Antibodies have become a major focus of therapeutic and diagnostic agents. Compared to small molecule drugs, the key advantages of using antibodies are based on more complicated interactions across antibody-antigen interface residues, which demand more systematic approaches to improve properties such as target binding affinity and specificity. Although the improvement in antibody-antigen binding affinity has been widely studied by introducing a single-point mutation at a time, replacing individual residues may not be enough to yield viable affinity changes and may lead to unwanted physicochemical property changes. There have been many efforts in developing in silico tools to guide rational antibody engineering, but most approaches are inaccurate when applied to antibody design, and have largely been limited by analysing single point mutations at a time.
Using a novel way to computationally encode the antibody-antigen interaction interface, using graph-based signature, we developed a highly accurate tool for analysing the consequence of point mutations on antigen binding affinity. In terms of predicting binding affinity changes upon single point mutations, the model outperformed available tools showing a Pearson's correlation of 0.76 on 10-fold cross validation, 0.67 on experimental blind tests and 0.72 on homology model blind tests. To overcome the gap in understanding whether single point mutation datasets can be used to predict binding affinity changes of multiple mutations, we curated a dataset of 334 mutations in antibodies with experimentally determined changes in binding affinity (100 stabilising, 200 destabilising). Our approach, outperformed other available tools on both 116 double/triple mutation dataset and 334 constructs showing Pearson and Spearman correlations of 0.90 and 0.81 and 0.73 and 0.56, respectively.
We have implemented our new approaches as web-servers that enables rapid and deep evaluations of specific mutations or systematic exploration of all possible combinations of single or a set of double and triple mutations across antibody-antigen interface residues. mCSM-AB2 and mmCSM-AB will help to guide rational antibody engineering by analysing the effects of introducing mutations. This user-friendly web-servers are freely available at http://biosig.unimelb.edu.au/mcsm_ab2 and http://biosig.unimelb.edu.au/mmcsm_ab.