UIDDA: a unified-input model-classifier combination framework for drug-disease association prediction

Drug-disease association (DDA) prediction methods differ substantially in data preprocessing, feature construction, model architecture, negative-sample treatment, and evaluation protocols, which limits fair comparison across studies. We developed UIDDA, a unified-input model-classifier benchmark that standardizes data processing and separates representation learning from downstream classification. UIDDA evaluates nine feature extraction models combined with four classification heads on four public DDA datasets. To prevent information leakage, association matrices, Gaussian interaction profile similarities, heterogeneous biomedical networks, and association-dependent representations were reconstructed independently within each training fold. Under the pair-level Random-U setting, FD-MSGL and AMDGT achieved the strongest overall performance, whereas DeepDR, HNetDNN, and LDSCNN showed lower average discrimination. Model selection produced substantially larger performance variation than classifier selection; the mean AUC difference between the best and worst models was 0.266, compared with 0.032 among classification heads. Dataset difficulty also varied, with the B-dataset achieving the highest overall mean AUC and the sparse T-dataset the lowest. The benchmark remained informative under alternative treatments of unobserved pairs: overall AUC/AUPR values were 0.724/0.732 for Random-U, 0.692/0.709 for Hard-U, and 0.675/0.694 for nnPU. To assess entity-level generalization, we considered three entity-held-out settings: drug-held-out, disease-held-out, and both-held-out, in which the test drugs, test diseases, or both entity types were excluded from training, respectively. The corresponding mean AUC values were 0.692, 0.614, and 0.552, showing that generalization became more difficult when unseen entities were introduced. Adding MolVis-inspired three-dimensional molecular features to FD-MSGL produced only small and dataset-dependent changes. A frozen top-10 retrospective literature assessment provided a limited qualitative plausibility check for ranked model outputs. UIDDA provides a standardized, leakage-controlled, and reproducible framework for separating the impact of feature representation and classification decisions and for comparing and evaluating DDA prediction methods. The results indicate that model selection is the main source of performance variation, and the conclusions depend on the underlying processing and evaluation settings.