RNA inverse folding stops reading only the backbone: treating molecular surface as a design constraint tops sequence recovery on all three benchmarks

RNA sequence design gains a computable design constraint: the target molecule's surface geometry enters generation alongside secondary structure and backbone topology. Earlier RNA inverse-folding methods, from early secondary-structure inverse folding to models incorporating 3D backbone topology, largely leave the molecular surface unrepresented, although surface geometry, charge distribution and hydrogen-bonding sites shape how RNA interacts with proteins, small molecules and other nucleic acids. Across the three benchmarks gRFold reaches sequence recovery of 49.65%, 51.67% and 45.00%, the highest among the methods compared in the table; on RNA-Puzzles its RNAfold and BPfold F1 are 0.626 and 0.580 and its Boltz-2 TM-score is 0.370.

Surface information does not act alone; it and backbone information each add a piece, and removing either branch changes the outcome. Removing the backbone branch drops recovery from 49.650%, 51.671% and 45.004% to 41.013%, 42.306% and 40.090% on the three benchmarks, with RNAfold F1 falling correspondingly from 0.497, 0.486 and 0.626 to 0.337, 0.355 and 0.415. Removing the surface branch affects recovery less but consistently affects 3D fidelity, with RhoFold TM-score falling from 0.297 to 0.255 and from 0.277 to 0.236 on the two RNAsolo benchmarks; the surface contribution is largest on RBP-Complex and smallest on RNA-Puzzles, where the two variants' F1 coincide at 0.626 and 0.580.

The model packs three heterogeneous inputs into one conditional representation, then generates nucleotides position by position. Secondary structure is a discrete dot-bracket string, the surface is an unordered point set, and the backbone is a residue-indexed tensor, so the three differ in length and structure and must be encoded separately before alignment. The surface point cloud is extracted with MSMS and octree-downsampled to 512 points, then encoded into 1024 dimensions by four equivariant graph-convolution layers plus frame-averaged attention; the backbone takes the P, C1' and C3' atoms of each residue into a graph encoded into 512 dimensions by three geometric vector perceptron layers; the fused representation drives a 24-layer autoregressive Transformer decoder.

At inference, folding agreement filters the output by ranking generated candidates against the target base-pairing pattern. The sequence-to-structure mapping is many-to-one, so a greedy decode may miss the target topology and multiple candidates must be compared. Inference draws 50 candidates by nucleus sampling, folds each with RNAfold, and keeps the one with the highest base-pair F1 against the target dot-bracket string.

Evaluation covers three computational benchmarks: RNAsolo RBP-Complex (283 sequences), RNAsolo Monomeric (278 sequences) and RNA-Puzzles (21 sequences), with sequence recovery, secondary-structure F1 from RNAfold and BPfold, and RMSD and TM-score after refolding with RhoFold and Boltz-2. The authors describe these as computational structural metrics rather than experimental validation, and generated sequences should be treated as computational candidates requiring expert review before downstream use.

The authors list several limitations: high-quality RNA 3D structures are limited, so generalization to rare families or unusual motifs may be constrained; the method presupposes an experimentally determined or reliably predicted tertiary structure, which relocates rather than removes part of the design difficulty; RMSD and TM-score inherit refolding-predictor error; and performance degrades beyond 200 nucleotides, although the compared methods share this degradation.