RNA 3D structure scorer that retrains the ARES equivariant network on a sequence-clustered PDB set to pick near-native conformations for new RNAs.
No providers recorded yet. Browse all providers
RNA structure predictors return many candidate conformations for one sequence, and a scoring function has to decide which to trust. ARES (Townshend et al., Science 2021) was an early deep-learning answer: an E(3)-equivariant network that reads atom types and coordinates and regresses each candidate's RMSD to the native structure. It was trained on only 18 RNAs, selected by hand and split from its test puzzles by publication year. When the ARES+ authors clustered those RNAs by sequence, five of ARES's training and three of its validation RNAs fell into the same cluster as fifteen RNA-Puzzles targets, so the original split let related sequences sit on both sides of the train–test boundary.
ARES+ keeps the ARES architecture and its RMSD-regression objective unchanged and retrains it on a larger, leakage-controlled dataset. Its purpose is to isolate how much of a learned scorer's accuracy comes from training data alone: with the network held constant, the new dataset raises top-1 success on RNA-Puzzles by 15 percentage points.
ARES+ was introduced in the same bioRxiv preprint as HomRank, posted in August 2025 by A*STAR — the Centre for Frontier AI Research at the Institute of High Performance Computing, the Bioinformatics Institute and the Genome Institute of Singapore — with the National University of Singapore and the University of Zagreb. It serves there as the retrained regression baseline against which HomRank's listwise ranking objective is measured.
The dataset starts from 190 non-redundant single-chain RNAs: 151 new PDB structures, the 18 ARES training and validation RNAs, and the 21 RNA-Puzzles targets with their Rosetta FARFAR2 decoys. Clustering on Biopython pairwise alignments yields CLS-0 (146 RNAs), CLS-1 (21) and CLS-2 (23). Training uses the 131 CLS-0 RNAs left after removing its RNA-Puzzles targets, about 131,000 structural candidates of 22–144 nucleotides; CLS-1 and CLS-2 form the TestCLS-1 and TestCLS-2 benchmarks. ARES+ is trained with ARES's mean-squared-error RMSD loss, so its score is a predicted RMSD for each candidate independently.
A near-native candidate is one within 2 Å RMSD of the experimental structure. On RNA-Puzzles, ARES+ ranks a near-native candidate first for 86% of targets, against 71% for ARES, and reaches 100% success within its top 20. On TestCLS-1, a cluster held out from training, its top-1 success is 71% against 33% for ARES. Retraining did not help on pooled predictions from six structure predictors, including trRosettaRNA and AlphaFold 3, for 84 RNAs: there ARES+ and ARES select nearly the same structures, and ARES+'s Spearman correlation with RMSD is negative (−0.258).
ARES+ scores candidate pools from sampling methods such as Rosetta FARFAR2 or molecular dynamics and returns the conformations most likely to be near-native, so a modeller can carry a short list forward. Its main use is as a reference point: it measures what a better training set buys an RMSD-regression scorer, separate from any change in objective. The same clustered dataset was later reused to benchmark SIRGE, a sequence-conditioned evaluator from the same group.
ARES+ supplies the ablation that makes the HomRank paper's argument legible: of the rise from ARES's 71% top-1 success on RNA-Puzzles to HomRank's 95%, the dataset accounts for 15 points and the ranking objective for the remaining 9. HomRank outperforms it on both benchmarks and is the paper's headline model. Neither the ARES+ checkpoint nor its training candidate sets have been released; the public HomRank repository holds only HomRank's ranking code and weights, and the curated data are available from the authors on request. The preprint has not yet been peer reviewed.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.