Learning meaningful biology with XAI: from evidence to prediction and design

 

Biomedical data provide unprecedented opportunities to understand complex biological systems, but predictive performance alone does not tell us whether a computational model has learned meaningful biology or produced knowledge that scientists and clinicians can use. My research focuses on closing this gap by building interpretable, biologically grounded AI methods and human-in-the-loop platforms that integrate diverse biomedical data, reveal what information drives their conclusions, and enable experts to interrogate and test what models learn. Across this work, I am particularly interested in how biological knowledge and context should inform models, how we rigorously distinguish meaningful biological signal from prediction alone, and how computational insights can support better scientific and clinical decision-making. Increasingly, my research asks how we can move beyond understanding and prediction toward action: using what models learn to determine what to test next, prioritize promising interventions, and ultimately design biological function. In this talk, I will review both my research and how I got to Microsoft Research from Brown. I will share a few insights I have found useful along my journey and look forward to a discussion.