Computational & AI Journal of Cheminformatics
A deep learning model links drug structure, cell-line gene signals and biological processes to predict synergistic anticancer combinations with interpretable attributions
- SGP-Syn encodes each drug with a language-model and a graph-based representation fused by token-to-atom attention, then combines the drug pair with cell-line expression to infer gene-level signals.
- Those signals pass through an ontology-structured neural network that yields hierarchical biological process representations, giving a structure-to-gene-to-process explanation chain.
- Retrospective tests on public combination screens show gains over leading methods, especially for unseen drugs and drug pairs, with only comparable results for unseen cell lines; attribution analyses suggest testable hypotheses but are not experimentally validated.
Synergy models usually look best under random splits, where the same drugs and cell lines leak across train and test, so the unseen-drug and unseen-pair gains are the part worth believing. But the weak result on unseen cell lines is what matters on a real project, since a new model system is where you most need a prediction. Before trusting the attributions, I would check whether the gene and process explanations survive randomized-ontology and shuffled-expression controls, then pick a few top-ranked combinations for a small dose-matrix test.
Paper: Journal of Cheminformatics ↗ Biologically informed Structure–Gene–Process integration for anticancer drug combination synergy prediction