Closing the gap between the biology and the clinic with a foundation model of immunology and inflammation

ICML 2025 Workshop FM4LS Submission8 Authors

Keywords: Machine Learning, ICML, Foundation Model, Computational Biology, Immunity and Inflammation, RNA-seq, Precision Medicine

TL;DR: We present EVA, the first foundation model for immunology and inflammation, which integrates transcriptomic and clinical data to predict drug efficacy and patient response in silico, enabling more scalable and personalized drug development.

Abstract: Drug development is a risky and lengthy process during which most compounds fail due to insufficient efficacy in the clinic. For this reason, there is an urgent need for scalable in silico approaches that can help identify promising drug candidates using preclinical data. We introduce EVA, the first foundation model for immunology and inflammation (I&I), specifically tailored to power such applications at the interface between transcriptomics (handling both bulk and single-cell data) and clinical modalities, pretrained on a large corpus of single-cell and bulk RNA-seq samples sourced from various clinical studies and databases. Throughout a retrospective in silico study conducted in ulcerative colitis, we illustrate how EVA can be fine-tuned to predict the clinical effect of new compounds in I&I patients, leveraging preclinical data from disease models and observational cohort data. We model the biological effect of the drug at the patient level as a transcriptomic perturbation in the primary organ, which can be extracted from EVA's latent gene representations. Using a secondary disease model, this perturbation can be transformed into a predicted change in disease activity. In addition to direct drug effect forecasting at the patient level, the pipeline output can also be used to stratify patients according to their expected drug responses, enabling the early identification of biomarkers of response in investigational treatments. This first-in-class study highlights how foundation models in computational biology can be harnessed to address modeling challenges in drug discovery, bridging the gap between molecular and clinical data, and paving the way for more effective and personalized treatments.

Submission Number: 8