Distilling Causal Signals for One-Shot Directed Evolution of Antibodies

Improving antibody binding to an antigen without antibody-antigen complex structures or antigen-specific training data is a central challenge in therapeutic protein design. We introduce AFFINITYENHANCER, a framework for one-shot antibody affinity improvement with strong generalization: given a single lead sequence, we propose variants that increase affinity without fine-tuning on the lead and without using antigen information, epitope/paratope labels, or the lead's structure in complex with the antigen. During training, AFFINITYENHANCER leverages a panantigen dataset of diverse binding environments (antigens) and constructs paired examples of related sequences with higher vs. lower measured binding. A shared, structure-aware module learns to transform low-affinity sequences toward highaffinity ones, distilling consistent, causal features associated with improved binding across environments. By combining pretrained sequence-structure embeddings with a sequence decoder, AFFINITYENHANCER generalizes to entirely unseen antibody seeds. Across multiple held-out internal and public leads, AFFINITYEN-HANCER concentrates mutations on the rim of the paratope, outperforms existing structure-conditioned and inpainting baselines, and achieves substantial in silico affinity gains in true one-shot experiments, despite never observing antigen-specific data at test time.[ https://github.com/prescient-design/AffinityEnhancer ] Published as a conference paper at ICLR 2026 250 residues). As a consequence, the resulting sets of designs can be suboptimal and fail to identify sufficient number of antibodies with the desired potency and drug-like properties. Figure 1 : One-shot affinity maturation of antibodies with AFFINITYENHANCER. A) The goal is to implicitly learn modes of affinity maturation by pairing a lower affinity antibody with a higher affinity one. B) Matched datasets are obtained by pairing antibodies against the same target/antigen from the SKEMPI 2.0 database. C) Architecture for AFFINITYENHANCER. D) Inference and validation pipeline for held-out-seed to determine whether sampled sequences are binders or not. Computational affinity maturation with machine learning offers an accelerated alternative to random or directed mutagenesis. However, the one-shot setting-where a model must propose improved variants from a single lead sequence without antigen context or fine-tuning-poses a key generalization challenge: the lead may be far from the training distribution in sequence and structural features. This challenge is compounded by the limited availability and diversity of paired antibody-antigen structures and affinity measurements, which impedes robust transfer to unseen targets (Hummer et al., 2023) . To bypass the challenges associated with explicitly modeling affinity, Tagasovska et al. (2024) proposed Property Enhancer (PropEn), a property-agnostic model which utilizes data matching to implicitly learn the direction of the gradient for a property of interest with the goal of proposing new optimized designs. It was previously demonstrated that this approach works for a range of tasks, including affinity maturation of antibodies. However, its effectiveness was only demonstrated in sequence-based models and in cases where a few hundred sequences related to the lead molecule we wish to optimize are already available in the training data, hence, not suitable to one-shot scenarios. In this work, we propose AFFINITYENHANCER, a model that goes beyond the PropEn framework, namely to the one-shot affinity maturation setup by leveraging structure information and introducing