How do we use AI to make biological discoveries? While AI is a powerful tool for scientific discovery, its successes are where conditions align with where AI already excels: ample data, abundant prior knowledge to turn into labels, and where problems are already defined to be computable. I develop AI methods that advance understanding of cellular genotype and phenotype, with a focus on developing biologically-grounded methods that can still uncover insights when research is exploratory, and data may be limited.

My research forms a circuit between several core principles:

• I precisely characterize limitations of current AI methods as applied to biology. My work has exposed that biological foundation models may not function well in zero-shot exploratory applications, learn universally relevant signal, or outperform bioinformatics baselines.
• I develop biologically-grounded methods that address these limitations. I’ve shown that focused self-supervised methods incorporating biological priors for genomes, proteins, and cell microscopy can unearth new insights. I also develop generative models. Within genomics, these models can design enhancers without prior knowledge of function and in cell microscopy, overcome experimental scaling limitations.
• I translate insights to widespread impact in science. The challenges I study are not exclusive to biology. I collaborate with domain experts to identify new areas of impact for the core insights I develop within computational biology, including linguistic study of low resource sign languages, AI models for pediatric healthcare, and probabilistic weather forecasting.