In Silico Drug Screening with Mechanistic Insight: A Cross-Attention Framework for Predicting Drug-Gene Interactions

Identifying how drug molecular structures interact with specific gene targets remains a key challenge in cancer precision medicine. Although large-scale in vitro drug screening datasets are available, most computational models fail to elucidate the molecular mechanisms underlying drug-gene interactions or inform mechanisms of action (MOA). Here, we present MIDI (Mechanism-Interpretable Drug-Gene Interaction), a novel, attention-guided, mechanism-aware deep learning framework designed to address this gap. MIDI is trained on 80% of cell lines from the Cancer Cell Line Encyclopedia (CCLE) and evaluated on the remaining 20%, demonstrating improved drug response prediction performance over existing benchmarks. Leveraging chemical structure information, MIDI incorporates a Graphformer-based self-attention mechanism to capture atomic-level interactions between small molecules and gene features. A cross-attention module further enables the model to map drug structures to candidate gene targets, supporting interpretable mechanistic inference. To simulate a realistic in silico drug screening scenario, we evaluated MIDI on 52 previously unseen drugs from the Genomics of Drug Sensitivity in Cancer (GDSC) dataset. MIDI consistently ranked true gene targets among the top candidates across thousands of possibilities. These results highlight the feasibility and utility of in silico drug screening with MIDI, enabling both predictive modeling and mechanistic exploration. By uncovering interpretable drug-gene relationships, MIDI provides a scalable, data-driven framework to accelerate therapeutic discovery and target prioritization in preclinical research.