abCAN: a practical and novel attention network for predicting mutant antibody affinity.

Area of Science:

• Computational Biology
• Structural Biology
• Immunoinformatics

Background:

• Accurate prediction of mutation effects on antibody-antigen interactions is crucial for developing effective antibody-based therapeutics.
• Existing methods often struggle to capture the complex interplay between structural and sequential features influencing binding affinity.

Purpose of the Study:

• To develop a novel deep learning model, abCAN, for predicting changes in antibody-antigen binding affinity due to mutations.
• To establish a new state-of-the-art benchmark for mutation effect prediction in antibody-antigen systems.

Main Methods:

• Developed abCAN, an attention network utilizing Progressive Encoding to integrate structural, residue-level, and sequential information.
• Employed attention mechanisms to prioritize interface residues.
• Trained the model on antibody-antigen complex structures and mutation data to predict binding affinity changes.

Main Results:

• abCAN achieved a root-mean-square error of 1.460 (kcal/mol) on a benchmark test set.
• A Pearson correlation coefficient of 0.731 was obtained, demonstrating high prediction accuracy.
• The model sets a new state-of-the-art performance for predicting mutation-induced affinity changes.

Conclusions:

• abCAN provides a practical and accurate method for predicting mutation effects on antibody-antigen binding affinity.
• The Progressive Encoding approach effectively captures complex interactions, advancing antibody engineering and drug design.
• The developed model and associated resources are publicly available to facilitate further research.