Multidimensional maturation of antibody variable domains with machine-learning assistance.

In antibody development, a mutagenesis approach has been widely used to improve affinity, but such mutations often compromise biophysical properties. Here, we combined molecular evolution with machine learning to simultaneously improve affinity and expression level of camelid heavy-chain antibody variable domains (VHHs). Using phage display and deep sequencing, we selected five residues in an anti-SARS-CoV-2 VHH for affinity maturation. We constructed training data using experimentally measured expression levels and target affinities of 117 variants with randomized residues. Machine-learning-predicted top-rank variants showed improved expression level and affinity compared to variants in the training data. Several variants achieved 50-70-fold stronger affinities in the pico-molar range and 4-5-fold higher expression level than wild-type. Furthermore, one variant showed 9.5°C improvement in thermal stability. These results highlight the utility of machine learning-assisted molecular evolution as a strategy for multidimensional optimization of antibody properties.