RNAprecis

At the resolutions where most large RNA structures are solved — roughly 2.5 to 3.5 Å — a density map reliably shows where each phosphate group sits and which way the glycosidic bonds point, but leaves the remaining backbone atoms undetermined. With six backbone torsion angles per residue plus a ribose pucker, a model builder facing ambiguous density has to guess, and those guesses drive geometric outliers in deposited structures. RNAprecis predicts full-detail backbone geometry for an RNA suite — the unit running from one ribose to the next — using only the sparse parameters that low-detail maps determine.

The method comes from the Felix-Bernstein-Institute for Mathematical Statistics in the Biosciences at the University of Göttingen, working with the Richardson lab at Duke University School of Medicine, whose RNA validation criteria are built into MolProbity and Phenix. A preprint appeared in February 2025 and the peer-reviewed version was published in PLOS Computational Biology.

RNAprecis is learned but not neural. Its training is a single unsupervised pass: a modified MINT-AGE algorithm partitions a curated gold-standard set of high-quality suites into conformer clusters, and those cluster parameters are then frozen. New structures — including suites originally modeled as conformational outliers — are classified against that fixed library by Bayesian posterior probability, with no refitting. That places it alongside CaCoFold-R3D, a fixed probabilistic grammar rather than a network, and apart from deep-learning map interpreters such as DeepCryoRNA, which build coordinates directly from cryo-EM density.

#Key Features

• Uses only what low-detail maps resolve: Inputs are the phosphorus atom, the two flanking C1′ atoms, and the two glycosidic bond directions — features visible at 2.5–3.5 Å — rather than backbone atoms that are not.
• Bayesian posteriors in place of distances: Assignment returns a posterior probability across conformer classes instead of a nearest-cluster distance, so each call carries a confidence and sparsely populated classes are not swamped by common ones.
• Small clusters survive: The modified MINT-AGE clustering recovers conformer classes represented by only a handful of suites, which threshold-based clustering folds into larger neighbours.
• New conformer distinctions: Clustering split the established 0a conformer into two groups seen in different structural contexts, a distinction with consequences for biochemical interpretation.

#Technical Details

The gold standard comprises 4,148 curated suites from high-quality, low-redundancy RNA chains, partitioned by the pucker pair of the two flanking riboses — P33, P32, P23, and P22, where 3 and 2 denote C3′-endo and C2′-endo. P33 holds 3,529 suites, 3,183 of them the ubiquitous 1a conformer; P32 holds 265, P23 216, and P22 138. Modified MINT-AGE — mode hunting followed by torus principal component analysis and adaptive clustering — returns 11, 15, 16, and 10 clusters for those four sets, against the 51 reference suite conformer classes with at least three members. Each suite is embedded in MUCCSS coordinates: two pseudo-bond distances from the phosphorus to each C1′, the opening angle between them, and the two glycosidic bond directions as points on unit spheres.

Evaluation used low-detail suites deposited as conformational outliers and later redetermined at higher resolution. RNAprecis recovered the correct conformer class for 80% of P22 cases, 79% of P32, 67% of P33, and 56% of P23. Remaining errors split between a wrong conformer within the correct pucker pair and an incorrect pucker assignment, the latter running 6–14%. The implementation is Python and lives in a public GitLab repository hosted by GWDG; the curated residue set the clusters are learned from is deposited on Zenodo under CC BY 4.0. Licensing there is mixed: the vendored PNDS torus-PCA and mode-hunting modules carry GPL-3.0 headers and a copy of the licence text, while the RNAprecis-specific files carry none and there is no top-level licence, so the method's own code grants no explicit reuse rights.

#Applications

The direct audience is crystallographers and cryo-EM structural biologists building RNA models in the 2.5–3.5 Å range, where the backbone is the least trustworthy part of a deposited coordinate file. Given the parameters a map does support, RNAprecis proposes a full-detail conformation and a probability for it, turning an arbitrary modeling choice into a scored one. The same output serves validation: a suite flagged as a conformational outlier can be checked against the posterior to separate a genuine rare conformation from a modeling error, extending the sugar-pucker validation already in MolProbity and Phenix to the backbone as a whole.

#Impact

Suite conformer nomenclature is the standard descriptive vocabulary for RNA backbone geometry, but it has been usable only where the backbone is directly visible. By deriving assignments from parameters that survive at lower resolution, RNAprecis extends that vocabulary into the range where most ribosome, riboswitch, and ribonucleoprotein structures are determined, and its clustering contributed a previously unrecognised conformer split back to the library. Being statistical rather than neural, it is cheap to run and its assignments are interpretable. The published artifact is a research codebase reproducing the paper's analyses rather than a packaged tool, so integration into refinement software is still ahead of it, and roughly one in ten outlier test cases is assigned the wrong ribose pucker.