SIRGE

An RNA folding pipeline returns a thousand conformations for one sequence and no indication of which is right. The usual answer is a learned scoring function that reads geometry alone and asks whether the arrangement looks like plausible RNA. But two candidates can be equally plausible as generic RNA geometry while only one is compatible with the particular sequence expected to adopt it: helix conformational preferences shift with which Watson–Crick pair sits where. SIRGE — Sequence-Informed RNA Geometric Evaluator — makes that compatibility the quantity scored.

A candidate's heavy atoms are encoded by an SE(3)-equivariant transformer and pooled into one invariant representation per nucleotide; each nucleotide's structural representation is then summed with the contextual embedding a frozen RNA language model produces for the same position in the sequence. RiNALMo supplies those embeddings and is never fine-tuned, so SIRGE inherits sequence statistics learned from 36 million non-coding RNA sequences instead of recovering them from ranked structures. Because every candidate for a target shares one sequence, a sequence-only term would shift all scores equally and change no ordering; sequence can act only through its interaction with each candidate's geometry.

SIRGE comes from a collaboration centred on A*STAR — its Centre for Frontier AI Research, Bioinformatics Institute and Genome Institute of Singapore — with Aalborg University, Imperial College London, Statens Serum Institut and the University of Copenhagen. It belongs to the same class of RNA structure quality assessment models as RNArank, and takes its benchmark and its homogeneous-batch training scheme from HomRank, an earlier evaluator from the same group.

#Key Features

• Reference-free scoring: The score needs only the candidate's coordinates and the RNA sequence, so unlike RMSD it applies to freshly predicted structures for which no experimental reference exists.
• Frozen sequence encoder: RiNALMo's 1,280-dimensional per-nucleotide embeddings enter through a layer-normalised learned projection and are never updated, preserving the pretrained representation and keeping the trainable model small.
• Conditioning that starts from nothing: That projection is initialised to zero, so at the start of training SIRGE is exactly its structure-only variant, SIRGE-SO, and any difference between the two is attributable to sequence conditioning.
• Listwise ranking supervision: A ListNet objective over candidate lists drawn from a single RNA supervises within-RNA RMSD order rather than absolute RMSD, so the signal is invariant to sequence length.
• Largest gains on poor candidates: Conditioning helps more as candidate quality falls, the regime in which geometry alone carries least signal.

#Technical Details

An EquiformerV2 atom encoder represents each candidate as a graph of heavy atoms, each connected to its 32 nearest neighbours, over three blocks with spherical harmonics to degree two. Atom features are rotated into nucleotide-local frames anchored at C4′ and pooled by attention into a 64-dimensional nucleotide representation, to which the projected 1,280→64 RiNALMo embedding is added. Four dense invariant point attention blocks then propagate information across nucleotides before a 64→256→1 MLP, with a parallel atom path adding mean-pooled scalar atom features back as a residual.

Training uses the HomRank benchmark: 190 non-redundant single-chain RNA targets — 151 from the PDB, 18 from the ARES set and 21 RNA-Puzzles — clustered by sequence similarity, giving 131 training targets with 131,000 candidates, 23 validation targets, and 21 held-out test targets with 152,917 candidates. The split is at the target level, so no test sequence, native structure or candidate appears in training. Over five epochs of AdamW training, SIRGE reaches a global Kendall-τ of 0.255 ± 0.008 on the held-out cluster, against 0.209 for SIRGE-SO, 0.146 for HomRank, 0.135 for lociPARSE and 0.130 for ARES. Top-1 retrieval is 31.7% against HomRank's 23.2% and a 10.0% random baseline, Top-3 ranking 31.3%, and Top-1 rises to 50.1% on candidate lists in the 2–5 Å band. All figures are means over three training seeds.

#Applications

The task is model selection at the end of an RNA structure pipeline: molecular dynamics sampling, Rosetta FARFAR2 runs and neural structure predictors all emit large candidate pools, and the choice of which conformation to carry into docking or experimental follow-up is made without the answer in hand. Results are reported in 2–5, 5–10 and 10–15 Å bands, the quality range in which that choice is actually made among computationally generated structures. Applied as a fixed checkpoint, SIRGE runs on new RNAs without refitting.

#Impact

SIRGE's contribution is a claim about where the remaining headroom in RNA structure evaluation lies. Evaluators have grown more expressive geometrically — voxel CNNs, equivariant graph networks, invariant point attention — while leaving untouched the sequence the geometry is supposed to realise. The matched comparison against SIRGE-SO isolates that: the geometric backbone supplies most of the improvement over prior evaluators, and pretrained nucleotide embeddings add a further, consistent increment in every quality band and on every metric reported. The authors present this as early work on a single held-out cluster of 21 targets, and no code or weights accompany the preprint, so independent replication is not currently possible.