Improved Docking of Protein Models by a Combination of Alphafold2 and ClusPro
Usman Ghani, Israel DestaiD, Akhil Jindal, Omeir KhaniD, George S. JonesiD, Nasser Hashemi, Sergey Kotelnikov, Dzmitry PadhornyiD, Sándor VajdaiD, Dima KozakoviD
Abstract
It has been demonstrated earlier that the neural network based program AlphaFold2 can be used to dock proteins given the two sequences separated by a gap as the input. The protocol presented here combines AlphaFold2 with the physics based docking program ClusPro. The monomers of the model generated by AlphaFold2 are separated, re-docked using ClusPro, and the resulting 10 models are refined by AlphaFold2. Finally, the five original AlphaFold2 models are added to the 10 AlphaFold2 refined ClusPro models, and the 15 models are ranked by their predicted aligned error (PAE) values obtained by AlphaFold2. The protocol is applied to two benchmark sets of complexes, the first based on the established protein-protein docking benchmark, and the second consisting of only structures released after May 2018, the cut-off date for training AlphaFold2. It is shown that the quality of the initial AlphaFold2 models improves with each additional step of the protocol. In particular, adding the AlphaFold2 refined ClusPro models to the AlphaFold2 models increases the success rate by 23% in the top 5 predictions, whereas considering the 10 models obtained by the combined protocol increases the success rate to close to 40%. The improvement is similar for the second benchmark that includes only complexes distinct from the proteins used for training the neural network.
02 /Specification & Classification
Click any tag to browse matching papers- Citations
- 68
- Assay
- Research Focus
- Approach
In-depth summary
Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (27 Jul 2022) · CC BY 4.0
Splitting AlphaFold2 complex models, re-docking with ClusPro, AlphaFold2 refinement and predicted-aligned-error ranking raised successful docking-benchmark targets from 88 (five AlphaFold2 models) to 123 (top 10) of 204.
The paper cites earlier work showing that AlphaFold2 and RoseTTaFold can build protein complexes from gap-separated sequences, and that AlphaFold2 and the physics-based PIPER-FlexPepDock succeed on largely different peptide targets. This protocol, which starts from sequence alone, links AlphaFold2 to ClusPro rigid-body docking at three points: re-docking its monomers, using the docked poses as templates, and ranking all models together by predicted aligned error (PAE).
Key results
ShownOn Benchmark 1, the combined protocol solved the most targets. It pools five original AlphaFold2 models with 10 AlphaFold2-refined ClusPro models and keeps the 10 with the lowest interface PAE.
123 of 204 targets had an acceptable-or-better model (DockQ ≥ 0.23), with an average DockQ of 0.415, which is 39.8% more successful targets than the five original AlphaFold2 models (88). The paper's stated 37.6% gain over ClusPro top 10 does not match its own counts (123 vs 98). The abstract reports a 23% top-5 gain from adding the refined ClusPro models; the results give no top-5 count for the pooled set. · Abstract; 2.5. Assessing the accuracy of docked structures; 3.1. AlphaFold2 refinement improves predictions
ShownRefining the 10 ClusPro poses with AlphaFold2, using each pose as a template, beat AlphaFold2 alone at top 5 and raised interface quality. AlphaFold2 alone still did better at top 1.
Top 5: 96 successful targets vs 88 for AlphaFold2 alone. All 10 refined models: 109 vs 98 for unrefined ClusPro top 10. Refinement improved average DockQ by 58.3%. However, the refined set alone gave fewer high-quality interfaces than the five AlphaFold2 models (Table S1). · 3.1. AlphaFold2 refinement improves predictions; Tables
ShownOn Benchmark 1, AlphaFold2 alone beat ClusPro at top 1 and top 5. ClusPro's 10 models covered more targets, but its interfaces were of lower quality.
ClusPro top 10: 98 successful targets vs 88 for the five AlphaFold2 models. ClusPro gave far fewer high-quality interfaces (Figure 1). Top-1 and top-5 counts are not given in the text. · 3.1. AlphaFold2 refinement improves predictions; Figure legends
SuggestedOn 17 template-free heterodimers released after the AlphaFold2 training cut-off, the number of successful targets did not rise from the ClusPro models to the AlphaFold2-refined or pooled models, but refinement improved interface quality.
Average DockQ rose from 0.422 for ClusPro top 10 to 0.561 for the AlphaFold2-refined ClusPro models (Table S2). At top 5, AlphaFold2 alone succeeded on fewer targets than ClusPro but gave more high-quality interfaces (Figure 2). · 3.2. Refinement improvement persists for new…; Tables
Methods, data and reproducibility
AlphaFold2 (pTM parameters, MMseqs2 multiple sequence alignments, chains joined by a 200-residue gap, no templates) built five complex models. The model with the lowest interface PAE was split and re-docked with ClusPro, and each of the 10 ClusPro models seeded one templated AlphaFold2 run. All 15 models were ranked by average interface PAE and scored against native structures with DockQ.
- Benchmark 1: protein-protein docking benchmark version 5 (BM5, Weng group): 40 antibody-antigen, 88 enzyme-containing and 102 other complexes · n = 230 complexes; results for 204 (34 antibody-antigen, 170 non-antibody)
- Benchmark 2: heterodimers released in the PDB after May 2018 with no homologous template found by ClusPro TBM · n = 17 complexes · PDB 5ZNG, 6A6I, 6GS2, 6H4B, 6IF2, 6II6, 6ONO, 6PNQ, 6Q76, 6U08, 6ZBK, 7AYE, 7D2T, 7M5F, 7N10, 7NLJ, 7P8K
- Methods & tools
- AlphaFold2, ClusPro, ClusPro TBM, PIPER, MMseqs2, DockQ, CHARMM, Protein Data Bank, PISA
- Reproducibility gaps
- No code, scripts or model files are linked. ClusPro was run through its server. Unstructured tails were trimmed by hand before docking. The choice of pTM model 1 for refinement rests on data not shown. Limited GPU resources meant only 204 of the 230 BM5 targets produced results.
Limitations (6)
- BM5 component proteins were in AlphaFold2's training data (PDB structures released before May 2018), so Benchmark 1 may overestimate AlphaFold2 performance. · stated by the authors
- Limited GPU resources meant results were obtained for only 204 of the 230 BM5 complexes. · stated by the authors
- ClusPro is rigid-body docking and does not model conformational change on binding, which lowers the quality of its interfaces. · stated by the authors
- On some targets where AlphaFold2 alone did better, refining the ClusPro models lost high-quality interfaces. The original AlphaFold2 models were added back to the pool to compensate. · stated by the authors
- Benchmark 2 has only 17 complexes, and its success counts did not rise from the ClusPro models to the refined or pooled models. The abstract's claim of a similar improvement therefore rests on DockQ gains alone. · noted in this summary
- The headline gains compare the 10 best pooled models with only five AlphaFold2 models, so part of the gain may come from scoring more candidates. · noted in this summary
Who should read:Read this if you model protein-protein complexes from sequence and want a recipe for pairing AlphaFold2 with rigid-body docking and ranking the candidates by interface PAE.
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
- Benchmarking AlphaFold for protein complex modeling reveals accuracy determinants
Rui Ming Yin, Brandon Y. Feng, … Brian G. Pierce·Protein SciencePeer-reviewed·Jul 13, 2022
- Improved prediction of protein-protein interactions using AlphaFold2
Patrick Bryant, Gabriele Pozzati, Arne Elofsson·Nature CommunicationsPeer-reviewed·Mar 10, 2022
- Reliable protein-protein docking with AlphaFold, Rosetta and replica-exchange
A. Harmalkar, S. Lyskov, J. J. Gray·bioRxiv·Jul 28, 2023
- ProteinDock: A physics-informed layer to improve protein-protein docking reliability
Gowrish Rajagopal, Soren C. Spina, … Blaise R. Kimmel·bioRxiv·Jul 20, 2026
- Adapting Co-Folding Models for Structure-Based Protein-Protein Docking Through Flow Matching
Da Zhuan Xu, Lee‐Shin Chu, Jeffrey J. Gray·PRX LifePeer-reviewed·Jul 8, 2026
- Harnessing protein folding neural networks for peptide-protein docking
Tomer Tsaban, Julia K. Varga, … Ora Schueler‐Furman·Nature CommunicationsPeer-reviewed·Jan 10, 2022