds that we've developed at Deep Genomics. Now there's chemistry, of course, that's needed to get that nucleotide sequence into your cells, and the chemistry is super important. But once that nucleotide sequence is in your cells, a digital sequence of information, you can change that sequence to target any gene in the genome. Change the sequence one way, you can increase the amount of protein produced by that gene. Change the sequence another way, and you can decrease the amount of protein produced. Change it even another way, and you can alter the type of protein that's produced. So that for the first time in human history, we have digital medicines. They're programmable, and it's offered us an amazing toolkit, of course, for drugging any gene in the genome in any way we want. But also, it connects the dots between artificial intelligence and information technology and drug development. And that's what's really important for all of us here today. Now, at Deep Genomics, the way we're approaching this is, of course, biology is complex, and so the way to crack it is to use artificial intelligence and huge amounts of data. And what's key here is that in RNA biology, that's what you need for RNA medicines, there's more data than in any other area of biology. So if you look at cell signaling, not as much data as what we have in RNA biology. That's true for pathophysiology, it's true for systems biology. RNA biology has more data than any other area. In fact, we have over 100 petabytes of data, just a huge amount of data. And the amount of data we have has grown 1 million-fold since 2002. So I'm sure you all already— you all kind of know biology is complex. RNA biology is complex too. And humans— there's no way humans could pore over all of this data to find the good the good mechanisms and the good molecules they could use to treat patients. And what's needed is artificial intelligence. And that's what deep genomics is all about, is the intersection of those three. And this enables us— and I'll talk about this in just a moment— to program the best RNA therapies for almost any gene and any genetic condition. I'm going to give you a quick 1-minute review or introduction, if that's the case, to RNA biologies, just to kind of show you why it's so relevant to, to the kinds of concepts we're all used to thinking about when it comes to digital information. So this is the central dogma of biology. We go from a gene, gene gets transcribed into pre-mRNA, and that pre-mRNA gets processed, manipulated to produce mRNA, which then gets translated to protein. Proteins are the things that do all the work in our cells. What's key here is that these sequences are digital sequences. They're sequences of letters. Also, when it comes to the RNA molecules, pre-mRNA getting processed into mRNA, The actions that take place, the processes that occur are quite well understood. So we know what the wiring diagram or the block diagram looks like for this area of biology. That is not true for other areas like pathophysiology or systems biology. We know a lot about the wiring diagram. Furthermore, the operations that take place are digital in nature. So that RNA sequence, certain segments are cut out and removed, other segments are added. Other segments are spliced together. Because of all these digital operations and this well-known wiring diagram, RNA biology is really a good fit to artificial intelligence. We've got digital genetic targets, which are mutations. We've got this digital RNA biology. And the third piece is, as I just told you before, these therapies are digital. They're a sequence of letters. All right. I'm going to quickly just show you a few of the machine learning predictors we have. At Deep Genomics, we have over 40 predictors, and the number is growing. And each of these predictors is carefully engineered engineered, derived from dozens of other experiments that were conducted by our machine learning and AI researchers to find out which models could really help drug development. So these are the, the high-quality predictors, if you like. We can use these predictors to identify novel genes to drug. We can use them to design therapies. We can also use them to predict the safety and toxicity of those therapies. We can even use them to identify new patient populations, so that— so to help figure out which patients we can drug with our compounds. On the right-hand side, I'm showing you just a couple examples of the kinds of predictions we can make. We can take an RNA sequence— again, that's a sequence of letters that might be in you or a loved one that has a mutation in it called a variant— and we can run that through our predictor, and it'll predict how that variant is going to cause something to go wrong, such as cause exon skipping. We can also evaluate a molecule to determine whether it will recover exon skipping. So that's how powerful these predictors are. They can make these kinds of predictions. And I'm going to give you a— so that's kind of the, the inputs, if you like. We build machine learning predictors that can make those kinds of predictions. I'm going to give you a snapshot of a proof point in terms of an actual therapy. When we, we think about a life-saving therapy that was developed about 3 years ago for spinal muscular atrophy, it's called Spinraza. That therapy took about 15 years to develop. We were able to identify the target the exon, and also design the therapy in one afternoon on a computer. So the top part of this slide shows how we use a target predictor to evaluate all of the exons in the genome and identify which ones could be good drug targets for rare disorders. When we looked at the top 10% of the predictions, this exon for spinal muscular atrophy was in the list in the top 10%. We also found a lot of other exons that we can drug and are in our pipeline right now, novel ones. On the bottom part of the slide, what you can see is we then and looked at that exon, and we evaluated over 2,000 potential therapies, all of this in the computer. And on the lower right, you can see the therapies are lined up on the x-axis, and the y-axis is the predicted effect, the predicted benefit of that therapy for that target. So the peaks correspond to potentially good therapies. You can see that one of those peaks is Spinraza, this drug for spinal muscular atrophy that saved thousands of lives. And the other peaks are other compounds which which are novel. And when we took those into the wet lab and tested them, they work better than Spinraza. I'll just give you another glimpse into what we're doing at Deep Genomics in terms of our predictive systems. So our DeepBind predictor is able to tell us when a DNA molecule or RNA molecule will bind to a protein. You probably heard the recent news about DeepMind coming up with their AlphaFold algorithm for predicting protein folding, where for us in RNA biology, protein folding is not so crucial. What's more important for us is understanding how RNA behaves. So that's what this is all about, is figuring out how RNA behaves. And we published this work originally in 2015, but have made a lot of improvements. So we take all sorts of high-throughput datasets, could be Celex data or other datasets, put them through our deep learning pipeline, produce predictors, and our predictors can then be used to diagnose mutations, can be used to develop molecules that can save patients' lives. I won't jump into the details, but I, but I just wanted to show you a slide that will be reminiscent We take our sequences, the sequence go through convolutions like a convolutional neural network. These produce scans, then we have rectification steps and pooling steps. We produce features. The features are then used to derive a prediction. The prediction is compared with the target, and we feed the information back to improve the predictor, of course, over time. Just wanted to kind of map the kinds of problems we're working on to things that you're familiar with. Second example is using transfer learning. So once we've built these RNA biology machine learning models, if you like, and these are models that really understand RNA biology better than humans do. Once we've built those models, we can then add a little bit more data in and use transfer learning to make those models really good at a specific task. And so, for example, we can use an RNA biology model that understands how proteins interact with RNA, how RNA folds, how exons are spliced, and all of those different things. We can then add some examples of molecules, therapeutic molecules, and corresponding reactions in the cell, add that data in, and then improve the model so now it can predict how well a new therapeutic candidate will work. And the next slide here is just showing you how that pans out. So on the left, what you're seeing is without any fine-tuning, without any transfer learning, our model is able to predict the efficacy of completely novel compounds with a correlation of 0.46. That's actually quite impressive considering that the model had never before seen any molecules. It's only trained on general RNA biology data. When we then do transfer learning, you give it some additional information about some molecules and their effects on the cells, we're able to get a 0.81 correlation in terms of predicting the effects of compounds. How beneficial is this in practice? Hugely beneficial. So this slide here actually shows you how we compete with other companies by using our AI systems. On the left, what you can see is the percentage of targetable genes that we can access because of our AI predictors. And roughly, we can target twice as many genes as our competitors can because of this technology. And on the right, it shows you that we can drug each of those genes in more ways than other companies. And so, for example, there might be one mechanism you hit in a certain gene, there might be another mechanism you can hit in that same gene to produce a therapeutic benefit. If other companies wanted to catch up with us in terms of the number of different ways we can drug a target, they would have to conduct 15 times more experiments. And that's a lot of work. These experiments are difficult, so that would take a long time. So I'm going to finish up and then ask for Q&A. I'm going to finish up by just kind of giving you— stepping back again and saying, okay, the high level, how does our company work? So you get it, this is what AI and these predictors— and how does the company operate? How do we realize more and more benefits as time goes on? So this is basically showing that the company is a big flywheel. So lots of positive feedback loops between our AI researchers, our molecular biologists, our chemists, people who do animal studies, our drug developers. At the lower left, you can see that first of all, we train our predictors and create new AI systems. Those are those 40 machine learning predictors I mentioned. We then use them to make predictions across the entire genome and for all sorts of different genetic conditions. And we've actually made over 2 billion predictions so far. We've looked at every gene in the genome, over 300,000 pathogenic mutations, and we've tested over 200 million potential novel molecules as therapeutic molecules. Married with that is our experimental facility. Over half of the company works in the wet lab. And so that— the combination of prediction, of course, goes hand in hand with doing experiments to see if we were right and also to see when we're wrong. We made over a million experimental measurements so far. These are efficacy and safety data points for 250 genes and over 20,000 novel molecules. All of that data is then fed back into our AI systems, so the more we do, the better we get. Also out of that experimental work comes our therapeutic programs. We have 10 in our preclinical pipeline today, and we expect to have 4 in the clinic in 2 years. So this is the real deal. We're making novel molecules. These are new chemical entities, not repurposing drugs, but making novel molecules that are going into humans to save lives. This slide here just gives you a sense of our pipeline. So we've got CNS disorders, including some neurodegenerative disorders like frontotemporal dementia, some infantile disorders like Niemann-Pick type C disease, pediatric epilepsy, Parkinson's disease is another one we're working on. And then we also have metabolic programs like refractory gout and Wilson disease. That's a disease where you get copper accumulation in your body. So I'll finish up by just saying that we're building a world-leading team to achieve the 3-year goals shown below. And if you're interested in joining us, please go to the, the Lever site. That includes building new AI systems. These AI systems we built, they're opening up completely new terrains for us to investigate, enabling us to discover drugs we couldn't discover without these AI systems. The next couple seconds, I'll just talk about productivity, which is not possible without these AI predictors. We're going to screen 100 genes for novel targets. We expect to identify 80 packages which could constitute 80 different drug development programs. We're going to take 28 into our preclinical pipeline ourselves, which is again amazing productivity, and we'll have 4 in the clinic within the next 3 years. Another big part of this is expanding our partnership activities so that we can access the clinical development pipelines of our partners and also bring in new datasets that we can use to derive new therapeutic targets and, of course, to train our AI systems. All right. Thank you, everyone.

Thank you, Brendan. That was fantastic. A couple of questions. So clearly there are numerous diseases that would benefit from the work of deep genomics. I can think of Huntington's, cystic fibrosis. You spoke about specific target and therapeutic predictors that you use. Can you talk a little bit about what, you know, which diseases to focus on bringing to the next level like you have with Wilson's disease and what other variables might be relevant in that decision-making?

Yeah, so it's a great question. You know, biology is complex and disease therefore is complex as well. We've tried to break it down into two different ways. One of them is the complexity of the disease. So on one end of the spectrum is rare Mendelian disorders, so monogenic, single gene. The other end of the spectrum is more complex disease. So things like neuropsychiatric disorders, for example, schizophrenia. Or autism would be another example that's very complex. And by complex, I mean it involves many different genes. And for any one patient, there could be a subset of genes that are active. Yeah, what we found in deep genomics is by focusing on those rare Mendelian disorders, first of all, we built machine learning systems that enable us to really make good progress on those more complex diseases. The other aspect is kind of the diseases, the tissues that we can get our molecules into. So right now our molecules are best suited to central nervous system tissues. That would be the brain and CNS, and also metabolic. So that would be liver cells, kidney as well, and also the eye. So those are the tissues we can get these molecules into very effectively right now. But as time goes on, we will develop new delivery systems, or we will license or acquire delivery systems that are developed by others that we can access other tissues. But our focus right now is metabolic and central nervous system, and within central nervous system, it could be developmental, so that's babies, or it could be degenerative, so people like Alzheimer's, for example, or Parkinson's is an example of what we're looking at too.

Fantastic. So a question from Mohammed: Do you focus specifically or exclusively on human genome data, and if yes, any specific demographics, thinking of how your findings can be generalized to a wider target audience?

Yeah, so we do use human genome data, but we also use mouse genome and non-human primate genome data because we do animal studies, for example, in those animals. What's really cool about RNA biology is that it's pretty conserved across species. So the way in which your, your RNA is transcribed or transpliced, for example, or the way DNA is transcribed into RNA, is pretty similar between humans and mice and monkeys, for example. And so that means we can look at these different kinds of datasets, and we can learn— our machine learning systems can learn RNA biology from these diverse species. And that helps us with things like, first of all, of course, finding human drug targets, but then it helps us with doing things like designing animal models. So if we want to insert a human gene into a mouse, we want to know how it will behave when it's in the mouse. And to do that, you kind of need to make a prediction for kind of like a human-mouse hybrid, right? So it's neither human nor mouse. And what's great about RNA biology is because it's conserved across these species, we can make accurate predictions that help us do that. The— in terms of the subsets of data, the other thing is, but because we're learning A lot of what we do is train these RNA biology machine learning systems to understand RNA biology. Once they've been trained, they can be applied to different subsets of genetic populations. Okay, so they're kind of like universal models. They're agnostic to the genetics. As I said, not only do they generate— generalize between different subsets of human populations, they generalize across species. So that enables us to really tease apart all these different genetic subsets in an unbiased way. Which is very valuable.

Excellent. I know that you have a meeting in 2 minutes, so last question for you. In terms of data infrastructure and workflows, how has Deep Genomics dealt with gathering, storing, and sharing this genetic data?

That's a great question, and there's, there's 2 aspects to it. First of all, there's the public datasets, and then the second of all, the internal datasets. But we built a system that we call Nexus which incorporates these two different kinds of datasets. This is the genetics types of datasets we're talking about. But it also— Nexus also incorporates a wide range of other kinds of data that we generate internally. I mentioned before animal studies. We put thousands of thousands of compounds into animals and obtain different readouts. Nexus combines that animal data with the genetics data. Also, we have a large number of robots conducting experiments in different cell models. These would be different cell types derived from human cells. It could be cancer cell lines or it could be induced pluripotent stem cells. Huge amount of data generated by those robots conducting millions of experiments. And all of that data is also fed into Nexus. And what's critical about Nexus is that it's also a portal that enables different stakeholders in the company to access that data and view it easily. The reason that's important is, for example, a biologist who's about to conduct an experiment in the wet lab, they want to check the data from a certain perspective, maybe a visual perspective. Whereas a machine learning researcher, they might want to download a giant dataset with a billion data points. And so it's not about visualization, they need a programmable interface so they can easily access the data. Somebody who's preparing a patent, for example, for Wilson disease, they want to access all of the data in a, in yet a different way so they can put the data into a patent and submit the patent. Are people submitting papers to regulators? Regulators require documentation, of course, and the documentation they require requires us to look at the data again from yet another perspective. And so what's great about Nexus is it allows us to, first of all, collect all these very diverse data sources, put them together in a structured and integrated way, and then extract the data from different perspectives for different use cases.

Sounds like we could all use a Nexus in our lives. Thanks so much for your time today, Brendan.