AI in Science and Medicine

Good morning and welcome to the Gardiner International Symposium in this Gardiner Science Week. The future of AI in science and medicine, obviously a very important and hot topic for us all to consider. I'm Janet Rossant and I'm President and Scientific Director of the Gardiner Foundation, a great job. Before we begin, I'd like to take a moment to acknowledge that MaRS is located in Toronto on the traditional territory of many nations, including the Mississaugas of the Credit, the Anishinaabe, the Chippewa, Haudenosaunee, and the Wendat peoples, and is now home to many diverse First Nations, Inuit, and Métis peoples. This territory is subject to the Dish of One Spoon Wampum Belt Covenant, an agreement to peaceably share and care for the Great Lakes region. As I said, this event is part of Gardiner Science Week, which is an annual week-long celebration of our Gardiner Award laureates who are all gathering here in Toronto here today and to really discuss and celebrate the impact of their research on human health. I'd like to thank our long-term partners, the Government of Canada through CIHR, the Government of Quebec, and the Government of Alberta. And a special thank to all our Gardiner Science Week sponsors who are on this slide who really made these events possible. Our platinum sponsors, the University Health Network, the Kremble Foundation, and TELUS Health particularly. Our gold sponsors listed on the slide, University of Toronto, London Drugs, McGill University, Moderna Canada, Women's College Hospital, Sunnybrook Research Institute, CAMH, McMaster, and IDRC, and all our other sponsors and donors and volunteers who've helped make the events here in Toronto, but also across the country, really exciting and successful this week. Without all their support, we wouldn't have this work of network this week. Of networking, engagement, and education. Today's symposium topic is really inspired by the work of the 2023 Canada Gairdner International Award laureates, Dr. Demis Hassabis and John Jumper. And you're going to hear from them and some of the world's top researchers in the field of AI as it relates to science and medicine. And now I'd like to introduce our morning chair, Kate Williams, who is Scientific Director of the Cremble Foundation. One of our major and long-term sponsors. Kate.

Thank you, Janet, and welcome and good morning, everyone. As Janet said, my name is Kate Williams. I'm the scientific director at the Krembla Foundation. We've had a longstanding partnership with the Gardiner Foundation, and we are thrilled to help support this year's international symposium on the future of AI in science and medicine. I think we would be hard-pressed to find any industry that isn't exploring or using AI, and of course, science and medicine are no exception. And I think the attendance here today and online is testament to the excitement and curiosity and interest in this particular area of science and medicine. And I'm personally looking forward to a fascinating session. So we have 6 speakers here today. We'll have a short break about halfway through. After each talk, we will have about 5 minutes for questions. I encourage everyone to submit them via the Slido link, which will be up throughout the presentation, and we'll have an opportunity to go through those questions afterwards, time permitting. So, without further delay, I'd like to introduce our first speaker, Dr. Cheryl Arrowsmith. Dr. Arrowsmith is a senior scientist at the Princess Margaret Cancer Center, professor in the Department of Medical Biophysics at the University of Toronto, and the chief scientist of the SGC. Her research focuses on the structural and chemical biology of chromatin and epigenetic regulatory factors especially as it relates to cancer and drug discovery. In partnership with major pharmaceutical companies, she leads the SGC's International Open Science Program that is developing and distributing unencumbered chemical probes that support the discovery of new medicines. Please join me in welcoming Dr. Cheryl Arrowsmith to the stage for her talk entitled From AlphaFold to Computational Drug Design: Open Experimental Data to Enable AI in Drug Discovery.

Good morning, everyone. I'm very pleased to be here. Thank you to the organizers, and it's such a pleasure to share in the celebration of the developers of AlphaFold. As a structural biologist, this is a revolution in our field and something we've worked towards for many, many years. Myself on the data generation side, protein structure side. Let's see. Today I would like to build on this wonderful achievement and look a little bit in the past about some of the factors that we think helped set up AlphaFold for success, and also how we can learn from that and move forward in the future. So as I said, AlphaFold was was a real breakthrough in the field. It's democratized structural biology so that most scientists can use it now instead of being limited to the experimentalists like ourselves. So we think that the next big challenge in the field, one of the next big challenges, probably as challenging as the protein folding problem, is being able to predict a small drug-like molecule that will bind to a given protein, predicted at will to use for drug discovery and/or for tools to modulate that protein and understand biology. So how will we get there to be able to do that purely computationally? So I will talk about lessons learned from the structural biology community and AlphaFold success. A pilot project that we've undertaken to assess AI-assisted protein compound interaction screening methods, a community benchmarking program called CASH that is modeled after the program that was used to validate protein structure prediction, and then our efforts to enable AI methods with open data in AI-ready format to facilitate the community to use AI in the future. So what does success look like? In the case of AlphaFold, it was participation in this longstanding 28-year program called CASP, which every 2 years would have a competition to assess groups that would predict 3D protein structure from a set of 3D structures that were not yet released or revealed to the world. And what you see here is that the AlphaFold 2 in CASP competition number 14, this is a measure of success score, and you can see it's miles above the other competitors. And so for many years, the plot looked somewhat like this, where the best groups were a little bit better than the rest, And this was really a remarkable breakthrough. So we participated in that whole field in the sense that structural genomics groups were meant to— our goal was to do high-throughput structural biology, not only to understand 3D protein structures, but some of the structural genomics groups, particularly the one in the US, the Protein Structure Initiative by the NIH, was meant to populate the protein database with a variety of 3D protein structures that were each very novel, that didn't look like anything else that was already in there. And so the SGC was among the largest contributors to the PDB of any one group, and the greatest single contributor to the CASP competitions for a number of years. But we then turned around a little over 10 years ago to using the proteins that we made in order to solve their structures. We had a large fridge, if you will, or freezer of recombinant proteins and knew how to make them. And so we started using them in partnership with industry and academic partners to identify what we call chemical probes or bioactive chemical tools to study the function of those proteins. And now going forward, what we want to do, we want to ramp up that process, try to make it more efficient using machine learning learning and AI methods and use a data-driven approach to chemical probe and drug discovery. And we're doing this in partnership with industry, and all our data and output is open science, public domain, unencumbered for the world to use. So we would like to— we can't do it all ourselves, but we would like to catalyze the community. So what is a chemical probe and why are we working on them? Our definition is a drug-like small molecule that selectively modulates the activity of a specific protein in cells and sometimes even in model organisms. It's among the most, or arguably the most useful and impactful research tool that one can have. It's the first step in a drug discovery program, and it's the counterpart of genetic knockdown, but you can use it much more easily and at will and in in a temporal manner, for example. So we've made a lot of chemical probes over the years. Over the last 10 years, about 200 of them. Even now, industry is starting to donate chemical probes from their internal drug discovery programs, and these are all available to the community. Hundreds of thousands— or tens of thousands have been distributed. Many, many publications, over 12,000, using these compounds by others in addition to us. And inspiring over 50 clinical trials from the, from the understanding of the targets that was generated from those compounds. Nevertheless, if you look at this plot here, which plots human proteins ranked at 20,000 or so, ranked by the number of publications that on that gene or protein product, you can see the vast majority of publication in the last over the last 20 years, in fact, since the human genome was sequenced, is still focused on a very small proportion of the human genome. Now we know the structures of most of these because of AlphaFold, which is wonderful, but nevertheless, not much research is focused on them. Unbiased genomics tells us though, from genome-wide association studies in the OMIM database, of the mammalian inherited disease-related proteins, that most of the proteome is associated with disease in some way. So we believe that if we had a pharmacological modulator to every human protein, that it would be absolutely transformational, and you could probe these— the function of these proteins. You could validate whether they are drug targets or not, and it would revolutionize the way we perform science and biology, we think. In order to try to do that, we've launched an initiative called TARGET2035, whose goal is to find a pharmacological modulator for most human proteins by the year 2035. Again, it would be transformational. It would vastly accelerate drug discovery because that's the first step in drug discovery to a protein target. And it would fulfill the promise of precision medicine. However, currently each chemical probe we estimate and our industry partners estimate that it costs between $2 to $5 million to generate, takes a couple of years. And so this is just not feasible with current technologies and methods. Requires a massive advance in technology, particularly we think artificial intelligence, deep learning, machine learning can help with many aspects of this. And it's going to require global collaboration, data sharing in particular, and providing the data that can help train machine learning algorithms. So the challenges in this effort are enumerated in many venues, including commentaries and and editorials, and I've underlined here the key aspects that we also recognize and are trying to address. For example, the— one needs independent verification of predictions made by computational methods.

For—

in the chemistry sphere, and particularly protein-ligand interaction, accurate— there's a paucity of good datasets out there for training of algorithms. And when something's predicted, you know, it needs to be validated experimentally. And, you know, there's different complementary expertise between industry and academia. Most of drug discovery has taken place in industry so far, and we need to work even more effectively together. So we started a pilot project about 2, almost 3 years ago. In the middle of a pandemic to try to assess the value of different screening methods for what we call hit finding. So that's doing a medium-throughput screen or ultra-high-throughput screen going from the left to the right over here of methods where you take a recombinant protein and you test what small drug-like small molecules interact with it. And, and so it's a variety of methods. I'm going to focus on the two here that use AI in some, in some way. DNA encoded library screening, so-called DEL, using machine learning to analyze the output to identify compounds that are commercially available that one could purchase, because we can't test, you know, all billions of the compounds that are screened in the initial screen. And then a purely computational strategy that computational chemists and AI groups and companies can try to predict what's going to bind to a given protein. Okay. And we focused this pilot project on a protein family that's understudied, so part of— a large part of the the dark proteome. These are protein-protein interaction domains that bring together large protein complexes like the one in the upper right. You can see these donut-shaped molecules, there's 3 of them in this complex, actually 4 of them in this complex, holding it together through protein-protein interactions and also DNA-protein interaction here. We were amongst the first to show that this protein family was druggable. These are two targets that are involved in epigenetic regulation, and the drug-like compounds bind in the central pocket here and compete with the protein that they're normally supposed to interact with. So this is the family that we've been working on, these so-called WDR proteins. It's very large, several hundred proteins, very much associated with various diseases, particularly cancer. So the DEL machine learning pilot project that we carried out is shown here annotated on the phylogenetic tree of the protein family. We cloned several hundred proteins. We purified many of them. We solved structures, experimental structures of 21 or so of them. They can all be predicted pretty very accurately by AlphaFold at the fold level, but the details of the side chains that are in the central pocket here are— one needs experimental data to accurately dock small molecules to them. And then we did a DNA-encoded library screen on a subset of these. We took 16 WDR proteins. One of them is WDR5 I showed you on the previous screen that we know is a druggable member of the family, and then we used two non-WDR proteins just to ensure we weren't biasing this collection too much. The DNA encoded library screen was performed at XChem, which is a leader in this field. Machine learning was done by a company called Zebi AI that uses AI to analyze the DEL output with participation from scientists at Google. And we— they predicted on the order of 3,000 or so compounds that would bind to one of the 18 protein targets, which we purchased and then tested at the SGC using a number of biophysical assays, primarily surface plasmon resonance, but a number of other experimental methods. So this was actually quite a bit of work on our part. So what did we find? Out of 3,000 compounds, we identified a spread over all these targets. Each target had a subset of compounds predicted for it. About a third of them, a third of the targets, we could identify at least one low micromolar, so the dissociation constant for the compound from the protein is in the low micromolar range, which is an okay starting point for a chemical probe program. There were 3 more that had very weak, probably not good enough to start with, but detectable. We solved the 3D structure of several of these experimentally, and so this was encouraging, and we think— and we only had to assess about 3,000 compounds over 18 proteins, which is much, much smaller amount than the typical high-throughput screen, which costs over $1 million and takes a long time and a lot of protein and a lot of reagents. We also used purely computational methods. So a company called Atomwise did some computational predictions for us using their their platform, their algorithm, and we identified again— now these are even weaker than the experimental predicted compounds, but they're measurable. And a local company called Cyclica, now part of Recursion, also had one success out of three. And even though it's pretty weak, 70 micromolar, we could co-crystallize it and show it's binding in the desired pocket. So these are weak but drug-like, potent, more potent than fragments, which is a common screening methodology used in some quarters. And the compounds are progressible. Their medicinal chemists say they look okay to try to make, convert further into a drug. And they're the first in class. They're novel compounds for each protein. This encouraged us to— I was actually quite surprised because the computational methods in the past on the proteins I worked on had really not worked. And so the field had seemed to have made some progress. But how do you know? As I mentioned in the beginning, how do you know when you're successful? So we thought we needed a competition, a benchmarking competition similar to the CASP CASP project that I mentioned in the first slide where every— there's regular competitions for revealing which group has made the best predictions. And so my colleague Matthew Shapiro, in collaboration with Alad Edwards and also John Moult, who ran the CASP competitions for protein folding, have established this regular competition called CASH. For critical assessment of computational hit-finding experiments because they all have to be— this is a prospective competition where all the predictions have to be confirmed experimentally. The first competition is almost over with. The data is all over with. The results are being analyzed and they're going to be released in December. The target was a WDR domain from the LRRK2 protein, which is a Parkinson's disease target. And the Michael J. Fox Foundation funded this competition. 35 groups applied to participate, but there's a selection process, and 23 were selected. It was viewed that their approach was reasonable. The first round involved 2,000 predicted compounds. Each group gets to predict or suggest 100 compounds. We tested all of them for binding to LRRK2 using surface plasmin resonance assays. 73 compounds showed weak binding. This information was provided back to the groups that had a successful prediction, and they were given the opportunity to try to improve that, the strength and the drug-like nature of those compounds. In a second round because a lot of the first round was so weak, could have been just chance perhaps. If there's progress in the second round, then it suggests that the algorithm is really working and progress is being made. 4 of the 5 top-performing methods did use artificial intelligence in their algorithm. And as I said, the data will be released soon. So there's 4 of these challenges running right now. And they're going to generate a lot of data because we have to— all the predictions we have to confirm experimentally. And most of the compounds do not bind, right? So there's a lot of negative data and some positive data. In addition— I'll go ahead and— these are systematic self-consistent datasets that can be compared across— consistently amongst one another. We're also ramping up our DEL screening platform in collaboration with Exchem and HitGen, two companies that are working with us on that, and an affinity selection mass spectrometry platform that will also create a lot of experimental data, both positive and negative, of compounds that bind to proteins. And we're hoping that this will provide some of the— open, accessible data that the AI community needs to develop their algorithms. So, where do we put this data? How do we make it available to everyone? Benjamin Head-Caines from Princess Margaret University Health Network and University of Toronto is— this is his space of data science, data, how do you access and store and disseminate good quality data. So he's helping us put together a database called AIR-CHEK, Artificial Intelligence Ready Chemical Knowledge Base. So it's going to have terabytes of data in it hosted on cloud servers available with no restriction. We want to have a low barrier to access. And the data will be formatted specifically for data scientists to develop their algorithms. And we encourage them at the same time to make their machine learning models public so that the whole community can learn over time. So our vision for the future is that chemical probe discovery and drug discovery can be transitioned from a largely empirical and experimental exercise, which is costly, slow, and not always successful, to a more computationally and data-driven method using AI. We think that AI is the method that's really going to help transform this. And so the typical pipeline for drug discovery goes like this. You identify your targets. You identify a small molecule that binds to your protein of interest. To start with, and we think that this is a key area and the area we're focusing on that can really advance this to a faster, less expensive, and more robust method. But then, of course, this is just the beginning of drug discovery. You have to develop that molecule into a non-toxic, specific, bioactive compound that's going to modulate the disease of interest. And that takes a lot of medicinal chemistry and an iterative project process going from new compounds synthesized to biological assays and back to analyzing the data in a cyclical manner. And we're working with the Acceleration Consortium here at the University of Toronto who are developing these methods to facilitate this process using automation and AI. And other methodologies. These are so-called self-driving labs. And the goal is to make public domain chemical probes for everyone to use. The learnings from this process are going to be useful for anyone doing drug discovery, and then we hope we can now democratize, if you will, the, the process of drug discovery. Companies can, of course, make drugs better, maybe this will lower the cost of making a new drug, discovering a new drug. If we know more about the biology, maybe they will fail less often in the clinic, and perhaps the cost of drugs will also go down. So I want to acknowledge all the wonderful scientists at the SGC, and our partners, and our collaborators, too many to mention in one slide, but this is the wonderful group at the SGC Toronto. Currently that I work with, and for the projects that I talked about today, I want to acknowledge in particular Matthieu Shapiro, who leads the CASH program, and Alad Edwards, who is the inspiration for many of these ideas, all the funders and our advisors from various universities and companies who help us. Thank you very much. Happy to answer a question or two.

So we do have time for a few questions. They are rolling in. I encourage anyone in the audience to submit them. We have 5 or 6 minutes here. Just to get this started, so how do you prioritize protein targets in TARGET2035?

It's a mixture of the individual interest of the scientists participating in it. TARGET2035 is a global effort. Anybody who has a target of interest, we hope, would, you know, sort of join the effort and partner with us to maybe do— we help do an assay, et cetera. So the target choice will— will be related to the biology and/or the disease of the person or the group of interest. So, for example, Michael J. Fox funded the first CASC cash competition because that target was a very important Parkinson's disease target. We think this is a way for groups working in rare disease, for example, where there isn't a lot of resources to bring on a given target or a disease, to join in this collaborative effort and get research done on their target, and maybe even a starting point for a drug discovery program. So it varies is the answer.

I guess on a similar line of questioning, how are the cash challenges determined?

We have an advisory group that's made up of people from the pharmaceutical industry and also academics working in the field, and there's a team that evaluates the results in the— it evaluates the results and says, yes, that's a good molecule, the data support that this is a good prediction, and then there's also committee that decides what the challenge should be. So each challenge is a little bit different. The first one, the LRRK2, there was already a 3D structure. There was no known compounds that bound to it. The next one is SARS-CoV-2 RNA helicase, where there are small molecule fragments that are known to bind to that protein in a crystal structure, but no drug-like molecules. And so, and the next one, and there's two more. And so each one is a slightly different challenge in the drug discovery process.

Yeah.

Can you elaborate on how you engage pharmaceutical companies and incentivize open science at the same time?

Right, right. So we're a pre-competitive consortium. So we work in areas where the problem is too large for any one organization to solve the problem themselves. So in the beginning in structural biology, the human genome had been sequenced, and, you know, all the pharma companies were thinking— the structural biologists were thinking, oh my gosh, we have to solve all these new drug targets. So the concept— that's how the SGC was founded, pool the— pool resources together and solve all these new structures and make them public. Then when we started doing chemical probes, similarly work in areas of, here's an interesting target, maybe it's not mission critical to our company, but we're interested in it. Let's work with the SGC and make a chemical probe and see if it really is a good drug target by using the chemical probe to test it in disease models.

We have a couple more minutes, so I'll keep going as long as the questions come in. This one is a bit more specific. WDR is a very large family. How do you deal with target specificity at the R&D commercialization levels? Which ones to select, which ones to invest?

The WDR is a very large family, 350 or so proteins. They all look the same, like a donut, at the fold level, but the middle of the donut, which is the canonical binding site is not conserved. Every one of them is different. So it's not like a kinase where you have similar— a similar binding site that causes trouble for selectivity. So it appears to— each WDR is a new problem in and of itself.

The negative results are also very important. Are those failings available for learning?

Yes, they will be. In AIR-CHECK, they will be. Yeah. OK.

Still got a bit of time. So you mentioned that probes take 1 to 3 years and $2 to $5 million to develop. Do you have a prediction on how much AI could reduce the time and cost to development of a chemical probe?

I can't predict that. I'm the data generator. But what we would love to have is to be able to computationally predict molecules that bind to your target and identify a binder maybe within a month, and then computationally or AI-guided medicinal chemistry that could then improve that compound, make it cell-permeable, get into cells and modulate the target maybe within several months. But that's very aspirational. So a challenge to the AI folks in the audience.

One more. How do you take into account the rapid transitional states of small molecules and how their interactions impact protein structure?

That is a very important aspect of how proteins function. Crystal structures and even AlphaFold structures are snapshots or averages of a protein conformation. The data that I talked about that we're generating is on equilibrium conditions of the protein. It's exploring all its different states in the conditions under which we're capturing small molecules to bind to. Those are two separate things. Then to understand how the small molecule binds to the protein, we often need experimental 3D structure determination because there's many cases where the compound or the drug binds and induces a change in the protein conformation that's not apparent before the compound bound.

Great.

Thank you so much.

Thank you, everyone.

Our next speaker is Dr. Bo Wang. Dr. Bo Wang is jointly appointed in the departments of Laboratory Medicine and Pathobiology and Computer Science, and obtained his PhD from the Department of Computer Science at Stanford University. His work focused on developing algorithms for solving problems in computational biology with an emphasis on cancer subtype prediction and single-cell analysis. He co-leads the UHN AI Hub and is Canada's first hospital chief artificial intelligence scientist. Dr. Wang holds a Canada CIFAR artificial intelligence chair from the Vector Institute. He has multiple first author papers in journals such as Nature Methods and Nature Communications, and he's expanding the AI education in LMP and has recently launched a new grad course on basic principles of machine learning in biomedical research. Please join me in welcoming Dr. Bo Wang to the stage for his talk entitled Building a Foundation Model for Single Cell Omics Using Generative AI.

All right.

Hello, everyone. Very honored to be here. Today I will talk about how to build a foundation model for single cell genomics. So, first, why do I care about single-cell omics? As we all know, cells are the basic unit of life. However, cells can be very heterogeneous. They can come from different cell types. For example, we have T cell, B cell. They may have different cell functions. They may come from, you know, healthy tissues or even diseased tissues. So, there are multiple levels of heterogeneity. Within cells. And because of that, single-cell sequencing technology has emerged to be one of the most important sequencing technologies for the past 5 to 10 years to understand the heterogeneity of cells. And why do I care about foundation models? First of all, what are foundation models? This is a term you probably hear a lot, particularly post-ChatGPT era. You probably heard a lot of such terms through social media. However, on the academic setting, this term was first coined in a seminal paper co-authored by many Stanford researchers. You can— I show the paper on the slide. You can see there's probably 100 authors. And really, the core idea of foundation model is that we can train these foundation models through massive amount of unlabeled data on the internet, for example, such as text, images, speech, you name it. And then once the foundation model is well trained, we can adapt this foundation model to many downstream tasks with few-shot or even zero-shot learning. So this paper is about 300 pages long, so too long, don't read. I summarize the key features. To build good foundation models as ABCD. So here, A, algorithms. Most of foundation models are based on transformer-based deep learning architecture, which I will briefly talk about later on. B, business. The idea of foundation model, as I mentioned before, is one foundation model can enable multiple tasks by transfer learning. And C, computing. Most of foundation model needs specialized hardware such as GPU or even TPUs. And the data, and most of foundation models do require massive amount of data during pre-training. And this prompted me to think about the following question. Since I'm a researcher in single-cell genomics, this is my PhD thesis when I was in Stanford. So I asked myself a question, can we build a foundation model for single-cell data? The motivation is quite clear to me. While texts are made of words, cells can be characterized by genes. So this analogy really inspired us to explore foundation model for single cells. Although I do— I'm not claiming that biology and languages share the same mechanism, but because there's enough technical similarities, this motivated us to explore these directions. But back then, we thought maybe this is a crazy idea. So let's see whether we have all the key features, ABCD, as I outlined before. So A, algorithm. As I mentioned, most of foundation models are based on attention. This was first proposed by a seminal paper in 2018 by Google researchers. It's called Transformer. Basically, you have— very briefly, you have encoder, decoder. Encoder takes the input data and projects into a latent space. And the decoder takes the latent space and tries to trying to reconstruct another form of output. So if you stack enough encoders, you have BERT type of model, which again, invented by Google. And also we have similar models in single-cell genomics, such as scBERT, scFormer, GenFormer, which is recently published in Nature. On another angle, if you stack enough decoders, you will have GPT. Type of model such as ChatGPT by OpenAI. And really inspired by GPT type of model, we have our own SCGPT, which will— this is the main topic of this talk. Okay, business. We have lots of application tasks in single-cell omics. For our work, we validated on 6 different tasks such as cell type clustering, batch correction, multi-omic integration, cell type annotation, gene regulatory network, and perturbation prediction. So, there's no shortage of variety of applications of business in single-cell omic. See, community, not much to say, but we do realize that many AI models need lots of GPUs. Thanks to— shout out to Vector Institute and Compute Canada. They do help us a lot with GPU access. Okay, D, data. Everyone says data is the new oil for AI, and thanks to Human Cell Atlas, which is an international initiative founded by CZI, and we have access to lots of single-cell data, well-curated. And specifically in this paper, we took 33 million healthy cells from cell-by-gene census from CZI, And these 33 million cells cover mostly 8 major tissues such as heart, blood, brain, lung, and kidney, and et cetera. And this covers a huge variety by covering almost 500 different studies and datasets. So thanks to this initiative, we start to have lots of data to start to really think about the foundation model for single cells. So looks like we have all the key elements. That's when we start to really think about, can we build SCGPT? So briefly, SCGPT consists of two steps. First step is generative pre-training, and the second step is fine-tuning towards many downstream tasks. And I will walk you through these two steps one by one. So let's look at the first step, generative pre-training in SCGPT. As I mentioned, what we did is we collected 33 million human healthy cells covering major tissues such as heart, blood. And then we took these 33 million cells and then trained a GPT-style model. And we used about 16 A100 GPUs and trained for 2 weeks. And then we have the model. We didn't— we weren't sure whether the model is well trained. So what we did is we randomly subsampled 30, 30,000 human cells and then draw a UMAP, which I show here— sorry— which I show above. And then we do find a cell type hierarchy. So that gives us confidence that this pre-trained model was well trained. And dive deeper into more technical aspects of STGBD. The input to STGBT is 3 sets. First is basically gene token—

gene tokens.

Basically, we tokenize all the gene identifiers. The second input is gene expression values. What we did is we binned those gene expression values into 55 bins and then tokenize these bins as well. The third input to scGPT, we call it condition tokens. And this design is really innovative in the sense that it gives us lots of flexibility flexibility to incorporate the metadata for human cells. For example, we can use these condition tokens to denote whether which gene were perturbed, if it's perturb-seq, and we can use these token— condition tokens to indicate which batch it comes from, which gene sequencing technology it comes from. So, this design of condition tokens really give us lots of flexibility to incorporate prior knowledges. And once we have the inputs to SCGPT, the algorithm is nothing novel really. It's a stack of transformer blocks. As I mentioned, this is a major type of algorithm we use to train foundation models. And the key innovative in SCGPT is the design of the generative pre-training. So as we know, in GPT-type of models, we train the model in an autoregressive manner because language has orders, which makes sense. However, when it comes to single-cell omics, the order of genes does not matter that much. So how do we train it in an autoregressive way is that we use the cell embedding, basically the COS token, in STGBD as a cell embedding, and then we calculated the attention of every gene to this particular cell, basically indicating the importance of genes that is specific to this particular cell. And once we have this order of importance, we basically have a set of orders of genes, and then we chunk these set of genes into different chunks, and then we treat them as ordered chunks of genes, and then we train them in an autoregressive way. Basically, we mask some future chunks of genes. We're trying to predict these masked genes using the previous predicted genes. Certainly, there's lots of nuance in terms of how to select hyperparameter, how to select these chunks of genes, which I won't talk about due to time limit. With this design of causal masking in SCGPT, we have a way to train SCGPT in an autoregressive way. So now we have a well-trained foundation model, SCGPT. Now we want to talk about the second step, which is fine-tuning. So there's lots of innovation we made in terms of design loss function for fine-tuning. For example, when we have multiple batches of single-cell RNA-seq data, how do we design loss functions to do fine-tuning? So we designed two types of loss functions. One is kind of a semi-supervised way, which we enforce the cells in neighbors have similar embeddings. And the second type of loss function, we call it adversary loss, which basically we're trying to force the model to not distinguish which batch this cell comes from. So that we remove the batch effect. So once we design the loss function, the way to fine-tune the model is quite straightforward. We basically use a standard pipeline such as backprop with SGD, for example. And then we did that. And in order to test the efficacy of SCGPT, we collected multiple human blood cells, PBMC 10K, which is one of the popular blood single-cell data from 10x. Then we compare how good it preserves the cell types and how good it removes the batch effect. We reported the average BIO, which is the average value of MI, RI, and ASW, which is metrics to see how well you preserve the cell heterogeneity. You will find that AcidGBT is through a bit of fine-tuning, can well preserve the cell type heterogeneity while removing the batch effect. And it outperformed many, many specialist approaches such as SRAv3, SVI, and Harmony. These are specialized design machine learning algorithm specifically for batch integration. We then move ahead to do multi-omic integration where we not only have RNA-seq, we also have other type of omics such as ATAC-seq. And we, again, we test it on PBMC human cells, and we find that similarly we outperform specialist approach such as SRED v4 and scGLUE. But note that we didn't have a foundation model for ATAC-seq. We only have foundation model for RNA-seq. What we did is simply mapped the bins from ATAC-seq to genes. So, this is a very rough way to map ATAC-seq to RNA-seq, but we didn't have a foundation model for ATAC-seq due to data limitations. And then we even move ahead to test a more challenging setting we call Moisaq integration setup in the sense that we don't have paired RNA-seq and ATAC-seq. What we have is unpaired RNA-seq, ATAC-seq, and even proteomic as well. And we find that, again, tested on PBMC, we find that scGBD outperformed many specialist approaches such as scMoMAT in integrating different omics. And then we move ahead to test cell type annotations. So, we collect 3 datasets covering, you know, immune cells, covering cancer cells, and we report the accuracy F1 value of cell type annotation, and you will find that Again, STGBT outperformed many specialist approaches that is specially designed for cell type annotations. Also, many people are wondering, can you just simply do zero-shot when you have new data to map the cell types? We did that. What we did is we took the COVID-19 dataset which were never seen before during pre-training, and then we run zero-shot and we actually find very high accuracy in terms of reference mapping for cell types. And because of this success, we are working again with Human Cell Atlas to incorporate this functionality into Cell by Gene portal so that when biologists have their own data, they simply run our scGBD with zero-shot, which is very lightweighted, and then get all the cell type annotation in seconds. And we also find a very interesting phenomenon we call a scaling law of scGPT in single cells. What we did is quite simple. So, we randomly sampled 30,000, 300,000, 3 million, 33 million cells from Human Cell Atlas, and then we construct scGPT and then report the accuracy on cell type annotations. And we find two phenomena. One is, which is not very surprising, the more cells you pre-train on, the better accuracy it is. Another interesting observation is that once you surpass 3 million cells, the performance can start to diminish. So this gives us lots of insights about how many cells you need, what type of cells you need during pre-training. And we also test on perturbation response, again, by reporting the accuracy on a few common database, and we find that SCGPT again outperformed many specialist approaches such as GEARs in gene perturbation predictions. A very good byproduct of SCGPT is that not only we have embeddings of cells, we also have embedding of genes. So gene embeddings really serve as a good vector to understand some of the biology and interpret SCGPT. What we did is we perform differential pathway analysis in comparison to traditional pathway analysis, just co-expression network. We find that through gene embeddings output by SCGPT, we can actually find a lot more relevant differential pathways in comparison to just simple co-expression network. We also did a very interesting experiment. That we build SCGPT per organ. For example, as I mentioned, for Cell Atlas, we have 8 major organs. Some organ has more cells, some have less. And then we want to test which foundation model is better. Is that the organ-specific foundation model better or the human healthy cell, the general foundation model better? We test it on COVID-19 because we want to test out-of-domain generalization. And we find very interesting phenomena is that if you take the foundation model trained on blood and the lungs, the accuracy on COVID-19 data usually is higher than the one trained on irrelevant organs such as, for example, brain or pancreas. However, the overall human health foundation model also serve as a very good default models to start with. And realizing, as I mentioned, realizing not every wet lab, not every biologist have access to GPUs. So in order to enable access of scGBD to everyone, we built scGBD Hub in collaboration with Amazon, particularly the cloud providers, AWS. And we built this scGBD Hub I hope I can play the video. Yes. So because of this portal, basically biologists now can simply log in this portal and upload their data. And then they choose which foundation model to start with, either lung or kidney. And then you set a few hyperparameters. And then you submit your job. And then you go out for a coffee. And then once you finish coffee, you come back. The results is sent to your emails, and then you can download the results to visualize, for example, to analyze your biological data. And also, we only published this scGB hub for about 2 months. We already have thousands of wet labs sign up to use this scGBD. All right. I think we're running out of time. So, talk about lots of advantages of scGBD, I also want to acknowledge some of the limitations of scGB. The first limitation I see is that scGB is mostly limited to RNA-seq due to data limitations. We— although we did multi-omics integration of ATAC-seq, but we actually don't have any foundation model for ATAC-seq. But, however, we are working with Human Cell Atlas to start to accumulate and collect lots of ATAC-seq data, so hopefully we can build another foundation model for text. And also, again, we don't think SGD is causal. It's still guilt by association, particularly all the analysis we did for gene regulatory network. And also, pre-training is very compute intensive. However, the good news is the fine-tuning is very lightweighted. And also, the way we hope we can move forward with foundation model in single-cell genomic community is that we can move from hypothesis-driven approach to data-driven approach, but this data-driven approach can also incorporate some of the expertise priors. This is what we hope scGB is leading. Also, many current practice in single cell start with very small single lab validations with large data consortium such as Human Cell Atlas and scGBD. We hope we can move to large scale data and multi-center validations for any biological findings you have. And also, currently for many single-cell pipeline, it is stepwise. All the steps are designed based on different requirements, and then the input to the next step requires the output from the previous step. What we hope scGB can bring the change to is a general use framework and carefully design end-to-end, even optimized with specialized hardware such as GPUs. In order to achieve those goals, we do think collaboration is key. We need collaboration between computer scientists like myself and with many biologists to really work together to find the novel biology or hypothesis. All right, with that, I want to thank all the key contributors, my students, Haotian and Chloe, and also Connor and Gary for help with STGBD Hub. And certainly most of the work got done by my lab. I'm just the one who speaks here. All right. Thank you very much. Happy to take questions.

[APPLAUSE] Okay.

I encourage everyone to submit your questions. We do have a few already. What do you think of the recent work from Kedurski et al. showing "Poor performance on zero-shot tasks for SCGPT and GeneFormer. Do we always need fine-tuning?" Yeah, very good question.

So I think I also responded on Twitter that zero-shot kind of validation is not really showing the purpose of foundation models, as also what we showed on our paper. And most of tasks do require fine-tuning. However, we find that zero-shot on reference mapping is actually very powerful already. That's because Human Cell Atlas has detailed annotations on the 33 million cells. This is an area we hope we can improve upon. Currently, SCV is very tiny. We only have 100 million parameters and trained only on 33 million. There's a long way to go to really improves the zero-shot capacity.

What type of genomic data is currently not available or not large enough but would enable future breakthroughs in your field?

Yeah, very good question. So I think most of common modality is RNA-seq. As I mentioned, that's why we have this SEGBD, mostly trained on RNA-seq. ATAC-seq start to emerge a lot more. However, in my view, I think proteinomics, will serve as one of the key players in the future to really bridge the current RNA to clinic side of clinical integration.

We actually have quite a few coming in. Have you considered non-autoregressive methods for generation during pre-training?

Yeah, very good question. So there's already a few other type of models that are using basically encoder-only models, such as Geneformer. And actually a few external validations already verified that SCGPT, which is trained in an autoregressive manner, outperforms Geneformer in almost every task we can find. So this gives us confidence that through this autoregressive training does improve the capacity. Certainly there's many other factors playing a role here, but we have we were very interested in this autoregressive training and we find that it is more powerful than just finding embeddings of cells. Because in order to predict next genes, you actually need to understand the data distribution within single-cell RNA-seq data.

Can you elaborate more on what the main reason that fine-tuning gives better results compared to training from scratch?

Yeah, very good question. One reason we start to speculate is that some of the ground truths were obtained through very simple kind of linear methods on that specific data. And then if you didn't see the data and find differential genes, it's very hard to actually get correct particular cell types. So that's why fine-tuning, at least you kind of let the model see the data once gives you a lot of advantages to get better accuracy. Another reason is just the current foundation model is not powerful enough. As I mentioned, it's still very small scale. So maybe once this model is better trained with more data, with more parameters, it can improve the zero-shot.

So as UHN's first chief AI scientist, where do you think Toronto is leading the way in the AI health healthcare space, and how can we compete internationally?

A very good question. I think Toronto is a very unique place in this, particularly for healthcare, because we have the most diverse, you know, populations. And, you know, bias towards certain groups is a big issue in AI for healthcare. And also, University Health Network is, which is one of the largest research hospitals in the world, have access to millions of patient data. So this gives us lots of resource to really train unbiased AI model for healthcare. And also, we have the kind of fast track to actually validate some of the models and also implement the model into clinic and see how it performs with respect to different settings. So this really gives us advantage in terms of AI development and AI deployment.

Can you elaborate on why fine-tuned GPT outperforms traditional pathway analysis methods? More pathways identified does not necessarily mean more accuracy.

That's a great question. So we actually did a very fine-grained comparison of what type of pathway did scGB find, what type of pathway did co-expression network find.

And as you—

yeah, as I shown in the slides, the difference mainly lies in that scGB find a lot of immune-related pathways. These are important because that dataset was immune cells, and co-expression often find many housekeeping genes or pathways, which are not super exciting for biological discoveries because co-expression is very noisy when you have, you know, for example, 20,000 genes.

We have a couple more minutes, so they're still coming in, which is great. Thank you, everyone. What kind of applications in health do you foresee for SCGPT?

So, so far we mostly focusing on biological discovery. We haven't we don't really see a great pathway to clinical integration of seqPT yet. That's because we don't have much patient-level single-cell data. Because, again, single-cell data is very costly. It costs a few thousand dollars just to get a sample sequenced. So hopefully as sequencing cost decreases, I think the trend is very clear, we're going to see more and more patient-level single-cell data. That's when I see great opportunity for foundation model like Geneformer, SCGVD can play a role in clinical applications.

3 million cells seems to be enough for cell type labeling. For rare cell types, would you recommend to stratify downsampling or what's their minimum number of cells?

Yeah, amazing question. So unfortunately, I don't have a great answer about what's the minimum required number of cells, what we find is that for major organs, 3 million is a great threshold. For rare cell types, it's always very hard because it's very rare and some of the cell type annotation is not very accurate. We're working with Cell Atlas to find what a great threshold for defining rare cell types. Unfortunately, there's no consensus within this community yet. Also, it varies depending on which field you are studying. There's always a joke saying if you put 5 immune cell researchers together, they always fight because there's no consensus about the definition of cell types. Hopefully with more accumulated studies from the field and with consortiums such as Human Cell Atlas, people start to build guidelines, principles to define these different cell types.

I've got time for one more. Yeah. Oh my gosh, there are so many for you. If you were to pinpoint a single challenge and a corresponding strategy for fine-tuning a pre-trained model, what would they be?

Very good question. I think almost always I would suggest to do fine-tuning at this stage. And then you have to design specific loss functions for your own tasks. We provide a few examples in that paper. Different tasks do require different loss function design to really reflect what you need. Basically, you need to ask the question, what what outcome you want from SCGPT for this particular task. And once you have a clear answer to that question, you can design your loss function. The rest of fine-tuning is quite straightforward. For zero-shot, again, pre-training, it's not very powerful yet due to data limit.

That's all the time we have. Thank you so much. And sorry we couldn't answer all the questions.

Thank you, everyone.

Our next speaker is, and I apologize if I get your name wrong because we didn't have a chance to meet before, is Dr. Duena Precup. Dr. Precup is an associate professor at McGill University and conducts fundamental research on reinforcement reinforcement learning, working in particular on AI applications in areas that have a social impact such as healthcare. She's interested in machine decision-making in situations where uncertainty is high. She's a senior fellow of the Canadian Institute for Advanced Research, fellow of the Association for the Advancement of AI, and she also heads the Montreal office of Google's DeepMind. Please join me in welcoming Dr. Precup to the stage for her talk entitled Reinforcement Learning for Therapeutics.

Thank you very much for the introduction. So I will talk today about reinforcement learning. This is a little bit of a different learning paradigm from self-supervised learning, which we've talked about before. I do split my time actually between Miguel and Google DeepMind, but the work that I'm going to talk about today is carried out at McGill. In particular, was spearheaded by two of my PhD students, Shimona Basu and Emmanuel Bengio. Now reinforcement learning is about learning from interaction with an environment, and it's about learning how to choose actions, right? And it's a paradigm that's been inspired actually from animal learning theory, where you can have an animal that interacts with its environment, takes actions, and and receives a numerical reward signal that's sort of conveying how good those actions have been, and animals are very good at learning this kind of association between the numerical reward and their actions, even if this reward is delayed. And so we think about digital agents as learning in a similar paradigm where they receive observations from an environment, they can choose actions, and they receive numerical rewards. And part of the appeal of this, as Rich Sutton has said in a 1995 1992 paper is that it's really in a sense a whole AI problem in a microcosm. Now the question I'm going to try and address today is whether reinforcement learning can actually help therapeutics, right? And part of this is to try and understand how we can formulate some problems in terms of reinforcement learning. What are the observations or the states of the system? What are the actions that can be chosen? What's the reward signal? We can think about where do we get the training from, right? Is there a simulator for our problem or are there good datasets that we can leverage? And then I'm also going to talk a little bit about whether current reinforcement learning algorithms are actually adequate for this problem and if not, sort of what's missing. And I'm going to do this by looking at two rather different case studies. And the first one that I'm going to talk about is reinforcement learning for precision drug dosing. This is work that we published at AAAI this this year. And just to sort of situate the problem a little bit, one of the current paradoxes in drug development is that clinical trials provide evidence of efficacy and safety of medications by looking at populations. And at the same time, physicians actually treat individual patients. And these patients could be rather different from the population. And in fact, when you have a randomized control trial, you oftentimes select patients for this trial in a way that's not reflective of the population as a whole. So, for example, participants that are in certain age ranges might be excluded. They might be excluded because they are pregnant or because they are taking certain medications and so on. So, there's lots of exclusion criteria, which mean that in a lot of randomized control trials, we actually do not have a representative representative population sample. And we would like to have a way to take what's recommended from a trial and then actually try to adapt to this in order to sort of avoid suboptimal dosing, right? So suboptimal dosing also often affects subpopulations. For example, women are more likely to receive doses that are too high because maybe they're not part of these clinical trials. So precision drug dosing is a problem that is individually tailoring the dose of a drug in order to get the greatest benefit and the least risk for each patient. And in this particular case study, we looked at whether we could use reinforcement learning for precision drug dosing. And so the idea is, can we actually adapt a treatment strategy to a particular individual by looking at the way that their body responds to the drug and trying to sort of adapt the dosing as well as the timing of the doses for that. Now, there's a challenge for people who know about reinforcement learning, which is that in reinforcement learning, we think of an agent that takes an action and the effect of that action is immediate. Whereas when we are administering a drug, actually there's an effect that is delayed and prolonged. So there's some time until the effect sort of kicks in and then it sort of peters out over time. And so in order to sort of take that into account, we need to modify the framework of reinforcement learning algorithms a little bit in order to account for this, which is sort of a feature of the problem at hand. And in order to do this, we actually looked at the pharmacology literature and some of the assumptions that are captured there. And there's a particular way that we capture action effects that's based on this literature. The details are not really important. What I want to show is a sort of an application of this in glucose control. Okay? And why we chose glucose control? Because this is a case where there's a sort of widely accepted simulator developed in Italy that many groups have been using over time. And so we can actually train a system 100% safe in simulation and kind of examine its effects, the effects of the algorithms, and run comparisons. The other reason for this is that glucose control typically you have one variable that you mainly look at, which is insulin levels, and so it's kind of a simplified setup compared to many cases where we have lots of different variables that we need to take into account. So, we're going to assume that we observe this blood glucose level, we have an action which is to administer a discretized insulin dose, and we have a reward function. Now, the reward function is kind of interesting to look at. Don't look at the details, just look at the size of it. Okay. Why is it so big? Well, because actually it's very tricky to figure out what's the right way to reward these systems. This has been a consistent finding in all of the applications that we've tried to do, that the reward function actually has to be sort of crafted with care. And again, sort of talking to the people in the field and trying to understand what they take into account and then trying to somehow formalize that in the math. Now, what methods are we looking at? We're looking at simple threshold policies. So that means, you know, just whenever the insulin is— whenever the glucose is above a certain level, we're just going to administer an insulin dose. This is sort of a heuristic. We're looking at deep Q network, which is a standard reinforcement learning algorithm that was actually pioneered at DeepMind early on. And then we're looking at our method, which is effective DQN. It's a way of incorporating as part of the system's observation its sort of estimate of the effects of the action. And these estimates are sort of geared based on the information that we're getting from this particular patient. And ADRQN is another way of incorporating history, but it doesn't necessarily think about the delayed effects of actions. It's just sort of uses the usual reinforcement learning assumption that actions sort of have their effects being instantaneous. And so we can look at the results here. On the sort of the rightmost graph, you see performance, so high is good, and on the other graph you see computational cost, right, of this method, and so low is good. And so what we can see here is that the method sort of in terms of quantitative measurements is obtaining very similar results to the other method that's looking at sort of the history of actions, but at a much lower computational cost. And the cost sort of stays low regardless of how far back we're looking at this history of actions. But the other thing that's really important in medical applications is actually to look qualitatively. Like, what are these policies actually doing? Because we can measure rewards. Right, but we also want to know, are there any particular really bad side effects, right? Are there any catastrophic events and so on? And so this is a little bit of a more complicated graph. At the top, we have blood glucose levels. In the middle, we have sort of what the policy is doing when it's administering insulin. And then on the bottom of each of these sets of panels is the food intake of this particular person. Simulated person. And so, one thing that we can see is that first of all, there are in all cases except for our method, there are some bad hypoglycemia events. So these are cases where the top graph goes below a sort of low threshold of glucose, and that's really bad because that basically means you've killed the patient, okay? DQN, which is this sort of standard algorithm, learns very quickly to just administer a flat low rate of insulin regardless of anything that's happening, okay? And so, that's not very adaptive, and also it means that we're basically giving too much. A little bit all the time, it sort of accumulates to too much. ADRQN, which is this algorithm that sort of existed in literature before, treats around meal times and whenever the blood glucose spikes, but treats very aggressively. Okay? And that's partly because it doesn't take into account the fact that once you administer a dose, it is going to have an effect over some period of time. And so, it just wants to give as much as possible right now because it thinks that effect is just going to diminish. Our algorithm gives sort of constant, sort of spaced out smaller doses over time, and so it does not end up with hypoglycemia events. It does have hyperglycemia events, which are not great either. Okay. And that is partly because the reward function still is not that great. And so that's something that still requires improvement. So just to conclude sort of this part of the talk, in our opinion, reinforcement learning can provide some effective learning of precision dosing policies, but the reward design is crucial as well as incorporating specifics of the problem, in this case of drug dynamics. Also, the sample efficiency of reinforcement learning definitely needs to be improved. So in this case, we simulate patients, therefore we can give them really, really, really long lives. But actually, this is a problem. We really need to get better algorithms as well as better simulators. This is a simulator that's accepted in the community. It's still quite simplified in its assumptions. And evaluating the results is also problematic right now. We're still evaluating in simulation. There's a big step that needs needs to be thought out about how do you then take that into the actual real world. So now I want to show you a very different case study where reinforcement learning is used not to tailor treatments and doses, but actually for the problem of drug discovery, so finding new molecules. And this was work that was spearheaded by Emmanuel Bengio while he was a PhD student in our lab. So here we're sort of focusing on small molecule drug discovery. This is where we want to find molecules that bind to particular protein targets. And this is basically just a huge search problem. There is a space of molecules. We need to sort of go through the space and find candidates that are promising. And one way to think about it is that we have sort of two things that we need to handle. One is how do we search efficiently? So which molecules do we actually generate to look at? The space is something like 10 to the 16 to the 10 to the 20, and most of the molecules here are bad. They either don't bind, they're toxic, they can't be manufactured, all kinds of problems. And so basically it's a needle in a haystack kind of problem. And reinforcement learning in principle, it has been used in order to find strategies for this type of problem. This is why we were interested in this. The other aspect is how do we actually tell if a molecule is good or not to bind to a particular target? And so sort of traditionally, this has been done by sending batches to a lab and then having that lab look at them. This is a very lengthy process. More recently, there are noisy physics simulators that can be used. It's still expensive, but we can sort of approximate some goodness scores, for example, a binding score. And so that's what we're mainly going to use in this work. So the way that we think about this is we're going to generate molecules. We have a cheap reward predictor that we train from data. We're going to sort of then train a generator that kind of searches through the space of molecules. If we have a set of candidates that's looking promising, a slate, we are then going to really evaluate it. Maybe by using a real lab, a chemistry lab, or maybe by using the sort of more expensive simulation. And then we're going to get that data back and incorporate its outcome into the reward predictor. And so the reward predictor is trained here. So how do we formulate drug discovery as a reinforcement learning problem? Well, we need to think about what are the states. The states are just configurations of molecules, and there's a standard way to represent this digitally. We think of the actions as things that we tack on to these molecules or possibly that we can remove from these molecules. And then at some point, we end up with something that we think is interesting and then we evaluate it that gives us a reward. Okay? And so in this case, our reward is going to be an approximation of the docking score that we have trained from this data. In principle, we can just apply reinforcement learning. The problem is that reinforcement learning traditionally is trained to to find a good way of choosing actions, which means that for a particular problem, we would find one candidate, one molecule, and then every time that we run the system, we would find the same one. And so what part of the goal of this research project was to understand how we can sufficiently randomize and provide a diversity of candidates that are actually all potentially good. There are other methods that are geared towards this. Markov chain Monte Carlo, for those who know, is one a particular example of such a method, but it's pretty slow and actually can also get stuck. And of course, generative models, like we've discussed in the previous talk, actually these are also sort of more recently a really good way of constructing all kinds of data, including possibly molecules. But in our case, we really want to be more efficient. And so we don't just want to construct interesting high-reward molecules, we also want to avoid things that are low reward. So we published in 2021 a paper called GFlowNet. It's a generative framework for discrete objects that is essentially understanding how to sample the space in a way that we get molecules in proportion to how good their reward is. Okay? And so just to give you a sort of a picture of of what this is like. We have a graph. Imagine that each sort of node in this graph is actually one of these molecules. Each edge is maybe adding something to this molecule. And we want to sort of traverse this graph in such a way that we end up at the square nodes, the sort of terminal nodes, in proportion to how good they are. Okay? And so one way to think about it is that if you had the whole graph in mind, You could send these particles down and you want them to pipe through these edges in such a way that we end up at the end nodes in proportion to the rewards. And then you can actually write down the math of this for a directed acyclic graph, which is what we have here. And we end up with an equation which made me very happy because it looks like a reinforcement learning equation. And so we have a big arsenal of tools that we can use to solve approximately these kinds of problems. Now, this is sort of the nice picture, and it's got some nice theorems in the paper and so on. The practice is a lot messier because actually if I just take that objective function and try to optimize it, it turns out that it's not very well behaved. And so we have to do some nitty-gritty details on the reward specification and on the objective function specification in order to make it well behaved. But with this, we can actually really get these sort of systems, which are essentially deep networks, to sample the space of molecules pretty effectively. And so here, we are looking at the problem where we pre-train a reward function on 300,000 molecules. And we are looking at how many modes— that means sort of different kinds of good molecules— are we able to find. As we sample more and more states in the system. And so the blue curve there is the GFlowNet approach. We're comparing it against a more traditional reinforcement learning approach called PPO and against what was at the time a state-of-the-art most sort of molecule generation program called MARS. And so in the rightmost graph, you can see that we find a lot of rewarding modes a lot quicker. So that also saves some computation. On the other graph, we're sort of looking at the average reward of a slate of molecules that we're getting. Right? And so again, what we're seeing is that we are obtaining good slates of molecules quicker. Right? And so this would sort of speed up the process of actually examining these. We have also looked at the problem of essentially doing batch active learning where we put the system in a loop with retraining the reward. And so we make a slate of candidates. We sort of ship it. We get a better reward estimate, we retrain the reward model. And in this case also, our approach performs better than sort of the state-of-the-art program. And we're looking in this case both at picking sort of 10 molecules at a time as a slate, also at picking 100 molecules at a time as a slate. And this is because sometimes we can only work with the lab and sometimes we can work digitally. And then, of course, we can evaluate more at a time. So this is sort of a generative model for discrete objects based on a reinforcement learning-like learning objective. It's a general setup really for solving black box optimization problems. It allows exploration, and the interesting thing is that unlike in reinforcement learning where eventually the way of moving through the sort of space would become deterministic, here we have sort of stochasticity that never goes away, and that's by design. It's easy to incorporate information about uncertainty in the reward function. And so for example, if we knew how bad our docking score estimates are, if we had some error margins, we could incorporate that in the process. And in general, it's a framework that can be used for causal modeling or discovery data explanations. And there's actually a large group of students at Mila that's sort of exploring various aspects of this problem. Again, I want to emphasize we need to design the reward function appropriately. And so here, it was more in terms of the numerics of it, right? But that's really a big part of making this kind of system work. So in my opinion, reinforcement learning actually has a lot of potential in terms of solving therapeutics problems, whether that is treatment design or drug design. The reward function is really crucial, as well as finding a good way to incorporate the the specifics of the problem. And I'm going to echo something that Professor Wang said before, which is we really need collaboration, right? A big key of making these kinds of projects work is actually to have an interdisciplinary team where there has to be a discussion between the computational scientists and the medical personnel or the biology personnel to understand how the specifics can be incorporated. From my point of view, it's also interesting because it's a great opportunity to improve existing algorithms and to actually, for example, understand if we can get reinforcement learning to work in a better sample efficiency mode. So thank you very much for your attention and happy to take questions.

Thank you so much. So we just have one question come in. You've left people speechless, so I encourage everyone to submit your questions, or I can give the mic if anyone wants to do it in person. So, could you use something like RLHF for optimal patient treatment?

That's a really interesting question. So, just reinforcement learning from human feedback is basically this kind of methodology which looks at two solutions and decides sort of, you know, you have a patient decide which one is better. And then we incorporate that information into our reinforcement learning algorithm. A place where this could be really interesting actually is where we can have trade-offs, right? So if you have medications, they have side effects. Some patients would prefer more medication despite the side effects. Others really mind the side effects. And so that's a case where I can imagine something like this being taken into account in order to kind of provide something on a Pareto front that is adapted to the patient. It's, as far as I know, it's not really something that's actually been done, but perhaps also because this whole field of using reinforcement learning for medical treatment is quite new. It, you know, a lot of the work that's being done has been more supervised learning, unsupervised learning, self-supervised, these kinds of methods that don't actually require action.

What experimental approaches are you taking to validate the AI models you've built?

That's a really great question. And again, as I mentioned, it's one of the challenges. So one way that we do it is right now we are confined to simulation. And so we sort of— one way to sort of approach the problem is to train on some simulations and test on others and have a wide enough space of these simulations that allows us to have diversity in that kind of process. Ideally, we would eventually test in the population, but that is very hard to do, right? And we need to really make sure that that safety is respected. In the case of the molecule discovery problem, we are actually working with some collaborators in the chemistry department at McGill, as well as some people in the Faculty of Medicine. And so they are helping us validate some of the molecules. And eventually, again, I think that a lot of this is going to end up being an interdisciplinary effort of trying both to develop the tools better as well as to understand how to validate that scale.

Can reward function be learned from the existing data?

Yes, whenever we have enough data that is reflective. And so this is exactly what we are doing, in fact, in the sort of molecule discovery case. In the diabetes case, we did not attempt to learn a reward function from the data because we sort of— we went into it basically thinking this is actually pretty straightforward. We're going to do the simplest thing that you do in reinforcement learning, which is you reward how long the patient is living. And we quickly found out that this is just not something that is going to work out. And so then, you know, Sumana, the student sort of leading the project, spent a lot of time actually trying to understand how to make this reward function reflective of various considerations, right, including not just how, you know, whether you've exceeded a bad threshold or not, but how close you are to it. In the end, I don't think that it was, you know, in hindsight looking at the reward function, it's clear what it's doing, but it was a lengthy process. And it, in my experience, it has been sort of repeated every time that we've tried to do an actual application. And I'm not entirely sure that this can actually be automated with the amounts of data that we currently have. If we had more data, that would be great.

What regulatory roadblocks must be overcome to allow personalized dosing? Is this limited to clinical trials, to post-trials after approval, or can it help both?

I think in the long run, probably we will figure out better ways of doing things. Right now, it's still, you know, the practice is still randomized controlled trials. A lot of this comes into play really when you have treatment that extends over a long period of time, so more situations like chronic diseases and so on. And so I think, I know that there was a lot of interest at the FDA in finding sort of new ways of validating some of these sort of treatments. Obviously, there always has to be medical personnel in the loop. I think if there was a way, for example, to deploy some of these strategies in a limited context and get feedback from the patients or from the medical personnel about how well they are doing, this would really go a long way. But we, you know, in terms of validation, p-values are very clear. This other kind of sort of statistics that are involved are a lot less clear.

How do you view reinforcement learning as a domain for black box exploration and optimization?

So I think reinforcement learning really has a lot of potential to improve, in fact, black box exploration and optimization. That's kind of where this sort of GFlowNet technology is going. And that's partly because there is a lot of methodology that we have figured out in reinforcement learning that can be successfully used in the setting, and that's sort of, you know, what we're piggybacking on right now in sort of this like little GFlowNet student community that's been developing is actually trying to bring some of the sort of standard ideas in reinforcement learning, you know, experience replay, prioritization, hindsight learning, and so on to this kind of methodology. So I think we're going to see more of that sort of coming into play.

Do you think that RL can be used as part of the fine-tuning process for transformer-based generative models?

It is actually used in that respect, certainly in the language generation, text generation community. That is, reinforcement learning from human feedback is sort of the standard way of doing fine-tuning.

Are you participating in the CASH competition for your small molecule discovery program?

We have not done that, no. Maybe it's something that we should consider. That's a good point.

What is enough data in the case of RL? How much of the space does the model need to see to make accurate predictions?

It really depends on the problem and whether we also have any kind of preexisting data to train on. To give a sort of a sense, in the diabetes simulation, we have millions of time steps, right, that we're taking into account. And so for a particular patient, you know, that means multiple lifetimes for that patient. So in order to really train a good policy. Now, one way to make this better is to leverage pre-existing data and start not from scratch, but start with a policy that's mimicking what's done in that data. So for example, if we had, let's say, a standard protocol for treatment, we could sort of start with that and then perturb around that rather than starting from scratch. That's not what we did in this particular project.

We have time for probably two more. How is the environment simulated to yield the reward from patients in precision drug discovery?

So what we do here in precision drug dosing is we run the simulator and we have our reward function on top. And so the reward function is looking at the state of the patient looking basically at the blood glucose level and the history of that blood glucose level in order to evaluate the outcome. In general, we can take a simulator and interface it with a reward function that has access to the sort of the part of the simulator that would be visible to a physician treating that patient. One last question related to—

is my mic—

OK. One last question related to an earlier one. How receptive do you think decision-making bodies would be to simulation outcomes? For example, RCTs are considered gold standard for data outcomes and decision-making.

So I really think that simulators are not enough, right? And so this is where I can see two paths. And I don't know where this is going to end up. One path I can imagine is having validated simulators, right? So that, you know, if the simulator is considered good enough, one can then move into a clinical trial phase with that. Another path is basically to sort of look at problems again where maybe the stakes are not so high, right, and where there can be a blend of validation and simulation and then also validation in real life. The fact of the matter is reinforcement learning systems, you really understand what's happening with them when you actually sort of observe the outcomes of them acting in the real world. And so that's been one of the challenges And one of the reasons, frankly, why I've only come to use reinforcement learning for medicine in the last couple of years, even though I've done sort of medical applications for maybe more than a decade, it's been more in the sort of traditional supervised setting.

Thank you so much. Thank you for a great talk. And thank you again to our audience for fantastic questions.

Fantastic. So hello everyone, I'm Leah Cowan, Vice President Research and Innovation and Strategic Initiatives at the University of Toronto, and thrilled to be here for such a fantastic event and delighted to chair the rest of the session for this morning. So it is my pleasure to introduce Dr. Artem Babien, who is the winner of the 2023 Gairdner Early Early Career Investigator Competition. So every year, the Gardiner Foundation invites early career investigators from across Canada to submit an application to present their research as part of the Gardiner Science Week events. Dr. Babien was chosen by Drs. Hassabis and Jumper. Dr. Babien is an assistant professor in the Department of Molecular Genetics and the Donnelly Centre at the University of Toronto. Dr. Bebian's research interests are very broad. They lie in bioinformatics, computational biology, cancer, epigenetics, epigenomics, evolution, phylogenetics, functional genomics, systems biology— I'm about halfway through— genome analysis, sequencing, global health, non-coding DNA and RNA, infectious disease, microbiology, microbiome, translation, post-transcriptional regulation, and viruses. Exciting. So buckle up, we are very excited for this talk, and please join me in welcoming Dr. Artem Babien to the stage. Thank you.

Hello everyone, thank you for, thank you for this wonderful opportunity to speak. You know, I'm deeply humbled that I can share some of this research that I've been passionately working on for the last few years with you. So the title of my talk today is called Entering the Platinum Age of Virus Discovery. And so like many projects in virology, this began around March of 2020, March 2nd of 2020. It was very obvious even back then that this was going to be a society-changing event. I'm talking about COVID-19, of course, and I could see that the global research community was spending billions of dollars on genomic surveillance. We were watching as this virus, through its sequencing, was unfolding in front of us in real time. I decided that I had a responsibility as a scientist to do my part and help in any way that I could, and that was through computational analysis. And so I decided to ask a simple question, not where is this virus going, but where is this virus coming from? What are the proximal and evolutionary origins of SARS-CoV-2, one of 44 known coronaviruses at the time. And this isn't just an academic question. SARS-CoV-2 was predicted to infect over 80 species of mammal biochemically in February of 2020. And the concern at the time was that if this virus entered one of these animal reservoirs, it would be able to mutate or recombine with another coronavirus and then spill back into humans. And as we were pushing towards getting herd immunity and vaccine immunity, that type of change could essentially destroy all the herd immunity we were so strongly pushing for. That was actually quite prophetic. SARS-CoV-2 has now been reported to infect over 24 species of mammal, including stable deer-to-deer transmission in whitetail deer in Ontario and Michigan, meaning that that has now entered a a new animal reservoir and is likely to be there forever. And this actually kind of ties back to a bigger point. It's important to understand that viruses are not just limited to humans. We are part of a greater ecosystem. The health of the animals and the environment around us is directly tied to our human health. We now have a case study how a bat virus in China is causing trillions of dollars of economic damage in America. And so where do you go looking for viruses when all you have is a laptop? Well, I decided to turn to public sequencing data. I hypothesized that there would be coronaviruses already sequenced in publicly available data that were missed. And so there's this database called the Sequence Read Archive. This is the global collection of public sequencing data. Every scientist When you're publishing your paper, that raw data gets deposited into the SRA and is available for free use. By happenstance, in February of 2020, all of the data on the Sequence Read Archive was mirrored onto Amazon Cloud and Google Cloud as part of a major infrastructure project at the NIH. And so to put that into perspective, how much data are we talking about, this is volumes of data that have been growing exponentially since 2008 when the database was started, and we are currently in excess of 60 petabases of sequencing data. So that's one nucleotide followed by 15 zeros. Another way to think about that is that if you were to generate that volume of sequencing data, you would have to run machines for $3.7 to $14.9 billion, and that is just the direct sequencing cost. That does not account for sample acquisition or personnel costs. So we can conservatively say there is a $50 billion resource available if we can analyze it. So this seemed like an almost impossible task. And so— oh, sorry, but before I get to that, if you come away from this talk with just one idea, I want you to conceptualize the Sequence Read Archive as the modern Library of Alexandria for genetics. We have incredibly rich genetic information from all over the world, from tumor samples taken here at SickKids Hospital to ant mound in the Amazon basin, all the way to anal swabs of penguins taken in Antarctica and everything in between. And so to analyze this enormous volume of data, we had to create a new cloud-native computing architecture which we called Seratus. I'm not going to get into too much of the technical details, but suffice That's to say, we aggressively cost-optimized how we are doing this analysis, this bioinformatic alignment problem, and the main problem that we had to overcome was actually not how do we use more CPUs, is how do we move data fast enough? This is called I/O. And so essentially, once we solved that I/O problem, we could scale horizontally and just fire up 22,000 CPUs and churn through incredibly large volumes of data quickly. How quickly? Using Seratus, we can process over 1 million sequencing libraries per day. The cost of an analysis per library is under 1 US penny, under half a cent per dataset. To put that into perspective, we are now able to analyze sequencing data faster than the world is producing that sequencing data. This is a fundamental shift in computing capacity. And so, you know, all of a sudden we built a really, really big hammer. We decided we're going to look for a bigger nail. We weren't just going to do every coronavirus, we were going to look for every RNA virus, full stop. And so to do that, we focused on this one gene called the RNA-dependent RNA polymerase. You can think of this as the beating heart of a viral genome. All RNA viruses by definition share it, but this is a very rapidly evolving gene. The outside, the protein sequences, changes incredibly rapidly. So to be able to organize that volume of data, we had to invent a molecular barcoding system which we call palm prints that focuses on the catalytic core of the RDRP, these A, B, and C motifs. So that way all this sequence essentially turns into a barcode and we could know where RNA viruses are and which one they are. So we went to public sequencing databases and So this is now by the end of 2020, and we collected and curated every RNA virus that had ever been defined, and that gave us 15,000 species units. Then we took that, we took those sequences and we applied Seratus and we searched every single RNA sequencing dataset that we could get our hands on with no bias about what we were looking for. We searched 5.7 million sequences. Sequencing libraries or 10.2 petabases of sequencing data. Using that, we discovered over 130,000 novel species of RNA viruses in 11 days. So let me reframe that. It took the global virology community 130 years to find the first 15,000 RNA viruses, and we found the next 130,000 in under 2 weeks. And so now what was a computationally intensive and difficult problem, essentially you can now find new viruses on our web interface by doing a simple database query. We can just go to coronaviruses, highlight all the samples that have a coronavirus under 90% identity, the new species, and then do a focused analysis of those datasets. And the beauty of doing a large unbiased screen is you end up in places you entirely did not expect. You expect to go. And we found 9 new species of coronaviruses in aquatic vertebrates— so the axolotl, seahorse, fugu fish. And as we started to interrogate these coronaviruses, these divergent coronaviruses, we kept running into this problem where we were getting the RDRP and the non-structural genes of the genome, but we weren't able to find the spike and the structural parts. And after a bit of kind of bioinformatic sleuthing, we were able to identify that in these diverged coronaviruses, these structural genes occur on a separate molecule. So if you open up your virology textbook, it'll tell you that coronaviruses are one linear molecule, but these coronaviruses violate that definition where they occur on two molecules, similar to how influenza viruses are organized, segmented genomes. This actually now ties into a much broader story of where the field is moving. The global— the World Health Organization recently put out a white paper defining how global genomic surveillance is going to be an important public health measure all over the world, especially in resource-limited settings. And one of the key things that they identified was that there is a high technical and financial cost to doing genomic surveillance. So now here in Toronto, the mandate of my lab is to overcome that barrier. We want to create state-of-the-art computational tools, develop them freely, distribute them, and so that they can be applied meaningfully in resource-scarce settings where they can do the most good. And that probably ties into a greater theme with all of these AI talks. We are entering an era of enormous volumes of data, but we are still inhibited by translating this enormous volume of data into meaningful knowledge. So you can think of that as a publication, as a Wikipedia page, as a figure, something that is human interpretable. Because that base of knowledge, that's going to give us that little bit of wisdom. So when the next pandemic rolls around, we know how to respond better. We can make better decisions to save more lives. And so I really view kind of the objective of my lab is really at that interface of data and knowledge. We need to create the knowledge systems for the world to use. And so you can now reconceptualize the Sequence Read Archive as a planetary passive viral surveillance network, where as sequencing data is growing exponentially faster than Moore's Law for a trivial cost of something like $80,000 a year, we can monitor all of that data. And if a pathogen that presents a with a danger to people arises, we can identify it quickly, cheaply, and then act sooner so that we can mitigate future outbreaks or pandemics. And so you might say, okay, well, if the data is growing exponentially, are you expecting to find another 150,000 viruses in another year? Well, no, that actually underestimates what we're— what's going to happen. That's because if you take a look at that green curve right there, that's Diamond, that's the algorithm than we used, as viruses diverge from a reference sequence, so at 100%, by the time that virus diverges 50%, we have a less than 50% probability of detecting that with Seratus. So that means that if you take a look at the histogram of our data, as the new viruses dip down, that dip down is completely artificial. It's a technical signal, it's not biological, which means that what we did was Ceratus is the tip of an enormous iceberg of viral biodiversity. We have only begun our work here. And so one way to address this is kind of the classical PSI-BLAST method. Well, we ran a search, we found new sequences, let's repeat and build out and explore into this dark matter of RNA viruses like an onion. I was always taught that the best scientists are lazy, and so I thought, well, why not spend a lot of effort, try to find a few representatives of these these dark RNA viruses, and then we can, using those representative sequences, repeat Seratus and explode the space of viral biodiversity. And that's essentially what we've done. Thanks to the folks at AlphaFold— shout out, you guys did a great job there— we took 50,000 sequencing libraries, we clustered them into 39 million open reading frames, and we just threw AlphaFold at them. Because AlphaFold is able to do this search for that same palm print molecular signature in the space of structures. So we can go behind the veil of what was the previous technical limits for identifying novel viruses. And then we took those representative sequences, the database has grown, and we have now reprocessed 7.5 million sequencing runs. And so as you can see in our 2022 study, we found that in those 130,000, those are 3 other major studies in virus discovery. And currently we have found now over a quarter million new RNA viruses, putting the total count to over half a million RNA viruses characterized. And this isn't to say like, oh, pat myself on the back, we did a great job, we took the known virome and we exploded it. I think that is kind of a bit of a short-sighted view. The real perspective to have here is that we have not even begun. If you take even all of this information that we found and you compare that to the global genetic diversity of the RNA virome out there, we have not even scratched the surface. This is why I think we are entering the most exciting era of virus discovery in virology that has ever happened before. And so if you combine this exponential data growth with improvements in data analysis, together We are on track to characterizing 100 million RNA viruses by 2030. But more importantly, it's not about just this number of getting more and more viruses. It's about asking, what's the next one that's going to matter? What's the next one that could potentially cause a pandemic? Which one is causing Alzheimer's disease or IBD or some other disease of unknown etiology? If we have now data analysis capacity that was completely impossible just 2 years ago, we need to fundamentally ask the most important questions that we can come up with. Where can we do the most good by finding that next one? And so thank you for your time. So I've now opened, as of last year, the Laboratory for RNA-Based Life Forms across the street at the Donnelly. If anyone is interested in AI and kind of pushing the frontier of virus discovery, be sure to find me in chat. I'd love to pick your brain. I am hiring. Thank you.

Fantastic. I don't see how anyone couldn't be interested, so beware. Fantastic talk, truly inspiring. So I'd encourage folks to submit their questions to the Slido platform, and they're starting to come in. So here's your first question. In which kind of data did you discover more viruses?

So there's kind of like a balance. The volume of viruses, the most that we find are in what are called environmental samples. So soil, water, air, places where like people stick a swab where, you know, you wouldn't normally in a laboratory. The places where we found the most meaningful viruses are in RNA-seq, classical transcriptome data. So the axolotl coronavirus, for example, that was from two different studies in America where they were studying limb regeneration in axolotls. Unbeknownst to them, their animals were infected by this axolotl coronavirus. They would have never have looked for that virus, because you have to specifically look for viral sequences.. And so all of these virus— not all, but I'd say a good chunk of the kind of meaningful viruses are completely incidental findings. This is why this data sharing in the Sequence Read Archive is so important, is we can use other people's data and look at it in a completely new light.

Excellent.

Thank you. So next question: Do you think that SARS-CoV-2 was present in other mammals before humans?

Well, it must have been present in some. So I think phylogenetically it clearly is related to these beta coronaviruses, sarbicoviruses that are predominantly found in bats. So the question is, is there another intermediate, or how did it get out of this kind of bat reservoir, I think is a bit open-ended. But absolutely, it is 100% a zoonotic virus at some respect.

Excellent. Thank you.

So looking back, that was that question. Next, we're going to look forward. So how do you predict which virus could become pathogenic to humans in the future?

So that's another fantastic question. So there's about 300,000. The estimate is that there's about 300,000 viruses that have the potential to infect mammals. There are 160 known viruses that can infect humans— RNA viruses, I should say. There probably is a very large space of viruses that are in, say, a mammal or bird reservoir that could infect humans given the opportunity, right? And then it's not real. I don't think that we'll be able to just pluck out and be like, oh, these are the ones that for sure will cause the next pandemic. What we have to do is survey kind of the entire space so that we at least are aware of what's out there. That way if a child shows up with a fever of unknown origin here at SickKids Hospital, we can create the computational tool to link that patient's blood, the virus in that patient's blood, to a camel sampled in sub-Saharan Africa in 2012. And that should be done in 30 seconds and cost half a cent.

Fantastic. So another question here is, all this sequence data a good fit for large language models?

There has been some success with large language models that work and operate directly on sequence space. I actually think that there's probably going to be more meaningful development of AI technologies when you don't just look at, say, like the k-mer representation of sequences, but are able to add in additional data such as structure from AlphaFold, I think. That's more meaningful.

In many ways.

Excellent.

So another question about sort of geopolitics, if you want to tackle that one. So how does sort of data accessibility, maybe we'll frame it as, you know, influence your ability to do this kind of work?

I, I think it's a, it's a central part of the work, right? So all of the data that we produce, like, is immediately released into the public domain. Uh, so we are reusing public domain public domain data and we are creating public domain analyzed data. And I think that that's important because we're kind of the stepping stone to allowing virology to enter this computational field. But the only reason that this project is possible is because of this standard of ethics of data sharing that was developed during the Human Genome Project. So the Toronto Protocols, the INSDC, so the policy that regulates GenBank sequences and SRA sequences. Having that very strong ethic of data sharing as part of the culture of science is enabling this type of research. And I think it also has enabled things like the PDB to be turned into AlphaFold. And really the entire AI revolution in biology is being powered by open access data.

Fantastic. So another question here. How do you think you can leverage the kind of knowledge and insights you're generating to guide treatment for viral infections?

So that's something that we actually have been exploring because we're essentially doing a very detailed characterization of this one enzyme, the RNA-dependent RNA polymerase. And there are antiviral targets to the RDRP, right? So that's how I think molnupiravir operates. So this data that we're creating Also, we're hoping to translate that into, you know, maybe we should pitch it to cash or something, right? Like, take these sequences and then can we find broad-spectrum antivirals so that, say, if another pandemic agent or, you know, another outbreak happens, not just in humans, in veterinary or plants, that we already have a precomputed arsenal of antivirals that will work on these obscure and highly diverged viruses.

Fantastic.

I'm going to just take a quick look at Janet because we could carry this conversation on easily for another hour. Time for a break. So Artem, I'd like to thank you for a phenomenal talk and note that there are dozens and dozens and dozens more questions for you. So people can find you in the break. And I encourage everyone to keep the conversation going. We will reconvene at 11:10 is the schedule. So we are going to stick to that schedule. So enjoy your 8 minutes of freedom and we will see you back for another stimulating session. Take care. Welcome back. We know that there's lots of exciting conversations to be had on so many fronts, but we also appreciate that there's great talks coming forward in this session and you all want to hear them and we are going to make that happen. So, everyone please come take your seats and we will be ready to move forward with our next talk. I am absolutely delighted to have the opportunity to introduce Dr. Elaine Nsueci, an associate professor in the Department of Global Health at Boston University School of Public Health. She also leads the Racial Data Tracker program at Boston University Center for Antiracist Research. She is a data science faculty Fellow and was a founding faculty member of the Boston University Faculty of Computing and Data Sciences. She currently co-leads the Artificial Intelligence Machine Learning Consortium to Advance Health Equity and Researcher Diversity AIM-AHEAD program at the National Institutes of Health through the Intergovernmental Personnel Act Mobility Program. Dr. Nswezi completed her PhD in computational epidemiology from the Genetics Bioinformatics and Computational Biology program at Virginia Tech. Her PhD dissertation was in sensitivity analysis and forecasting in network epidemiology models. This was at the Network Dynamics and Simulation Sciences Lab at Virginia Tech Biocomplexity Institute. After postdoctoral associate position at Harvard Medical School and Boston Children's Hospital, she joined the faculty of the Institute for Health Metrics and Evaluation at the University of Washington. And I am thrilled to hand the stage over to Dr. Elaine Nsoisi.

Good morning, everyone. It's a pleasure to be here to talk to you about AI and health equity. I want to thank the organizers for inviting me to give this talk. My talk is a little bit different from the ones that you've heard this morning. And I'm going to start with a story. So in 2021, the Boston University School of Public Health and the Boston University Center for Antiracist Research held a symposium on antiracism as public health policy. One of the speakers at that symposium was a legal scholar and a professor called Patricia Williams. And she talked about an experience that she had at a doctor's visit. So she was visiting the doctor to talk about osteopenia and osteoporosis care. And the doctor was asking her a series of questions. He was using an algorithm to generate the questions. And she knew this because she studies algorithms. And so the doctor was open with her about that fact. And so the doctor was asking questions and at some point the doctor the doctor stopped asking questions. And so she asked the doctor to change her race in the algorithm from Black to white. The doctor did that, and suddenly the algorithm had more questions. It asked her about her medical history, asked her if she had any broken bones, asked her about her family history. And this is a pretty good example of what happens when algorithms are trained with data that's either biased or data that is not representative of the different populations that we're trying to study. So you can think about this at an individual level, but if you think about it at a much broader level, which is what happens when we train AI models and apply them to entire populations, the impact can be very significant. So this is what motivates my work. I'm going to start off by talking to you all, defining what health equity is and how I think about health equity. So health equity is assurance of the conditions for optimal health for all people. Achieving health equity requires valuing all individuals and populations equally, recognizing and rectifying historical injustices, and providing resources according to need. Health disparities will be eliminated when health equity is achieved. I really love this graphic because I think it captures it so beautifully. A lot of the time when I talk to people people about equity, they think about equality. And equality is quite important. But when we talk about equality, we mean giving everyone the same thing. So everyone needs some kind of vehicle to move around. We all— we give them all the same bicycle, the same size, whether you're small, you're big, you get the same bicycle. When we talk about equity, we have to think about the needs of the individuals or the needs of the populations that we're trying to serve. We have to think about circumstances and the conditions that impact them and provide resources to them so they can achieve optimal health based on what their needs are. So different individuals will get different types of resources. And it can be really challenging to implement equity because sometimes it means someone gets more than another person. So if someone has a lot less and a lot more obstacles to achieve optimal optimal health, they might need a lot more resources than someone else who already has those resources and can achieve optimal health. So a lot of my work involves thinking about how we can achieve equity, how can we advance health equity. So in the space of AI, in the space of global health, I think as well there is the need for representation. There's a need for bringing individuals voices that are not typically represented in this space. And when we talk about AI specifically, we talk about a lack of diversity in data, but also a lack of diversity in the researchers that develop the models that we use in everyday practice. And diversity in data, depending on what kinds of datasets you're talking about, can involve having some populations not represented in the data, but it could also involve representing people in the datasets in a way that is very biased. And so, being able to identify that and deal with those biases is a challenge, but it's quite important for us to achieve health equity. And also, communities that are most impacted by AI when AI doesn't work the way that it's supposed to work are usually the ones that are left out of this conversation. So, we want to think about what are the risks risk that AI might perpetuate when the models or the datasets that we're using to train the AI models are quite biased. In a lot of cases, the algorithms that we develop, if we don't have this type of representation in data and people, means that some people benefit, others don't benefit. We also tend to have limited understanding of health disparities and inequities. So when we assume— so for example, I was born in I was born and raised in Cameroon, but I've lived in the U.S. for many years now. So I can't really go to a village or a community in Cameroon and say, I am going to solve your health problems, because I don't really understand the context in the way that it needed to be understood. So it's very important for us to allow individuals who experience health inequities to drive the research when we're trying to solve the problems that are affecting them. And lastly, there's limited understanding of historical and present-day policies that perpetuate inequities. So when I told the story about the osteopenia and osteoporosis experience of Patricia Williams to a class, the professor told me that when she was in medical school, what she learned was that it was mostly white women who had osteoporosis. So if you're bringing that context to your model, then you can see why that would be a problem and why she would have the experience that she had. So probably the data that was used used to train the model mostly represented white women and didn't really have a good representation of black women. There's a lot of work that's happening, especially on the African continent. I wanted to give some examples and also talk a little bit about In My Head. So there are a few examples here on the slide. There is over here, this is Data Science Africa, Data Sciences Network, Deep Learning Daba. Deep Learning Daba currently is one of the largest, or actually the largest AI conference that is being led by Africans for Africans. It is held on the continent, across the continent in different cities every year. It's quite exciting to see different individuals come in and present research that they are doing to address problems in their own communities. Data Scientists Network is based in Nigeria and also has similar aims to create capacity in data science and AI. And Data Science Africa, which was started in 2015, is also focused similarly. We have summer schools and workshops across the continent. So to talk a little bit about Aim Ahead, Aim Ahead is at the National Institutes of Health in the U.S. The project aims to enhance participation and representation of researchers and communities that are that are currently underrepresented in the development of AI models. He also wants to empower these communities to use AI to address health disparities and inequities and to improve the capabilities of AI by starting with electronic health workers' data and extending to other types of data, including social determinants of health data. And also to mention Data Science Africa, Data Science Africa was started in 2017. 2015, as I mentioned, by researchers in Kenya and from Cambridge as well. So it was a joint collaboration, and it's been running ever since. There are summer schools and workshops being held across the continent. It's mostly focused on training individuals in data science. And then for the workshops, you have people coming from across the continent to present research that they're working on in their various communities. It offers research grants, offers fellowships, basically focused on creating capability across the continent. So moving on to my research, so this is another way in which I think about health equity, so how can I advance health equity using research? My research for the last 10, 11 years is mostly focused on how do we use data that we are constantly producing using digital devices for public health purposes. So how can we use this to improve public health surveillance? So I've done everything from using data from crowdsourcing websites. So these are websites or mobile apps that you will get. It will ask you questions about whether you experience a symptom of a particular disease, or if you go out to a restaurant and you eat and you think that you were infected because of the food that you consumed at that restaurant, you could go to a website. I don't know if this is still up, actually, iwaspoisoned.com, and you can report that. And usually they look at groups of people reporting instead of an individual. So if it's just you, maybe it's something you ate at home 2 days ago and you had the incubation period and now you have symptoms. But if 5 people go out to a restaurant, they all eat and get sick, then that's probably an indication that they got infected at the restaurant. We've done a lot of work with research data where we try to understand how we can use that to improve infectious disease surveillance, trying to get signals a lot earlier than we would usually get from traditional surveillance systems. So, just to go back again to the crowdsourcing website, so one thing that we do, especially with foodborne illnesses, is to develop a platform. So, we worked with local departments of health where we mined information specifically from Twitter where if someone said, I ate somewhere, I got sick, and they posted it on Twitter, we'll capture that information and we'll provide it to the local department of health. And then they can follow up. They can ask that individual for more information, do more investigation, and in that way could detect outbreaks that were not being reported to the department of health. We have done a lot of work around consumer reviews as well. So, specifically around foodborne illnesses unsafe food products. So when people eat and they go to websites like Yelp and write reviews, we capture that data and use it for that purpose. We've also looked at data from Amazon. So when people buy food products that are unsafe, we mine that data and look at— we specifically created an AI system that could detect unsafe food products using that. Well, the idea was that if we can detect unsafe food products sooner, then the FDA can take actions instead of waiting for weeks and weeks after a lot of people have been affected. Another good example is mobile phone data. So specifically during the pandemic, there was a lot of conversation around how different people were responding to social distancing policies. And we wanted to look at whether there was any association with socioeconomic factors. So were there people that were responding a lot more because they could actually respond. So for example, some people were already staying at home before you even had social distancing policies in some regions, me for example. And so we wanted to see if the socioeconomic factors were a strong predictor of this. And so we looked at different neighborhoods in the US. So we got millions of mobile phone records or mobile phone information. So this is de-identified information where we could tell where individuals were located in different neighborhoods in the US. And then we looked at whether individuals were moving around in different time periods at the beginning of the pandemic. And what we saw was that if you look at higher income neighborhoods, those individuals were already social distancing before we even had social distancing policy in place. While lower income neighborhoods, those individuals could not really most of the individuals were not social distancing because a lot of them were in essential jobs. So they actually had to be working during that time. So they were more likely to be exposed and also more likely to get COVID. One last example I want to mention, this is around social determinants of health. So we've been looking at remote sensing and place data to try to understand how we could better capture data on social determinants of health. So social determinants of health are the conditions in which we live, play, work. So these are the factors that affect our daily lives. We collected about 150,000 satellite images from Google. And the goal was to try to test if by just looking at those images for different neighborhoods, we could capture indicators of those neighborhoods in terms of access, what do people have access to, and could that predict health outcomes? So we looked at obesity specifically and we used AI to process the images to extract indicators. This is work that has been done in the past, but in what has been done in the past is that individuals, well, researchers will look at specific indicators. So they'll look at the presence of fast food restaurants, for example, or the presence of green from space or the presence of specific other types of indicators that have an association with health. And so we wanted to create a comprehensive indicator of each neighborhood and then use that to predict health outcomes. And we did this for obesity. And it did quite well, so well that there were a lot of ridiculous news stories about how you could predict how fat different people were in their neighborhoods from space. A lot of— some of them were quite offensive, but basically the goal of the project was to see if we could use AI to create more comprehensive indicators of neighborhoods so that individuals can use these to study what different individuals in different types of neighborhoods are exposed to and how that can impact their health outcomes. So what my work is currently focused on is to use AI, as I mentioned, to improve measurements of structural determinants of health, to look at historical policies and structural mechanisms and how those continue to influence present-day health outcomes, and to study and propose policies that can transform systems and structures to improve health outcomes for everyone. So just to give you an example about what we're doing around historical injustices or historical policies. So one of the projects that I worked on during the pandemic is a racial data tracker. This is available online. And the goal of the racial data tracker was to collect, analyze, and disseminate data on racial disparities that points to the structural nature of racism. And by doing this, we want to be able to link this back to policies that we can then study and propose policy correctives so that we can actually create change in communities in the US. So we're doing— the project is focused on using visualization and narrative, making data more easily downloadable for everyday people, for advocates, for policymakers, and others who are interested in using this kind of data for change. One research project, just to give an example of the research that is coming out of this, is a redlining project that we did to basically look at— I don't know how many people know about redlining in the audience. I see a few hands. So during the 1930s, the US government created color-coded maps to delineate areas that were deemed risky for investment based on very overtly racist criteria. So those were the 4 different criteria that are listed there. A and B were mostly for affluent and predominantly white neighborhoods, while C and D were mostly for predominantly members of racial and ethnic minority groups. So in some cities, you could not see a single black person living in a neighborhood that was not classified as D. And there's been a lot of research today that shows that these policies continue to impact health disparities, including things like preterm births, asthma, and mortality. And this is kind of how the data looks like. So if you look at the different cities that were included, I think there were about 200 cities that were included across the US, and those were the 4 different criteria as I just mentioned. So we worked with a colleague from the University of Maryland, Quynh Nguyen. She had processed about 164 million street view images to to extract information on these different aspects of neighborhoods. So things like the presence of crosswalks, the presence of green space, the presence of cars and what kinds of cars and so forth. And we took this information and looked at how this information related to different topographical information. So we looked at structures of neighborhoods, specifically how different racial and ethnic groups were present in different neighborhoods and then also looked at those government codings, so the HLSA codings from A to D and how those related to present-day distribution of resources. So the first question we wanted to answer was whether there were disparities in both environment resources when we looked at race and ethnic compositions of neighborhoods. The second question or the second The main goal was to quantify how the built environment mediates the association between neighborhood racial composition and specific health outcomes. And so the first step that we took was to classify those different neighborhoods based on data. So we have predominantly white, predominantly black, predominantly minoritized, other racial groups other than white, and then we had unclassified neighborhoods. And this is what our data looked like. We had about 59 census census tracts. So census tracts have on average about 3,000 people. And these are the built environment resources we're interested in, things like dilapidated buildings, green spaces. Green spaces have been studied quite a lot. The other ones on the list have not been studied very much. So we looked at crosswalks, non-single family homes, and single lane roads. So our major findings were that if we looked at those 59 in 1,000 census tracts compared with other neighborhoods, predominantly white neighborhoods, still had better access to resources that have been shown to encourage health. So things like green spaces, but also fewer of those resources that have been shown to promote poor health. So things like dilapidated buildings. And then when you moved at the gradings from A to D, the gradings I showed you before, poorer, the conditions of the neighborhood got worse. So you compare the resources in neighborhoods that were graded A to resources that were in neighborhoods that were graded D, there was significant differences between what they had access to, especially for resources that tend to promote health. We also noted that the most significant mediator in our models, so when we looked at health outcomes and racial composition of neighborhoods was on single-family homes. So what this basically told us was that the type and quality of home is quite important, especially for the health outcomes that we're looking at, sleep quality and asthma. So why does this matter? This is a question I get all the time. It's very important because nothing that's happening today is just happening now. A lot of what we experience today has been shaped by history. So historical policies continue to impact our lives today. If we don't measure things, they can't change. And sometimes we have to measure things over and over and over again. And if you ever work in advocacy, you have to see the same thing over and over and over again before change can happen. And policies created those disparities. I believe that policies can also be used to to correct this disparity. So what kinds of policies do we need? And can we create those policies so we can create change for different populations? And I just want to end by saying that data and algorithms can be used to create a more equitable or inequitable world. And we all have the power to do that. Thank you. Fantastic.

Fantastic.

Thank you so much. So opening the floor to questions again, and you'll remember the protocol for entering your questions right on the screen. So excellent. We have some great questions coming in already. So I can start with the first one. So this one was triggered by your comment on the satellite measuring obesity or predicting obesity from space. So the question really is broader though. It's how do you couple causation from correlation? Triggered by that thought.

Yeah, that's a very good question and one that comes up quite a lot. It's actually quite challenging in this space of research to do causal research. We have done— developed some models where we try to create causality, but it's actually quite challenging to show causality in some of these cases because there's so many different factors that contribute to the outcomes that we see. And the thing about disparities or systems that impact disparities that we see is that systems tend to overlay on top of each other. So when you think about things like, let's say for me as a Black woman, there's so many different factors that I have to deal with. So there is, there are factors around sexism and racism and sometimes ageism. Based on how people— how old people think I am. And so, those different factors can impact my health in different ways. So, uncoupling those different factors and actually linking them to a single one is quite challenging, but it's something that we're thinking about in the work that we're doing. Fantastic.

Thank you. So, another question has to do with single-family homes in particular and why that might imply fewer environmental exposures. So just thinking about condos and different modalities that might not be single-family homes. So if you could just reflect on that one.

Yeah, so single-family homes in this particular case were quite important, I think, because individuals are not typically exposed to the same things that you have in a apartment buildings, for example. So those might— those in some cases can be very noisy, which can affect sleep quality. But in other cases, in the types of buildings that we were capturing in this data, where you have things like dilapidated buildings that have been associated with things like asthma. So in those neighborhoods, you also not— you don't only have the fact that you don't have single-family homes, but you also have homes that people should not be living in. In the first place.

Fantastic.

So another question pertains to your earlier framing with the great work happening in Africa. And if you could maybe reflect on what are some of the key projects you're sort of most excited about and engaged with on the African continent.

There's so many, so many, many, many exciting projects. There's a lot of work going on around Developing datasets for African languages, that's one that there's a lot of funding that has been coming to that. A large community of individuals are working on that. There are lots of startups that are coming out of that space. There's also a lot of work around disease detection and specifically how do you use AI in medical settings to improve how diseases are detected or diagnosed. So there's a lot of work around that space as well. And then there's also work just around education. So how do we use AI to develop better educational tools, especially in rural settings where there might not be enough resources for different individuals who are interested in learning about health?

Fantastic.

So here's another one, thinking about how we go to from sort of research and innovation to impact. So the question is, are there organizations outside of academia that you engage with actively to reduce inequality and sort of mobilize your discoveries?

In the past, it's mostly been local departments of health. So working with local departments of health to develop tools that they can use. Currently, I'm having more conversations with individuals who are active activists and working in the space and how we can better do research that they can then use. So instead of just doing research that is not easy for them to translate in something that can be used to create change, how can we do research and produce research in a way that can be useful for them in the long term?

Terrific.

So maybe we'll have just a couple more questions.

Here.

So one has to do with representation in datasets, which I think is something you highlighted earlier on. So if there's particular groups, for example, African populations that are— have lower representation in particular kinds of datasets, how would we best address this moving forward?

I think one way is creating more data. And I think that's work that's already being done, but I think is work that definitely needs needs a lot more investment because there are researchers who are interested in creating datasets to be used in AI models, but they do need the resources to be able to create that data. Another one is around incentives. I've heard stories of people who've created datasets and then you have the big tech companies come in and take that data and use it to train their models without giving them any kind of compensation or even acknowledging the work. And in some cases, coming back to them and then asking them to apply for grants so then they can consider funding them, which is insulting because they already used the data. So, you know I can do the work. You've already seen the work I've produced. And now you're asking me to come in and apply so then you can evaluate whether I am qualified to get funding to do this work. And so, I think there's that imbalance in terms of power. A lot of the power and privilege exists in the Global North. And the Global South is not always given that power and always given that seat at the table. And I think there are opportunities for us to do better.

Fantastic. So, we'll close with one more question. And I'll note there's lots more interest and lots more opportunity for discussion. But the question is, what are some effective interventions, or maybe you could phrase frame one that you think is particularly impactful to reduce health disparities?

That's a very big question. I can't think of one example that was very contextualized, and I heard someone else— this was another presentation and I really liked it. So they were trying to develop an app to prevent cardiovascular vascular disease in African American men in the US. And the way the researcher went about it was engaging with these communities from the beginning. So instead of just creating an app and then expecting them to use them, they had conversations with them from the beginning by asking them, what would you like an app to look like? What is the kind of app that you would use on a daily basis? And then they got into the design process, gave them ideas of the kinds of things that they would like to see in an app. The app was developed, brought back to them again, and they evaluated it, gave it feedback, and it was improved. And they actually used that app. And so I think that's one way of going about reducing health disparities instead of just creating solutions that we think will work and then expecting different populations that are being impacted by health disparities to use those solutions. Co-creating those solutions with them, I think, will have a significant impact in reducing health disparities.

Fantastic. So again, lots more interest in further discussion. And folks can find you at subsequent moments in the program. But for now, let's all thank Dr. Lehman Tsvisi for a great talk.

Hello.

I wanted to follow up Demis's incredible talk by thinking more about what will or what do AI breakthroughs in science really look like. And we think that AlphaFold really has been a revolution in a certain area of science. And what does it really mean in practice? What do we do faster? What do we not do faster? How can we think about these new technologies that we're going to develop? How are they going to empower scientists in detail? And want to look at it from the point of view of something that in a certain sense is very narrow, AlphaFold, that it implements, it turns the left into the right. It's something that takes a protein sequence, it runs a series of computational steps, and ends up with a protein structure. And in fact, protein structures are rare but not extremely rare. After all, we were able to train a deep learning system, right, that there's about 200,000 known structures. So, what does kind of abundance of the right mean for science? How are people using this? And I think it's important as we think about it to say also that AlphaFold really produces two things, and I should really update this slide. It doesn't just produce an answer, but it produces really calibrated belief in the correctness of this answer and of different parts of the answer. And in fact, a lot of the work that went in that went into both the database, that went into the paper, and really went into the systems that we actually use is that these models can know when they're right. In fact, that you really can— we're not totally trapped by the notions of, you know, hallucination, which have been very, very challenging within language models. But with AlphaFold, typically the answer will be right or it will be reported as uncertain. And in fact, we even report multiple measures of uncertainty because an answer itself, a protein structure, isn't just right or wrong, but it's going to be used by scientists who are going to look at these structures, develop hypotheses, and then do experiments based on those hypotheses. And within a protein structure, there are many, many things that you can learn from it, and you need to know for each of those facts whether the model is reporting that confidently or hesitantly. And we have, though, been successful, that this seems to really, really work, that these confidences really do give not a perfect but a very, very useful indication of when you should believe such a result. And so because of that, this process that we're doing of, in essence, amplifying experimental data, right, taking sparse gathered experimental data, each with its own particular purpose, schematically represented on the left, and turning it into this enormous abundance of data, does it in some sort of reliable way that we can then use as the basis of further science. But I think it's really, really important to say that on the right side are not drugs that treat specific conditions. They're a particular piece of scientific information, protein structures. How are they integrated into the wider processes by which we ultimately understand biology develop drugs, and ultimately help patients. And so I think it's really interesting to look at a couple case studies. And I should say that all of these are external work, and I'm very proud to say that we had nothing to do with them, that our system is so useful that people do really impactful science that lands in very high-tier journals, and then we read about it at the same time as everyone else. And one of the first kind of things that we noticed is that when we had trained AlphaFold it was always much easier to predict small proteins, not something as large as, for example, the left, the structure of the nuclear pore, so hundreds and hundreds of protein chains far beyond what AlphaFold can do. And we thought maybe, okay, that's going to be in the future work section of our paper, but it turned out to be the present due to the clever work of Jan Kaczynski-Martenbeck and a number of other scientists who said, well, actually, let's just couple AlphaFold fold with other data that we have experimentally, in this case, cryo-electron tomography, which produces extremely relevant in situ images of important parts, in this case, the nuclear pore, which controls what can go in and out of the nucleus. Now, that provides very low-resolution imagery, but we can figure out where in the— in that image what kind of regions correspond to the shapes of proteins that AlphaFold has found. And so, we can, in essence, take a model that is very low resolution, 15 angstroms for the specialists in the audience, and we can convert that into a near-atomic model. And we're really doing two things as we do this. We're bringing much, much more utility to the cryo-electron tomography data. You can understand better the detailed molecular interactions that make the nuclear pore, that control its function, but you're also simultaneously validating the AlphaFold structures, that they agree or don't agree with these densities, or the predicted interactions agree or don't agree. And so what we're finding is that there's quite a lot of hybrid work where it enables new types of structure determination and understanding of large cellular complexes and resolving a lot of the paradoxes in experimental data so that we can come to detailed conclusions, so that we can understand these big machines that make up the cell. But, okay, that's still kind of structure determination. It's clearly within AlphaFold's wheelhouse. But we've started to see maybe even more interesting types of uses. In this case, take a protein that was known to be involved in replication in vertebrates, Donsen, but whose function wasn't really understood. And you can see the prediction in the AlphaFold database. I don't think I have a laser here, but you can see it in the top left in subpanel A. And you— and a lot of the interesting biology is actually in the disordered loop, the little piece of spaghetti that's wrapping around the ordered bit of the structure. So, you might say, well, the AlphaFold database didn't have much to say about this protein, but we also have AlphaFold Multimer, which predicts protein-protein interactions. And the scientist said, "Well, we know about 70 proteins involved in DNA replication that maybe interact with this protein." So they simply ran AlphaFold using that speed that Demis mentioned, ran AlphaFold Multimer against all of these proteins and found a number of very specific interactions with parts of Donsan that you can see in panels B through F that were all attaching to different parts of Donsan. In fact, could put this together into a unified model of this interaction complex that Donsen scaffolds and in fact could confirm all this. And the important thing is you could say, "Well, this is just hypothesis," but they could confirm it with relatively direct biochemical experiments, right? A molecular gel on the bottom right. And so in a certain sense, in fact, what is being confirmed here is not the structures. Certainly, you can't pull a structure out of the ink on the bottom right. But what's being confirmed is that the interesting things that are found agree with experiment, which gives us confidence in the prediction, gives us models. In fact, there was parallel structure determination work that confirmed a large number of details in a separate study. And then this is giving us more confidence, in fact, and the authors went back and explored and said, well, what if we didn't even know these 70 proteins? What if we we scanned against the whole 20,000 human protein coding repertoire, and in fact, with a little bit of work and a little bit of intervention, we're able to pull very similar findings and say that we're really at the moment at which, even though we built the system for structure determination, that it can be really, really useful in finding new interactions, new components, in a sense, new questions that we should be understanding, and it's all kind of driven in a way that certainly was not in our expertise, but also was not in our intention. It wasn't part of the PDB, doesn't have interaction discovery. But when you train really powerful deep learning systems, Ken called them foundation models, they start to learn, they start to get aligned with the underlying problem, to learn something about the physics, the biology, the evolution that is useful in many, many contexts. And in this case, without even fine-tuning or retraining. Now, that's kind of interesting, but can we go even more in the reverse? And I think this has been a very exciting kind of area of research empowered by the AlphaFold database. In this case is another cryo-electron tomography image, a density of protein material from sperm— I think this is actually from mouse— within the flagellum. So, trying to to understand how mobility is controlled at a molecular level for sperm, and there were a number of proteins that were simply not known. And this is, for example, the blue, this is in essence what people knew. There is something that, something that looks like this, that looks like maybe multiple proteins in a line, and they said, "Well, how are we going to figure out what this protein is?" Well, we have AlphaFold database, so they simply scanned the entire mouse proteome and said, "Which proteins fit in that shape?" And so they in fact found a couple, I believe, tech proteins that fit nicely within this density that the AlphaFold predicted structure was concordant with this, but how did they know that they're right? And so they said, "Well, then that induces certain hypotheses that if we look at," I believe it was mass spec— oh, sorry, mass mass spectrometry or protein identification in extracts, in different cellular extracts, then they found these proteins upregulated in a way concordant with AlphaFold. So this gives us kind of belief that we've in fact gone from shape, gone from some unknown— there's something here and it looks about like this— to finding exactly what that is because we have this complete repertoire of mouse protein structures. And so we're seeing kind of new types of discovery that as we take higher and higher resolution pictures of the cell through techniques like cryo-electron tomography, we can use the bottom-up knowledge from AlphaFold in order to generate really, really interesting hypotheses and then confirm those hypotheses. And this is a fact throughout kind of science that if you have very specific hypotheses, you can often come up with much, much easier ways to confirm those hypotheses than if you have a large set of hypotheses to choose from. So, we're really seeing now running AlphaFold in reverse because of AlphaFold database, from shape to protein. We're also seeing really another case in which— and this is a study of Midnolin. This was a protein that was almost unstudied before this paper came out, and it was discovered to have an important role in degrading certain classes of proteins, I believe it's the intermediate early genes that are involved in the cell cycle. And since it was pretty much unstudied, of course, no one had taken the year or more to determine its structure. And the authors had run AlphaFold and found— and sorry. And so somehow it recognizes hundreds of proteins and drags them off to the proteasome to be chewed up and discarded. After all, you know, as the cell has many cycles, sometimes it wants a protein and then doesn't on a protein. So this is one of the mechanisms by which it gets rid of it. And they found if they upregulated this particular protein, Midnolin, that there were hundreds of genes, I think they found about 500, whose abundance was strongly decreased by upregulating Midnolin. And they said, maybe, maybe this is degraded, but maybe there are many other causes. So they took all 500 of these proteins and they put them through AlphaFold Multimer together with Midnolin. And what they found for 200 of them is that the exact same type of thing happened. So IRF4 beta is a particular protein that is regulated by Midnolin, and they found that if they folded them together with AlphaFold Multimer, that this particular— a small bit of IRF4 beta beta is trapped between two different parts of Midnolin. So Ketch2 and Ketch1 are both parts of Midnolin, and that together forms the superstructure where you can see the five strands together. And that suggests that maybe actually what Midnolin is doing is it's recognizing proteins by whether or not they can complete these catch domains. And so they then go and do just the really obvious experiment. Well, what happens if I get rid of that little small section of 4-beta? And of course, that dramatically reduces or eliminates the amount by which this protein is degraded in the cell, so that strongly suggests, in fact, this is recognized. Of the 200 hits that they had, they tried, I think, 10 of them at random, and in every one, deleting that little region that AlphaFold Multimer was predicting was binding between the catch domains would abolish the degradation of these proteins. So we have a really, really strong signal both that this structural model of recognition is right, we have information that we can use to think about how mutations might affect that recognition, and we have really now a belief that we understand much, much more about the biology. And I think that it was already a very high-quality paper without the AlphaFold results, but we would simply say Mindlin is a important, somehow it knocks these proteins down. But because of AlphaFold, we get to skip 2 or 3 steps and say, well, it does it just like this, and this is how mutations in midnolin or mutations in these proteins might affect. There was actually one really beautiful case where eliminating one of these strands didn't actually reduce the degradation of this protein but did not eliminate it, but then they found that AlphaFold multimer models predicted binding of midnolin in 2 different different places, and so eliminating both of those sites did abolish the degradation. And so we're really getting at the point at which we can convert, in a sense, biochemical orthogonal insights into structural insights and skip steps. And I think we should start to think of this, and we'll see this more, that we're skipping steps in the process, and then we can confirm the hypotheses that we've developed based on this. That each of these induce really, really specific hypotheses that we can understand, that we can think about, that we can test, and we're making science a bit faster. We're making some science tremendously faster, but overall we're making science a bit faster, and we'll keep just seeing this snowball. Another really beautiful case of this was some protein engineering done by the Jang Lab at MIT. And this is a particularly interesting protein called the contractile injection system that— I can't remember the system in which it does it, but it in essence injects proteins into cells. And somehow the feet of this protein that looks a little bit like a lunar lander or a rocket— the feet of the rocket recognize specific cell types before injecting. And this was in, I think, a plant pathogen system. Or something else, and people wanted to say, well, can this work in mammalian cells? How can we retarget this? How can we turn this from an oddity of biology into a really powerful system of biotechnology, much like CRISPR was converted from, oh, that's an interesting bacterial immune system, to one of the most important technologies in manipulating cells and manipulating our genome? And they found after about 100 experiments, they said they were unable to figure out how to modify this protein, even which parts of it should they modify because they didn't have a structure. The structure was drawn later. Which parts of this should we modify in order to retarget this? And then they looked at AlphaFold predictions and almost immediately said, "Oh, well, that region at the bottom must be the feet and I bet I can replace the exactly this with a design protein, a DARPin, in order to target. And they did that and almost immediately produced a system that they could target to arbitrary cell types and deliver drugs including Cas9 for gene therapy. And they demonstrated this in mice where they were able to do targeted delivery into a certain, I believe, cell in the mouse brain. And you can see it because they delivered a fluorescent protein and are able to verify that they did in fact— or delivered, I think, did gene editing to deliver a fluorescent protein. And so we're really starting to see this is a new mechanism of protein engineering. And I think again, it's something in which it's incredible that AlphaFold was able to play a part, that really the author said it was very important at unlocking a key step, and also used their creativity. That it's not, you know, AlphaFold delivers drugs, it's that AlphaFold enables the creativity of others to deliver drugs, that accelerates these key steps of science, that helps us also move from instead of saying, well, let us find someone else who can finally do structure determination and then we can continue, lets us skip that step. Now, I think there's some lessons here in general about what will AI-driven kind of scientific revolution what do the solutions really look like? I think one answer is almost certainly we're going to get dramatically better at some problems, orders and orders of magnitude better at some problems. When you get orders and orders of magnitude better and faster at certain problems, you find new ways to use that hammer. People are going to be incredibly diverse and creative, and we're seeing this across what I think is really the foundation model phenomenon, is that if you can train one really, really powerful model that really understands a system or a question, then you can use it in many ways. I think we will experimentally confirm not the predictions themselves in most cases. We'll confirm really the hypotheses that people have from them. And scientists always work with incomplete data. They always work from hypothesis. Ask any scientist, "Do you trust every paper in the literature?" and they'll laugh a little before they answer. And so really scientists are used to dealing with uncertainty, but we confirm ultimately what we wanted to learn after looking at the structure. Computational and AI science is really good when it leads scientists to make better experiments, to more efficiently use their time. I think, and maybe I'll step out a little bit on a limb here, that trust in these systems will come from them being able to predict their uncertainty from really good confidence measures not actually interpretability. You don't need interpretability, you don't need to perfectly understand everything. We don't need a grand unified theory of physics in order to do proper engineering. What we really need is to know when we know and when we don't. And sometimes we develop those intuitions as humans, and sometimes we are aided by reports of uncertainty such as we provide from AlphaFold, and I think this will be very, very important and will let us use the most state-of-the-art models, but will force us to think really carefully about how do we accurately measure our uncertainty, how are those durable across different usages, how do we think about them in different contexts, etc. I think also we will see science and medicine will be driven by the areas that are really very compatible with AI discovery, and I think kind of cellular biology is one of those wonderful areas in which there's a lot of really interesting data, there's a lot of different related systems that have an underlying physical basis, and I think this will drive a lot of the improvements in health because we'll be better able to understand biology, we will do drug discovery, our clinical trials will ultimately succeed more often because we understand biology better, right? Ignorance makes it very, very hard for us to do really great drug discovery. And I think the progress will be uneven because some things will be better suited to AI algorithms and we will lean into those areas and we will find ways to leverage the places where the data is on our side and try and step around the problems where it isn't as best we can. And so this will change the nature of how scientists work in different fields where these AI-driven advancements occur. Now, there we go. Finally, I think I want to just say cells are complicated, right? So this is an artist's rendition of the JCVI minimal cell, right? And it's not so minimal. I think it's 473 genes. Cells are really, really complex, and it's extraordinarily hard to study any tiny subpart of this image. You know, many great careers, I believe many worthy Gairdner Prize winners have studied a tiny amount of a tiny part of biology, and we will have— oh, geez— we will have the tools we need to understand them, but we will have to build them, and it will have to be a conversation between experiment and computation. Thank you.

Fantastic.

So totally inspiring talks. Thank you, Demi, for coming back up. And we have lots of questions, and I will keep an eye on our fearless leader at the front who's going to give me signals of how best to navigate given the level of interest and the timeline in our program. But brilliant, you know, brilliant work, and thank you. So lots of different questions. I'm going to go through some different ones. So we've clearly got some structural people in the audience here. So I'll put out a few questions and you can answer them not at a super deeply technical level maybe, but to keep the general— I'll follow up later. Yes, exactly. So first question that lots of interest was supporting was, how does AlphaFold deal with protein flexibility dynamics, changes in conformation, et cetera?

So, this is kind of— so, for— actually, I will tell the story. You know, I used to work more directly in protein dynamics and I do remember telling my boss Tobin, you know, if he believes that, you know, proteins are points or static rigid objects, I'm not sure we can be friends. Of course, he didn't. But— so proteins move and yet we have to stop them from moving to take their picture. And so what we do find actually, what a lot of scientists have found is a lot of correlates of the protein dynamics in AlphaFold. Both the more mobile regions have much lower confidence, as expected. How can you know where they're going to be if they're moving? We also see new methods being developed that basically change the inputs to AlphaFold without retraining that start to capture the kind of landscape of multiple modes. And I think this is an area in which it's really probing how much does AlphaFold understand, not just like what will appear in a crystal But how much does it understand protein physics? And I think the answer is becoming clearer that it understands a fair amount and that it's able to drive multiple states. But we have to figure out how to induce those states. And it's something that I hope will get better over time. But yeah, proteins move and I think people are starting to understand how that motion might be captured in AlphaFold.

Fantastic.

Thank you. So, we'll stay on the general theme here and ask maybe a question about post-translational modification. How do we think about that? Is it integrated? What do we do?

[Speaker:Joshua Elkington] So there's been— I guess I'll quickly go— I mean, there's been some really interesting studies of both cofactors and— so this is different things. After the protein is made, it gets changed because protein's made of atoms. The cell likes to do lots of chemistry on atoms, so there's lots of changes to proteins. What we do find in some cases is that the AlphaFold structure obviously only reports the structure of the canonical amino acids in a protein, but will often leave space for the post-translational modifications, or especially cofactors like heme. But we don't really have a way to say, well, how is the structure going to be different if I have a post-translational modification, or how is the phosphorylated form different than the non-phosphorylated? And that has to be a subject of future research.

[Speaker:Joshua Elkington] If I just— I'll give a more general answer to both those questions. So obviously with AlphaFold, we've kind of dealt with the static picture, as John was saying, of proteins. But of course, in biology, it's a hugely dynamical system. So you can think of what we're trying to do next in both at Google DeepMind and also isomorphic is go up that interaction stack. So first of all, pairwise interactions, proteins and proteins, proteins with ligands, proteins with nucleic acids. I think we're making great progress with that. Then eventually, you might imagine modeling a whole pathway, maybe like the TOR pathway. And then eventually, that's how we can maybe build up to something like a virtual cell, like John flashed up at the end, which of course would, you know, hugely complex interactive system. But I think over the next maybe decade, I think we can get there.

Fantastic.

Thank you.

So maybe building on these questions and taking it a little forward-looking, this question asks whether AlphaFold can predict structures that don't exist. And I think one of the earlier slides highlighted sort of building new enzymes with new functionalities. So, maybe you could reflect on that.

So, what we have found, first of all, is that AlphaFold has been quite important. Some of the ideas inspired, but also the system itself in designed proteins. And there's been a real revolution in protein design in the past 2 years. One of the most interesting kind of facts is that when people have designed proteins, they found trying to predict the structure with AlphaFold and seeing whether they got the answer they expected is extraordinarily powerful at figuring out which of their designs are real. And so they— there have been reports of 10 times higher success rates in protein design by screening through AlphaFold. So pretty much no designed proteins get made anymore that AlphaFold doesn't like. So it has turned out to be very useful in the verification step of designing new proteins. And I think we'll see more developments of the ideas and technologies that will continue this protein design revolution that we're in.

I think it's pretty interesting to also think about the reverse of— I mean, people in protein design are using this— the reverse of AlphaFold, which is sort of, you know, here's a structure you would like, what is the amino acid sequence that would create that structure? And I think that's, you know, clearly works and will work. Even better. I think the missing part of what we have for protein design is to go from function to structure. Like, what structure should you specify for the function that you're looking for? So that's actually the part that needs to be solved, in my opinion. And I think it's pretty, you know, easy-ish now to go from a structure to the sequence by doing the reverse, you know, AlphaFold in reverse, effectively.

Terrific. So we're now getting two more questions. So, we are going to take a big leap into a slightly different space. And there was a question that just highlighted the remarkable innovation at a really early age. And as we're at this sort of moment of transformation and how we compute and think about the world, maybe you could reflect on how do we make sure we're supporting the next generation? So, we're going to shift gears here and think about how do we support, you know, youth? How do we support this sort of transformation and how we think about the world around us?

Yeah, well, I mean, it depends on what age we're talking about, right? So if we're talking about the kind of undergraduate level, you know, going thinking about grad school, I would say what we need a lot more of, and I've been trying to encourage philanthropists to do this in my own sort of philanthropy, is to encourage multidisciplinary work, right? So grants for multidisciplinary PhDs, combining different fields together, obviously including machine learning, but actually getting expert in the other domain. And I think that was what was critical for AlphaFold actually, was bringing together world-class sort of machine learning people and engineering, but also people who knew structural biology such as John and chemistry and also biophysics. So in the end, I think we needed an incredibly multidisciplinary, talented team to crack something like AlphaFold. And I think that's going to be more important in future when thinking about where to apply AI.

Fantastic.

Thank you. And so maybe as a— did you want to jump in on that, John?

I just wanted to say also that it's actually the easiest time for young people to get involved, that a lot of these tools that we're thinking of are available. You're not just the AlphaFold database, but AlphaFold Multimer and others. So it's a golden age to try and really work with the same tools as specialists because computation has gotten powerful, because you don't need both the incredible investment and the safety protocols needed to work in a lab. So it's an incredible time, I think, for young people to play with the same kind of questions that working scientists are.

Fantastic.

Thank you so much. So maybe we'll close on a last question that just kind of highlights clearly remarkable power and promise of the technology, and I think you highlighted highlighted the importance of responsible AI, and these discussions are obviously incredibly timely. So, just curious if you could reflect on sort of how you think we need to be moving forward in a thoughtful way to ensure the kinds of amazing platforms you're building are utilized in an ethical way.

Well, look, I mean, there's many components to answering that, and we're still finding our way as a field. But I think the obvious ones are to involve many, many different inputs and voices into the designs of these systems, and especially the communities that these systems might affect. I think we need much better testing and evaluation protocols, so benchmarks for testing these learning systems, which are pretty complex. And in a lot of cases, they're kind of black box-like, but they don't need to be. I think we could understand these systems a lot better, and we We need a lot more work on analysis and research on analysis techniques and methods. So I think that's coming, something we're pushing a lot forward. And we just announced today a fund, a safety fund, with some of the other leading labs to give grants to academia and others who want to research these questions. So I think there's a lot of work on that. But then, you know, it also goes into societal questions about how do we want to, you control sort of access to these systems. And, you know, we're big fans of open source science and open science. Of course, that's the way things progress. But one also has to think about bad actors getting hold of these systems and repurposing them for bad ends. So, they are general purpose technology. So, how do we enable all the good use cases but restrict and mitigate against the bad use cases and the risks? And so there's lots of huge questions. And in fact, next week in the UK, there's a convening of an AI safety summit, international AI safety summit, with many of the leading technologists and also heads of government are going to be there to discuss exactly these questions. And I think it's great as a first step towards international dialogue on these points.

Fantastic.

Well, please join me in thanking again Demis Hassabis and John Dumphy. And over to Janet.

Oh, it was so exciting. I forgot I had to finish the session. So thanks, Leah. And thanks, Demis and John and everyone else, all the other speakers and all of you. Great questions from the audience. And I hope you had a great time. We have not finished Gardiner Science Week, and I want to invite you all back tomorrow morning in the same place, 8:30 this time, so you'll have to get up a little earlier, for the Laureate Lectures, where we'll hear from all of this year's Canada Gardiner awardees of their life story to the point of receiving the award and the impact of the research that they've done on many different aspects. So thank you, and we'll We'll see you tomorrow.