Research

Whole-slide pathology images are enormous and only a small part of the tissue matters. The lab builds models that pick the informative regions, turn them into a slide-level representation, and report how much evidence supports each prediction. Current work screens H&E-stained lung cancer slides for ALK rearrangements, so cases can be prioritized before molecular testing.

Gene expression, DNA methylation and copy number carry different parts of the same story, and pathology images carry another. The lab's models organize these around biological pathways and the interactions between them, so a survival prediction can be read as a pathway effect rather than a black-box score.

The same idea applies at molecular scale: predict enzyme function from sequence, cholesterol-binding sites from structure, or protein-compound interactions from positive examples alone, and point to the residues or atoms responsible. Much of this work is joint with the Luo Lab, which brings the molecular dynamics.

Models trained on clinical data can behave differently for different groups of patients. The lab looks at where those differences come from and what they cost, from race-specific dementia care costs in Medicare and Medicaid records to sex-common and sex-specific signals in transcriptomic data.