Currently there are no successful bioinformatics methods to predict antibody:antigen binding sites. This is at least partially due to the fact that there is a limited amount of large scale mapping of antibody binding sites available. Antibody cross-blocking is a method in which two antibodies compete against each other for binding to the same antigen, which has been used for decades to identify antibodies binding to the same region and thereby provides a mapping of antibody binding sites at low resolution. With high-throughput antibody production becoming mainstream it becomes possible to scale up cross-blocking experiments and enumerate all binding sites for an antigen. The analysis of antibody cross-blocking has typically been done manually leading to inconsistencies in defining binding sites. To apply cross-blocking analysis to larger datasets, a robust computational approach is needed to standardize the analysis. Here we show a method incorporating non-negative matrix factorization, k-means clustering, and hierarchical clustering to analyze cross-blocking data and extract both (1) antibody groups that show identical binding patterns and (2) the number of distinct antigenic sites occupied by the antibodies. We apply our approach to previously published experimental data, and compare our identified number of binding sites and antibody groups to the published analyses that often do not make this distinction.
