AbICL Enhances Antibody Affinity Ranking with In-Context Learning

The discovery of therapeutic antibodies relies heavily on accurately ranking antibody candidates by their binding affinity to specific antigens. Current methods often treat affinity comparisons in isolation, overlooking valuable contextual information from other labeled comparisons. This research introduces AbICL, an In-Context Learning (ICL) framework designed to address this limitation by exploiting available experimental affinity data to infer antigen-specific ranking patterns. AbICL combines a pre-trained structural encoder with a context ranking head and employs an episodic meta-training strategy. This allows the model to adapt to new target antigens at test time by leveraging support demonstrations without requiring gradient updates. Experiments on the AbRank benchmark show that AbICL consistently outperforms existing ranking baselines across various data splits and evaluation metrics. The framework proves particularly effective in challenging scenarios involving distribution shifts and fine-grained affinity discrimination, highlighting the significant potential of ICL for advancing therapeutic antibody discovery.