RNA language model that locates the four cut sites defining a mature miRNA duplex on its precursor hairpin and ranks artificial miRNA scaffolds.
No providers recorded yet. Browse all providers
A miRNA precursor is a hairpin of a hundred-odd nucleotides, and the mature product that actually silences transcripts is a ~22-nt stretch cut out of one arm of it. Which 22 nucleotides matters enormously: in animals, target recognition runs through the 5' seed, so a one-nucleotide shift in where Drosha or Dicer cuts changes which genes the miRNA represses. Establishing those cut sites experimentally means small RNA sequencing, RACE, or CLIP assays, and the computational alternatives — PHDcleav, LBSizeCleav, DeepMirCut, DiCleave — were built and benchmarked largely on human data, leaving cross-species annotation thin.
miRstring, from Xiang Yu's group at Shanghai Jiao Tong University, treats the problem as locating four boundaries at once: the start and end of the 5p mature segment and the start and end of the 3p segment. Sequence is encoded by the pretrained ERNIE-RNA backbone; the RNAfold-predicted secondary structure is encoded by a trainable embedding and injected through a residual cross-attention block in which sequence states query the structural representation. A shared transformer encoder then learns the boundary cues common across taxa, after which representations are routed by super-clade label to dedicated tail encoders and prediction heads. The four boundaries are read out as masked position classification over the valid precursor coordinates.
The clade routing is what makes the model workable across kingdoms rather than one species at a time: animal and plant precursors are processed by different machinery (Drosha–DGCR8 then Dicer versus nuclear DCL1 with HYL1 and SERRATE), and the model gives each its own refinement stage without splitting into separate networks. The published work refines and validates the Metazoa and Viridiplantae branches; the Viruses and Others branches carry deliberately simpler linear heads and remain sparsely supervised.
Training used 77,708 precursors from 414 species, integrated from miRBase, MirGeneDB and PmiREN, capped at 250 nt and required to carry an annotated mature product; negatives for the precursor-validity head came from RNAcentral biotypes plus curated pseudo-hairpins. Training ran in four stages under AdamW with cosine warmup: the language backbone frozen while the rest trained on miRBase, then unfrozen for joint refinement with balanced animal–plant mini-batches, then the shared trunk frozen while the Metazoa and Viridiplantae branches were refined separately. Model selection used a family-held-out split, the stricter of the two protocols. On that benchmark miRstring reached mean family-level IoU of 0.789, center MAE of 2.664 nt and length MAE of 1.590 nt, cutting center error by 53.0% against DeepMirCut and 77.3% against DiCleave. On strictly held-out species IoU was 0.965 (metazoans) and 0.880 (plants), with center MAE of 0.396 and 1.464 nt. Ablations put ERNIE-RNA ahead of a non-pretrained encoder and of RiNALMo and mRNABERT as backbones.
Genome annotators working in newly sequenced or sparsely curated organisms can infer mature products from precursor candidates directly, with calibrated confidence usable for selective prediction. Plant biologists gain the more immediate payoff: in Nicotiana benthamiana, model-selected MIR166 and MIR157 scaffolds knocked down PDS transcript by roughly 60% against about 40% for the conventional MIR319 scaffold, and MIR535 and MIR157 reduced CHLH by about 70% against 36.2%.
Most precursor-to-mature predictors have been human-centric and evaluated in distribution; miRstring's contribution is the demonstration that duplex boundaries transfer to unseen families and species when sequence, structure and lineage context are modeled together, and that boundary confidence is informative enough to drive design. Attention concentrating on paired stem positions near the cut sites, rather than on loops or flanks, gives the predictions a mechanistic reading that can nominate sequence–structure determinants of cleavage choice. The work is a preprint awaiting peer review, the structural input is a single minimum-free-energy fold rather than an ensemble, and the wet-lab validation covers plant scaffolds only.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.