Speaker
Description
There is significant interest in using machine learning to develop “AI virtual cell” models that can serve as computational proxies for wet lab experiments. The central requirements for such models are inherently causal: they must accurately predict the effects of interventions such as gene knockouts or drug perturbations on cellular state, and they must support the generation of testable hypotheses about the mechanisms underlying observed responses. I will argue that while the abundance of modern biological datasets now enables us to leverage large-scale machine learning, paradoxically, we remain in a “small data” regime if we truly want to characterise the effects of perturbations in unseen contexts, such as cell types, perturbation combinations, and concentration regimes. For these harder generalisation tasks, what matters is the number of unique (combinations of) perturbations and contexts that we observe. I will illustrate this through a series of recent works that show both the effectiveness of large-scale generative models on some metrics, and how we can use relatively simple algorithms by appropriately leveraging the causal structure of biological interactions or the sparsity of intervention targets to improve predictions of unseen perturbations.
