Drug-induced liver injury is one of the hardest safety problems in drug development. It is a leading reason drugs are pulled from the market, and it often shows up late, after a compound has already absorbed years of work. A signal you can read early, from biology rather than from a clinical surprise, is worth a great deal.
A team at the Department of Medical Informatics at Erasmus MC Rotterdam has now published such a test, and the substrate they used is the Euretos knowledge graph, the same harmonised evidence foundation that Disease Atlas is built on. The paper appeared in Chemical Research in Toxicology (American Chemical Society) on 8 July 2026. It is worth reading closely, because it is an independent, peer-reviewed result rather than a claim from the people who built the data.
The question the authors set was general: can you classify the semantic relations between a drug and an adverse reaction well enough to predict the reaction? They picked drug-induced liver injury, or DILI, as the worked case, because it has a clean public benchmark to score against.
Their method walks the graph. For a given drug and a given adverse reaction, they collect two-hop paths that connect the two through an intermediate biological concept, usually a gene or a protein. They group these paths into what they call path bundles, and they treat each bundle as the basic unit of analysis. A context-aware, four-level statistical model then reads the structure of those bundles and decides whether the drug is likely to cause the reaction. The context matters: the same intermediate concept can carry different weight depending on what surrounds it in the graph, and the model is built to account for that.
Scored against the FDA DILIrank dataset of DILI-positive and DILI-negative drugs, the four-level model reached an AUC of 0.850 and a Matthews correlation coefficient of 0.639. On average it assigned a predicted probability of 0.641 to drugs that do cause liver injury and 0.402 to drugs that do not, so the separation is real and not a knife-edge.
Two comparisons make the result more interesting than a single benchmark number. First, the model correctly reclassified 29 of 37 drugs that an earlier structure-based QSAR study had gotten wrong. Second, on a 130-compound subset of the test set used by DeepDILI, a deep-learning DILI classifier, combining the graph approach with a molecular-descriptor QSAR model rescued DILI-positive drugs that the deep-learning method missed on its own. The biology-derived signal and the chemistry-derived signal were catching different things, and together they did better than either alone.
The authors released their code and data at github.com/mi-erasmusmc/multilevel_dili, so the work is reproducible rather than asserted.
Most claims about what a biomedical data resource can do come from the people who sell it. This one does not. A group at a top European medical centre took the graph, built their own model on it, ran it against an FDA benchmark, compared it against two independent methods from the toxicology literature, and published the result in a society journal. That is the kind of evidence that survives a skeptical reading.
The finding underneath the numbers is the part worth sitting with. A structure-only model has no way to see that a compound perturbs a pathway the liver depends on. A graph that connects the drug to that pathway, and the pathway to hepatotoxicity, does. The rescued cases were not noise. They were drugs whose risk lived in the biology, not in the molecular structure, and the biology was readable in the graph.
Disease Atlas draws on the same evidence foundation, and safety is one of the dimensions it scores. For any candidate, the safety layer predicts adverse-event risk organ by organ, including the liver, alongside genetic essentiality and expression context. So the two lines of work sit next to each other rather than on top of each other. The Erasmus MC paper tests the graph's hepatotoxicity signal at the level of a drug and its adverse reactions. Disease Atlas carries that kind of signal into per-candidate, per-organ risk scoring for target decisions. They are complementary, and the paper is independent evidence that the foundation both rely on holds real toxicological signal.
For a researcher weighing a candidate, the practical version of this is simple. Liver risk is one of the axes on which a candidate can be ranked down, and that axis is not guesswork. It rests on a foundation that an outside group has now shown can discriminate hepatotoxic from non-hepatotoxic drugs against an FDA benchmark.
The result that will age well here is not the AUC. It is the demonstration that structure and biology are not redundant. A model reading chemical structure alone missed drugs whose liver risk was written into their biological context, and a model reading that context caught them. Reading disease and drug biology at the level of the mechanisms involved, rather than at the level of the molecule alone, changes which calls you get right. That is the premise Disease Atlas is built on, and it is good to see it tested by someone with no reason to flatter it.
Reference: Vos R, van Mulligen EM, Kors JA. Context-Aware Multilevel Classification of Semantic Relations in Drug-Adverse Drug Reaction (ADR) Networks: Predicting Drug-Induced Liver Injury (DILI) as a Case Study. Chemical Research in Toxicology, 2026. DOI: 10.1021/acs.chemrestox.5c00252. Code and data: github.com/mi-erasmusmc/multilevel_dili
.svg)
