Target-biology and interactome-derived signatures predict target-level associations with safety-related drug attrition

Clinical drug development suffers from high rates of toxicity-related failure despite the use of compound-centric preclinical safety screening, with approximately one-third of all clinical failures attributable to safety concerns. An ability to prioritize early-stage drug development programs toward those with a lower probability of causing clinical toxicity would improve drug development success rates. Here, we propose a target-centric framework that integrates network medicine principles with target biology features to predict operational target-level labels associated with safety-related drug attrition. From a set of 3,696 drugs with widely launched or safety-related termination outcomes, we curated 541 non-overlapping target labels, comprising 302 safety-liability-associated targets and 239 widely launched-associated targets. We engineered target-level features encoding both biological properties and human interactome (HI) topology and trained a gradient boosting classifier to predict the safety-liability-associated label. The model achieved a held-out test ROC AUC of 0.712. These results suggest that target biology and interactome context contain signal associated with safety-related clinical attrition and may support early target prioritization when used alongside compound-centric safety assessments.