Spectral Graph Features for Reference-free RNA 3D Quality Assessment

Motivation

Existing RNA 3D structure quality assessment (QA) methods rely on local geometric descriptors or statistical potentials that evaluate atomic-level contacts but are blind to global topological coherence. This creates a critical failure mode—structures that are “locally correct but globally wrong”—where well-formed local helices mask misplaced domains and incorrect overall packing.

Results

We introduce SpecRNA-QA, a lightweight RNA QA method based on multi-scale graph-Laplacian features of inter-nucleotide contact networks. In CASP16 leave-one-out cross-validation, it achieves median per-target Spearman ρ = 0.69 (target-clustered bootstrap 95% CI [0.64, 0.73]) versus 0.47 for an internal geometry baseline—a +0.22 gap that is significant at p = 1.2 × 10 ⁻¹⁰ (one-sided Wilcoxon signed-rank) and reflects a per-target win rate of 93%. The gain is concentrated on large, multi-domain RNAs, where global coherence is poorly captured by local descriptors. In a contextual comparison with established statistical potentials, local energy-based scores remain strongest on compact RNAs, while SpecRNA-QA yields the strongest signal we observed on targets longer than 200 nt; within the single-threaded runtime budget used here, the strongest local-energy comparator, rsRNASP, timed out on 22 of 26 large targets, and we report an explicit paired head-to-head on the four commonly scored targets in Section 4.2. A training-free heuristic variant further shows that the spectral prior carries intrinsic quality information even in the absence of labeled QA data.

Availability

SpecRNA-QA is available as a Python package at https://github.com/yudabitrends/specrnaq .

Contact

ybi3@gsu.edu

Supplementary information

Supplementary data are available online.

Key Points

• SpecRNA-QA uses multi-scale graph-Laplacian spectra to score global RNA fold coherence that local geometric descriptors and local statistical potentials can miss.

• The method uncovers a size-dependent division of labor: on compact RNAs that can be scored exhaustively, atom-level statistical potentials such as rsRNASP remain strongest, whereas on >200 nt RNAs—where the strongest local comparator times out on most targets under the single-threaded runtime budget used here—SpecRNA-QA provides the strongest signal we observed.

• Heat-kernel traces at intermediate diffusion times emerge as the most discriminative spectral features and form an interpretable bridge between local packing and long-range tertiary organization.

• A training-free heuristic variant of SpecRNA-QA retains informative spectral signal without any labeled QA data, supporting the interpretation of the learned model as amplifying a real structural signal rather than overfitting one.