THBKG: A Temporal Biomedical Knowledge Graph for Decision-Aligned Clinical Advancement Prediction

Inadequate target–disease linkage accounts for 40–50% of Phase II efficacy failures, making it critical to anticipate which therapeutic programmes will advance. THBKG (Temporal Heterogeneous Biomedical Knowledge Graph) is designed to reconstruct the evidence profile of a target–disease pair as it existed when the pair entered the clinic, which no prior knowledge graph supports. The graph contains 110,396 entities and 11.1M edges across nineteen relation types, with each edge carrying the year its evidence changed, so a pair's profile can be recovered at any past decision point.

The authors define a decision-aligned benchmark that predicts, for a target–disease pair entering Phase II, whether it advances to Phase III using only evidence datable before that decision. Graph propagation over THBKG outperforms every direct-evidence reference under the same protocol, reaching a relative success of 4.3–4.5 at the top ten pairs per therapeutic area. The improvement concentrates on the 72.8% of pairs with no direct target–disease evidence at their decision point, where a direct-edge model has nothing to read; the graph-based encoders still rank five- to sixfold above chance by propagating over intervening biology. A path-based explainer adapted to the decision-time subgraph decomposes each prediction into the evidence landscape behind the hypothesis, enabling explainable prediction. The authors release THBKG as a continually updated substrate for retrospective validation of therapeutic target hypotheses.