MuSIC

RNA-binding proteins (RBPs) control splicing, transport, translation, and transcript stability, and the maps of where they bind come almost entirely from CLIP-seq experiments in human cells. Coverage elsewhere is thin: the POSTAR3 compendium holds binding data for 65 RBPs in yeast, 45 in mouse, and only five or six each in fly, worm, and Arabidopsis thaliana. Established predictors such as DeepBind, PrismNet, and HDRNet are trained and evaluated inside the single species whose CLIP data they consume, and accuracy falls sharply when a human-trained model is asked about another organism.

MuSIC (Multi-Species RBP–RNA Interactions using Conservation) attacks that transfer problem with an idea borrowed from style transfer in computer vision. Rather than training against hard binary bound/unbound labels, it softens the label distribution by an amount set by how conserved the RBP itself is between the source and target species. A highly conserved protein such as AGO2 keeps a nearly hard label; a weakly conserved one such as SERBP1 gets a much softer target, discouraging the network from overfitting human-specific signal. The work comes from Lei Sun's group at Shandong University with collaborators at Peking University, Fudan University, and City University of Hong Kong, and was posted to bioRxiv in September 2025 and revised in February 2026.

#Key Features

• Conservation-weighted label smoothing: The smoothing coefficient is the RBP's conservation score, derived by PCA over six similarity measures spanning full-length and RNA-binding-domain sequence identity, TM-score, and pLDDT.
• Gradient weight adaptation: The hard-label and smoothed-label losses are reweighted each step by the ratio of current to previous loss, resolving conflicts between their gradient directions instead of fixing a static balance.
• Sequence and structure input: Each 200-nt window is encoded as one-hot nucleotides plus RNAfold-predicted paired/unpaired state; experimentally measured in vivo RNA structures gave little further gain and are scarce elsewhere.
• Cross-species binding atlas: 2,046 models — 186 RBPs across 11 species from human to yeast — yield predicted peaks and motifs for organisms with no CLIP data of their own.
• Calibrated confidence: A regression between RBP conservation and observed AUC lets predictions be filtered at an estimated accuracy threshold, so distantly related species are flagged rather than silently trusted.

#Technical Details

Inputs are 200-nt RNA windows encoded as a 6 × 200 tensor (four one-hot sequence channels plus two structure channels). The network is a convolutional block (3 × 3 kernels, 16 channels, batch norm, ReLU), a squeeze-and-excitation block that recalibrates channel responses, 2D and 1D residual blocks, average pooling, and a sigmoid-activated fully connected output. Training uses Adam at a learning rate of 0.001, batch size 64, and a positive-class weight of 2. Peaks come from POSTAR3: the top 5,000 per dataset, extended 200 nt on each side, against 5,000 randomly sampled genomic negatives, yielding 262 datasets covering 186 RBPs. On within-species prediction MuSIC reaches 0.88 AUC against 0.87 for HDRNet, 0.86 for PrismNet, 0.72 for GraphProt, and 0.67 for DeepBind. Training on human and validating on mouse and zebrafish across 19 datasets, the gap widens: 0.85 AUC versus 0.73, 0.73, 0.70, and 0.71 respectively. Lengthening the input window from 101 nt to 200 nt lifted cross-species AUC from 0.55 to 0.75; 400 nt was worse than 200. RBP structures were taken from UniProt and filled in with AlphaFold2 where missing. The reference implementation is MIT-licensed and adds pretrained RiNALMo RNA embeddings and ProtT5 RBP embeddings to the preprocessing pipeline; it ships code and example benchmark datasets but no trained checkpoints, so users retrain locally or use the hosted webserver.

#Applications

The immediate use is generating RBP-binding hypotheses and recovering binding motifs in non-model organisms — plants, fly, frog, chicken — where CLIP experiments are impractical. The second is variant interpretation across the human–mouse boundary. The authors scored 32,207 human synonymous SNVs and their 85,599 homologous mouse positions, predicting 59,811 mouse variants that disrupt RBP binding, and confirmed three PUM2-disrupting variants by in vitro RNA pull-down and in vivo POND-qPCR. That makes MuSIC useful for deciding whether a mouse model reproduces a human variant's post-transcriptional effect.

#Impact

MuSIC's contribution is less a new backbone than a demonstration that a measurable biological quantity — protein conservation — can be injected into the loss function as label uncertainty, an idea that generalizes to other cross-species transfer problems. The resulting catalog of predicted peaks for 186 RBPs in 11 species is the broadest such resource to date, and the downstream analysis surfaced a specific signal: 42.5% of the RBPs whose binding is perturbed by variants are enriched for ubiquitination and protein-degradation pathways, with TARDBP prominent among them. Confidence degrades in fly, A. thaliana, and yeast, precisely where predictions would be most valuable; each RBP–species pair requires its own trained model rather than a shared backbone; and the work remains a preprint awaiting peer review.