AlphaFold 3 Prediction Accuracy and Reliability across RNA Structures and Protein-RNA Interfaces
- 1University of Freiburg
Abstract
AlphaFold 3 (AF3) models RNA and protein-RNA complexes, but when its predictions and confidence metrics can be trusted is unclear. We evaluated AF3 on more than 1,300 protein-RNA complexes and about 270 RNA-only systems, and compared post-cutoff RNAs with pre-cutoff candidates selected by sequence and fold similarity. On structurally novel RNAs, AF3 largely preserves the protein fold, while RNA-fold and interface accuracy drop sharply with little change in predicted RNA confidence score. Input ablations show that protein alignments and templates from RNA-bound proteins govern complex assembly, whereas RNA alignments matter most for single-chain and long RNAs. The most frequent failed interface retains protein-side contacts but places an often-misfolded RNA in the wrong pose. RNA interface confidence only partly distinguishes this error from correct placements. Against CLIP-seq, SELEX and RNAcompete motifs, AF3 usually places the protein against the experimentally observed motif-bearing RNA segment, even where the interface fails, but often not in the native pose. In protein-RNA complexes released in 2026, a score combining AF3 confidence with RNA familiarity and template use ranked RNA chains with accurate folds and correctly placed interfaces better than AF3's ranking score.
Related
- General prediction of T cell receptor antigen specificity from sequence using AlphaFold 3
- From pose estimation accuracy to motor biomarker reliability: a reliability-centred evaluation framework for markerless movement analysis
- Beyond prediction accuracy by integrating GWAASB into wheat stability assessment across environments in Saudi Arabia
- A low-rank decoder bottleneck bounds reliability in foundation-model perturbation prediction
- Neural prediction decorrelation reveals that adversarial robustness substantially improves DNN prediction accuracy across the entire human auditory cortex
BibTeX
@misc{tennie2026alphafold,
title = {{AlphaFold 3 Prediction Accuracy and Reliability across RNA Structures and Protein-RNA Interfaces}},
author = {I. K. Tennie and M. Uhl and J. K. H. Franke and R. Backofen and F. Runge},
howpublished = {bioRxiv},
year = {2026},
month = nov,
doi = {10.64898/2026.11.02.763690},
url = {https://www.biorxiv.org/content/10.64898/2026.11.02.763690v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - AlphaFold 3 Prediction Accuracy and Reliability across RNA Structures and Protein-RNA Interfaces
AU - I. K. Tennie
AU - M. Uhl
AU - J. K. H. Franke
AU - R. Backofen
AU - F. Runge
PY - 2026
DA - 2026/11/02
PB - bioRxiv
DO - 10.64898/2026.11.02.763690
UR - https://www.biorxiv.org/content/10.64898/2026.11.02.763690v1
AB - AlphaFold 3 (AF3) models RNA and protein-RNA complexes, but when its predictions and confidence metrics can be trusted is unclear. We evaluated AF3 on more than 1,300 protein-RNA complexes and about 270 RNA-only systems, and compared post-cutoff RNAs with pre-cutoff candidates selected by sequence and fold similarity. On structurally novel RNAs, AF3 largely preserves the protein fold, while RNA-fold and interface accuracy drop sharply with little change in predicted RNA confidence score. Input ablations show that protein alignments and templates from RNA-bound proteins govern complex assembly, whereas RNA alignments matter most for single-chain and long RNAs. The most frequent failed interface retains protein-side contacts but places an often-misfolded RNA in the wrong pose. RNA interface confidence only partly distinguishes this error from correct placements. Against CLIP-seq, SELEX and RNAcompete motifs, AF3 usually places the protein against the experimentally observed motif-bearing RNA segment, even where the interface fails, but often not in the native pose. In protein-RNA complexes released in 2026, a score combining AF3 confidence with RNA familiarity and template use ranked RNA chains with accurate folds and correctly placed interfaces better than AF3's ranking score.
ER -