Data splits, predictions and code for "Mechanism-aware features, not architecture, determine drug–drug interaction severity prediction under drug-disjoint evaluation"

Supporting materials for a controlled study of drug–drug interaction severity prediction under drug-disjoint evaluation. Includes the locked random and leave-drug-out partitions, raw per-pair predicted probabilities for every model, protocol and seed, pair-level mechanism and transcriptomic feature matrices, the fitted PCA projection, paired DeLong test outputs, calibration and decision-rule analyses, the mechanism ablation and annotation-noise study, all figures, and every analysis script. All tables and figures in the associated paper can be regenerated from the deposited predictions without retraining any model. The partitions were written to disk before any model was trained and must not be regenerated. Interaction records derive from DrugBank (academic licence) and severity labels from DDInter. The deposit contains DrugBank identifiers and derived labels only; molecular structures, fingerprints and DrugBank annotation records are excluded in accordance with the DrugBank academic licence. Holders of a DrugBank licence can regenerate them using the deposited preparation scripts. Academic access is available on registration at https://go.drugbank.com. The raw LINCS L1000 data is available from GEO under accession GSE92742. The analysis code is additionally available under the MIT License.