RNArank

RNA three-dimensional structure prediction has improved rapidly, but the models it produces are uneven in quality: for the same sequence, different pipelines can yield structures that differ by many angstroms. Deciding which candidate to trust is a separate problem from generating candidates, and it determines whether a predicted structure is usable for mechanism, design, or experimental interpretation. Estimation of model accuracy is well established for proteins; for RNA it has lagged, and the confidence scores emitted by end-to-end predictors describe only their own outputs.

RNArank, developed by the Yang Lab at Shandong University, is a deep learning method for both local (per-nucleotide) and global quality assessment of predicted RNA 3D structures. It was posted as a preprint in July 2025 and published in Communications Biology in January 2026. Rather than regressing an accuracy score directly from the input coordinates, RNArank first predicts two intermediate two-dimensional maps — an inter-nucleotide contact map and a distance deviation map — and then derives local and global lDDT estimates from them, a spatially resolved supervision signal that also makes the per-residue confidence interpretable.

RNArank is method-agnostic: it takes any RNA structure in PDB or mmCIF format and scores it, regardless of how it was produced. That distinguishes it from the built-in self-assessment of predictors such as AlphaFold 3, which cannot rank models from other sources, and it supplies the selection step downstream of generation pipelines like the Yang Lab's own trRosettaRNA.

#Key Features

• Intermediate 2D map prediction: A contact map and a distance deviation map are predicted first and then converted into per-nucleotide and whole-structure accuracy estimates, rather than regressing a score end-to-end.
• Multi-modal feature encoding: Inputs combine 1D sequence and backbone torsion features, 2D inter-nucleotide distances and Rosetta energy terms, and a voxelized 3D description of each nucleotide's local atomic environment.
• RNA-adapted lDDT: Predictions target lDDT-RNA, a variant of the protein lDDT metric with distance thresholds widened to 1, 2, 4, and 6 Å and an inclusion radius of 30 Å to match RNA's larger geometric scale.
• Method-agnostic scoring: Any modeled RNA structure can be assessed without access to the generating method, making cross-method ranking possible.
• Practical throughput: A typical 100–300 nucleotide structure is scored in under six seconds, with peak usage below 3 GB of CPU and 5 GB of GPU memory.

#Technical Details

The network is a Y-shaped residual architecture: 50 dilated ResNet blocks build a shared representation, which then splits into two branches of 50 blocks each, one predicting contacts and one predicting distance deviations. Feature channels number 67 (1D), 58 (2D), and 3 (voxelized 3D), and the released system averages two independently trained networks. Training used the trRosettaRNA dataset restricted to RNAs resolved before January 2022 — 1,635 chains after filtering — with roughly 200,000 decoys generated by five strategies (deep learning prediction, SimRNA folding, backbone perturbation, molecular dynamics, and Rosetta fragment assembly), then clustered to keep about 120 diverse, quality-stratified models per RNA.

On the independent T24 benchmark (24 post-cutoff targets, ~2,837 decoys), RNArank reached Spearman ρ of 0.949 against global lDDT-RNA versus 0.423 for the next-best method, 0.936 per-target versus 0.675, an AUC of 0.951 for identifying models above lDDT-RNA 75, and a top-1 selected accuracy of 83.6 versus 76.9. On the 12 CASP15 RNA targets it achieved ρ = 0.706 versus 0.511, and on 42 CASP16 targets (~6,500 submitted models, 58–833 nucleotides) it found at least one true top-5 model within its top 10 for 42.9% of targets versus 35.7% for lociPARSE. Baselines included cgRNASP, rsRNASP, DFIRE-RNA, RNA3DCNN, and lociPARSE; ablations showed the voxelized 3D features contribute most.

#Applications

The immediate use is model selection: given a pool of candidate structures from one or several predictors, RNArank ranks them and reports which regions of the chosen model are reliable. Per-nucleotide scores flag loops and junctions that should not be over-interpreted when a structure is used for docking, restraint generation, or hypothesis building, and the same signal can filter candidates in RNA design and refinement loops. Because it needs only coordinates, it also serves as an assessment tool for community experiments and as a triage step for groups that cannot adjudicate between predictors themselves; a web server makes it accessible to experimental RNA biologists.

#Impact

RNArank addresses a gap in the RNA structure toolchain, where generation has outpaced evaluation, and its blind CASP15 and CASP16 results show the gains hold on targets released after the training cutoff rather than only on curated test splits. Its limitations are stated plainly by the authors: performance degrades for RNAs that form complicated multimers or bind proteins, dynamic RNAs with multiple conformations remain difficult, and on CASP16 its per-target RMSD correlation of 0.722 trailed DFIRE-RNA's 0.751. The code is Apache-2.0 licensed with pretrained weights distributed by the authors, so the method can be embedded directly in prediction and design pipelines.