Predicting viral sensitivity to antibodies using genetic sequences and antibody similarities

For genetically variable pathogens such as human immunodeficiency virus (HIV)-1, individual viral isolates can differ dramatically in their sensitivity to antibodies. The ability to predict which viruses will be sensitive and which will be resistant to a specific antibody could aid in the design of antibody therapies and help illuminate resistance evolution. Due to the enormous number of possible combinations, it is not possible to experimentally measure neutralization values for all pairs of viruses and antibodies. Here, we developed a simple and interpretable method called grouped neutralization learning (GNL) to predict neutralization values by leveraging viral genetic sequences and similarities in neutralization profiles between antibodies. The trained model is interpretable and can identify key mutations that impact viral sensitivity. Our method compares favorably to state-of-the-art approaches and is robust to model parameter assumptions. GNL can predict neutralization values for viral sequences without observed neutralization measurements, an important capability for assessing antibody coverage in populations whose viral diversity is genetically characterized. We also demonstrate that GNL can successfully transfer knowledge between independent data sets, allowing rapid estimates of viral sensitivity based on prior knowledge.Author summary: Antibodies play a central role in controlling viral infections, yet viral strains can differ widely in their sensitivity to a given antibody. Quantifying this sensitivity is critical for understanding immune escape, interpreting experimental and clinical studies, and designing effective antibody-based therapies. However, virus neutralization assays are labor-intensive and inherently sparse: only a small fraction of the vast space of possible virus-antibody combinations can be measured experimentally. Here, we present a computational framework, grouped neutralization learning (GNL), for predicting viral sensitivity across large virus-antibody panels using incomplete neutralization data and viral genetic sequences. The central idea of GNL is to share information across antibodies with similar neutralization profiles, allowing sparsely measured antibodies to benefit from related, more extensively characterized ones. Building on prior insights that neutralization data exhibit strong low-rank structure, we integrate antibody-level information sharing with sequence-based prediction and low-rank matrix refinement to produce robust and accurate neutralization predictions. Across large HIV-1 neutralization datasets, GNL performs favorably compared to state-of-the-art methods, particularly in when data is limited. Our model is interpretable, identifying specific viral mutations that strongly influence antibody sensitivity and recovering patterns consistent with known resistance mechanisms. Importantly, GNL can generate sensitivity estimates even for viral sequences without any measured neutralization data. Collectively, this work provides a scalable and interpretable framework for mapping virus-antibody interactions from partial data. While developed and validated using HIV-1, our approach is broadly applicable to other rapidly evolving pathogens for which genetic sequences and sparse neutralization measurements are available.