Informed training set design enables efficient machine learning-assisted directed protein evolution

Abstract

Due to screening limitations, in directed evolution (DE) of proteins it is rarely feasible to fully evaluate combinatorial mutant libraries made by mutagenesis at multiple sites. Instead, DE often involves a single-step greedy optimization in which the mutation in the highest-fitness variant identified in each round of single-site mutagenesis is fixed. However, because the effects of a mutation can depend on the presence or absence of other mutations, the efficiency and effectiveness of a single-step greedy walk is influenced by both the starting variant and the order in which beneficial mutations are identified-the process is path-dependent. We recently demonstrated a pathindependent machine learning-assisted approach to directed evolution (MLDE) that allows in silico screening of full combinatorial libraries made by simultaneous saturation mutagenesis, thus explicitly capturing the effects of cooperative mutations and bypassing the path-dependence that can limit greedy optimization. Here, we thoroughly investigate and optimize an MLDE workflow by testing a number of design considerations of the MLDE pipeline. Specifically, we (1) test the effects of different encoding strategies on MLDE efficiency, (2) integrate new models and a training procedure more amenable to protein engineering tasks, and (3) incorporate training set design strategies to avoid information-poor low-fitness protein variants ("holes") in the training data. When applied to an epistatic, hole-filled, four-site combinatorial fitness landscape of protein G domain B1 (GB1), the resulting focused training MLDE (ftMLDE) protocol achieved the global fitness maximum up to 92% of the time at a total screening burden of 470 variants. In contrast, minimal-screening-burden single-step greedy optimization over the GB1 fitness landscape reached the global maximum just 1.2% of the time; ftMLDE matching this minimal screening burden (80 total variants) achieved the global optimum up to 9.6% of the time with a 49% higher expected maximum fitness achieved. To facilitate further development of MLDE, we present the MLDE software package (https://github.com/fhalab/MLDE), which is designed for use by protein engineers without computational or machine learning expertise..

ECNet is an evolutionary context-integrated deep learning framework for protein engineering

“…11). This implies that the design of more effective training data should be taken into account when developing ML algorithms to assist protein engineering, especially when the experimental test budget is limited 62 .…”

Heterogeneity of the GFP fitness landscape and data-driven protein design

“…Experimental survey of the fitness landscape of a protein of interest is increasingly used in protein engineering to discover novel sequences with specific functions (Bryant et al, 2021b; Romero and Arnold, 2009; Russ et al, 2020; Wittmann et al, 2021). While this approach remains challenging for proteins with a function that cannot be easily ascertained in a high-throughput manner (Romero and Arnold, 2009), it is likely to be more widely used in the future due to technological advances of experimental (Romero and Arnold, 2009) and analytical (Wittmann et al, 2021; Wu et al, 2019) tools. Our description of heterogeneity of fitness peaks of orthologous GFPs suggests some practical considerations for such surveys of other proteins.…”

“…Originally, the fitness landscape was introduced to describe the relationship between fitness and the entire genome (de Visser and Krug, 2014; Wright, 1932). Over time, the usefulness of the concept of the fitness landscape led to the adaptation of this term to describe the relationship between protein function and its protein-coding gene sequence (Biswas et al, 2021; Ogden et al, 2019; Romero and Arnold, 2009; Wittmann et al, 2021; Zheng et al, 2020). Absolute knowledge of the fitness landscape would reveal the phenotypes conferred by any arbitrary genotype (de Visser and Krug, 2014; Ferretti et al, 2018; Fragata et al, 2019), with immense and obvious practical implications (Alley et al, 2019; Bryant et al, 2021a; Hirabayashi and Arai, 2019; Kemble et al, 2019; Wrenbeck et al, 2017; Wu et al, 2019).…”

“…However, how different shapes of fitness peaks may be distributed in genotype space has not been explored (Chan et al, 2017; Kemble et al, 2019). Furthermore, the exploration of the fitness landscape of specific proteins is one of the approaches in protein engineering (Bryant et al, 2021b; Romero and Arnold, 2009; Russ et al, 2020; Wittmann et al, 2021). Such studies explore the fitness landscape of the protein of interest through deep mutational scan of a known protein sequence.…”

Structure-Functional Analysis of Human Cytochrome P450 2C8 Using Directed Evolution

“…Recently, many artificial intelligence techniques have been developed, and they have been widely applied to various relationship analyses and used for the repositioning of FDA-approved drugs [31][32][33]. Therefore, cooperative work with artificial intelligence, particularly in the initial screening for the relationship and in the analysis of the structurefunction relationship, should be pursued to elaborate the directed evolution approach to CYP2C8 in the future.…”