ACCELERATING ANTIBODY DESIGN WITH ACTIVE LEARNING

Abstract

Discovering therapeutic antibody starts by screening antibody library of phage-displayed, transgenic mouse or human B cells. The coverage of those kinds of libraries in the entire antibody sequence space is small; thus, the result highly depends on the quality of the library. Exploring sequence space by mutating a template antibody is also impossible to even with the state-of-the-art screening methods because of the cost. Deep learning helped with its pattern recognition nature to predict target binding, which is only applied to HCDR3 because the number of data deep learning needs increases exponentially. We construct a sequence generation model with transfer learning and active learning to leverage deep learning even in data deficiency. With only six thousands data, the generative model finds nine binding antibody sequences at least per antigen with novel HCDR3.