Machine Learning for Protein Engineering

Engineering of highly active and diverse nuclease enzymes by combining machine learning and ultra-high-throughput screening

“…ML4-MBO-DNN (Methods) was the strongest-performing member of the portfolio of methods in ML4 in terms of hit rate (see Supplementary Table 1). It had similar size as HR4 (1356 vs. 1540 variants) and significantly outperformed HR4 at finding hits above the A73R level (chi-square test; P=1.3×10 -9 ): 52 ± 7 of the 1356 ML4-MBO-DNN variants (hit-rate 3.9% ± 0.5%; standard deviations estimated via bootstrapping) had greater activity than A73R, compared to only 7 ± 4 of the 1540 HR4 variants (hit-rate 0.5% ± 0.2%). For discovering variants with activity greater than the WT, ML4-MBO-DNN significantly outperformed HR4 (chi-square test; p < 1e-15).…”

“…The ability to engineer proteins has revolutionized applications in industry and therapeutics [1][2][3][4][5][6] . Generally, a protein engineering campaign can be divided into two stages [7][8][9] : discovery first finds a candidate protein that performs a desired function at a non-zero level of activity and then optimization improves its attributes, such as the binding strength of an antibody 10 or the catalytic activity 11 , thermostability 12,13 , or stereoselectivity of an enzyme 14 . Directed evolution (DE) is the standard technique for protein optimization, where a pool of genotypes is iteratively improved using in-vitro selection and mutagenesis [15][16][17][18] .…”

“…To address these limitations, ML-guided directed evolution (MLDE), involving rounds of data collection, modeling, and model-guided sequence generation, has emerged as a potential alternative to DE 9,[21][22][23][24] . However, there is limited work directly comparing the techniques across multiple rounds of high-throughput experiments, an increasingly-common setting due to advancements in automation 25 , microfluidics 26,27 , display systems [27][28][29][30] , multiplexing 31 , and continuous evolution 32,33 .…”

“…While DE alternates between hit selection and mutagenesis, ML-guided directed evolution (MLDE) alternates between data collection, modeling, and suggesting new genotypes using the model 30 . A number of recent surveys discuss the state of MLDE, much of which is powered by advancements in deep learning methods for generating protein sequences and predicting protein fitness 9,31–34 . Both MLDE and DE benefit from the deeper exploration of the fitness landscape afforded by higher screening throughput, yet there is limited prior work directly comparing the two approaches across multiple rounds of high-throughput experiments.…”

Latent-based Directed Evolution accelerated by Gradient Ascent for Protein Sequence Design

“…Directed evolution (DE) is a conventional approach in protein engineering that aims to search for global maximal protein sequences from a large database of unlabeled candidates S with minimal experimental validation [38], [39]. DE employs an accelerated cycle of mutation and selection, which iteratively generates a pool of protein variants and selects those having improved desired properties as the next generation.…”