GraFSyn: An Interpretable Deep Learning Framework for Anticancer Drug Synergy via Graphlet Fingerprints

Clinical Snapshot

Key Findings

• P Value: Not reported

• Effect Size: ROC-AUC: 0.972 (Merck), 0.823 (AstraZeneca); PR-AUC: 0.912 (Merck), 0.906 (AstraZeneca) — reported as superior to all representative baseline methods

• Primary Outcome: Binary anticancer drug synergy classification on two benchmark datasets

• Nnt Or Sensitivity: Sensitivity and specificity not individually reported; PR-AUC of 0.906–0.912 suggests strong positive predictive performance, particularly relevant given likely class imbalance in synergy datasets. No NNT applicable (computational study)

• Confidence Interval: Not reported

Clinical Application

Feasibility as a research tool for computational oncology and drug discovery laboratories is plausible, contingent on open-source code availability and computational infrastructure requirements. Clinical deployment would require prospective experimental validation, regulatory consideration, and integration into drug development pipelines — none of which have been demonstrated. GraFSyn has no direct current applicability to Australian clinical practice, PBS listings, or TGA regulatory pathways. It is a preclinical computational tool. In the Australian context, relevance lies in: (1) potential use by translational oncology research groups (e.g., ACRF-funded centres, WEHI, Garvan Institute) for hypothesis generation in drug combination screening; (2) alignment with the NHMRC's strategic priority of accelerating cancer drug discovery; (3) future relevance to TGA evaluation of combination therapies if predictions are experimentally validated and enter clinical development. RACGP guidelines on cancer management are not applicable at this stage. Oncologists and clinical pharmacologists should be aware of this class of tool as a future component of precision oncology pipelines, but should not alter prescribing practice based on in silico predictions alone. Not directly applicable to any patient population at this stage. The tool is designed for preclinical drug combination screening in oncology research settings, potentially applicable to cancer cell lines representing haematological and solid tumour types present in the Merck and AstraZeneca datasets.

Abstract

Predicting drug synergy is important for accelerating the discovery of effective anticancer combination therapies. Synergy intrinsically depends on the precise interactions of key chemical substructures within specific cellular environments. However, current molecular graph-based computational methods typically rely on implicit atom-level feature aggregation, which may obscure the topological representation of critical chemical substructures and limit structural traceability. Therefore, we present Graphlet Fingerprint-based Synergy prediction (GraFSyn), a deep learning framework for anticancer drug synergy prediction that uses graphlet fingerprints to encode drugs as explicit connected substructure units, preserving predefined chemical substructures and their topological identity. We further introduce a Dynamic Multi-Scale Convolution (DMSC) module to learn informative representations from high-dimensional and sparse graphlet features. The framework also includes an interaction module to capture context-dependent interactions between drug substructures and cell line gene expression. On the Merck and AstraZeneca benchmark data sets, GraFSyn achieved ROC-AUC/PR-AUC values of 0.972/0.912 and 0.823/0.906, respectively, outperforming representative baseline methods. In addition, attributed signals can be mapped back to specific pharmacophoric regions, supporting substructure-level analysis of synergistic interactions. In general, GraFSyn provides an accurate and structurally traceable approach for anticancer drug combination screening.

References

1. 1.Xia, W., Tian, Y., Zhou, S., Du, H., Xiao, M., Ge, Z., & He, X. (2026). GraFSyn: An interpretable deep learning framework for anticancer drug synergy via graphlet fingerprints. Journal of Chemical Information and Modeling. Advance online publication. https://doi.org/10.1021/acs.jcim.6c00458