henhaobio
Sign in

Proteins Structure Function and BioinformaticsBiophysicsPeer-reviewed journal article

Assessment of Nucleic Acid Structure Prediction in CASP16

Rachael C. KretschiD, Alissa M. HummeriD, Shujun HeiD, Rongqing YuaniD, Jing ZhangiD, Thomas KaragianesiD, Qian CongiD, Andriy KryshtafovychiD, Rhiju DasiD

PDFDOIPreprintCodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

Consistently accurate 3D nucleic acid structure prediction would facilitate studies of the diverse RNA and DNA molecules underlying life. In CASP16, blind predictions for 42 targets canvassing a full array of nucleic acid functions, from dopamine binding by DNA to formation of elaborate RNA nanocages, were submitted by 65 groups from 46 different labs worldwide. In contrast to concurrent protein structure predictions, performance on nucleic acids was generally poor, with no predictions of previously unseen natural RNA structures achieving TM-scores above 0.8. Even though automated server performance has improved, all top-performing groups were human expert predictors: Vfold, GuangzhouRNA-human, and KiharaLab. Good performance on one template-free modeling target (OLE RNA) and accurate global secondary structure prediction suggested that structural information can be extracted from multiple sequence alignments. However, 3D accuracy generally appeared to depend on the availability of closely related 3D structure templates, and predictions still did not achieve consistent recovery of pseudoknots, singlet Watson-Crick-Franklin pairs, non-canonical pairs, or tertiary motifs like A-minor interactions. For the first time, blind predictions of nucleic acid interactions with small molecules, proteins, and other nucleic acids could be assessed in CASP16. As with nucleic acid monomers, prediction accuracy for nucleic acid complexes was generally poor unless 3D templates were available. Accounting for template availability, there has not been a notable increase in nucleic acid modeling accuracy between previous blind challenges and CASP16.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (16 Oct 2025) · CC BY 4.0

Across 42 CASP16 nucleic acid targets, accuracy stayed tied to template availability; human experts (Vfold first) led servers and AlphaFold 3, though servers now improve on templates on average.

Nucleic acid structure prediction has trailed protein prediction in RNA-Puzzles and CASP, and CASP15 showed deep-learning RNA methods falling behind human experts. CASP16 for the first time posed DNA structure and RNA multimer challenges and scored nucleic acid interfaces with proteins and small molecules, and this assessment calibrates progress against the best template available before each target.

Key results

  1. ShownMonomer 3D accuracy tracked how close the best prior PDB template was; no unseen natural RNA was predicted above TM-score 0.8, and OLE RNA (R1285), with a deep multiple sequence alignment (MSA), was the notable template-free success.

    Acceptable folds (TM-score > 0.45) for 21 of 36 monomer targets, only 2 of them without a template (TM-align < 0.45). Target clustering split a poorly predicted group (2 of 18 with templates) from a well predicted one (17 of 18). R1285's closest earlier structure was a ribosome (PDB 7ZW0, TM-align = 0.365), but its MSA had effective sequence count Neff = 235. · Abstract; 3.2 | Metrics to Assess NA Monomer Structure…; 3.3 | Difficulty and Performance of NA Monomer…

  2. ShownThe three top-ranked groups (Vfold, KiharaLab, GuangzhouRNA-human) were human expert predictors; Vfold led overall and in both the template-based and template-free monomer subsets.

    Summed Z-scores across four categories put Vfold first, KiharaLab second and GuangzhouRNA-human third. The AlphaFold 3 server was outranked by 8 groups overall and 12 in the monomer category, ranking 8th on deep-MSA vs 19th on shallow-MSA monomer targets; it does no nucleic acid template search. Vfold's first place and the top-five monomer groups held under direct TM-score or lDDT scoring. · Abstract; 3.4 | Ranking Group Performance on NA Monomer…; 3.10 | NA Structure Prediction Current State and…

  3. ShownCanonical secondary structure was well predicted, but pseudoknotted (crossed) and singlet pairs, non-canonical pairs and tertiary motifs such as A-minor interactions, platforms and UA-handles were often missed.

    CASP16 groups (servers included) outperformed sequence-only packages (e.g., ViennaRNA) on canonical pairs, with accuracy only weakly tied to template or MSA depth. Crossed-pair F1 rose with MSA depth; singlet F1 depended on template and length. Non-canonical pair F1 averaged below 0.5. In motif prediction, the AlphaFold 3 server and Yang-Server matched the top group within error. · 3.5 | Base Pair Prediction Accuracy; 3.6 | RNA Motif Prediction Accuracy

  4. ShownNucleic acid complexes were rarely predicted accurately without templates: RNA multimer stoichiometry, symmetry and interfaces were mostly wrong, and good nucleic acid-protein and nucleic acid-ligand interfaces appeared only for templated targets.

    About 10% of 1,072 round 1 multimer models lacked meaningful inter-chain contacts, and only R1290 (sequence-identical templates) got an accurate interface among RNA-only multimers; in hybrids only simple, perfectly complementary double helices were predicted well. Accurate nucleic acid-protein interfaces only in three Fab-RNA loop complexes. Ligands were well placed only in three templated ZTP riboswitches. · 3.7 | Assessment of NA-NA Interfaces and…; 3.8 | Assessment of NA-Protein Hybrid Targets; 3.9 | NA-Ligand Targets

  5. ShownCalibrated by best-template TM-align, accuracy has not risen appreciably from RNA-Puzzles and CASP15 to CASP16, but server predictions moved from below to above template quality on average.

    Mean gain of best prediction over best template TM-align: 0.027 ± 0.014 before vs 0.060 ± 0.016 in CASP16 (p = 0.13, t-test; four CASP15 designed nanostructures excluded). Servers: –0.067 ± 0.012 before vs 0.022 ± 0.017 (p = 7.7 x 10^−5). Independent structure pairs at 2–5 Å agree at TM-align 0.8 to 1.0, yet only three targets lacking a >0.8 template (R1221s2, R1224s2, R1289) beat 0.8. · 3.10 | NA Structure Prediction Current State and…

Methods, data and reproducibility

Assessors scored blind CASP16 models of 42 nucleic acid targets in four categories (monomers, RNA multimers, nucleic acid-protein hybrids, nucleic acid-ligand) with TM-score, GDT_TS, local distance difference test (lDDT) and interface scores, ranked groups by summed Z-scores, and related accuracy to best-template TM-align, multiple sequence alignment (MSA) depth and base-pair and motif F1-scores.

  • CASP16 nucleic acid-containing blind targets from 22 experimental groups (36 monomer, 11 RNA multimer, 18 hybrid and 5 RNA-ligand assessments) · n = 42 targets
  • Blind models from CASP16 predictor groups (up to five per target), including the AlphaFold 3 server baseline submitted as group TS304 · n = 65 groups from 46 labs
  • Round 1 RNA multimer predictions · n = 1,072 predictions from 32 groups
  • RNA-containing PDB entries released before November 11, 2024, searched as candidate templates
  • RNA-Puzzles and CASP15 targets and predictions used for the historical comparison
  • Curated pairs of RNA structures solved independently by different groups (including ROOL, raiA and OLE RNA) to estimate experimental precision
Methods & tools
US-align, OpenStructure, Local-Global Alignment, rna-tools (ClaRNA), rMSA, DSSR, Rosetta rna_motif, AnAnaS, AlphaFold 3 server, OpenKnot Score Pipeline, ViennaRNA, CONTRAfold, EternaFold, RNAstructure, NUPACK, IPknot, SPOT-RNA, RNet, RNAcentral, RFam and NCBI Nucleotide databases
Software & environment
US-align version 20220511; DSSR v1.9.9-2020feb06; rMSA with RNAcentral v20.0, RFam v14.7 and NCBI Nucleotide (2022-10-03); versioned benchmark packages (e.g., ViennaRNA 2.4.16, RNAstructure v6.4, NUPACK 3.0.6). No container or requirements file stated.
Reproducibility gaps
No hardware or runtime reported. 182 models could not be processed by Rosetta rna_motif and some failed DSSR, so they were left out of the motif and base-pair analyses. The ranking metric weights appear as an equation not rendered in the provided text.
Limitations (6)
  • The ligand ranking used only three ZTP riboswitch targets (R1261, R1263, R1264), too few for a comprehensive view of nucleic acid-ligand prediction. · stated by the authors
  • Non-canonical pairs were matched by nucleotide identity only; base-pair edges were not scored because different software tools assign edges inconsistently. · stated by the authors
  • Motif matching set aside chain mapping, which was acceptable for the dimeric and symmetric homo-oligomeric targets here but will need handling for more complex assemblies. · stated by the authors
  • A single experimental reference may not capture every conformation a nucleic acid adopts in solution; scoring took the best of five models partly to compensate. · stated by the authors
  • Method traits (templates, MSAs, literature use) were labelled manually from abstracts and talks, so links between strategy and performance are correlational and may be incomplete. · noted in this summary
  • The historical comparison spans different target sets and measures difficulty by best-template TM-align alone, without MSA depth, so the non-significant change (p = 0.13) has limited power to rule out modest progress. · noted in this summary

Read this if you build or apply RNA or DNA 3D structure predictors, including AlphaFold 3, and want a template-calibrated account of where they still fail: tertiary motifs, multimer interfaces and template-free folds.

Citations
51
FWCI
15.24
Field percentile
top 1%

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISKretsch '25Das '23Nithin '26Si '26Lee '25Wang '25
85% semantic matchBiophysics
Rachael C. Kretsch, Reinhard Albrecht, Ebbe Sloth Andersen+30 more·Proteins Structure Function and Bioinformatics·Sep 4, 2025· 23 citations