Modeling antibody recognition across SARS-CoV-2 and SARS-CoV-1 using epitope-informed transfer learning

Antiviral antibody discovery is constrained by the enormous combinatorial space of antibody paratopes and viral epitopes, while novel approaches to search for these vast possibilities such as machine learning models for antibody affinity prediction typically require large target-specific datasets that are rarely available for emerging pathogens. To address this problem, we investigated the ability of transfer learning to reduce the required training data and still create a robust predictive model of antibody affinities. We show that antibody-affinity prediction can be locally transferred from a data-rich viral target to a related data-limited viral target by combining transfer learning, epitope-centered antigen encoding, antibody CDR descriptors, and targeted experimental validation. Using SARS-CoV-2 receptor-binding-domain (RBD) interaction data as the source domain, we developed dual-encoder neural-network models that capture functional determinants of antibody recognition and supports prediction of SARS-CoV-1 variant:antibody affinities despite limited direct training data. High-throughput ELISA measurements generated compact validation datasets spanning heavy-chain-only antibodies (HCAbs) and conventional antibodies, enabling iterative model refinement through similarity-guided, model-informed data expansion. Across expanded withheld-antibody evaluations, the final models showed improved but antibody-dependent predictive performance, with mean Pearson correlations of 0.71 ± 0.12 for SARS-CoV-2 and 0.63 ± 0.09 for SARS-CoV-1, and best-performing holdouts reaching 0.85 and 0.78, respectively. These results demonstrate that existing immunological datasets can support feature-constrained transfer learning for data-efficient prioritization of antibody:antigen interactions across closely related Sarbecoviruses, particularly when conserved epitope regions can be aligned and limited target-specific measurements are available for calibration.