ProRB

Ask three questions about the same protein–RNA pair — how tightly do they bind, which residues and nucleotides touch, and what other RNA would bind this protein — and the field hands you three unrelated tools, each with its own inputs, its own training set and its own notion of what a binding event is. The answers do not have to agree, and usually cannot be checked against one another. ProRB, from the Institute of Image Processing and Pattern Recognition at Shanghai Jiao Tong University, collapses the three into one pretrained representation, so that affinity, interface and design are read off the same learned picture of how a protein and an RNA relate.

The mechanism is a protein-conditioned RNA latent space. A transformer decoder takes protein embeddings as memory and RNA embeddings as target, producing an RNA representation that already carries information about the partner protein. Pretraining forces that synthetic representation to agree with the native one by minimizing a KL divergence between them, and simultaneously reconstructs masked tokens from both modalities on the fused embedding. The alignment objective is what makes generation possible at all: once a protein sequence can be pushed into RNA representation space, sampling an RNA from it is the same operation as decoding.

ProRB is deliberately structure-free. It never sees atomic coordinates or a multiple sequence alignment, which is the trade its authors make explicitly — worse coverage of induced-fit and conformational effects, in exchange for inference that runs in under a second on a single consumer GPU where a structure predictor takes hundreds to thousands of seconds.

#Key Features

• One representation, three heads: An outer-product regressor predicts binding free energy from the full residue–nucleotide interaction matrix, a shared lightweight classifier scores every protein residue and every RNA nucleotide on the same probability scale, and a cross-attention decoder generates RNA autoregressively.
• Adaptive cross-modal fusion: Native RNA-FM embeddings and the protein-conditioned decoder output are blended by a learned coefficient α, then refined by eight-head self-attention within each modality and eight-head cross-attention between them.
• Interpretable attention: Cross-attention weights concentrate on RNAcompete-derived binding motifs rather than spreading across the sequence, tested over 37 complexes from 11 RBP families carrying 52 FIMO-confident motif instances.
• Constrained generation: Decoding masks every non-canonical token to −∞, applies temperature scaling and top-4 multinomial sampling, so no degenerate base can ever be emitted.

#Technical Details

Protein sequences are embedded by ESMC at 960 dimensions and RNA by RNA-FM at 640; both encoders stay frozen, and both are projected to a shared 64-dimensional space. Pretraining runs on PRI30K — 30,006 protein–RNA pairs drawn from a maximum connected subgraph over 5,909 BioLiP2 PDB complexes — with 15% BERT-style masking, for 40 epochs on an RTX 4090. On the PRA201 affinity benchmark under 20% PSI-CD-HIT sequence identity, ProRB reaches RMSE 1.3766 ± 0.3649 and Pearson r 0.5347 ± 0.0741, roughly doubling the correlation of the structure-based CoPRA. On Q-BioLip at 40% CD-HIT identity it reaches AUROC 0.684 for protein residues and 0.757 for RNA nucleotides. Against AlphaFold 3 on nine highly divergent FoldBench cases — filtered from 70 after removing six training overlaps — it holds a slight edge on protein-side residues while remaining behind on the RNA side, a comparison the authors describe as illustrative given the sample size.

#Applications

The practical target is RNA aptamer and therapeutic design. Given an RBP of interest, ProRB proposes candidate binders, ranks them by a score combining perplexity with motif content, and narrows a large sampling space to a handful worth folding or synthesizing. Generated sequences show GC content of 53.05% and mean folding free energy of −25.53 kcal/mol, close to native binders and well separated from length-matched random controls. The interface head serves a different audience: groups designing reporter constructs or point mutations to disrupt a specific protein–RNA contact get per-position scores on both partners without needing a solved complex.

#Impact

ProRB argues that for protein–RNA recognition — which is largely dictated by short linear motifs encoded in primary sequence — the geometric machinery that dominates structure prediction is not the only route to a usable interface map. That the same pretrained body supports regression, per-position classification and conditional generation is evidence the shared representation captures binding logic rather than three task-specific shortcuts. The honest limits are stated in the paper: the FoldBench head-to-head rests on nine cases, the RNA-side interface still trails a structure-based predictor, and every design remains computational, ranked by predicted confidence rather than measured by any binding assay.