ALLM-Ab: Active Learning-Driven Antibody Optimization Using Fine-Tuned Protein Language Models

Area of Science:

• Biotechnology
• Computational Biology
• Immunology

Background:

• Antibody engineering faces challenges in optimizing binding affinity while preserving developability.
• Protein language models offer potential for predicting antibody sequence fitness.

Purpose of the Study:

• To introduce ALLM-Ab, an active learning framework for accelerated antibody sequence optimization.
• To leverage fine-tuned protein language models for efficient candidate sequence generation.
• To integrate antibody developability metrics into the optimization process.

Main Methods:

• Utilized parameter-efficient fine-tuning (low-rank adaptation) of protein language models.
• Employed a learning-to-rank strategy for mutant fitness assessment.
• Integrated a multiobjective optimization scheme including developability metrics.
• Validated using deep mutational scanning data and online active learning trials.

Main Results:

• ALLM-Ab accurately assesses mutant fitness and generates candidate sequences efficiently.
• The framework successfully balances improved binding affinity with therapeutic antibody-like properties.
• Demonstrated expedited discovery of high-affinity antibody variants compared to baseline methods.
• Preserved critical antibody developability metrics during optimization.

Conclusions:

• ALLM-Ab provides an efficient and reliable strategy for antibody design.
• The framework has the potential to significantly reduce therapeutic development costs.
• This approach advances the field of antibody engineering and drug discovery.