| Nature provides a wealth of proteins with untapped potential.However,in the actual technical application process,a large number of proteins do not meet the required multi-dimensional complex requirements.In addition,the natural evolution cycle of proteins is long and the direction is blind and random,so people have researched and developed directed evolution technology.Directed evolution micmics the principle of natural evolution in the laboratory without relying on structure and mechanism information and mutates and screens iteratively in order to selecting protein mutants with optimal functional properties.However,directed evolution is highly dependent on highthroughput screening methods and the protein search space is too large and ideal protein sequences are scarce in the vast sequence space,it is unrealistic to exhaustively search through experimental methods.In recent years,machine learning(ML)approaches are be used to guide protein directed evolution,which can accelerate the process by learning to map protein sequences to functions.Although machine learning methods are powerful,they still encounter huge challenges in the face of high-dimensional feature representation and inefficient search methods for protein sequences.To address these issues,we aim to efficiently search the protein search space using Bayesian optimization methods.Bayesian optimization is designed for expensive black-box derivative-free global optomization.It is best-suited for optimization over continuous domains of low dimensions,and tolerates stochastic noise in function evaluations.It builds a surrogate for the objective and quantifies the uncertainty in that surrogate using Gaussian process regression,and then uses an acquisition function defined from this surrogate to decide where to sample.In order to improve the efficiency of the Bayesian global optimization algorithm,we adopts the experimental design method to select the initial training sample set,which has achieved the purpose of maximizing the sample information.For the high-dimensional feature representation of protein sequences,we adopts a low-dimensional encoding strategy based on functional values,which can make protein sequence recommendation.In order to improve the robustness of the optimization algorithm,under the premise of using a robust Gaussian process to eliminate outliers,Bayesian optimization is used to iteratively search for the target protein mutant with the optimal functional value step by step.This paper aims to use an optimization framework that combines outlier diagnosis with Bayesian optimization method to guide the directed evolution of proteins.In order to get rich and sufficient experimental verification,this paper uses robust Bayesian optimization method to search on five protein data sets,which proves that compared with classical Bayesian optimization,the method in this paper has better search effect and faster convergence speed.In this paper,a robust Bayesian optimization method and an efficient optimization framework are used to help the laboratory better find the optimal protein mutants,reduce the work of laboratory screening,greatly save time,and also bring new ideas for better guidance of protein directed evolution in the future. |
