AbDist: a lightweight, distance-based model for antibody affinity prediction as an interpretable benchmark for...

Area of Science:

• Biochemistry
• Computational Biology
• Immunology

Background:

• Complex models for antibody affinity prediction exist.
• Simple distance-based models show promise in related fields like T-cell receptor epitope prediction.
• These simpler models offer advantages in interpretability and computational efficiency.

Purpose of the Study:

• To develop a novel distance-based model for antibody affinity prediction.
• To assess the effectiveness of local sequence environments for featurization.
• To provide an interpretable and efficient tool for antibody engineering.

Main Methods:

• Developed AbDist, a distance-based computational model.
• Utilized fragments around mutation sites for sequence featurization.
• Applied AbDist to classification and regression tasks on public datasets.

Main Results:

• AbDist performance matches state-of-the-art machine learning models.
• The model demonstrates that local sequence environments are sufficient for effective featurization.
• AbDist is computationally efficient and interpretable.

Conclusions:

• AbDist is a viable, efficient, and interpretable alternative for antibody affinity prediction.
• The model is well-suited for data-sparse, early-stage antibody engineering.
• It shares limitations in out-of-distribution generalization with current models.