Multi-omics feature engineering driven by biomedical foundation models improves drug response prediction for inflammatory bowel disease patients

Abstract Using biomedical foundation models (FMs) for inference on small cohorts, represents a promising and practical route to advance drug response biomarker discovery and target identification. Here, we demonstrate this via an innovative data-driven inference workflow, using a fine-tuned, biomedical FM. We study multi-omics (genomic, transcriptomic) data and predict pharmacological responses, both from surgical diseased tissue of inflammatory bowel disease (IBD) patients. We use FM inference to inform feature selection and feature engineering strategies, where FM-derived features provide advantage for predicting IBD patient drug response and target identification. Firstly, calculating drug-target binding affinity (BA), enabling prioritisation of protein/gene targets and associated SNPs for drugs of interest. Secondly, using patient SNPs to mutate reference proteins and assess impact on drug BA. Thirdly, building strategies to fuse BAs and transcriptomics. Additionally, we created an open-source Model Context Protocol server, making our FM inference example accessible to the community via AI agents and natural language prompts.