GPDRP: a multimodal framework for drug response prediction with graph transformer.

Area of Science:

• Computational biology
• Bioinformatics
• Drug discovery

Background:

• Drug response prediction (DRP) is crucial in personalized medicine.
• Existing DRP methods often use simplified drug representations (strings) and overlook gene pathway interactions.
• A more sophisticated approach is needed to model drug molecules and their biological effects.

Purpose of the Study:

• To develop a multimodal deep learning model for accurate drug response prediction.
• To represent drugs using molecular graphs and cell lines via gene pathway activity.
• To integrate these representations for improved predictive performance.

Main Methods:

• Proposed the drug Graph and gene Pathway based Drug response prediction method (GPDRP).
• Utilized Graph Neural Networks (GNN) with Graph Transformers for molecular graph processing.
• Employed deep neural networks for gene pathway activity analysis.
• Integrated predictions using fully connected layers.

Main Results:

• The Graph Transformer-based approach within GPDRP demonstrated superior performance.
• GPDRP outperformed existing models on hundreds of cancer cell lines using bulk RNA-sequencing data.
• The model showed strong generalizability and applicability on novel drug-cell line pairs and xenografts.

Conclusions:

• GPDRP offers a powerful and interpretable method for drug response prediction.
• Incorporating gene pathways enhances the model's ability to understand drug effects.
• The findings support the use of molecular graphs and pathway data in computational personalized medicine.