In this inaugural keynote, Professor Andreas Vothoghe, co‑director of the Cambridge Centre for AI in Medicine, outlines the Centre’s mission to transform biomedicine by developing machine‑learning (ML) and AI methods that go beyond clustering and prediction toward generative, causal, and mechanistic models. He highlights systemic challenges that hinder ML adoption in biology, such as information loss from poorly defined problems and insufficient integration of existing biomedical knowledge. Vothoghe then showcases four research pillars where the Centre applies AI: (1) drug discovery, featuring ML‑guided fragment‑based antibiotic design and permeability prediction to overcome bacterial uptake barriers; (2) structural biology, where interpretable deep learning predicts mutation effects on protein function and improves ligand‑binding forecasts; (3) functional genomics, proposing a graph‑based framework that transforms DNA sequences into protein‑activity graphs to infer causality and epistatic interactions; and (4) multi‑omics integration, leveraging phenomics, proteomics, metabolomics, transcriptomics, and epigenetics to discover diagnostic/prognostic patterns, predict individual treatment responses, and repurpose drugs. He cites concrete contributions from Centre members (e.g., José Miguel, Vita, Aaron, Tom Dundell, Chris Abell, Anthony Coyne, Tom Landell, Sarah Tightman, Eoin McKinney) and emphasizes partnerships with AstraZeneca and GSK aimed at delivering revolutionary, not evolutionary, advances in discovery science.
Chapters
Hello, my name is Andreas Vothoghe, Professor of Respiratory Biology at the University of Cambridge and co-director of the Cambridge Centre for AI in Medicine. And like Mihaila, I'd like to welcome you all to this inaugural event and launch of the centre and hope that I can share with you some of our enthusiasm and excitement and ambition for the potential for AI to transform biomedicine. As you will have heard in Mahala's talk, there are 4 main pillars of activity within the Centre. And while Mahala focused on precision medicine and clinical trials in her talk, I'd like to share with you over the next few minutes the exciting potential of applying machine learning methods to biomedical discovery and to the analysis of large data sets. I think it's fair to say that there is enormous unfulfilled potential for the application of machine learning in biomedicine, driven in large part by information loss, by which I mean from the biomedical scientist's perspective, a failure to properly specify the problem, and a lack of understanding about the enormous potential for novel machine learning methods. And conversely, from the machine learning aspect, a failure to appreciate the importance of incorporating existing biomedical knowledge and to understand the sometimes complex relationship between biological features. So our mission within the Cambridge Centre for AI in Medicine are to develop methods in machine learning and AI that can act as discovery tools for biomedicine, to move away from the existing focus on clustering and prediction, to generate methods that can reveal informative generative models accurately predict causal relationships, provide deep mechanistic understanding, and thus real-world solutions. Conversely, we're convinced that biomedicine can act as an important motivator and catalyst for advances in methods development in machine learning. So the move from current focuses on simple or simplistic models and abstract features to developing methods and frameworks that can incorporate and encode prior biological knowledge accurately, but to embed complex feature relationships and deliver actionable intelligence. As examples, I'd like to focus on 4 areas of biomedical research: drug discovery, structural biology, functional genomics, and the analysis of multi-omics datasets. So in drug discovery, José Miguel within the Centre has generated exciting and accurate machine learning methods to enhance the design of new drugs. So he's been able to encode the 3D structure of chemicals and importantly define which chemical synthesis routes are possible, and then to use these methods to reimagine how to develop drugs in an effective and economical way. In my group, we've been focused on understanding how machine learning can drive antibiotic discovery. This is work from Vita and Aaron. And over the last 5 or 6 years, we've been working with Tom Dundell and before his death, Chris Abel, to apply fragment-based drug discovery to generate novel antibiotics. Now, fragment-based drug discovery was pioneered in Cambridge by Tom and Chris 20 years ago, and its premise is straightforward: that there is a huge, enormous complexity in drug-like chemical space, and that in order to reduce this complexity, we can break down drugs and complex chemicals into their building blocks, or fragments, and then you only need a few hundred fragments to screen against your target of interest, define low-affinity binding hits, and then take these fragments and either grow them, or merge them, or link them into drug-like molecules. As you can see here, this can be very effective. Within a few weeks, we've been able to go from fragment hits to a drug-like molecule that has activity against a number of important mycobacterial pathogens. However, this is the exception, and the problem of using structure-guided antibiotic discovery is that activity is currently unpredictable. I'm grateful for Anthony Coyne in Chris Abell's group for sharing this example, where we've gone from the original fragment hit on the left, through a series of structure-guided chemical elaboration steps, and we've been able to improve the binding affinity to its target by over 400-fold. During that period, there has been no change in the ability of these compounds to kill the bacteria. And this really illustrates the greatest biological hurdle to novel antibiotic discovery. There are other financial hurdles as well, but the greatest biological hurdle is that we don't understand the rules governing which chemicals get into bacteria. And in fact, slightly more complicated, that although some chemicals will get in and stay in, others will be prevented from, from entering, can't get through the cell membrane, Others will get in and get out very— and be pushed out very quickly through efflux pumps, and yet more will be internalized and then rapidly degraded or metabolized. And so to address this challenge, we wondered whether machine learning might provide a framework to direct chemical elaboration. So firstly, we derived experimentally the chemical permeability properties of a large compound library, and as you can see, only a small Proportion of the chemicals will get into bacteria, and I've just illustrated some of their physical chemical properties. But the machine learning predictor can allow us to focus our efforts on chemical elaboration that will maintain permeability, improve retention, and prevent metabolism to optimally end up with a drug-like molecule with improved activity. My second example is in structural biology, and this work is driven by Tom Landell, who you'll hear from later today. He was one of the early pioneers of applying machine learning to computational structural biology problems, and we've been very excited by the potential for novel machine learning methods, particularly interpretable deep learning methods, to two problems. The first is the ability to accurately predict the impact of a mutation on protein function. And obviously this has huge impacts for a number of areas of biomedicine, including cancer, antimicrobial resistance, and functional genomics. Secondly, the ability to improve the prediction of ligand binding to proteins. And we think new methods will allow a massive improvement in our ability to optimally elaborate fragments, so from the initial fragment hit to a drug-like molecule, and improve in silico screening of compounds. So more from Tom later. The third area I wanted to describe is functional genomics, which is what my group spends a lot of its time doing, and I've just characterized the current approach in this slide. So the starting point is the DNA sequence, Conventional statistical genetic analyses are carried out in order to see whether there is an association of a particular mutation with a phenotypic trait, and we hope that this provides medical and biological insight into some pathway. The vast majority of existing machine learning effort has been focused on this step after the output from the conventional statistical analysis. And my feeling is that there is a huge potential for reimagining this scheme. That is to say, to develop a machine learning framework that can take the original DNA sequence, to model each of the mutations in terms of the impact it has on protein function, to leverage existing transcriptional proteomic analysis to give an idea of protein abundance, and to use this information to generate a holistic protein activity graph, which can then be used to infer causality in terms of function. Thus, each genome can be represented by a discrete series of protein graphs encoding, you see, the activity of each protein type important for a specific outcome, and then mapped onto a series of phenotypic traits. There are enormous advantages of using this method to encode existing biological knowledge. So we can form a framework for the generation of— in the creation of generative models. It is a concise way of reducing feature complexity, and it provides an explicit representation of pathways and epistatic interactions, that is to say, gene-gene or protein-protein interactions. And we're very fortunate that we can now leverage enormous advances in machine learning methods to allow us to encode complex graph structures accurately, unambiguously, so-called injectively, into embedding space. And moreover, we can actually embed whole graphs in a rational framework that will allow us to then employ machine learning methods. Finally, I'd like to describe advances that have been made possible thanks to machine learning and AI in multi-omic analysis. That is to say, the combination of multi-dimensional phenotypic data, so-called phenomics, proteomics, metabolomics, transcriptomics, epigenetics, genomics. So there are enormous challenges in accurate data integration, but enormous opportunities in terms of using these complex datasets to identify patterns with prognostic or diagnostic usefulness, the ability to infer causal relationships between changes seen at the cellular or molecular level, to use these methods to create generative models for disease processes, to predict treatment effects in specific individuals, and perhaps most excitingly, to use them to find new treatments. So what do I mean by that? Let me give you an example that you can create or curate a large data set of the cellular responses of a variety of different diseases and the variety of the different drugs. And also, you can begin to generate models that would allow you to take a particular cellular response and predict outcome. And by doing this, it allows you to select a specific drug based on the clinical— the cellular response, and then use that as a therapy for a specific disease. So within the group, we have two scientists who've been working on problems in this area: Sarah Tightman, who you will hear from later today, and Eoin McKinney. And Eoin's work has focused on both infectious and non-infectious diseases and used a variety of different machine learning methods to predict clinical outcome based on programs of different gene expression, to screen in silico and in vitro for repurposed drugs based on their predicted cellular responses, and demonstrated efficacy in vivo at improving vaccine responses. And finally, he's used machine learning to leverage large multi-omic datasets and to extract actionable information about fundamental disease processes. So in summary then, I hope I've shown you that the Cambridge Centre for AI in Medicine is going to deliver revolution and not evolution in biomedicine, and we're really excited to be working in partnership with AstraZeneca and GSK to transform discovery science.
