Standardized Annotation of CASP16 RNA Predictions Using DSSR

Overview & Background

The 16th Critical Assessment of Structure Prediction (CASP16) evaluated blind predictions for both monomeric RNAs and RNA–RNA multimeric complexes. In a study published in Bioinformatics, Chandran Nithin, Smita P. Pilla, and Sebastian Kmiecik (Laboratory of Computational Biology, University of Warsaw) examined their group's (LCBio, group G189) CASP16 submissions—which ranked first in the official CASP16 best-of-submitted-models assessment for RNA–RNA multimers—to diagnose where modeling accuracy degrades across the RNA structural hierarchy. Rather than relying solely on global superposition metrics (TM-score, RMSD), the authors established an automated evaluation pipeline centered on DSSR (Dissecting the Spatial Structure of RNA; Lu et al., 2015) to benchmark local, modular, and tertiary structural features.

Methodological & Parameter Details

To ensure objective comparisons between experimental reference structures and predicted 3D models, the authors relied on DSSR for consistent feature extraction:

"Structural features were extracted from both reference and predicted RNA structures using DSSR (Lu et al. 2015) to ensure consistent annotation across all models."

As detailed in Supplementary Methods S3.2:

"All structural features used in the analysis were derived from reference and predicted RNA structures using DSSR, which provided a standardized annotation of base pairs, stacking interactions, secondary structure elements, junctions, and tertiary motifs."

Using DSSR outputs, the authors computed F1-scores across key structural tiers:

• Base Pairs: Canonical Watson–Crick pairs were evaluated by strict residue correspondence. Non-Watson–Crick (nWC) pairs were scored on a 6-point rubric based on DSSR's Leontis–Westhof classifications (residue identity, cis/trans orientation, and interacting edges).
• Base Stacking: Pairwise contacts from DSSR-annotated stacking segments were scored using residue identity, inter-planar distance (≤ 0.5 Å deviation), and base-ring normal angles (≤ 15° deviation).
• Secondary Structure & Junctions: Secondary structure elements—including hairpins, bulges, internal loops, and both helices (defined by DSSR via base-stacking interactions regardless of pairing type or backbone continuity) and stems (helices consisting strictly of canonical Watson–Crick/wobble pairs with a continuous backbone)—were matched at ≥ 50% residue overlap. Multi-helix junctions were evaluated via a topology-gated framework assessing branch count, helix connectivity, terminal base pairs, unpaired linkers, and coaxial stacking.
• Tertiary Motifs & Interfaces: A-minor interactions, ribose zippers, K-turns, kissing loops, pseudoknots, and multiplets were scored by residue overlap, while multimer interfaces were quantified by inter-chain base-pair recovery.

Step-by-Step Workflow

1. Consensus 2D Restraints: Secondary structures were predicted using thermodynamic, probabilistic, pseudoknot-capable, and alignment-based tools (RNAfold, RNAstructure, CentroidFold, IPknot, RNAalifold, ScanFold 2.0), then combined into dot-bracket restraints.
2. 3D Modeling & Refinement: Monomers were generated via deep learning (AlphaFold 3, DeepFoldRNA, RhoFold+, NuFold, trRosettaRNA) and physics/fragment-based tools (SimRNA, FARFAR2, RNAComposer, Vfold2, BriQ), refined with QRNAS, and filtered for entanglements with RNAspider.
3. Multimer Assembly: Higher-order complexes were modeled by screening AlphaFold 3 across candidate stoichiometries and ionic/cofactor conditions, supplemented by manual assembly in PyMOL.
4. DSSR Benchmarking: Representative models were processed with DSSR alongside experimental references to quantify feature recovery (F1-scores) across the structural hierarchy.

Conclusion & Paper References

Using DSSR's standardized annotations, Nithin et al. showed that while current prediction pipelines reliably recover Watson–Crick base pairs (monomer F1 ≈ 0.84), accuracy drops sharply for non-canonical pairs, stacking geometry, multi-helix junctions, and tertiary motifs—errors that compound across RNA–RNA multimer interfaces.

• Primary Publication: Nithin, C., Pilla, S. P., & Kmiecik, S. (2026). Hierarchical breakdown of RNA structure prediction in CASP16: from reliable local helices to speculative multimer assembly. Bioinformatics, 42(10), btag689. https://doi.org/10.1093/bioinformatics/btag689
• DSSR: Lu, X.-J., Bussemaker, H. J., & Olson, W. K. (2015). DSSR: an integrated software tool for dissecting the spatial structure of RNA. Nucleic Acids Research, 43(21), e142. https://doi.org/10.1093/nar/gkv716