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

Abstract

Many complex models for antibody affinity prediction have been developed and successfully deployed. Recent results for T-cell receptor epitope prediction have shown, that even simple distance-based models can achieve a similar performance while requiring less parameters, being more easily interpretable and faster to compute. Encouraged by these results AbDist, a new distance-based model, was developed for antibody affinity prediction. It uses fragments around mutation sites to calculate distances between antibody sequences, demonstrating that a local environment alone suffices as an effective featurization. AbDist was used to perform classification and regression tasks on multiple disjunct public datasets. Its performance matches state-of-the-art machine-learning (ML) models. AbDist is interpretable, computationally efficient, and well suited for data-sparse, early-stage antibody engineering workflows, while sharing the limited out-of-distribution generalization common to current models. AbDist is available as an open-source, publicly accessible tool.