MicroRNAs, or miRNAs, are small RNA molecules that help control how genes are expressed. They do this by binding to messenger RNAs, or mRNAs, preventing them from producing proteins or triggering their breakdown. Because microRNAs regulate many biological processes, understanding exactly where they bind is important for studying diseases and developing RNA based therapies.

Researchers from McGill University and Mila-Quebéc AI Institute have developed a new artificial intelligence framework called MiRformer that improves the prediction of interactions between microRNAs and messenger RNAs directly from their genetic sequences.

Many existing computational tools rely on manually selected features or struggle to analyze long messenger RNA sequences. MiRformer instead uses a transformer based deep learning architecture that processes both microRNA and messenger RNA sequences simultaneously. A specialized sliding window attention mechanism allows the model to efficiently analyze long RNA molecules while maintaining nucleotide level precision.

MiRformer Overview

Schematic of the MiRformer framework shown in three panels. Panel a shows the model architecture: miRNA and mRNA sequences pass through convolutional tokenization layers into separate transformer encoders, then interact via sliding-window cross-attention with LSE pooling, producing embeddings used for binding prediction, seed localization, and cleavage site prediction heads. Panel b illustrates three attention mechanisms: full self-attention on the short miRNA where every base attends to all others, sliding-window self-attention on the long mRNA where each base attends only within a 2w window, and sliding-window cross-attention where miRNA embeddings scan across mRNA windows. Panel c depicts the three training tasks: binary target prediction, seed start and end position regression, and cleavage site localization on the mRNA sequence.

(a) MiRformer leverages convolutional tokenization layers to extract seqeunce features from miRNA and mRNA, and encodes the features via a Dual-Transformer-Encoder Architecture (DTEA). The resulting embeddings are then fed into two dedicated transformer encoders for process miRNA and mRNA. The two embeddings are then fused via cross-attention that emulates the mRNA-mRNA nucleotide-level sequence recognition. (b) Attention mechanisms. Full self attention encodes miRNA sequences by attending every base of the input miRNA sequence to itself and other bases. Sliding-window attention encodes mRNA sequences by attending only bases within a 2w window. Cross attention fuses both miRNA and mRNA sequence embeddings by scanning a 2w window across the mRNA embedding. (c) Training tasks: target prediction, seed region recognition and degradation event prediction.

The researchers evaluated MiRformer using multiple benchmark datasets and experimental Human Degradome sequencing data. The framework achieved state of the art performance in predicting whether microRNAs interact with messenger RNAs, identifying where those interactions occur, and locating cleavage sites where messenger RNAs are cut and degraded.

An important advantage of MiRformer is that it provides interpretable results. Rather than functioning as a black box, the model highlights the regions of RNA that drive its predictions. The attention maps consistently identified biologically meaningful microRNA seed regions, providing confidence that the model is capturing genuine molecular interactions instead of statistical patterns.

When the researchers analyzed more than 13,000 microRNA and messenger RNA pairs, they found that predicted binding sites frequently overlapped with predicted cleavage sites. This finding supports the idea that many microRNAs regulate gene expression by directing messenger RNA degradation.

As RNA sequencing continues to generate increasingly large and complex datasets, artificial intelligence tools such as MiRformer will become increasingly valuable. By improving the identification of microRNA targets and their functional binding sites, these computational approaches can accelerate research into gene regulation while supporting the development of future RNA therapeutics.

Availability and implementation: Python code and datasets are publicly available at https://github.com/li-lab-mcgill/miRformer.

Gu J, Chen C, Li Y. (2026) MiRformer: a dual-transformer-encoder framework for predicting microRNA-mRNA interactions from paired sequences. Bioinformatics 42(Supplement_1):btag222. [article]

MicroRNAs, or miRNAs, are small RNA molecules that help control how genes are expressed. They do this by binding to messenger RNAs, or mRNAs, preventing them from producing proteins or triggering their breakdown. Because microRNAs regulate many biological processes, understanding exactly where they bind is important for studying diseases and developing RNA based therapies.

Researchers from McGill University and Mila-Quebéc AI Institute have developed a new artificial intelligence framework called MiRformer that improves the prediction of interactions between microRNAs and messenger RNAs directly from their genetic sequences.

Many existing computational tools rely on manually selected features or struggle to analyze long messenger RNA sequences. MiRformer instead uses a transformer based deep learning architecture that processes both microRNA and messenger RNA sequences simultaneously. A specialized sliding window attention mechanism allows the model to efficiently analyze long RNA molecules while maintaining nucleotide level precision.

MiRformer Overview

Schematic of the MiRformer framework shown in three panels. Panel a shows the model architecture: miRNA and mRNA sequences pass through convolutional tokenization layers into separate transformer encoders, then interact via sliding-window cross-attention with LSE pooling, producing embeddings used for binding prediction, seed localization, and cleavage site prediction heads. Panel b illustrates three attention mechanisms: full self-attention on the short miRNA where every base attends to all others, sliding-window self-attention on the long mRNA where each base attends only within a 2w window, and sliding-window cross-attention where miRNA embeddings scan across mRNA windows. Panel c depicts the three training tasks: binary target prediction, seed start and end position regression, and cleavage site localization on the mRNA sequence.

(a) MiRformer leverages convolutional tokenization layers to extract seqeunce features from miRNA and mRNA, and encodes the features via a Dual-Transformer-Encoder Architecture (DTEA). The resulting embeddings are then fed into two dedicated transformer encoders for process miRNA and mRNA. The two embeddings are then fused via cross-attention that emulates the mRNA-mRNA nucleotide-level sequence recognition. (b) Attention mechanisms. Full self attention encodes miRNA sequences by attending every base of the input miRNA sequence to itself and other bases. Sliding-window attention encodes mRNA sequences by attending only bases within a 2w window. Cross attention fuses both miRNA and mRNA sequence embeddings by scanning a 2w window across the mRNA embedding. (c) Training tasks: target prediction, seed region recognition and degradation event prediction.

The researchers evaluated MiRformer using multiple benchmark datasets and experimental Human Degradome sequencing data. The framework achieved state of the art performance in predicting whether microRNAs interact with messenger RNAs, identifying where those interactions occur, and locating cleavage sites where messenger RNAs are cut and degraded.

An important advantage of MiRformer is that it provides interpretable results. Rather than functioning as a black box, the model highlights the regions of RNA that drive its predictions. The attention maps consistently identified biologically meaningful microRNA seed regions, providing confidence that the model is capturing genuine molecular interactions instead of statistical patterns.

When the researchers analyzed more than 13,000 microRNA and messenger RNA pairs, they found that predicted binding sites frequently overlapped with predicted cleavage sites. This finding supports the idea that many microRNAs regulate gene expression by directing messenger RNA degradation.

As RNA sequencing continues to generate increasingly large and complex datasets, artificial intelligence tools such as MiRformer will become increasingly valuable. By improving the identification of microRNA targets and their functional binding sites, these computational approaches can accelerate research into gene regulation while supporting the development of future RNA therapeutics.

Availability and implementation: Python code and datasets are publicly available at https://github.com/li-lab-mcgill/miRformer.

Gu J, Chen C, Li Y. (2026) MiRformer: a dual-transformer-encoder framework for predicting microRNA-mRNA interactions from paired sequences. Bioinformatics 42(Supplement_1):btag222. [article]

Submit a Post to the Blog

SUBMIT CONTENT

Subscribe to the RNA-Seq Blog

RNA-Seq Products & Services