Active Learning framework for Antibody Binding Affinity Prediction

Abstract

Antibody engineering campaigns must prioritise variants from large combinatorial libraries under limited assay budgets. We present a Gaussian process (GP) active-learning benchmark for antibody-antigen binding affinity prediction across eight public AbBiBench assays spanning influenza hemagglutinin antibodies CR6261 and CR9114 and AAYL-series scFv libraries against the SARS-CoV-2 Spike HR2 peptide. The framework embeds variable heavy and light (VH/VL) sequences with ESM-2 650M, trains GP surrogates over compressed protein-language-model representations, and evaluates kernel, acquisition, representation, classical featurisation, and labelling-budget choices under a unified budget of 10\% of each full pool with a label-blind seed and a fixed held-out evaluation split. Across the benchmark, Tanimoto+UCB has the highest mean Recall@2\% point estimate among tested kernel-acquisition configurations, achieving \textbf{57.8\%} mean Recall@2\% across five random seeds (68.3\% when scored against the acquisition-eligible pool), compared with 9.6\% for random selection -- a gap that is significant by a paired Wilcoxon test across datasets ($p=0.0078$); its point estimate is not significantly different from RQ+UCB or Tanimoto+EI at five seeds. We adopt the continuous Tanimoto kernel of \citet{gessner2024}. The same configuration recovers the largest fractions of top binders in the influenza libraries (74.2\% on \texttt{3gbn\_h1}, 82.7\% on \texttt{3gbn\_h9}, and 84.9\% on \texttt{4fqi\_h3}) while exposing harder AAYL combinatorial landscapes where recall remains lower and more variable. An additional three-seed representation ablation, run under the original protocol and not repeated after the evaluation corrections, ranked ESM-2 650M highest (mean Recall@2\% 0.673 $\pm$ 0.28) among ProtBert, AntiBERTy, AbLang2, ProGen2, and classical encodings. On BindingGYM, the same configuration matches or exceeds ALLM-Ab on two of three datasets using the 100-sample held-out test metric. Structure-mapped Integrated Gradients and KernelSHAP analyses recover CDR-region attribution patterns consistent with contact-derived paratope geometry, though threshold-free (AUROC) analysis shows this qualitative agreement is not matched by strong residue-level discrimination. These results support Tanimoto-kernel GP active learning as a simple, reproducible strategy for label-efficient antibody affinity optimisation with qualitative structural validation.