Expanding all-α-helical protein space through rational computational design

Abstract

Abstract De novo protein design is advancing rapidly 1,2 . This is being driven by AI to generate protein backbones, sequences, and structural models 3–7 . As a result, de novo designed proteins are becoming larger and more complex 8–10 , and increasingly explore new protein structures 11,12 . By contrast, natural proteins have evolved structural and functional complexity by modular combination of recurring protein domains 13 . Approximately 25% of these natural domains are mostly α-helical structures 14 . Here we show how these can be expanded using rational computational design. Following the domain classification scheme CATH 15 , we build complex all-α de novo proteins hierarchically using sequence-to-structure relationships for helix-helix interactions, systematic rules to connect helices, computational tools to design loops, and in silico evaluation. The pipeline starts with a target architecture of free-standing helices. These are connected into a topology by considering local arrangements of helical bundles using understood sequence-to-structure relationships for helix packing. Single-chain sequences are completed using template- and AI-based methods. Finally, AlphaFold models are assessed to give small numbers of designs for experimental validation. We test 31 designs for 14 different architectures and 25 topologies. 75% of these express as stable, monomeric, water-soluble proteins; and >30% yield X-ray crystal structures matching the designs to atomic accuracy and with new-to-nature structures. Finally, several of the scaffolds are functionalised through one-shot designs to deliver ion, small-molecule and protein binders.