Hello!

How do we use AI to make biological discoveries? While AI is an increasingly powerful tool for scientific discovery, most of its successes are in fields where ample prior knowledge exists to pose well-defined prediction problems. However, many biological domains are inherently exploratory, and characterized by a lack of complete knowledge. My research focuses on figuring out how we can still use AI to empower biological discovery in these settings.

Specifically, I focus on several core themes:

• Foundation models and self-supervised representation learning: These methods can learn from unlabeled data at scale, potentially unearthing unsupervised insights in yet-to-be-characterized data. I design biologically-motivated and principles-first methods.
• Generalization and robustness: For foundation models to be useful in discovery, they must learn biological signal. I design evaluations that understand what models learn and scale on, and methods that ensure that models learn biology, not irreproducible technical variation.
• Aligning with downstream use-cases: Applying models for discovery in less-characterized disciplines means we need to interact with them differently. These include both technical methods like zero-shot evaluation, interpretability, and generative models, but also developing an understanding of overlooked applications.