A calibrated heterogeneous-network ensemble framework for predicting disease-associated non-coding RNAs, with leakage-safe evaluation and label-provenance analysis

Motivation: Computational prediction of non-coding RNA (ncRNA)-disease associations is dominated by network-propagation and matrix-completion methods whose reported accuracies are often close to perfect, a pattern that frequently reflects evaluation leakage and unexamined positive-label provenance rather than biological signal. Results: We present HetRWR-Ens, a framework that integrates miRNA, long non-coding RNA (lncRNA) and circular RNA (circRNA) associations within a single heterogeneous network (6,297 nodes; 24,495 associations across 52 diseases), combines Random Walk with Restart (RWR) propagation with Gaussian interaction profile (GIP) similarity features, and calibrates a gradient boosting-random forest soft-voting ensemble. The evaluation protocol is leakage-safe: within every cross-validation fold, GIP kernels, the propagation graph and all derived features are rebuilt from training-fold associations only, training features are additionally cross-fitted so that no sample's target edge contributes to its own feature vector, probabilities are calibrated by Platt scaling on held-out inner-block predictions, and decision thresholds are tuned on training data only. Under a stratified 10-fold protocol, the calibrated ensemble attained a mean AUROC of 0.942 and AUPRC of 0.744, with F1 of 0.711 and MCC of 0.665 at fold-tuned thresholds; it outperformed logistic regression, kernel-matrix baselines and pure propagation scores (all p = 0.002, Wilcoxon over 10 folds) and the degree-popularity baseline (p = 0.006), and matched or exceeded a single gradient boosting model in ranking while remaining well calibrated (Brier score 0.083). Beyond predictive performance, a label-provenance analysis showed that 98.4% of lncRNA and 100% of circRNA positive labels derive from computational predictions rather than experiments, that the model largely reproduces such prediction-derived labels (pooled out-of-fold AUROC 0.983 for lncRNA) while transferring imperfectly to experimentally supported lncRNA associations (AUROC 0.814 when experimental labels are seen in training; 0.452 when trained exclusively on computational labels), and that diseases whose labels are dominated by dense computational predictions show degraded per-disease discrimination. Case-study candidates for psoriasis, pancreatic ductal adenocarcinoma and glioma were prioritised by rank and validated against recent literature: seven of the fifteen top-five candidates, including hsa-mir-21-psoriasis and hsa_circ_0001649-pancreatic ductal adenocarcinoma, have direct published support.