De Novo Design Of Protein Structures With Designated Topology

The problem of de novo protein design is often decomposed into two sub-problems:one is to design a sequence for a given backbone structure,and the other is to design from scratch(de novo)a backbone structure that meets design specifications while having high designability.Compared with methods for sequence design that have been developed for many years,methods for structure design are not yet well-developed and existing approaches are of only limited capabilities.Currently,there are two main routes for designing backbone structures:one is to construct new backbone structures by assembling fragments from natural backbones,and the other is to sample and optimize backbone structures without using specific sequence information.The latter may be achieved with a side chain-independent energy function of backbone structure.The protein backbone design model SCUBA(Side Chain-Unknown Backbone Arrangement)adopts the second route.Previously,computational protocols for using SCUBA to generate designable backbones of regular,relatively simple topologies have been developed and experimentally tested.To design proteins of more complex topology,whether the exiting protocols are appropriate,or they require further optimization,and how,remain to be investigated through computational tests and experimental validations.Here we explore and optimize the computational workflow of using SCUBA together with the sequence design program ABACUS2(a backbone-based amino acid usage survey)to design proteins of irregular or complex topologies.First,in this paper,we explore the feasibility of using SCUBA together with the sequence design model ABACUS2 for de novo design of specified naturally existing and non-existing complex protein folds.Using the de novo design of the proteins of the DNA clamp fold of the SCOP classification system as an example,we established the following workflow:simplify the target fold as a sketch of a regularized backbone framework;build initial backbone structures conform to the simplified framework using a geometric method;sample and optime the backbone structure using simulations driven by the SCUBA energy function;further optimization of loops in the backbone structure;select amino acid sequences for the designed backbone;predict structures for the designed sequences by using AlphaFold2 and compare the predicted and the designed structures.In developing the computational workflow,we constantly turn to experimental characterization of design results,and improve the computational protocols based on the experimental results.Based on the experimental results on 8 sequences from the first round design,we introduced further loop optimization steps into the design protocol.The second round of experiments on 33 designed sequences suggested improved properties of the designs,and protein crystals were obtained for a designed protein Clamp2-9.Based on these results,we further attempted to optimize the treatment of packing interactions in the sequence design steps.This leads to improved protein expression and solubility in the third round of experiments on 26 designed sequences.A protein crystal for the designed protein 9βα3-24 was obtained.Finally,we experimentally tested the most recently developed,deep-learning based protein structure design method SCUBA-D and the sequence design method ABACUS-R.For 16 experimentally tested designs,a number of protein crystals are obtained.These results suggest that deep learning-based methods can significantly improve the success rate of de novo protein design.