GraphSynergy: A Network‑Inspired Deep Learning Model for Predicting Anti‑Cancer Drug Combinations

GraphSynergy is a network‑inspired deep learning framework that predicts anti‑cancer drug combination efficacy by leveraging protein‑protein interaction (PPI) networks and graph convolutional networks (GCN).

The model defines a Therapy Score to estimate therapeutic benefit and a Toxicity Score to assess over‑exposure risk, combining them via a sigmoid function to output a synergy probability.

Key innovations include attention‑based neighborhood aggregation for protein nodes, multi‑layer embedding of drug, cell‑line, and protein features, and explicit modeling of local sub‑graph proximity (network proximity, separation score).

Training and evaluation use integrated datasets from DrugCombDB and Oncology‑Screen, comprising thousands of drug‑cell line triples, PPI edges, and drug‑protein associations.

GraphSynergy was benchmarked against seven baselines (Network Proximity, GraRep, DeepWalk, Node2Vec, DeepSynergy, GCN, KGNN) and consistently achieved higher AUC on both datasets, demonstrating the advantage of incorporating rich topological information.

Model variants explored the depth of the radiative field (optimal at two hops) and different embedding integration strategies, confirming that concatenating drug embeddings before interaction with cell‑line embeddings yields the best performance.

A case study on non‑small‑cell lung cancer (NSCLC) showed that the model could predict effective novel combinations (e.g., Pemetrexed + Gefitinib) that were later validated in clinical studies.

Future work aims to extend the single‑layer PPI network to multilayer biological networks, improve interpretability by tracing mechanistic pathways, and apply the approach to other cancer types such as prostate cancer.

The presentation concluded with a Q&A discussing data alignment, heterogeneous graph computation, and sub‑graph extraction challenges.