Combining AI Structure Prediction and Integrative Modeling for Nanobody-Antigen Complexes
Synopsis
This study evaluates state-of-the-art machine-learning-based methods for nanobody structure prediction and benchmarks various HADDOCK workflows for modeling nanobody-antigen interactions across different input nanobody ensembles and information scenarios, proposing an ensemble docking pipeline that starts from nanobody structural models predicted by AlphaFold2 and ImmuneBuilder and, provided some epitope information is available, achieves higher success rates than the AlphaFold baseline on all generated models.
Graphical representation of the nanobody-antigen complex prediction benchmark. First, nanobody conformations are generated using AlphaFold (AF2), AlphaFold2-Multimer (AF2-multi), and ImmuneBuilder (IB). They are then clustered based on the similarity of the CDR3 loop and passed to HADDOCK for ensemble-based information-driven docking. Three different information scenarios are benchmarked: True Interface (the true interface residues), Loose Interface (using the CDR loops and a broad area surrounding the real epitope), and Two-hit Interface (using the CDR loops and two main residues from the antigen obtained from alanine scanning). Active residues are represented in red and passive residues in green. Those residues are used to define ambiguous interaction restraints to guide the docking. The HADDOCK docking protocol consists of the following steps: topoaa (topology generation and building of any missing atoms), rigidbody (rigid body docking), seletop (selection of the top 200 ranked models), flexref (flexible refinement), emref (energy minimization), and clustfcc (clustering based on the fraction of common contacts). The results from the benchmark data set are evaluated based on CAPRI criteria, comparing the docking model with the experimental structure.
PubMedInterpretation
The study systematically evaluates machine-learning-based methods for nanobody structure prediction and benchmarks multiple HADDOCK workflows for nanobody-antigen complexes across different input nanobody ensembles and information scenarios. Prior work lacked a systematic comparison of nanobody-antigen modeling across different input ensembles and information conditions; this work places structure prediction methods and docking workflows within a single evaluation framework. Based on benchmarking of multiple HADDOCK workflows with different input nanobody ensembles and information scenarios, constituting a methodological comparison study.
The study proposes an ensemble docking pipeline that starts from nanobody structural models predicted by AlphaFold2 and ImmuneBuilder and achieves high success rates. The pipeline combines AI structure prediction with physics-based information-driven docking, offering an actionable integrative route for nanobody-antigen complex modeling. The ensemble docking pipeline using predicted structural models as input reports high success rates, though specific values are not given in the text.
Provided some epitope information is available, the pipeline achieves higher success rates than the AlphaFold baseline on all generated models. This indicates that the availability of epitope information is the condition under which the pipeline outperforms the AlphaFold baseline, offering a reference for information requirements in practical modeling. The conclusion is conditioned on epitope information being available and covers all generated models, but the text provides no specific success-rate values or sample sizes.
Perspective
This work targets three-dimensional modeling of nanobody-antigen complexes and applies to settings where AlphaFold2 or ImmuneBuilder predicted nanobody structural models can be obtained and partial epitope information is available; the proposed ensemble docking pipeline uses HADDOCK as the docking engine, and the evaluation is limited to the tested input nanobody ensembles and information scenarios.
The text is at the abstract level and does not provide specific success-rate values, the number of test systems, sample sizes, or statistical details, nor does it specify the form and source of the epitope information; readers who need to judge the pipeline's robustness across different antigen types or information conditions would still need to consult the figures and data in the original paper.