DeepICER: A deep learning framework for predicting compound-induced gene expression profiles

Area of Science:

• Computational biology
• Pharmacology
• Genomics

Background:

• Accurate prediction of drug-induced gene expression is vital for phenotype-based drug discovery.
• Existing computational methods face challenges with variable drug doses and treatment durations.
• There is a need for advanced models to predict gene expression changes under diverse perturbation conditions.

Purpose of the Study:

• To develop DeepICER, a novel computational model for predicting gene expression profiles induced by chemical perturbations.
• To enable accurate predictions across a wide range of drug doses and durations.
• To improve the flexibility and accuracy of gene expression prediction for drug discovery.

Main Methods:

• Developed DeepICER, a model employing a bilinear attention mechanism.
• The model integrates dose, duration, and basal gene expression data.
• Evaluated DeepICER's performance against existing computational methods.

Main Results:

• DeepICER demonstrated superior flexibility in handling any dose and duration compared to existing models.
• Achieved a 45.1% improvement in predictive performance.
• Experimental validation identified PD-166285 as a potent inhibitor, outperforming paclitaxel in A549 cells.

Conclusions:

• DeepICER offers a powerful and accessible tool for predicting drug-induced gene expression.
• The model advances drug repurposing and accelerates the drug discovery pipeline.
• DeepICER's online platform enhances accessibility for researchers worldwide.