Processed MTI Dataset for PAIR-Former (NeurIPS 2026 Anonymous Submission)

Processed pair-level miRNA-mRNA target interaction dataset (~420K pairs) used by the NeurIPS 2026 anonymous submission "PAIR-Former: Budgeted Relational Multi-Instance Learning for Functional miRNA Target Prediction". File: MTI_pair_random_split.txt (~785 MB, tab-separated, columns: miRNA_name, mRNA_name, CTS_sequence, label, additional features, split). Authors and affiliations will be revealed upon paper acceptance. We evaluated PAIR-FORMER on MTI-Synth (miRNA–Target Interaction Synthetic data), a large-scale benchmark constructed from TarBase v9. We first selected human miRNA–target interactions with experimental support from either Western blot or reporter assay experiments. Western blot assays validate miRNA regulation at the protein level by detecting changes in target protein abundance, while reporter assays provide evidence of direct targeting by measuring the effect of a miRNA on a reporter construct containing the candidate target sequence. After filtering, each target was linked to its Ensembl transcript ID, from which the corresponding mRNA or 3' UTR sequence was retrieved. The miRNA sequences were then paired with their experimentally validated target sequences to form positive MTI examples. This procedure yielded a large-scale sequence-level benchmark for testing whether PAIR-FORMER can generalize beyond small, class-balanced datasets.