henhaobio
Sign in

Proteins Structure Function and BioinformaticsBioinformaticsPeer-reviewed journal article

AlphaFold3 and Intrinsically Disordered Proteins: Reliable Monomer Prediction, Unpredictable Multimer Performance

Tuan Minh Dao, Sebastien Ghent, Vladimir Nikolaevich UverskyiD, Taseef RahmaniD

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

Abstract

AlphaFold3 represents a major advance in protein structure prediction, yet its performance on intrinsically disordered proteins remains uncharacterized. We present the first systematic evaluation of AF3 on disordered systems, revealing a striking dichotomy. For monomers, AF3's pLDDT scores reliably predict disorder (MCC: 0.693), matching AlphaFold2 and rivaling dedicated predictors. This consistency across fundamentally different architectures confirms that disorder prediction emerges from training data, not model design. For multimers, the picture grows complex. Despite comparable aggregate performance (mean DockQ: 0.563 vs 0.571), AF3 and AF2 achieve these results through fundamentally different mechanisms. Conventional structural features explain 58% of AF2's variance but only 42% of AF3's. Users cannot predict when AF3 will succeed or fail from interface properties alone. On disorder-to-order transitions (MFIB benchmark), both models perform equally well, successfully predicting final folded states. Yet seed variance analysis reveals AF3's failures are deterministic: the model converges to identical structures across independent runs, whether correct or incorrect, indicating rigid structural priors override available information. Our findings establish AF3 as reliable for the prediction of monomer disorder but unpredictable for multimers. Architectural innovation alone cannot overcome training data bias. Progress demands disorder-enriched datasets and ensemble sampling, not merely novel architectures.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (13 Oct 2026) · CC BY 4.0

AlphaFold3 pLDDT matches AlphaFold2 at calling disordered residues (Matthews correlation 0.693 vs 0.701); on disorder-containing complexes mean DockQ is similar, but interface features explain less variance (42% vs 58%).

Earlier work showed AlphaFold2 (AF2) pLDDT tracks intrinsic disorder and AlphaFold2-Multimer captures binding by intrinsically disordered regions (IDRs), but AlphaFold3 (AF3), which swaps the Evoformer for diffusion-based generation, had not been systematically tested on disordered systems. The paper benchmarks AF3 against AF2 on monomer disorder, on IDR-receptor complexes chosen to avoid training overlap and on folding-upon-binding complexes, then dissects cases where the two disagree.

Key results

  1. ShownAF3 pLDDT (disordered if below 70) classifies disordered residues on par with AF2 pLDDT and the best dedicated predictors, and well ahead of sequence-only tools such as fidpnn and espritz_d.

    DisProt-PDB, 478 proteins: Matthews correlation coefficient (MCC) 0.693 for AF3 vs 0.701 AF2, 0.698 spot_disorder2, 0.692 spot_disorder1, 0.611 ESMFold, 0.397 espritz_d, 0.390 fidpnn; AUC 0.90 vs 0.91 (AF2) and 0.87 (ESMFold); F-max 0.78 for both AlphaFold versions. Per-amino-acid disorder frequencies track ground truth at r = 0.985 for both. · Datasets; Disorder Prediction Evaluation; AlphaFold3 pLDDT is a reliable predictor of…; AlphaFold3 matches AlphaFold2 disorder…

  2. ShownOn 90 IDR-receptor dimers (B90), AF3 and AF2-Multimer reach near-identical interface accuracy, contact accuracy and interface confidence, and both lose accuracy to a similar degree as interface disorder rises.

    Mean DockQ 0.563 (AF3) vs 0.571 (AF2), medians 0.625 vs 0.642; contact accuracy 0.59 vs 0.58, with 51.1% vs 54.4% of complexes over-contacted; interface predicted aligned error (PAE) mean 7.86 vs 7.68 Å. Mean DockQ is 0.67 for both below 20% interface disorder and drops to 0.20 above 40%; intrinsically disordered protein (IDP) chains are predicted mostly helical (median helix 0.60-0.62). · Similar performances for AF2 and AF3 for…; Figure legends

  3. SuggestedRetrospective linear fits on interface disorder fraction and secondary-structure (DSSP) agreement explain less of AF3's DockQ variation than of AF2's, and disorder content dominates the fit only for AF2.

    Combined R² 0.421 (AF3) vs 0.577 (AF2); disorder alone R² 0.310 (r = −0.556) vs 0.492 (r = −0.701); DSSP alone 0.237 vs 0.275. Adding disorder to a DSSP-only fit gains ΔR² 0.184 (AF3) vs 0.302 (AF2), adding DSSP to disorder 0.111 vs 0.085. DSSP agreement is scored against the native structure, so the fits diagnose rather than forecast quality. · Structure Prediction and Analysis; Regression and Statistical Analysis; Conventional Structural Features Explain Less of…

  4. ShownOn 90 MFIB complexes, where disordered regions fold on binding, AF3 and AF2 give nearly identical accuracy distributions, yet neither model's pLDDT identifies in advance which molecular recognition features (MoRFs) will fold.

    Excellent/good/moderate/poor counts AF3 43/23/16/8 vs AF2 42/23/15/10; median interface RMSD about 1 Å for both (Mann-Whitney p = 0.428). MoRF calls at the optimal pLDDT cutoffs reach mean accuracy 1.98% (AF3) and 0.84% (AF2); mean bound vs unbound pLDDT shift −0.96 vs −2.07; Mean Confidence Weighted (MCW) scores vs pLDDT r = −0.106 (p = 0.322) and −0.161 (p = 0.129). · Molecular Recognition Feature Analysis; AF3 Performance on MFIB; Discussion; Figure legends

  5. SuggestedThe 13 most divergent complexes sort into four proposed failure modes, and repeated AF3 runs mostly land on the same outcome; where seeds split, only a minority of them match or beat AF2.

    Of 13 complexes with |ΔDockQ| > 0.3, AF2 wins every structured-IDP and diffusion-collapse case and AF3 every genuine-disorder and multiple sequence alignment (MSA) over-assignment case. With 30 AF3 seeds on six cases: 7TA3 and 8K00 std <0.04 (means 0.871 vs AF2 0.097; 0.062 vs 0.769); 8CZF, 8CZH and 8FCN bimodal, 7-17% of seeds matching or beating AF2; 8GTX mean 0.294 vs AF2 0.926. · High-Divergence Cases Reveal Four Distinct…; Seed Variance Analysis Reveals Mostly…; Discussion; Tables

Methods, data and reproducibility

AF3, AF2 and ESMFold pLDDT were thresholded (cutoff tuned over 60-80 for best MCC) to call disorder on DisProt-PDB and compared with 11 disorder predictors. AF3 and AF2-Multimer models of B90 and MFIB complexes were scored against native structures (DockQ v2; interface and ligand RMSD), then probed with linear regression on interface features, grouping of divergent cases and 30-seed AF3 reruns.

  • DisProt-PDB (Wilson et al. 2022; DisProt release 2018_11): residues labelled disordered from DisProt or ordered from PDB, unannotated residues masked · n = 478 protein targets
  • B90: IDR-receptor dimers from the Omidi et al. (2024) blind set, deposited after September 30, 2021 (97 eligible); 90 chosen for a roughly normal spread of AF2 DockQ · n = 90 complexes
  • Complexes from the Mutual Folding-Induced Binding (MFIB) 2.0 database, in which IDRs fold upon binding · n = 90 complexes
  • Complexes with |ΔDockQ| > 0.3 between AF2 and AF3, grouped into four failure-mode clusters · n = 13 complexes · PDB 8CZH, 8CZF, 8GTX, 7XV1, 8K00, 7UZU, 8FCN, 7NMI, 7OS1, 8F7Q, 7TA3, 8B8V, 7SJQ
  • Divergent cases rerun with 30 AF3 random seeds each · n = six cases · PDB 7TA3, 8K00, 8CZF, 8CZH, 8FCN, 8GTX
Methods & tools
AlphaFold3 (AlphaFold Server), AlphaFold2-Multimer v3 (ColabFold), ESMFold, DockQ v2, DSSP, PyMOL, ChimeraX matchmaker, DisProt, MFIB 2.0, 11 disorder predictors incl. spot_disorder1, spot_disorder2, RawMSA, fidpnn, espritz_d, Kyte-Doolittle hydrophobicity scale, scipy, numpy, pandas
Reagents & environment
AF3 via the AlphaFold Server; AlphaFold2-Multimer v3 via the ColabFold AlphaFold2.ipynb notebook with default parameters; DockQ v2; DisProt release 2018_11; Python (scipy, numpy, pandas, matplotlib, seaborn) with no library versions stated.
Reproducibility gaps
Analysis scripts are not mentioned (the repository is described as holding B90 IDs, curation criteria, predicted structures and metrics); AF3 server version and MFIB selection criteria are not given; Methods cite five extra seeds per case, but Results and Table S4 report 30.
Limitations (6)
  • The multimer benchmark comprises 90 complexes, which the authors acknowledge is modest. · stated by the authors
  • Main comparisons use only each complex's top-ranked model. · stated by the authors
  • The two interface features may miss other determinants of quality, and the regression only describes predictions already scored against experiment, so it cannot estimate DockQ for unsolved complexes. · stated by the authors
  • pLDDT-based disorder fractions deviate more in highly disordered proteins (RMSD 0.48-0.58) than in highly ordered ones (0.20-0.22), which the authors read as underestimation of extreme disorder. · stated by the authors
  • Seed-to-seed variation was examined for just six high-divergence cases, so how stable AF3 is across seeds on the rest of B90 is unknown. · noted in this summary
  • The claim that disorder sensitivity arises from training data rather than architecture rests on AF2 and AF3 agreeing; no experiment varies training data or architecture to test it. · noted in this summary

Read this if you use AlphaFold3 pLDDT to call disorder or model complexes with disordered partners and want evidence on when AF3 and AF2 diverge and how to flag doubtful AF3 interfaces.

Citations
5
FWCI
10.46
Field percentile
top 2%

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
THISVelasquez '26Wilson '22Brotzakis '25Neupane '26Eicholt '26Brotzakis '24
74% semantic matchBiophysics
Juan Velasquez, Sebastien Ghent, Vladimir Nikolaevich Uversky+1 more·bioRxiv·Jun 8, 2026