Deep learning‐based prediction of insect odorant receptor–ligand interactions for behavioral regulator discovery

BACKGROUND: Widespread chemical insecticide use causes resistance and ecological damage, driving need for sustainable pest control via olfaction-based behavioral regulators targeting odorant receptors (ORs). Conventional screening is costly and low-throughput. Deep learning (DL) offers solutions but faces challenges of scarce data, distinct OR-ligand binding properties, and requirements for explainability to guide regulator design. RESULTS: To explore effective representations of volatile compounds under a limited data setting, we developed three deep learning models based on molecular descriptors (Concat-Desc), molecular graphs (Concat-GNN), and Simplified Molecular Input Line Entry System (SMILES) sequences (Cross-Attn) for OR-ligand interaction prediction. Trained on 4566 experimental pairs from M2iOR database, all DL models outperformed quantitative structure-activity relationship (QSAR) baselines. Concat-Desc and Cross-Attn achieved superior performance (test area under the receiver operating characteristic curve (AUC) = 0.891-0.927; accuracy = 0.810-0.843). The explainability modules highlighted model-associated ligand features and protein sequence positions. CONCLUSION: Molecular descriptors and SMILES are effective for volatile compounds in data-limited scenarios, while graph representations require further optimization. Integrated explainability modules provide testable hypotheses for experimental validation and future design, establishing a computational framework for artificial intelligence (AI)-driven insect chemical ecology. Limitations include data scarcity and limited species coverage that future studies must address.