Towards the accurate modelling of antibody-antigen complexes from sequence using machine learning and...

Area of Science:

• Computational biology
• Structural biology
• Immunoinformatics

Background:

• Accurate antibody-antigen complex modeling is crucial for therapeutic antibody design.
• Experimental structures are often unavailable, necessitating computational approaches.
• Machine learning advances enable accurate prediction of antibody and antigen structures.

Purpose of the Study:

• To analyze the efficacy of protein-protein docking tools using machine learning-generated input structures.
• To evaluate information-driven docking protocols for therapeutic antibody design.

Main Methods:

• Utilized machine learning tools to generate an ensemble of antibody structures.
• Employed AlphaFold2 for predicting antigen structures.
• Applied HADDOCK for information-driven protein-protein docking.

Main Results:

• HADDOCK successfully generated accurate antibody-antigen complex models using ML-predicted structures.
• Targeted docking with knowledge of complementary determining regions and epitope information improved model quality.
• The protocol achieved high-quality models with reduced sampling, outperforming the ZDOCK baseline.

Conclusions:

• Machine learning-generated structures are viable inputs for protein-protein docking.
• Information-driven docking with HADDOCK offers a computationally efficient and accurate approach for therapeutic antibody design.
• This workflow enhances the design of novel antibody therapeutics.