Fig. 1: Overview of DML to guide the selection of antibodies targeting SARS-CoV-2 RBD.

a, For the construction of the high-distance Omicron BA.1 RBD library, short ssODNs are designed to possess mutations. The fragments are assembled into full-length RBD sequences using GGA and transformed into yeast. The resulting library is screened for antibody binding and escape using FACS. b, The sorted high-distance RBD variant library is deep sequenced, and the data are used to train ensemble deep learning models to predict ACE2 and antibody binding or non-binding (escape). Deep learning models are used to predict the breadth of antibody combinations as well as their binding to synthetic RBD variants and lineages. mAb, monoclonal antibody.