RNA 3D structure scoring function: an equivariant network that reads only atom coordinates and predicts each candidate model's RMSD from the native.
No providers recorded yet. Browse all providers
Sampling programs such as Rosetta FARFAR2 can generate thousands of candidate 3D structures for an RNA, and the hard part is deciding which of them is right. Hand-built scoring functions encode what their authors believe good RNA structure looks like — base-pair geometry, stacking, statistical contact potentials — and those beliefs are incomplete. The Atomic Rotationally Equivariant Scorer (ARES) takes the opposite approach: it is given only each atom's chemical element and 3D position, is told nothing about nucleotides, helices or hydrogen bonds, and learns to predict how far a candidate model sits, in RMSD, from the unknown true structure.
The obstacle to learning this directly is data: there are very few experimentally solved RNA structures. ARES addresses it with an architecture that builds the symmetries of physics into the network. Its layers are rotation- and translation-equivariant, so a motif is recognized the same way wherever and however it is oriented, and its position and orientation are passed up to later layers that assemble fine motifs such as base pairs into coarser ones such as helices. With that built in, ARES learns from 18 RNA structures.
ARES was developed at Stanford University by Raphael Townshend, Stephan Eismann, Andrew Watkins, Ramya Rangan, Masha Karelina, Rhiju Das and Ron Dror, spanning the Dror lab in Computer Science and the Das lab in Biochemistry, and was published in Science in August 2021. ARES is a scoring function, not a structure generator: it ranks candidates that a separate sampler produces.
The training set is 18 RNAs whose structures were published between 1994 and 2006. For each, 1,000 candidate models were generated with FARFAR2 without reference to the known structure, and the network was fit so its output matches each candidate's RMSD to the native. The penultimate layer has 256 nodes. Training code, a trained checkpoint, an ARES-specific adaptation of the e3nn library and the analysis notebooks are deposited on Zenodo, and all candidate models and predictions are in the Stanford Digital Repository.
Benchmark 1 covers RNA-Puzzles targets with structures published 2010–2017, each with at least 1,500 FARFAR2 candidates, 1% of them generated with restraints to the native so near-native models exist. The top-scoring candidate is within 2 Å RMSD for 62% of RNAs under ARES, against 43% for Rosetta (2020), 33% for RASP and 5% for 3dRNAscore; the top 10 contain a near-native model for 81% of RNAs, against 48%, 48% and 33%. Benchmark 2 holds 16 structurally complex RNAs with no near-native candidates, generated with experimentally determined Watson–Crick pairs but no templates; ARES selects models with lower median RMSD than each of six other scoring functions. In four blind RNA-Puzzles rounds — the adenovirus VA-I RNA and three T-box–tRNA complexes — FARFAR2 sampling with ARES selection produced the most accurate submission of any participant.
ARES is used as the selection step after conformational sampling: a modeller generates candidates with FARFAR2 or another sampler, scores them, and carries the lowest predicted RMSD models forward. A web server accepts PDB or Rosetta silent files of candidate models and returns one score per model. ARES has also become a standard baseline for RNA structure quality assessment, rescoring pooled predictions from deep-learning structure predictors.
ARES showed that an equivariant network can learn a useful atomic scoring function for RNA from a handful of structures, and it is a reference point for later RNA quality-assessment models. ARES+ keeps its architecture and retrains it on a larger, sequence-clustered PDB set, after finding that several ARES training RNAs share a sequence cluster with RNA-Puzzles targets; HomRank replaces its RMSD regression with a ranking objective, and SIRGE benchmarks against it. ARES depends on an external sampler: if no near-accurate candidate is generated, it cannot produce one. A 2023 erratum to the paper clarified that benchmark 2 candidates used experimentally determined Watson–Crick base pairs. The training code and the trained checkpoint ship in one Zenodo archive under a non-commercial license, so the weights are non-commercial too, and Stanford has filed a patent application on the work.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.