A Ranking-Focused Graph Attention Autoencoder for lncRNA–Disease Association Prediction: A Real Hub-Disease Reversal Across Three Databases

Abstract

Graph-attention networks and graph autoencoders dominate lncRNA-disease association (LDA) prediction, but the field evaluates them almost exclusively as classifiers (AUC, AUPR) even though they are deployed as rankers: a biologist wants the top candidates for one lncRNA or disease, not a calibrated score for every possible pair. Ranking metrics (MAP, NDCG, Precision@k) and cold-start generalisation are rarely reported together.

Results:

We propose RF-GAAE, a heterogeneous graph-attention autoencoder trained jointly with a listwise ranking loss and a cross-view contrastive regulariser, and evaluate it on three independent real databases (LncRNADisease v3.0: 824 lncRNAs, 211 diseases; Lnc2Cancer 3.0: 625 lncRNAs, 132 diseases; MNDR v3.0: 636 lncRNAs, 169 diseases). RF-GAAE wins decisively under lncRNA cold-start on every database (MAP up to 55% relatively higher than attention-only baselines), but is consistently the weakest of six graph-based variants under disease cold-start, where a plain graph-convolution autoencoder wins instead. This reversal is statistically significant (Wilcoxon p = 0.002, 10-fold) and replicates across all three databases. A reimplemented GraLTR-LDA-style baseline (ranking loss without attention) consistently traces the weakness to graph attention rather than to the ranking or contrastive objectives across all three databases; a targeted hub-disease intervention confirms this on two of three, with the third (MNDR) reported as an honest exception. A supplementary leakage check shows the standard evaluation protocol partly inflates the effect's absolute size on every database without changing which component causes it.

Conclusions:

Ranking-aware training closes a genuine evaluation gap in LDA prediction, but graph attention's benefit is conditional on disease-degree structure, not universal, and this study identifies the mechanism, and where it does and does not generalise, rather than only the correlation