Germline-aware deep learning models and benchmarks for predicting antibody VH-VL pairing.

Area of Science:

• Immunology
• Computational Biology
• Bioinformatics

Background:

• Antibody variable heavy (VH) and variable light (VL) chain pairing dictates antibody function.
• Experimental identification of productive VH-VL pairs is resource-intensive.
• Computational prediction methods are needed to accelerate antibody engineering.

Purpose of the Study:

• To develop a comprehensive framework for predicting compatible VH-VL pairs.
• To create a benchmark dataset with natural and synthetic negative pairs.
• To evaluate deep learning models trained with different negative sampling strategies.

Main Methods:

• Development of a novel benchmark dataset including natural and synthetic VH-VL pairs.
• Training and evaluation of three deep learning models using distinct negative sampling strategies (random, V-gene mismatching, V(D)J germline mismatching).
• Implementation of a BERT-based model for discriminating natural from synthetic pairs.

Main Results:

• The BERT-based model achieved over 90% accuracy in distinguishing natural from synthetic VH-VL pairs.
• V(D)J germline-informed negative sampling significantly enhanced model generalization.
• The developed framework provides reproducible baselines for computational antibody engineering.

Conclusions:

• A robust computational framework and benchmark dataset for VH-VL pairing prediction have been established.
• V(D)J germline information is critical for improving the biological interpretability and accuracy of predictive models.
• This work facilitates the development of efficient computational tools for antibody engineering.