STRAND

RNA-protein interactions drive transcriptional regulation, splicing, and protein synthesis, but they remain one of the harder targets for structure prediction. General-purpose complex predictors model proteins well and RNA less well, and the resulting complexes often place the two chains incorrectly relative to one another even when each chain is individually reasonable. STRAND (STructure Refinement of RNA-proteiN complexes via Diffusion) attacks that specific failure mode: rather than predicting a complex from scratch, it takes an existing predicted complex and denoises it into a better-docked pose.

STRAND was developed at the University of Freiburg by Mohsen Al-zeqri, Jörg K.H. Franke, and Frederic Runge, and posted as a preprint in July 2025. It adapts the diffusion-docking framework of DiffDock-PP, built for protein-protein docking, to RNA-protein complexes by giving the RNA its own node featurization and sequence embeddings. The model is trained once on experimentally determined complexes from the PDB and then applied at inference to arbitrary new predictions, which makes it independent of whichever tool generated the starting structure.

The authors demonstrate refinement on predictions from AlphaFold 3 and ProRNA3D-single. Because refinement is strictly additive — the original prediction can always be retained if the refined pose is worse — STRAND is positioned as a module to bolt onto an existing pipeline rather than a replacement predictor.

#Key Features

• Predictor-agnostic refinement: The reverse diffusion process is initialized with any existing RNA-protein complex structure, so the same trained model refines outputs from different upstream predictors without retraining.
• RNA-aware featurization: Proteins are represented by residue type, Cα coordinates, and ESM-2 embeddings, while RNA uses nucleotide type, phosphate coordinates, and RNA-FM embeddings, with separate embedding layers for each molecule type.
• Decomposed transformation study: Five separate models were trained over translation, rotation, torsion, and their combinations, isolating which degrees of freedom actually matter for complex accuracy.
• Protein as the diffused ligand: The RNA is fixed as the receptor and noise is applied to the protein regardless of relative chain length, which costs training time but improves stability when test-set lengths differ from training.
• Confidence-based pose ranking: A companion confidence model regresses ligand RMSD to rank the 40 sampled poses plus the original prediction, giving an automated selection path.

#Technical Details

STRAND uses a tensor-product SE(3)-equivariant score network with four convolution layers, 16 scalar and 4 vector node features, spherical harmonics up to lmax = 2, and 30 nearest neighbors, diffusing over the product space of 3D translations, SO(3) rotations, and a copy of SO(2) per rotatable bond. Training used every PDB entry containing both RNA and protein deposited before 30 September 2021 — AlphaFold 3's own training cutoff — preprocessed into interacting monomeric chains, with multimeric complexes added as augmentation. The five variants trained for roughly 6 to 72 hours each on a single A40 GPU with Adam; sequence embeddings cap inputs at 1022 residues.

On 35 X-ray-derived test complexes, AlphaFold 3 reaches a mean complex RMSD of 3.09 Å, which the translation-plus-rotation variant lowers to 2.97 Å while raising the fraction of predictions under 2 Å from 34.3% to 42.9%. On 46 non-X-ray complexes the mean improves from 10.86 Å to 10.41 Å. Gains over ProRNA3D-single are larger: across 63 X-ray complexes the mean falls from 12.02 Å to 11.19 Å, the sub-2 Å fraction rises from 4.8% to 14.3%, and 84% of individual targets improve. These headline figures use the best of 40 samples chosen against ground truth; with automated confidence-model selection the improvements are smaller and appear mainly on the harder non-X-ray set. Torsion-only refinement consistently underperforms, since torsion changes local geometry without correcting global misplacement.

#Applications

STRAND is intended as a post-processing stage in structural biology pipelines that model RNA-protein assemblies — ribonucleoprotein complexes, RNA-binding protein interfaces, and viral replication machinery — where a predictor already returns a plausible complex but the interface geometry is uncertain. Structural biologists preparing models for experimental validation, and computational groups benchmarking RNA-protein predictors, can run refinement over an existing set of predictions and keep whichever pose scores better, at no risk of degrading the starting model.

#Impact

STRAND's main contribution is diagnostic as much as methodological: the finding that simple rigid-body translation and rotation of an existing prediction recovers most of the available accuracy implies that current predictors systematically misplace the protein relative to the RNA rather than misfolding either chain. That observation points toward interface-aware training objectives for the next generation of RNA-protein predictors. The work is a preprint that has not been peer reviewed, refines backbone atoms only rather than full-atom structures, and reports results on benchmark sets of a few dozen complexes. The repository ships training and inference code along with trained score and confidence checkpoints, but carries no license file, which leaves reuse terms unsettled.