SynProtX: a large-scale proteomics-based deep learning model for predicting synergistic anticancer drug combinations.

Area of Science:

• Computational Biology
• Drug Discovery
• Oncology

Background:

• Drug combination therapy is crucial for overcoming cancer's molecular heterogeneity and improving treatment outcomes.
• Deep learning models accelerate drug combination discovery, overcoming limitations of traditional experimental methods.
• Integrating protein-level expression data offers a more accurate cellular behavior and drug response representation than gene expression alone.

Purpose of the Study:

• Introduce SynProtX, a deep learning model that integrates large-scale proteomics with deep neural networks (DNNs) and drug molecular structures with graph neural networks (GNNs).
• Enhance the accuracy and efficiency of predicting effective anticancer drug combinations.
• Provide a framework for personalized medicine by leveraging multiomics data.

Main Methods:

• Developed SynProtX, a model combining graph neural networks (GNNs) for drug molecular structures and deep neural networks (DNNs) for proteomics data.
• Utilized graph attention network architecture (SynProtX-GATFP) to integrate molecular graphs and fingerprints.
• Employed rigorous validation strategies including cold-start prediction (leave-drug-combination-out, leave-drug-out, leave-cell-line-out).

Main Results:

• SynProtX-GATFP demonstrated superior predictive performance on the FRIEDMAN dataset and achieved high accuracy across diverse cell lines and datasets.
• Incorporating protein expression data consistently improved predictive performance compared to gene expression-only models.
• The model successfully identified key cancer-associated proteins and molecular substructures, revealing mechanisms of drug synergy.

Conclusions:

• SynProtX effectively integrates proteomics and drug structure data for enhanced anticancer drug combination prediction.
• The model's robust validation and identification of synergistic mechanisms highlight its potential for clinical applicability and personalized medicine.
• Leveraging large-scale proteomics and multiomics data is a promising avenue for advancing anticancer drug design.