Abstract LB151: Deep learning - based protein therapeutics design to treat HER2+ cancers

Abstract

Abstract Background: HER2 amplification and activation occurs in approximately 15~25% of all breast cancers and is further indicated in other cancer types with high un-met medical needs, such as cholangiocarcinoma, ovarian and esophageal cancer. Further improvement of current HER2 targeting therapeutics on safety and efficacy remains to be desired. Deep learning - based computational approach, such as AlphaFold, has been successfully applied to the understanding of protein structure. However, its applications to protein therapeutic design remain largely underdeveloped. Method: We have developed an end-to-end protein design engine Apollo, which can be used to design target-binding protein therapeutics in silico. Apollo uses recurrent neural networks (RNNs), transformers and graph convolution neural networks (GCNs) to represent proteins, protein structures and protein-protein interactions. This is further improved by multi-head and hierarchical attention layers. Using generative AI, Apollo can either generate putative binding proteins de novo, or to optimize protein lead molecules. Candidate proteins are evaluated in silico for binding affinity and specificity, secondary and tertiary structure, isoelectric point, solubility, and other drug-like properties. Generated designs are iteratively improved by Apollo via a reinforcement learning loop. Results: Apollo has achieved high prediction accuracy by in silico evaluations, using metrics such as AUC and the confusion matrices. Using Apollo, we then designed mini protein binders of 50~100 aa to target HER2 extracellular domain (ECD). These binders can be produced with high yield and purity using mammalian protein expression systems. From de novo generation, the designed candidates are shown to bind to human HER2 with EC50 in the nanomolar range by ELISA. From optimizations based on lead molecules, the designed candidates have shown improved production yield and thermostability while maintaining binding affinity. Conclusions: We have shown the proof-of-concept success of using machine learning to design protein therapeutics. Citation Format: Ziwei Liang, Tom Murray, Max Bluestone, Chris Zoumadakis, Luke Martin, Alexey Vlasenko, Jeffrey LaFrence, John Lazar, Matthias Denecke, David Longo. Deep learning - based protein therapeutics design to treat HER2+ cancers [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr LB151.