UCSF

Patients differ in their response to clinical interventions. Both genetics and the environment contribute to the phenotype of a patient after a chemical intervention such as a drug. Well-validated interventions have large population effect sizes, meaning they are valuable for an individual on average. However, for complex conditions such as autoimmune disease and cancer, the precise outcomes of chemical interventions remain difficult to predict for an individual patient. Here we present methods for predicting chemical response and apply these methods to understand biological mechanisms. We discover a new biomarker associated with psoriasis patient response to the drug ustekinumab and use this biomarker to stratify patient populations at various clinical endpoints. Moving beyond association analyses, we develop a machine learning model that integrates chemical structures and gene expression information to predict cellular responses to hundreds of chemicals, and apply methods to investigate how this model works. Finally, we improve model generalizability by pretraining a new model on massive amounts of gene expression data and then applying it to downstream prediction tasks. This work contributes models that predict biological responses to chemicals and methods to interpret how these models work. We anticipate this work will advance precision medicine by improving the ability to predict how patients respond to drugs. Simultaneously, this work provides an in silico platform for screening new biological models against a diverse set of chemical probes. Lastly, our pretrained model may find broad applications in phenotype prediction tasks ranging from disease risk modeling to drug response prediction.