EpiBench: Can LLMs Understand Epitopes for Antibody Drug Discovery?

Epitopes—the specific regions on an antigen where an antibody binds—are critical for determining the therapeutic efficacy, functional blockade, and escape resistance of antibody drugs. While large language models (LLMs) have demonstrated impressive biomedical reasoning, it remains unclear whether they can infer these complex binding interfaces directly from primary protein sequences. This paper introduces EpiBench, a new benchmark designed to evaluate how well LLMs handle the sequence-based reasoning required across the antibody development lifecycle.

EpiBench consists of 1,609 curated samples derived from structural data, functional B-cell assays, and deep mutational scanning. Unlike previous benchmarks that focus on isolated prediction tasks, EpiBench simulates a multi-stage drug discovery workflow. It covers five key tasks:

1. Targetable Region Discovery: Identifying potential binding sites on an antigen.
2. Antibody-Conditioned Epitope Identification: Linking a specific antibody to its binding region.
3. Epitope Binning: Determining if two antibodies compete for the same binding site.
4. Functional Epitope Assessment: Distinguishing epitopes that drive therapeutic function from those that merely bind.
5. Antibody Escape Assessment: Predicting whether specific mutations allow an antigen to evade antibody recognition.

The benchmark uses a closed-book, sequence-only format, stripping away identifiers like PDB IDs or protein names to ensure models are reasoning from biological sequences rather than memorized metadata.

The authors evaluated nine general-purpose LLMs against these tasks. Results indicate that while LLMs can identify broadly relevant antigen regions, they struggle with the precision required for drug discovery. Specifically, models often fail to ground their predictions in the specific antibody-antigen context, leading to poor performance in tasks requiring functional or escape-related reasoning. The study reveals that "reasoning" capabilities in current LLMs do not consistently translate to better performance on these biological tasks, suggesting that current models lack the deep, sequence-aware understanding of protein interfaces necessary for reliable, autonomous antibody design.