Ab-VS: Evaluating Large Language Models for Virtual Antibody Screening via Antibody-Antigen Interaction Prediction

Abstract

We present Ab-VS-Bench, a new benchmark for evaluating large language models (LLMs) on antibody virtual screening (VS) tasks through natural language instructions. Unlike prior efforts that focus on structural annotation, binding prediction, or developability using frozen antibody models, Ab-VS-Bench targets the end-to-end VS workflow and leverages LLMs instruction-following capabilities. The benchmark comprises three core tasks inspired by small-molecule VS: (1) Scoring--predicting antibody-antigen binding affinity, (2) Ranking--ordering antibodies by affinity or thermostability, and (3) Screening--identifying high-affinity binders from large antibody libraries. We convert experimental datasets into instruction-tuning format and evaluate multiple model variants, including zero-shot LLMs, instruction-finetuned LLMs, and multimodal models enhanced with antibody-specific embeddings. Ab-VS-Bench provides a unified framework to benchmark LLMs for antibody discovery, aiming to catalyze progress in language-guided biotherapeutic design.