Accurate prediction of drug-induced gene expression profiles is crucial for phenotype-based drug discovery. Although computational methods have shown potential, they struggle with the complexities of varying doses and durations. To overcome these limitations, we developed DeepICER, a model that predicts gene expression profiles induced by chemical perturbations across any dose and duration. Utilizing a bilinear attention mechanism, DeepICER captures the interplay between dose, duration, and basal gene expression, enabling accurate predictions for novel compounds and cell lines. DeepICER outperforms existing models with superior flexibility in handling any dose and duration and accuracy, achieving a 45.1% improvement in predictive performance. Experimental validation confirmed that PD-166285, identified by DeepICER, exhibits stronger inhibitory effects on A549 cells compared to paclitaxel. To enhance accessibility, DeepICER is developed as an online platform, providing researchers with a tool to predict gene expression in compound-treated cells, thereby advancing drug repurposing and accelerating drug discovery.

Drug discovery  /  Chemical perturbations  /  Gene expression profile  /  Deep learning  /  Bilinear attention mechanism