(53c) Hallucinating Native-like Antibodies with Deep Learning

Antibodies recognize and bind an extremely diverse repertoire of antigens via 6 hypervariable loop regions known as the complementarity determining regions or the CDRs. A common goal in antibody engineering to generate a library of CDRs enriched in binders also known as affinity maturation. Experimental methods for affinity maturation are expensive, laborious, and time-consuming and rarely allow the efficient exploration of the full design space. Conventional computational methods are relatively inexpensive but marred by inaccurate scoring functions and slow and inefficient sampling of the full combinatorial sequence space. Deep learning (DL) models are transforming the field of protein structure-prediction, engineering, and design.¹ While several DL-based protein design methods have shown promise, specialized models for antibody design are highly desirable.² Our approach aims to fully leverage structural information which is becoming increasingly abundant in the era of highly accurate structure prediction models such as AlphaFold³ (all proteins, multimers) and DeepAb⁴ (antibodies only). Inspired by the hallucination frameworks that specifically leverage such structure prediction DL models, we propose the F_vHallucinator for generating antibody sequences conditioned on a target antibody structure with DeepAb.^2,5,6 On a benchmark set of 60 antibodies, the F_vHallucinator recovers over 50% of the wildtype CDR sequence on all six CDR loops. At the V_H-V_L interface, the F_vHallucinator designs amino acid substitutions that are highly enriched in human repertoire sequences. Furthermore, when compared to a large experimentally characterized library of CDR H3s of the anti-HER2 antibody,⁷ trastuzumab, the F_vHallucinator designs exhibit high sequence identity to known HER2-binders. We propose a pipeline that screens F_vHallucinator designs to obtain a virtually screened library enriched in binders for an antigen of interest. We apply this pipeline to the CDR H3 of the trastuzumab-HER2 complex to generate designs that retain the original binding mode and improve the binding affinity and interfacial properties. Thus, the F_vHallucinator pipeline enables fast and inexpensive generation of diverse, structure-conditioned antibody libraries enriched in binders.

1. Gao, W., Mahajan, S. P., Sulam, J. & Gray, J. J. Deep Learning in Protein Structural Modeling and Design. Patterns 1, 100142 (2020).
2. Norn, C. et al. Protein sequence design by explicit energy landscape optimization. doi:10.1101/2020.07.23.218917.
3. Jumper, J. et al. Highly accurate protein structure prediction with AlphaFold. Nature (2021) doi:10.1038/s41586-021-03819-2.
4. Ruffolo, J. A., Sulam, J. & Gray, J. J. Antibody structure prediction using interpretable deep learning. Patterns 3, 100406 (2022).
5. Anishchenko, I. et al. De novo protein design by deep network hallucination. Nature 1–19 (2020).
6. Wang, J. et al. Deep learning methods for designing proteins scaffolding functional sites. bioRxiv 2021.11.10.468128 (2021).
7. Mason, D. M. et al. Optimization of therapeutic antibodies by predicting antigen specificity from antibody sequence via deep learning. Nat. Biomed. Eng. (2021) doi:10.1038/s41551-021-00699-9.