Loading…
miRNA-target interaction model fusing five gated evidence experts, whose frozen representation transfers to siRNA efficacy prediction.
No providers recorded yet. Browse all providers
Small interfering RNAs and microRNAs are loaded into the same Argonaute protein and locate their targets the same way — by pairing a short guide against a complementary site in an mRNA. The data available for each is wildly asymmetric. Chimeric eCLIP and AGO-CLASH experiments have yielded millions of sequence-resolved miRNA–target interactions, while quantitative siRNA efficacy still comes from controlled perturbation assays: the OligoFormer study pooled nine published datasets and reached only 3,714 siRNAs against 75 mRNAs. DuplexFM treats that asymmetry as a transfer problem. It learns a representation of small-RNA target recognition from the abundant miRNA resource, then freezes that representation and reuses it for siRNA efficacy, where labels are scarce.
The model routes five complementary evidence sources through independently learned, sample-specific sigmoid gates. Two are mechanistic "core" experts — Bio23, a 23-feature description of canonical seed-site architecture, local AU context and Watson–Crick/G·U pairing, and IntaRNA21, 21 duplex-energetics and geometry features from IntaRNA v3.4.1. Three are learned plugins: an experimentally supervised mRNA-accessibility predictor, asymmetric target-to-guide cross-attention, and a score-matrix attention module over guide–target token compatibility. Each expert is projected into its own 32-dimensional slot, and the five gated slots concatenate into a 160-dimensional representation feeding a binary interaction classifier. Sequence encoding uses two frozen public backbones adapted with rank-16 LoRA: RNA-FM for the guide and mRNABERT for the target.
DuplexFM was developed by Baiming Chen and colleagues at the Chinese University of Hong Kong, Shenzhen with collaborators at Harbin Institute of Technology, Jinan University and XtalPi, and released in August 2026 with code, three source checkpoints and the transfer heads.
The source model was trained on the Manakov 2022 partition of the miRBench v7 benchmark — 2,094,107 training interactions with a 401,428-interaction miRNA-family-grouped validation split — using fused binary cross-entropy plus per-expert auxiliary losses. The accessibility predictor is a LoRA-adapted mRNABERT with FiLM conditioning and two residual Transformer layers, trained on 1,509,015 sequence views from human icSHAPE profiles (GSE145805, five cell lines) and reaching a held-out nucleotide-level Pearson correlation of 0.627 across 8.8 million positions. On miRBench v7, three independently seeded models averaged a macro average precision score of 0.873 ± 0.002, and their soft-voting ensemble reached 0.876, ahead of the 0.868 of miRBenchCNN-Manakov and highest on all four test partitions. Transferred to six siRNA settings, the frozen representation lifted macro Pearson correlation from 0.482 to 0.535 over a matched descriptor-only control, and to 0.577 as a three-head ensemble.
The practical use is triaging candidate siRNAs before synthesis and cell-based testing, and scanning transcripts for functional miRNA sites. Because the transfer head is tiny, a group holding a few hundred in-house efficacy measurements can fit one without touching the source model — the regime where the reported advantage is largest, at 5–10% training fractions. Inference requires IntaRNA on the path and a CUDA device, so it fits computational pipelines rather than a hosted web tool.
DuplexFM's contribution is evidence that miRNA interaction data carries transferable supervision for siRNA design, a link the two literatures have largely treated as separate. The result is honestly bounded: OligoFormer remained stronger on the Huesken and Takayuki datasets and matched or exceeded DuplexFM from 20% training data onward, and the parameter comparison covers adaptation cost only, since the frozen source encoders still run during feature generation. Accessibility is predicted within 50–60 nt windows, so long-range structure, RNA-binding proteins and modifications are outside its view, and icSHAPE is an in vivo proxy rather than a direct measure of unpaired probability. The preprint has not yet been peer reviewed, and the code and checkpoints carry a CC BY-NC 4.0 license that excludes commercial use.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.