Modeling Protein-Protein and Protein-Ligand Interactions by the ClusPro Team in CASP16 - PubMed (original) (raw)

. 2026 Jan;94(1):183-191.

doi: 10.1002/prot.70066. Epub 2025 Oct 20.

Ryota Ashizawa 1 2 3 4, Sergei Kotelnikov 1 2 4, Stan Xiaogang Li 1 2 3 4, Ernest Glukhov 1 2 3 4, Xin Cao 1 2 6, Maria Lazou 7, Ayse Bekar-Cesaretli 5, Derara Hailegeorgis 1 2 3 4, Veranika Averkava 1 2 3 4, Yimin Zhu 1 2 3 4, George Jones 1 2, Hao Yu 8, Dmytro Kalitin 1 2 4, Darya Stepanenko 1 2 4, Kushal Koirala 9 10, Taras Patsahan 11 12, Dmitri Beglov 7, Mark Lukin 13, Diane Joseph-McCarthy 5 7, Carlos Simmerling 2 14, Alexander Tropsha 9, Evangelos Coutsias 1 2, Ken A Dill 2 14, Dzmitry Padhorny 1 2 3 4, Sandor Vajda 5 7, Dima Kozakov 1 2 3 4

Affiliations

Modeling Protein-Protein and Protein-Ligand Interactions by the ClusPro Team in CASP16

Ryota Ashizawa et al. Proteins. 2026 Jan.

Abstract

In the CASP16 experiment, our team employed hybrid computational strategies to predict both protein-protein and protein-ligand complex structures. For protein-protein docking, we combined physics-based sampling-using ClusPro FFT docking and molecular dynamics-with AlphaFold (AF)-based sampling, followed by AF-based refinement. Our method produced numerous high-accuracy complex models, including cases where AF alone failed, underscoring the critical role of physics-based sampling alongside deep learning-based refinement. For protein-ligand docking, we integrated the ClusPro LigTBM template-based approach with a machine learning-based confidence model for rescoring. The method preserves conserved interaction fragments derived from homologous complexes, followed by local resampling using physics-based sampling and a diffusion model. Our template-based strategy achieved a mean lDDT-PLI of 0.69 across 233 targets, which was highly competitive. These results demonstrate that combining physics-based modeling with AI-driven refinement can significantly enhance the accuracy of both protein-protein and protein-ligand structure predictions.

© 2025 The Author(s). PROTEINS: Structure, Function, and Bioinformatics published by Wiley Periodicals LLC.

PubMed Disclaimer

Conflict of interest statement

The PIPER docking program, which is the base of ClusPro, has been licensed by Boston University to Acpharis Inc. Acpharis, in turn, offers commercial sublicenses of PIPER. D.K. and S.V. consult for Acpharis and own stock in the company, and D.B. is the acting CEO of the company. However, the PIPER software and the ClusPro server are free for non‐commercial use.

Figures

FIGURE 1

FIGURE 1

General overview of the methodology employed for modeling of protein multimers. Ligand and receptor denote the smaller and the larger component proteins (see text). For each of the multimeric targets, we generate a large number of multimer poses, using either deep learning (AlphaFold2/3 and their derivatives) or physics‐based (fast Fourier transform‐based docking and molecular dynamics) sampling. The models are pooled together and subjected to refinement and interface‐focused rescoring using the AlphaFold2 template‐based modeling option. Top‐scoring models are submitted as final predictions for each target.

FIGURE 2

FIGURE 2

General overview of the methodology employed for modeling of protein–ligand complexes.

FIGURE 3

FIGURE 3

Structural superpositions of predicted and experimental complexes for selected CASP16 targets. Predicted structures are shown in color, and crystal structures are shown in white. To simplify the figure, a protein of interest and its interface partners are displayed. (A–C) Antibody–antigen complexes (antibody in pink; antigens in cyan); (D) antibody–peptide complex (antibody in pink; peptide in cyan); (E) viral protein–translation factor complex (translation factor in pink; viral protein in cyan).

FIGURE 4

FIGURE 4

Top: Sample of accurate models obtained for ligands that bind to (A) Chymase, (B) Autotaxin, and (C) MPro. Bottom: Ligand RMSD distribution and success rates for the most accurate model in the top (D) 1 and (E) top 5 submissions, split by protein target. RMSD distributions for target ligands bound to Chymase (L1000) is displayed in green, Cathepsin G (L2000) is displayed in light blue, Autotaxin (L3000) is displayed in orange, and MPro (L4000) is displayed in gray. The number of ligands bound to each target protein is annotated above each respective violin plot. The dashed red line indicates an RMSD cutoff of 2.5 Å for a successful prediction.

References

    1. Pereira J., Simpkin A. J., Hartmann M. D., Rigden D. J., Keegan R. M., and Lupas A. N., “High‐Accuracy Protein Structure Prediction in CASP14,” Proteins: Structure, Function, and Bioinformatics 89, no. 12 (2021): 1687–1699. -PubMed
    1. Jumper J., Evans R., Pritzel A., et al., “Highly Accurate Protein Structure Prediction With AlphaFold,” Nature 596, no. 7873 (2021): 583–589. -PMC -PubMed
    1. Ko J. and Lee J., “Can AlphaFold2 Predict Protein‐Peptide Complex Structures Accurately?,” bioRxiv 2021.2007.453972, 2021, 10.1101/2021.07.27.453972. -DOI
    1. Tsaban T., Varga J. K., Avraham O., Ben‐Aharon Z., Khramushin A., and Schueler‐Furman O., “Harnessing Protein Folding Neural Networks for Peptide‐Protein Docking,” Nature Communications 13, no. 1 (2022): 176. -PMC -PubMed
    1. Bryant P., Pozzati G., and Elofsson A., “Improved Prediction of Protein‐Protein Interactions Using AlphaFold2,” Nature Communications 13, no. 1 (2022): 1265. -PMC -PubMed

MeSH terms

Substances

Grants and funding

LinkOut - more resources