Hallucinating structure-conditioned antibody libraries for target-specific binders

• DOI: 10.1101/2022.06.06.494991
• Publication Date: 2022-06-08T16:15:22Z
• Authors: Sai Pooja Mahajan; Jeffrey A. Ruffolo; Rahel Frick; Jeffrey J. Gray
• Keywords: Rational design; Protein Engineering

Abstract Antibodies are widely developed and used as therapeutics to treat cancer, infectious disease, inflammation. During development, initial leads routinely undergo additional engineering increase their target affinity. Experimental methods for affinity maturation expensive, laborious, time-consuming rarely allow the efficient exploration of relevant design space. Deep learning (DL) models transforming field protein design. While several DL-based have shown promise, antibody problem is distinct, specialized desirable. Inspired by hallucination frameworks that leverage accurate structure prediction DL models, we propose F v Hallucinator designing sequences, especially CDR loops, conditioned on an structure. Such a strategy generates targeted libraries retain conformation binder thereby mode binding epitope antigen. On benchmark set 60 antibodies, sequences resembling natural CDRs recapitulates perplexity canonical clusters. Furthermore, designs amino acid substitutions at V H -V L interface enriched in human repertoires therapeutic antibodies. We pipeline screens obtain library binders antigen interest. apply this H3 Trastuzumab-HER2 complex generate silico improving upon interfacial properties original antibody. Thus, enables generation inexpensive, diverse, maturation.