Mechanism-Aware Prediction of Tissue-Specific Drug Activity via Multi-Modal Biological Graphs

Abstract

Predicting how small molecules behave across human tissues is essential for targeted therapy development. While some existing models incorporate tissue identity, they treat it as a label—ignoring the underlying biological mechanisms that differentiate tissues. We present Expresso, a multi-modal architecture that predicts tissue-contextual molecular activity as measured by the assay by modeling how compounds interact with transcriptomic and pathway-level tissue context. Expresso constructs heterogeneous graphs from GTEx data, linking samples, genes, and pathways to reflect expression profiles and curated biological relationships. These graphs are encoded using a hierarchical GNN and fused with frozen molecular embeddings to produce context-aware predictions. A multi-task pretraining strategy—spanning gene recovery, tissue classification, and pathway-level contrastive learning—guides the model to learn mechanistically grounded representations. On nine tissues, Expresso improves mean AUC by up to 27.9 points over molecule-only baselines. Our results demonstrate that incorporating biological structure—as defined by the assay—yields more accurate and interpretable models for tissue-specific drug behavior in human cell-based in vitro assay systems.