Emily Fox

Monday, November 17, 2025

4:00 - 5:00 pm (refreshments at 3:30 pm)

Broad Auditorium (Merkin building, 415 Main St.) and virtually at broad.io/ewsc

Abstract:

While machine learning shows tremendous promise for extracting disease-relevant signals from rich, multimodal clinical and cellular datasets, predictive accuracy alone does not yield decision-ready science—establishing causality is critical.  This talk traces a path from pattern-finding to causal insight in two parts.

First, I will present our approach at insitro to therapeutic target discovery, which builds target conviction by triangulating two complementary lenses. On the clinical side, we leverage human genetics as experiments of nature to identify causal genetic drivers, exploiting ML-derived precision phenotypes imputed at scale from rich clinical data modalities. On the cellular side, we run perturbation screens in disease-relevant models to measure phenotypic consequences across high-content modalities including microscopy and omics. Because the biases of these ClinML and CellML lenses differ, cross-lens concordance holds promise in raising confidence in causal targets. I’ll highlight key sources of bias and how modeling choices and assay design chip away at them to refine target prioritization, including our work on multimodal embeddings and sparse autoencoders for aligning and interrogating ML-derived phenotypes.

Second, I’ll shift to my Stanford work on leveraging mechanistic knowledge as causal inductive bias.  When training blackbox models on observational data, high predictive accuracy can coexist with low causal validity—the counterfactual simulation of a glucose trajectory may respond nonsensically to a simulated insulin intervention.  We address this by encoding domain knowledge about sets of treatment effect rankings into a causal loss that, combined with standard predictive loss, biases learning toward causally valid models. Using type 1 diabetes as a case study, I'll show how hybrid models—combining mechanistic ODE dynamics with flexible neural network components—trained with our proposed hybrid loss achieve both state-of-the-art predictive performance and causal validity for the challenging task of modeling glucose dynamics post-exercise.

Biography:

Emily Fox is a Professor of Statistics and Computer Science at Stanford University and Chief Technical Advisor at insitro, where she previously served as SVP of AI and Machine Learning. Before joining Stanford, she was the Amazon Professor of Machine Learning at the University of Washington and led Apple’s Health AI team as a Distinguished Engineer.  Emily has been awarded a Presidential Early Career Award for Scientists and Engineers (PECASE), Sloan Research Fellowship, ONR Young Investigator award, and NSF CAREER award. Her research interests are in machine learning and causal AI for biohealth applications.

Questions? Email Amanda Ogden at aogden@broadinstitute.org.