Heterogeneous Relation-Aware Attention Improves Cold-Start miRNA–Disease Association Prediction

Abstract—Predicting associations between microRNAs (miRNAs) and diseases helps prioritise candidates for experimental
study. Most graph-learning approaches are evaluated solely
in the transductive context, where every miRNA already has
known associations. The tougher and more practically relevant cold-start setting, predicting associations for miRNAs with
no prior relationships, is rarely investigated. We study how
architectural complexity behaves across both settings on the
HMDD benchmark. We propose HetSeqFormer, a spectral graph
Transformer with heterogeneous relation-aware attention over
an association graph augmented with sequence-based miRNA–
miRNA and disease–disease similarity networks, and a learned
sequence branch. Under a leakage-free protocol we find that
the value of heterogeneity is strongly setting-dependent. In
the transductive setting it does not help: inductive baselines
(GraphSAGE, a graph Transformer) lead and a streamlined
variant of our model only matches DARSFormer. In the coldstart setting the ranking inverts: HetSeqFormer ranks first and
significantly outperforms six of seven baselines (all but GCN,
which it ties on AUC and leads on F1, 0.841 vs. 0.840), while
the transductive leaders GraphSAGE and the graph Transformer
fall to last. Ablations confirm that relation-aware attention—not
the sequence branch—drives the cold-start gain, by propagating
information through the similarity networks to association-free
miRNAs. These results identify when heterogeneity helps in
miRNA–disease prediction and highlight cold-start as the setting
where it matters.