Skip to content
DraconiticResearch
Search Sources API
190,860 studies31 sourcesUpdated 15 min ago
PreprintCC BYChemRxiv

Integrative Physics and Machine Learning-Based Optimal Binding Pose Generator for Protein-Ligand Complexes

  • Agastya P Bhati1,2,3,4
  • Jinyu Dong3,4
  • Sean Black5
  • Shunzhou Wan3,4
  • Vineeth Gutta6
  • Xibei Zhang3,4
  • Mateusz Bieniek3,4
  • Darren V. S. Green3,4
  • Sunita Chandrasekaran6
  • Eric Stahlberg5
  • Peter V. Coveney3,4
6 affiliations
  1. 1Wadhwani School of Data Science and Artificial Intelligence
  2. 2Indian Institute of Technology Madras
  3. 3Department of Chemistry
  4. 4University College London
  5. 5Frederick National Laboratory for Cancer Research
  6. 6University of Delaware
Open at sourcePDFDOICite

Abstract

Accurately identifying the optimal protein–ligand binding pose is crucial for reliable binding affinity prediction, as molecular dynamics–based approaches are highly sensitive to the initial structure. However, there is currently no efficient way to predict such poses. Indeed, we assess the capabilities of three proposed AI-based software solutions (AlphaFold3, Boltz and RoseTTAFold All-Atom) and demonstrate their substantial shortcomings in this context. Here, we present a combined physics-and machine-learning-based method that predicts optimal protein-ligand poses in their relaxed states, enabling rapid and cost-effective downstream estimation of binding affinities. We apply our approach to two targets, tankyrase 2 (TNKS2) and the 3C-like protease (3CLPro), using datasets of 10,000 molecules for each system. Evaluation on an external validation set of approximately 3,000 molecules demonstrates the method’s ability to generalize to previously unseen datasets, aligning with the practical needs of high-throughput drug discovery pipelines that iteratively evaluate new molecules. Our method substantially reduces the need for costly calculations by efficiently filtering out poses unlikely to yield favorable binding affinities upon relaxation.

Related

  1. Beyond Organic Molecules: MetalLipoDB, a Lipophilicity Dataset for Metal Complexes and Machine Learning Benchmarks
    Lev Krasnov, Dmitriy Malikov, Marina A. Kiseleva +3
    ChemRxivPreprintCC BYGeneral chemistryChemRxiv
  2. A physics-driven machine learning framework for predicting XRD patterns, magnetization, and magnetocaloric response in rare-earth-doped double perovskite oxides
    Sana Meftah, Emna Ammar Elhadjamor, Souhir Bouzidi +4
    RSC AdvancesPeer-reviewedCC BY-NCMaterials & polymersHAL
  3. Machine Learning-Enhanced Solid-State Nanopore Analysis for Single-Molecule Discrimination of Albumin and Aptamer-Bound Glycated Albumin Complexes
    Harit Pitakjakpipop, Ibrar Alam, Kiatnida Treerattrakoon +13
    ACS OmegaPeer-reviewedCC BY-NC-ND 4.0Full textGeneral chemistryCrossref
  4. Sensitized near-infrared luminescence in sulfur-coordinated lanthanide complexes for machine learning-enhanced thermometry
    Mikko Rautiainen, Emily Andreato, Essi Barkas +7
    ChemRxivPreprintCC BYFull textMaterials & polymersChemRxiv
  5. Integrative network toxicology and machine learning reveal the potential role of 5-hydroxymethylfurfural in exacerbating pulmonary fibrosis
    Qinghuan Yang, Fuli Li, Jiani Li +1
    PLOS OnePeer-reviewedCC BYFull textMolecular biology & gene regulationDevelopmental biologyPLOSEurope PMCPMC

Cite

BibTeXRIS
BibTeX
@misc{bhati2026integrative,
  title = {{Integrative Physics and Machine Learning-Based Optimal Binding Pose Generator for Protein-Ligand Complexes}},
  author = {Agastya P Bhati and Jinyu Dong and Sean Black and Shunzhou Wan and Vineeth Gutta and Xibei Zhang and Mateusz Bieniek and Darren V. S. Green and Sunita Chandrasekaran and Eric Stahlberg and Peter V. Coveney},
  howpublished = {ChemRxiv},
  year = {2026},
  month = sep,
  doi = {10.26434/chemrxiv.15009330/v1},
  url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009330/v1},
  note = {CC BY},
}
RIS
TY  - UNPB
TI  - Integrative Physics and Machine Learning-Based Optimal Binding Pose Generator for Protein-Ligand Complexes
AU  - Agastya P Bhati
AU  - Jinyu Dong
AU  - Sean Black
AU  - Shunzhou Wan
AU  - Vineeth Gutta
AU  - Xibei Zhang
AU  - Mateusz Bieniek
AU  - Darren V. S. Green
AU  - Sunita Chandrasekaran
AU  - Eric Stahlberg
AU  - Peter V. Coveney
PY  - 2026
DA  - 2026/09/23
PB  - ChemRxiv
DO  - 10.26434/chemrxiv.15009330/v1
UR  - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009330/v1
AB  - Accurately identifying the optimal protein–ligand binding pose is crucial for reliable binding affinity prediction, as molecular dynamics–based approaches are highly sensitive to the initial structure. However, there is currently no efficient way to predict such poses. Indeed, we assess the capabilities of three proposed AI-based software solutions (AlphaFold3, Boltz and RoseTTAFold All-Atom) and demonstrate their substantial shortcomings in this context. Here, we present a combined physics-and machine-learning-based method that predicts optimal protein-ligand poses in their relaxed states, enabling rapid and cost-effective downstream estimation of binding affinities. We apply our approach to two targets, tankyrase 2 (TNKS2) and the 3C-like protease (3CLPro), using datasets of 10,000 molecules for each system. Evaluation on an external validation set of approximately 3,000 molecules demonstrates the method’s ability to generalize to previously unseen datasets, aligning with the practical needs of high-throughput drug discovery pipelines that iteratively evaluate new molecules. Our method substantially reduces the need for costly calculations by efficiently filtering out poses unlikely to yield favorable binding affinities upon relaxation.
ER  -