MiRformer: A Unified Generative Framework for mRNA-Conditioned miRNA Synthesis and Interaction Prediction

MicroRNAs (miRNAs) are small non-coding RNAs that regulate gene expression by binding to target messenger RNAs (mRNAs), leading to mRNA degradation or translational repression. Accurate prediction of miRNA-mRNA interactions is critical for understanding post-transcriptional regulation and enabling RNA therapeutics. However, existing computational methods rely on handcrafted or indirect features, struggle to scale to kilobase-long mRNA sequences, or provide limited interpretability. We present MiRformer, a unified generative framework that models miRNA-mRNA interactions, pinpoints miRNA binding sites within long mRNA sequences, and synthesizes mRNA-conditioned miRNA sequences. MiRformer employs a dual-transformer encoder architecture to learn interaction patterns directly from raw miRNA-mRNA sequence pairs. To efficiently model long mRNA contexts, we incorporate a sliding-window attention mechanism that scales to kilobase-length sequences while preserving nucleotide-level resolution. MiRformer employees the pre-trained MiRformer encoder and trains a decoder to auregressively generate miRNA. MiRformer achieves state-of-the-art performance across multiple miRNA-mRNA prediction tasks, including interaction prediction, binding-site localization, and cleavage-site identification from experimental Human Degradome-seq data. Across all 500-nt mRNAs, MiR-former produces 5712 target-specific miRNA and 99.30% of these contain canonical seed regions. Beyond accuracy, MiRformer provides strong interpretability: attention patterns consistently highlight miRNA seed regions within 500-nt mRNA windows, revealing clear and biologically meaningful interaction signals. Our results demonstrate that MiRformer offers a scalable, accurate, and interpretable approach for modeling miRNA-mRNA interactions and synthesizing biologically plausible target-specific miRNA sequences. Code is available at https://github.com/li-lab-mcgill/miRformer.