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

The SynProtX-GATFP model, which combines molecular graphs and fingerprints through a graph attention network architecture, demonstrated superior predictive performance for the FRIEDMAN study dataset. We further evaluated its cell line-specific performance, which achieved accuracy across diverse tissue and study datasets. By incorporating protein expression data, the model consistently enhanced predictive performance over gene expression-only models, reflecting the functional state of cancer cells. The generalizability of SynProtX was rigorously validated using cold-start prediction, including leave-drug-combination-out, leave-drug-out, and leave-cell-line-out validation strategies, highlighting its robust performance and potential for clinical applicability. Additionally, SynProtX identified key cancer-associated proteins and molecular substructures, offering novel insights into the biological mechanisms underlying drug synergy. These findings highlight the potential of integrating large-scale proteomics and multiomics data to advance anticancer drug design and combination therapy strategies for personalized medicine. Availability and implementation: https://github.com/manbaritone/SynProtX.