Skip to main content
Back to timeline
arXivSource publication:

Drug synergy prediction stops choosing between molecules and networks: BiGSyn aligns both graph levels and reaches 94.37 AUC on the independent test

Synopsis

Drug synergy prediction has mostly been modeled at a single level, using either atom-level molecular graphs or a biological network of drugs, targets, cell lines and tissues, so cross-level signals were hard to fuse. BiGSyn uses heterogeneous graph transformers to learn representations from molecular graphs and biological subgraphs, then a bi-level contrastive objective pulls the two representations of the same drug together. On the Oncology-Screen and DrugCombDB benchmarks it beats six compared methods on independent-test metrics, with an AUC of 94.37 on Oncology-Screen.

Source-provided article image: BiGSyn: A Bi-Level Heterogeneous Graph Contrastive Learning Framework for Interpretable Drug Synergy Prediction
arXiv

Interpretation

Drug synergy prediction now draws on both atom-level molecular graphs and entity-level biological networks instead of modeling one level alone. Earlier approaches used only molecular graphs, losing the biological context in which drugs act, or only biological networks, losing fine-grained chemical structure. On the independent test sets of the Oncology-Screen and DrugCombDB benchmarks, BiGSyn exceeds six compared methods, DeepSynergy, DeepTraSynergy, GraphSynergy, KGANSynergy, SynergyGTN and GraphTranSynergy, on all six metrics ACC, Precision, Recall, AUC, AUPR and F1.

The two levels are learned with heterogeneous graph transformers, aligned by contrastive learning, concatenated into a fused drug representation and passed with cell-line features to a classifier that decides synergy. A heterogeneous graph transformer is a network that propagates information by attention over graphs with several node and edge types; the molecular graph has atoms as nodes and chemical bonds as edges, while the biological network holds drugs, target proteins, cell lines and tissues. Evaluation uses three kinds of five-fold cross-validation, splitting by random sample, by drug pair and by cell line, with 10% of samples held out as an independent test set; under the stricter drug-pair and cell-line splits BiGSyn still gives the best independent-test performance.

Perspective

The demonstrated scope is binary synergy classification on the Oncology-Screen and DrugCombDB benchmarks, with evaluation over sample-level, drug-pair-level and cell-line-level five-fold splits and a separate 10% independent test set.

Validation performance is slightly lower under the cell-line split, indicating that predicting synergy labels for unseen cell lines is harder; agreement between high-attention regions and known pharmacological sites is suggestive evidence, and candidate combinations in the case study were supported by a non-exhaustive literature search.