henhaobio
Sign in

Nature CommunicationsBioinformaticsPeer-reviewed journal article

Improved prediction of protein-protein interactions using AlphaFold2

Patrick BryantiD, Gabriele PozzatiiD, Arne ElofssoniD

PDFDOIPreprintCodeData

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 3 (15 Dec 2021) · CC BY 4.0

AlphaFold2, fed concatenated chains with optimized organism-paired alignments and pDockQ-ranked seeds, gives acceptable models for 63% of heterodimers; the same predicted DockQ score flags interacting pairs.

Docking methods remain unreliable: web servers reach up to 16% success on Benchmark 5, and the authors' earlier trRosetta-based Fold and Dock pipeline modelled only 7% of pairs correctly. This work tests whether single-chain AlphaFold2 with tuned multiple sequence alignments (MSAs) can fold and dock heterodimers, and whether its confidence outputs can rank models and identify genuine interactions.

Key results

  1. ShownGiving AlphaFold2 both its default MSAs and organism-paired MSAs substantially improves heterodimer docking success. Running several random seeds and ranking the models by pDockQ adds further gains.

    Development set: 39.4% (AF2+paired) vs 33.3% (AF2 MSAs). Test set success rate (DockQ≥0.23): 57.8% (856 of 1481) with AF2+paired vs 45.0% with default AF2 MSAs; block diagonalization+paired 58.4% (median DockQ 0.363). Ranking five seeded runs by pDockQ lifts these to 61.7% and 62.7%. · Identifying the best AlphaFold2 model; Test set performance; Model variation and ranking for the test set

  2. ShownThe AlphaFold2 pipeline beats RoseTTAFold, GRAMM, template-based TMdock and MDockPP, even though the last three were given bound-form structures. It trails AlphaFold-multimer, whose development data overlap this benchmark, which complicates the comparison.

    RoseTTAFold outperforms AF2 on only 14 test pairs; GRAMM and TMdock interface do so on 188 and 225 pairs; MDockPP (best CASP14-CAPRI method) reaches SR 24.2%; AlphaFold-multimer reaches SR 72.2% (median DockQ 0.560) vs 63% for this protocol. · Introduction; Test set performance

  3. ShownpDockQ is a sigmoid fitted to mean interface predicted lDDT (pLDDT) multiplied by the log of interface contacts. It estimates DockQ and separates acceptable from incorrect models better than any single confidence or interface metric.

    Test set AUC 0.95 for pDockQ vs 0.92 (interface Cβ contacts), 0.91 (interface residues), 0.88 (interface pLDDT) and 0.66 (whole-complex pLDDT); mean DockQ prediction error 0.11. Ranked models: AUC 0.93, so 31% of models are called acceptable at 1% error and 54% at 10%. · Distinguishing acceptable from incorrect models; Model variation and ranking for the test set; pDockQ

  4. ShownThe same pDockQ score separates truly interacting heterodimers from experimentally supported non-interacting pairs. Its main advantage over simpler metrics is at low false-positive rates.

    E. coli plus mammalian Negatome negatives: AUC 0.87 for pDockQ vs 0.82 (interface pLDDT), 0.84 (contacts), 0.85 (residues); TPR 51% at 1% FPR vs 27%, 18 and 13%; TPR 66% at 5% FPR. E. coli-only negatives (3989 pairs): AUC 0.87 vs 0.76 for a prior method. · Identifying interacting proteins; Non-interacting proteins

  5. ShownSuccess depends on taxonomy, the interface and the alignments. Prokaryotic and yeast pairs, larger or helical interfaces, and deeper paired MSAs with stronger direct coupling analysis (DCA) interface signal succeed more often.

    SR by kingdom: Archaea 84.5%, Bacteria 73.7%, Eukarya 61%, Virus 60%. E. coli 76.4% vs S. cerevisiae 66.2% and Homo sapiens 58.1%. Interfaces: helix 62%, sheet 59%, loop 53%; smallest to largest interface tertile 47 to 74%. SR rises with the paired-MSA number of effective sequences (Neff) and with DCA interface PPV. · Bacterial complexes are predicted more…; Findings and future prospects

Methods, data and reproducibility

Heterodimers were folded and docked with AlphaFold2 (model_1, 10 recycles, one ensemble, no templates, no relaxation). The two chains were joined by a 200-residue chain break and the model was given default AF2 MSAs, organism-paired HHblits MSAs or block-diagonalized MSAs, alone or combined. Models were scored with DockQ and ranked by pDockQ, a sigmoid of interface pLDDT × log(interface contacts).

  • Dockground benchmark 4 heterodimers (each chain ≥50 residues, <30% identity) used to develop the pipeline · mixed: 54% eukaryotic, 38% bacterial, 8% cross-kingdom · n = 216 complexes (219 before MSA memory failures)
  • Test set of heterodimers with known interfaces from a prior study: <30% identity, 1-5 Å resolution, unique Pfam domain pairs · human 25%, S. cerevisiae 10%, E. coli 5%, other Eukarya 30% · n = 1481 complexes (from 1,661)
  • CASP14 heteromeric targets folded in pairs · n = six heterodimers
  • PDB complexes deposited after 15 June 2021 with no HHpred hits for at least one chain (template-free test) · n = eight complexes
  • Non-interacting pairs from yeast two-hybrid and AP-MS evidence, taken from the same study as the positive test set · E. coli · n = 3989 pairs
  • Negatome manual stringent set of non-interacting pairs filtered against IntAct · mammals · n = 1964 pairs; 1705 or 1715 after filtering (sections differ)
  • Dockground template library (release 28-10-2020) for template-based docking with TMdock · n = 11756 protein complexes
  • Sequence databases for MSAs: BFD, MGnify v.2018_12, Uniref90 v.2020_01, uniclust30_2018_08, Uniprot v.2021_04
Methods & tools
AlphaFold2 (model_1, model_1_ptm), AlphaFold-multimer, RoseTTAFold, MDockPP (ITScorePP scoring), GRAMM (AACE18 scoring), TMdock template-based docking, HHblits (hh-suite), jackhmmer (HMMER3), CD-HIT 4.7, DockQ, DSSP, GaussDCA
Software & environment
FoldDock commit 2e4c96aa352338976260ece0646ceaaa75392dec (Apache 2.0); hh-suite v.3.0-beta.3; HHblits 3.1.0 with uniclust30_2018_08; MGnify v.2018_12; Uniref90 v.2020_01; CD-HIT 4.7; SciPy v.1.4.1; AF2 run without templates or relaxation
Assay conditions & compute
MSA timing on 16 CPU cores of an Intel Xeon E5-2690v4: AF2 MSA generation 7884 s vs 338 s average for single-chain HHblits searches (24 times speedup); m1-10-1 folding averaged 191 s on the same pairs (hardware unstated); 128 Gb RAM MSA node; 40 Gb GPU for RoseTTAFold.
Reproducibility gaps
Code (Apache 2.0) and data are openly deposited. Some complexes failed MSA generation or folding because of memory or tensor-size limits (AlphaFold-multimer folded 1458 of 1481), so reruns may need large-memory hardware.
Limitations (6)
  • Only the heterodimeric state was modelled. Homodimeric configurations, higher-order oligomers and larger biological assemblies that these chains may form were left out of scope. · stated by the authors
  • A proper test of interaction discrimination would compare interactions within individual organisms. Pairing an E. coli-only negative set with mostly eukaryotic positives likely inflates the reported AUC. · stated by the authors
  • AlphaFold-multimer's development data overlap this benchmark, so its 72.2% vs 63% advantage over this pipeline is hard to interpret directly. · stated by the authors
  • GRAMM, TMdock and MDockPP were given bound-form structures, which supplies near-ideal shape complementarity. Their performance here is therefore optimistic compared with docking unbound models. · stated by the authors
  • pDockQ's sigmoid and its reported mean error both come from the same test-set DockQ scores, with no held-out set for the fit; that error is also stated inconsistently (0.11, 0.14, 0.1). · noted in this summary
  • Evidence on unseen complexes rests on very small sets (four of six CASP14 pairs acceptable; SR 50% on eight template-free complexes deposited after 15 June 2021), so performance on genuinely novel complexes is only loosely estimated. · noted in this summary

Read this if you plan to use AlphaFold2 for heterodimer docking or interactome-scale screening and need guidance on MSA pairing and confidence-based ranking of complex models.

Citations
1,015
FWCI
70.92
Field percentile
top 1%

Abstract

Predicting the structure of interacting protein chains is a fundamental step towards understanding protein function. Unfortunately, no computational method can produce accurate structures of protein complexes. AlphaFold2, has shown unprecedented levels of accuracy in modelling single chain protein structures. Here, we apply AlphaFold2 for the prediction of heterodimeric protein complexes. We find that the AlphaFold2 protocol together with optimized multiple sequence alignments, generate models with acceptable quality (DockQ[≥]0.23) for 63% of the dimers. From the predicted interfaces we create a simple function to predict the DockQ score which distinguishes acceptable from incorrect models as well as interacting from non-interacting proteins with state-of-art accuracy. We find that, using the predicted DockQ scores, we can identify 51% of all interacting pairs at 1% FPR. The protocol can be found at: https://gitlab.com/ElofssonLab/FoldDock.

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
THISYin '22Evans '21Gao '21Danielsson '25Johansson-Å '22Chojnowski '25
82% semantic matchBioinformatics
Rui Ming Yin, Brandon Y. Feng, Amitabh Varshney+1 more·Protein Science·Jul 13, 2022· 335 citations