Order-Aware Deep Learning for Drug Combination Benefit Prediction in Cancer Cell Lines

Drug combination therapy has exhibited favorable effects in treating cancer patients, with less toxicity and adverse reactions compared to monotherapy. To accelerate the discovery of therapeutic drug combinations, numerous computational methods have been developed to predict drug synergy in cancer cell lines, typically modeling the task as binary classification (synergistic vs non-synergistic) or regression (continuous synergy scores). Yet, a recent study proposes categorizing drug combination benefits into multiple ordered classes (e.g., synergy, bliss additivity, independent actions) based on clinical activities, and suggests that drug combinations remain valuable if they reduce cancer cell viability, even without defined synergy. To distinguish various levels of combination benefits, we present a novel order-aware deep learning model, called OrderCombo. Specifically, OrderCombo extracts the drug representation via a pretrained chemical language model and the cell line representation via an omics-oriented linear network. Then, these representations are fused into a unified embedding for each drug-drug-cell line triplet, by leveraging a hybrid encoder that combines concatenation-based dependencies and attention-based interactions. Finally, an ordinal contrastive loss is designed to promote a discriminative embedding space and maintain class ordinality, thereby improving the predictions of drug combination benefits. We evaluate OrderCombo on a large-scale combination benefit dataset, and in silico results show that our method outperforms the state-of-the-art baselines in terms of prediction accuracy, while maintaining robust generalization to unseen drug pairs and cell lines. Substantial case studies further demonstrate OrderCombo's potential value in discovering novel anticancer drug combinations across different therapeutic levels.