Separating selection from mutation in antibody language models.

Area of Science:

• Immunoinformatics
• Computational Biology
• Protein Engineering

Background:

• Antibodies are crucial for adaptive immunity, encoded by V(D)J recombination and shaped by mutation and selection.
• Current antibody language models focus on amino acid sequences, implicitly modeling nucleotide mutations, which hinders accurate functional prediction.
• This implicit modeling degrades performance in predicting the functional impact of antibody mutations.

Purpose of the Study:

• To develop a novel framework that explicitly separates nucleotide-level mutation processes from amino acid selection effects in antibody modeling.
• To improve the accuracy of predicting functional properties of antibodies by exclusively quantifying selection effects.
• To create a more efficient and interpretable antibody language model.

Main Methods:

• A Deep Amino Acid Selection Model (DASM) framework was devised.
• The model explicitly factors out the nucleotide-level mutation process.
• Selection effects were fitted as a separate term from mutation processes.

Main Results:

• The DASM framework substantially improved performance on standard antibody functional prediction benchmarks.
• The model exclusively quantifies functional effects by separating selection from mutation.
• The DASM is an order of magnitude smaller and significantly faster to evaluate than existing models.

Conclusions:

• Explicitly modeling selection effects separate from mutation processes enhances antibody functional prediction.
• The DASM offers a more accurate, efficient, and interpretable approach to antibody engineering.
• This framework advances computational methods for understanding and designing antibody therapeutics.