A comparison of antibody-antigen complex sequence-to-structure prediction methods and their systematic biases

Area of Science:

• Structural Biology
• Immunology
• Computational Biology

Background:

• Accurate prediction of antibody-antigen complex structures is vital for understanding immune responses and developing antibody therapeutics.
• Recent advances in machine learning (ML) have significantly improved protein-protein interaction (PPI) prediction.

Purpose of the Study:

• To evaluate and compare six distinct methods for predicting antibody-antigen complex structures from amino acid sequences.
• To identify the most effective current computational approaches and understand their limitations.

Main Methods:

• Comparative analysis of six methods: AlphaFold-Multimer, RoseTTAFold, ClusPro, SnugDock, and AbAdapt.
• Evaluation based on prediction accuracy of antibody-antigen complex structures from sequence data.

Main Results:

• AlphaFold-Multimer demonstrated superior performance compared to other evaluated methods.
• Lower quality AlphaFold-Multimer models exhibited structural biases, with fewer common tertiary motifs (TERMs) at the interface compared to non-antibody structures in the Protein Data Bank (PDB).
• Higher quality models showed more PDB-like TERMs at the antibody-antigen interface.

Conclusions:

• While AlphaFold-Multimer is a leading method, significant improvements in predicting antibody-antigen complex structures are still needed.
• The scarcity of interfacial geometry data in the PDB may currently limit the effectiveness of ML-based prediction methods for antibody-antigen interactions.
• The commonness of interfacial TERMs correlates with prediction performance, highlighting a potential area for future method development.