Beyond single-protein de novo design: A generative algorithm for the NTF2-like superfamily

Abstract

Natural proteins evolved over billions of years to regulate cellular growth, ward off infection and capture and store solar energy. Proteins thus serve as the molecular basis for life. The promise of protein design is to use nature’s favorite toolbox to solve modern human problems without having to wait for the long and meandering path of natural selection. Protein structure determines function, so it is not surprising proteins sample a great variety of structures. Thus, it is reasonable to expect that designing proteins with functions not seen in nature would require access to comparable structural diversity, more specifically, diversity of active site structure. Despite the advances in de novo protein design, the systematic generation of proteins containing pockets that can harbor substrates has been lacking. The use of a natural small alpha-beta fold, the Nuclear Transport Factor 2-like (NTF2-like) fold, to design a high affinity small-molecule binding protein, and the diversity observed in that family, have posed the idea that significant pocket structural diversity could be derived from this relatively simple, small, alpha-beta fold. To explore this idea, we analyzed the structures of proteins belonging to the NTF2-like superfamily and other proteins with similar characteristics to understand the determinants of their structural diversity. The most salient feature of the NTF2-like superfamily is the curved beta-sheet that forms most of their pockets in its concave face. Curved beta sheets depart from the classic beta pleated structure displaying bulges, tight kinks and irregular bending and twisting. The first step towards de novo design of a large variety of NTF2-like proteins is devising principles for designing curved beta sheets. In this work, we demonstrate we can generate a number of different curved sheets in the context of NTF2-like proteins. We then use these principles, along with additional information from native NTF2-like proteins, to create a generative algorithm that can widely sample structural diversity while still producing physically realistic models. We show this algorithm can produce a large variety of NTF2-like proteins, and through cycles of large-scale design and validation, we increase the diversity and success rate of its output. As a proof of principle, we design a binder of the mycotoxin aflatoxin B1, which could serve as starting material for devising aflatoxin-chelating or degrading materials.