Toward Explainable Anticancer Compound Sensitivity Prediction via Multimodal Attention-Based Convolutional Encoders.

Area of Science:

• Computational biology
• Cheminformatics
• Cancer research

Background:

• Accurate prediction of anticancer compound sensitivity is crucial for personalized medicine and drug discovery.
• Existing models often lack interpretability or fail to integrate diverse data types effectively.

Purpose of the Study:

• To propose a novel, interpretable multimodal architecture for predicting anticancer compound sensitivity.
• To leverage compound structure, gene expression, and protein-protein interaction networks for enhanced prediction accuracy.

Main Methods:

• Developed a multimodal attention-based convolutional encoder integrating SMILES sequences, gene expression profiles, and protein-protein interaction networks.
• Utilized multiscale convolutional attention mechanisms for feature extraction and interpretation.
• Compared model performance against baseline models and state-of-the-art methods.

Main Results:

• The proposed model significantly outperformed baseline and state-of-the-art methods for multimodal drug sensitivity prediction (R² = 0.86, RMSE = 0.89).
• Attention weight analysis revealed biologically relevant gene enrichments (apoptotic processes) and correlation with chemical structure similarity.
• A case study demonstrated the model's ability to focus on informative genes and drug substructures for specific inhibitors.

Conclusions:

• The interpretable multimodal model shows strong potential for in silico prediction of anticancer compound efficacy.
• The approach supports personalized therapy development and candidate compound evaluation in de novo drug design.
• The model's generalizability and interpretability offer a valuable tool for advancing cancer drug discovery.