Lani Wu

Professor
University of California, San Francisco

​

Dr. Lani Wu is a Professor in the Department of Pharmaceutical Chemistry at the University of California, San Francisco (UCSF). Lani received degrees in pure mathematics from the National Taiwan University (BA) and the University of California at San Diego (MA and PhD). She has held several academic positions, including at Princeton University, Harvard University and the University of Southwestern Medical Center, as well as industrial positions at Microsoft, Rosetta Inpharmatics, and Google.
For nearly four decades, Dr. Wu has maintained a remarkable scientific partnership with Dr. Steven Altschuler, beginning when they met as mathematics PhD students at UC San Diego. Together, they have forged a uniquely collaborative career that has crossed traditional boundaries between pure mathematics, technology, and biomedical research.
Lani's mathematical journey began in differential geometry and topology as a student of Dr. Richard Hamilton. Her work in geometric heat flows helped resolve conjectures about singularity formation. In the mid-1990s, she pivoted to Microsoft Research, where she led teams developing early machine learning algorithms for speech recognition, computer vision, and search technology that helped shape our digital world. In 2000, seeking new frontiers, Lani transitioned to biomedical research, bringing her mathematical and computational expertise to some of medicine's most challenging problems. She pioneered the field of microscopy-based high-dimensional phenotypic profiling, revolutionizing how researchers visualize and analyze complex cellular behavior. Her interdisciplinary approach has led to fundamental discoveries in tissue patterning, cancer biology, and most recently, neurodegenerative diseases like ALS and Parkinson's Disease.
Currently Professor at UCSF's Department of Pharmaceutical Chemistry, Lani continues to tackle complex biological questions by integrating cutting-edge experimental techniques with mathematical modeling and interpretable machine learning. Her work has been recognized through innovations awards including multiple UCSF Catalyst Innovation Awards, and Program for Breakthrough Biomedical Research grants.
Throughout her journey, she has maintained her commitment to working as a true partner with Dr. Altschuler—finding important problems, brilliant collaborators, and pushing the boundaries of what's possible together.

 

Speech title & Synopsis

AI-Powered Alignment of High-Content Screening Data: Accelerating Drug Discovery Through Cross-Dataset Learning

Twenty years ago, we introduced high-content image-based phenotypic screens (HCS) (Perlman, Science, 2004). Our work demonstrated that cellular phenotypic profiles of compounds with similar mechanisms of action tend to cluster together. This enabled functional prediction of uncharacterized compounds to be performed via guilt-by-association comparison with well characterized “reference” compounds in the same dataset.
 
HCS is now a standard platform for screening large-scale compound libraries in academia and pharmaceutical industry. The wide-spread adoption of HCS has led to a rapidly growing number of datasets. However, unlike omics studies, which measure a consistent set of features (such as genes or proteins) and datasets can be combined to gain synergy, the highly customized nature of HCS (experimental, computational and reference drug choices) produces heterogenous phenotypic profiles that cannot be directly compared. A critical, long-standing challenge is how to integrate diverse—but currently isolated—HCS dataset resources.
 
To address this, we developed CLIPn, a contrastive deep-learning approach to align heterogeneous HCS resources. This AI-powered framework is designed to enable the cross-dataset “transitive” prediction, whereby the function of an uncharacterized compound screened in one dataset could be predicted through comparison with reference compounds profiled in other datasets. We applied CLIPn to 14 diverse HCS datasets generated using different experimental systems and computational pipelines over the past 20 years. By integrating these datasets, we predicted and experimentally validated functions for compounds that could not be characterized in the original, isolated HCS studies. Our work demonstrates, for the first time, that accurate “transitive” predictions can be made across diverse HCS profile resources.