Harnessing biomedical foundation models for genomic feature engineering to investigate patient drug response

Foundational models (FMs) have revolutionised AI. Pre-trained on vast quantities of data, FMs could be of value for inference on small cohorts e.g., zero-shot learning, but applications to multi-omics are nascent. Here we define a series of inference tasks that demonstrate the benefit of public multi-omic FMs (e.g., MAMMAL) for small genomics cohorts, not just for predictions, but for feature engineering tasks.

Our inference tasks aim to understand inter-patient variation in drug response. We study surgically resected diseased tissues from inflammatory bowel disease (IBD) patients. From these tissues REPROCELL Europe Ltd derived multi-omic (genomic, transcriptomic) data, and pharmacological responses (reduction in inflammatory cytokine TNFα with/without test drugs) during preclinical drug efficacy testing.

We showcase innovative and unique data driven workflows where FMs can provide discernible advantage for the inference of patient drug response from genomic data. For example, firstly, by calculating drug-target binding affinity (BA) for reference proteins enabling ranking and prioritisation of gene targets and associated SNPs for our drug of interest. Secondly, assessing the impact of patient SNPs on BA. In combination these approaches can guide biological discovery for target identification, derive new features for genes e.g., patient specific BAs, and guide feature selection for downstream prediction of drug response, even using classic Machine Learning. Downstream explainability distinguishes SNPs that are most influential for drug-target BA. We use our multi-threaded, FM-based, feature engineering approach to study inter-patient variation in drug response. We aim to expand contexts for FM use for non-experts.