- Sponsor
- Mind in Vitro
- Originating Calendar
- Siebel School Speakers Calendar
Abstract: Diseases are classified by clinical presentation, yet many share molecular mechanisms that no single study can reveal. Nitisinone, an herbicide that became a treatment first for tyrosinemia and later for alkaptonuria because the two diseases share a pathway, shows what such connections can offer. Finding them systematically requires analyzing hundreds of heterogeneous datasets consistently, work that has depended on bioinformaticians handling one dataset at a time. In this talk I describe how my group builds large language model agents to take on that work, and what becomes discoverable once they do.
I begin with GenoTEX, a benchmark of 1,384 expert-curated gene–trait analyses over 911 cohorts, and GenoMAS, a multi-agent system that writes, executes and reviews cohort-specific code and reaches 84% of expert accuracy. The rest of the talk turns to analyses across whole corpora. Comparing the transcriptomic signatures of 1,384 disease–condition pairs recovers clinical taxonomy, links diseases it places far apart, and shows conditions such as hypertension amplifying these links. Rondo builds regulatory networks for 663 patient subgroups and compares them more than 212,000 times. Disease identity is visible from one comparison, but shared regulators become reproducible only as comparisons accumulate. This singles out 222 disease pairs, including an enteric neural crest program shared by Parkinson’s disease and gastric cancer. Finally, an exhaustive comparison of all 231 pairs among 22 human tissues finds conserved protein co-abundance clusters even between unrelated organs. Of these clusters, 379 replicate in independent proteomes, and one links three white matter disease genes through liver and bone marrow.
Across these studies, agents make analysis cheap enough to repeat, and repetition turns a single observation into a quantity that can be tested. I close with the candidates that most need experimental validation.
Biography: Haohan Wang is an assistant professor in the School of Information Sciences at the University of Illinois Urbana-Champaign. His research focuses on the development of trustworthy machine learning methods for computational biology and healthcare applications, such as decoding the genomic language of Alzheimer's disease. In his work, he uses statistical analysis and deep learning methods, with an emphasis on data analysis using methods least influenced by spurious signals (features that are statistically associated with the target but not causal). In 2019, Wang was recognized as the Next Generation in Biomedicine by the Broad Institute of MIT and Harvard because of his contributions in dealing with confounding factors with deep learning. Wang earned his PhD in computer science through the Language Technologies Institute of Carnegie Mellon University.