CombDNF: Disease-specific drug combination predictions with network-based features on clinically validated data

Drug combinations are increasingly applied to treat a wide range of complex diseases. Drug action and thus also drug combination effects can differ between diseases, e.g., due to molecular differences. Therefore, disease-specific predictions are required for treatments. A plethora of methods based on cell-line screening data in cancer has been proposed. However, their extendability to other diseases is limited, as is their applicability in the clinical context due to the in-vivo-in-vitro gap. In contrast, only few approaches rely on clinically validated data. Here, we propose CombDNF, a novel machine-learning based method for disease-specific drug combination prediction on clinically validated data. CombDNF is trained and predicts both on clinically approved (effective) and clinically reported adverse drug combinations from a broad collection of data sources. It can cope with the highly imbalanced label distribution in drug combination data. Further, CombDNF leverages network-derived features based on drug target and disease gene relationships. To incorporate uncertainty of the network topology it relies on edge weights in the underlying network. We systematically evaluate CombDNF against available state-of-the-art methods in four diseases with different underlying mechanisms and ground truth data characteristics. CombDNF outperforms all state-of-the-art methods in all four diseases by at least 84% in the AUPR. This translates, for example, to an enrichment of effective drug combinations in the top ten hypertension-specific predictions of four, compared to one for the best competing method. In addition, network edge weighting by interaction confidence scores indeed yields improved predictions. Further, we find evidence for biological plausibility of our top ranked drug combinations. We make the training and evaluation pipeline for CombDNF available ready-to-use at https://github.com/DILiS-lab/CombDNF.