CARD

RNA inverse folding — finding a sequence that will fold into a specified tertiary structure — has followed protein design in treating the target structure as the whole problem. Tools such as gRNAde and its contemporaries encode an RNA backbone with a geometric network and decode a sequence from it. For a large share of functional RNAs that framing throws away the most important constraint. Aptamers, riboswitch components, and regulatory elements do their work bound to a protein, and the residues at that interface determine which sequences are viable. A design produced from RNA geometry alone is optimized for a fold, not for the complex the fold exists to form.

CARD — Complex-Aware tertiary structure-based RNA Design — from the Chinese University of Hong Kong, Shenzhen and Zelixir Biotech, adds the partner protein to the conditioning. Protein representations come from ESM-2 and are fused with the RNA structural encoding through a transformer the authors call CAFormer. Because a protein chain is mostly irrelevant to any given RNA contact, the fusion is preceded by a distance-aware filter that retains only the local protein features near the interaction site, keeping the conditioning signal from being diluted by the rest of the chain.

The design model is one half of the system. The other is an evaluation stage: candidate sequences are scored by a trained binding-affinity model and by structural agreement with the target, and only the top-ranked fraction is returned.

#Key Features

• Complex-level conditioning: The partner protein enters the design as learned language model features rather than as hand-built interface descriptors, so no contact map or docking step is required at inference.
• Distance-aware feature filtering: Only protein residues within a local neighbourhood of the RNA are passed to the fusion module, which is what makes a 650M-parameter protein encoder usable as a conditioning signal.
• CAFormer fusion: A six-block attention module merges RNA structural embeddings with the filtered protein representation before sequence decoding.
• High-affinity design pipeline: An ensemble affinity regressor filters generated candidates to the top 10–20%, with structural alignment as a second constraint, producing a shortlist rather than a raw sample set.

#Technical Details

Training and evaluation use PRI30K and PRA201, two curated protein-RNA collections. PRI30K is filtered to remove RNA chains sharing high similarity with the test set and to drop chains longer than 128 nucleotides, leaving 21,050 protein-RNA pairs over 2,309 unique RNA sequences; CD-HIT clusters RNA chains at 80% sequence identity so that evaluation splits are not contaminated by near duplicates. PRA201, 201 unique single-chain protein-RNA complexes assembled from PDBbind, PRBABv2, and ProNAB, serves as a blind test set. Protein features come from ESM-2 650M, with the local filtered size set to 64 residues and CAFormer using six attention blocks. Training runs 200 epochs on two A100 GPUs at batch size 48, taking roughly six hours. RhoDesign, an RNA-structure-only designer, serves as the base model. On the PRI30K test set CARD reaches 59.56% overall sequence recovery and 0.6002 macro F1, 5.6 points of recovery above RhoDesign's 54.00% and well clear of RDesign at 45.64%; on the PRA201 blind set it reaches 63.98% recovery and 0.6108 macro F1 against RhoDesign's 56.39%. Stripping the CAFormer entirely, leaving only the RNA structure encoder and decoder, drops macro F1 to 0.5677, and performance climbs monotonically as complex-aware attention blocks are added back. A case study on the 2LBS complex illustrates the full design-then-screen pipeline.

#Applications

The intended users are groups designing RNA that must bind a specific protein: aptamer engineering, synthetic riboswitches, and RNA components of engineered regulatory circuits. The affinity-screening stage matters here more than the raw recovery numbers, since a wet-lab campaign needs a ranked shortlist, not a distribution over sequences. The same conditioning is also useful for redesigning known RNA-protein interfaces — resampling a natural binder's sequence while holding the fold and the partner fixed.

#Impact

CARD's contribution is a clean demonstration that RNA inverse folding benefits from the same complex-level context that protein binder design already exploits, and that an off-the-shelf protein language model supplies it without bespoke interface features. The evidence is in-silico: recovery rates and affinity model scores, with no experimental validation of designed sequences reported. The preprint has not been peer reviewed, and no code or weights have been released, which leaves the results difficult to reproduce independently.