A Multi-View Sequence–Ensemble Fusion Model for Predicting Cancer-Associated microRNA–Gene Interactions

Abstract—microRNAs (miRNAs) are short non-coding RNAs
that post-transcriptionally regulate target genes and are strongly
implicated in oncogenesis. Determining whether a given miRNA–
gene interaction operates in a cancer or a non-cancer disease
context is valuable for prioritising candidates in functional
studies, yet it is non-trivial because the same miRNA frequently
participates in both settings. We study a curated dataset of
3,666 miRNA-gene-disease associations (2,004 cancer; 1,662 noncancer) and first show that the disease-name field is a definitional
label proxy that trivialises the task (accuracy 99.8%, ROC-AUC
1.000) without conferring any biological insight. We therefore
reformulate the problem as predicting cancer association from
molecular features only: the miRNA identity, its mature nucleotide
sequence, and the target gene. We propose mSEF (Multiview Sequence–Ensemble Fusion), which couples (i) a leakageaware task formulation, (ii) a multi-view feature representation
combining a k-mer sequence spectrum, engineered nucleotidecomposition descriptors, miRNA-family attributes, and crossfitted target encodings of high-cardinality categoricals, with (iii)
a weighted heterogeneous soft-voting ensemble of extremely randomised trees, a random forest, and histogram gradient boosting.
Under stratified 5-fold cross-validation, mSEF attains 90.6%
accuracy, 0.915 F1, 0.967 ROC-AUC, and 0.811 Matthews correlation coefficient (MCC), outperforming nine standard baselines,
and reaches 91.0% accuracy with 0.969 ROC-AUC on a held-out
test set. Under a stricter grouped protocol that withholds entire
miRNAs or genes, accuracy remains 80–83% (ROC-AUC 0.87–
0.92), quantifying generalization to unseen molecules. Ablation
shows that all three feature views contribute and that their fusion
is necessary for the best result. The study illustrates how careful
leakage control turns an apparently trivial classification problem
into a meaningful, sequence-driven prediction task.