Fig. 1: Overview of antibody library sorting, deep sequencing, and machine learning methods used to co-optimize the affinity and specificity of a therapeutic antibody.

A clinical-stage antibody (emibetuzumab) was mutated at eight positions in three heavy chain CDRs, and the antibody libraries were sorted using yeast surface display and magnetic- and fluorescence-activated cell sorting for high affinity and high and low levels of non-specific binding. The sorted libraries were deep sequenced, and the resulting antibody sequences were used to train models for predicting metrics correlated with antibody affinity and specificity (non-specific binding) using different types of molecular features. These features included antibody sequences encoded as binary vectors, physicochemical features, and deep learning features. The resulting models were used not only to predict the classification of antibody affinity and specificity (e.g., high or low affinity), but also continuous metrics correlated with each property to predict intraclass variability (e.g., high vs. very high affinity). The model predictions were also used to identify antibody mutants in the library at the Pareto frontier that maximize antibody affinity to different extents while minimizing tradeoffs due to reduced specificity (i.e., increased non-specific binding). Some of the models, which generalized to novel mutational space, were used to identify antibodies with even greater improvements in affinity and specificity than was possible in the experimentally sorted libraries.