henhaobio
Sign in

bioRxivBioinformaticsPreprint

Protein complex prediction with AlphaFold-Multimer

Richard EvansiD, Michael O’NeilliD, Alexander PritzeliD, Natasha Antropova, Andrew SenioriD, Tim GreeniD, Augustin ŽídekiD, Russ BatesiD, Sam BlackwelliD, Jason YimiD, Olaf RonnebergeriD, Sebastian W. BodensteiniD, Michal Zielinski, Alex Bridgland, Anna PotapenkoiD, Andrew CowieiD, Kathryn TunyasuvunakooliD, Rishub JainiD, Ellen ClancyiD, Pushmeet KohliiD, John JumperiD, Demis HassabisiD

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

Abstract

While the vast majority of well-structured single protein chains can now be predicted to high accuracy due to the recent AlphaFold [1] model, the prediction of multi-chain protein complexes remains a challenge in many cases. In this work, we demonstrate that an AlphaFold model trained specifically for multimeric inputs of known stoichiometry, which we call AlphaFold-Multimer, significantly increases accuracy of predicted multimeric interfaces over input-adapted single-chain AlphaFold while maintaining high intra-chain accuracy. On a benchmark dataset of 17 heterodimer proteins without templates (introduced in [2]) we achieve at least medium accuracy (DockQ [3] [≥] 0.49) on 13 targets and high accuracy (DockQ [≥] 0.8) on 7 targets, compared to 9 targets of at least medium accuracy and 4 of high accuracy for the previous state of the art system (an AlphaFold-based system from [2]). We also predict structures for a large dataset of 4,446 recent protein complexes, from which we score all non-redundant interfaces with low template identity. For heteromeric interfaces we successfully predict the interface (DockQ [≥] 0.23) in 70% of cases, and produce high accuracy predictions (DockQ [≥] 0.8) in 26% of cases, an improvement of +27 and +14 percentage points over the flexible linker modification of AlphaFold [4] respectively. For homomeric inter-faces we successfully predict the interface in 72% of cases, and produce high accuracy predictions in 36% of cases, an improvement of +8 and +7 percentage points respectively.

In-depth summary

Claude Opus 5.5 · xhigh effort · 17 Oct 2026 · from the full text of version 2 (10 Mar 2022)

AlphaFold trained on complexes, with cross-chain paired alignments and symmetry-aware losses, gets 70% of recent heteromeric interfaces acceptable or better (DockQ ≥ 0.23), 27 points above linker-joined AlphaFold.

Single-chain AlphaFold predicts most well-structured monomers accurately, but many complexes are still missed, and earlier workarounds only altered the input (residue-index gaps or flexible linkers) to a model trained on single chains. AlphaFold-Multimer instead changes the model and trains it directly on multi-chain structures of known stoichiometry, with paired multiple sequence alignments (MSAs), permutation-aware losses for identical chains and interface-biased cropping.

Key results

  1. ShownOn non-redundant, low-template-identity chain pairs from recent PDB complexes, AlphaFold-Multimer recovered most interfaces and clearly beat single-chain AlphaFold with a glycine-serine linker, with the larger gain on heteromers.

    2,609 chain pairs from 4,446 complexes. Heteromeric: DockQ ≥ 0.23 in 70% and ≥ 0.8 in 26% of cases (+27 and +14 percentage points vs AlphaFold-Linker); homomeric: 72% and 36% (+8 and +7 points). Mean DockQ gain of the 5×1 setting over AlphaFold-Linker: +0.2 (heteromeric) and +0.06 (homomeric); 5×5 sampling added about +0.01. · Abstract; 4. Results; 4.2. Recent-PDB-Multimers; 7.8. Datasets

  2. SuggestedOn Benchmark 2, template-free heterodimers released after the training cutoff with no homologous complexes in the PDB, AlphaFold-Multimer produced more medium- and high-accuracy models than the best previous AlphaFold-based system.

    17 targets: DockQ ≥ 0.49 on 13 and ≥ 0.8 on 7, vs 9 and 4 for the previous best system; mean DockQ 0.63 vs 0.49 for the ClusPro-refined AlphaFold system. Docking baseline scores were taken from the earlier publication; bootstrapped 95% intervals are in a figure not included here. · Abstract; 4. Results; 4.1. Benchmark 2

  3. ShownThe new interface confidence score (ipTM, an interface version of predicted TM-score, mixed with pTM for ranking) tracks real interface quality: most high-confidence models were accurate, and a wrong complex (PDB 6QF7) showed high predicted error between chains.

    Based on the distribution of DockQ against predicted interface TM-score on Recent-PDB-Multimers; the text gives no correlation coefficient or calibration statistic. · 2.8. Model Confidence; 4.2. Recent-PDB-Multimers; 5. Discussion; 7.9. Interface pTM

  4. ShownFor heteromer chains, AlphaFold-Multimer gave better single-chain structures than AlphaFold when each chain was folded within its full complex, but worse ones when chains were folded alone; AlphaFold-Linker showed no such effect.

    Per-chain accuracy compared for AlphaFold-Multimer (5×1, whole complex folded then chains extracted), AlphaFold-Multimer run on isolated chains, AlphaFold and AlphaFold-Linker; the text reports the direction only, without values. · 4.2. Recent-PDB-Multimers

  5. ShownAccuracy on recent complexes showed little dependence on how similar the targets were to training structures.

    Similarity defined as identity × coverage of the least well-matched chain against training templates; the authors report fairly little correlation with model quality but give no statistic in this text. · 4.2. Recent-PDB-Multimers

Methods, data and reproducibility

AlphaFold was adapted to multi-chain inputs (chain-aware positional encodings, species-paired MSAs, greedy permutation alignment of identical chains, frame aligned point error clamped at 10 Å within and 30 Å between chains, a chain centre-of-mass loss, mixed contiguous and interface cropping), trained on PDB complexes up to 2018-04-30, and scored by DockQ against AlphaFold variants and ClusPro docking.

  • Training set: PDB biological assemblies released up to 2018-04-30, sampled through 40% identity chain clusters (MMSeqs2), plus an MGnify-derived self-distillation set
  • Benchmark 2: heterodimers released after the training cutoff with no homologous complexes in the PDB, predicted without templates · n = 17 heterodimers
  • Recent-PDB-Multimers: PDB complexes released 2018-04-30 to 2021-08-02 with more than one and fewer than 9 chains and under 1,536 residues, clustered for redundancy · n = 4,446 complexes; 2,609 in-contact chain pairs after filtering (5,287 before)
  • Sequence databases: unclustered UniProt 2020_05 for species-based MSA pairing; Uniref90 MSAs and PDB seqres for per-chain template search
Methods & tools
AlphaFold, AlphaFold-Multimer, JackHMMER v3.3, HMMER v3.3 (hmmsearch, hmmbuild), MMSeqs2, UniProt, MGnify, ColabFold (AlphaFold-Gap baseline), ClusPro, DockQ, TPUv3
Reagents & environment
Genetic search with JackHMMER v3.3 on UniProt 2020_05 and HMMER v3.3 for templates (PDB seqres downloaded 2020-05-14); no container, dependency list or model-code versions are given in this text.
Assay conditions & compute
Training: roughly 10M samples over 2 weeks on 128 TPUv3 cores (batch size 1 per core), then two days of fine-tuning; inference times are said to be similar to those of an earlier cited report (the reference is missing from this text), with no figures here.
Reproducibility gaps
The discussion says source code, weights and inference scripts are available, but the link is missing from this text. Per-target scores and DockQ category breakdowns are in supplementary tables not included here.
Limitations (6)
  • Multimer templates and self-distillation on multimer predictions were not yet implemented. · stated by the authors
  • Heteromeric interfaces are predicted less well than homomeric ones, which the authors attribute to weaker and harder-to-access co-evolutionary signal across different chains. · stated by the authors
  • Anecdotally, the model generally fails to predict antibody binding. · stated by the authors
  • When heteromer chains are folded alone (wrong stoichiometry), AlphaFold-Multimer is less accurate than single-chain AlphaFold. · stated by the authors
  • Stoichiometry must be supplied as input, and robustness to wrong or unknown stoichiometry is not tested beyond the isolated-chain comparison. · noted in this summary
  • The comparison with docking pipelines rests on 17 Benchmark 2 targets with borrowed baseline scores; the 2,609-pair evaluation compares only with AlphaFold-Linker. · noted in this summary

Read this if you predict protein complex structures, use ipTM-style confidence to triage models, or want the specific training changes needed to adapt AlphaFold to multiple chains.

Citations
3,798

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
THISZhu '23Gao '21Chen '22Oda '22Yin '22Harmalkar '23
90% semantic matchBioinformatics
Wensi Zhu, Aditi Shenoy, Petras J. Kundrotas+1 more·Bioinformatics·Jul 1, 2023· 199 citations