Machine learning-assisted directed protein evolution with combinatorial libraries

Abstract

To reduce experimental effort associated with directed protein evolution and to explore the sequence space encoded by mutating multiple positions simultaneously, we incorporate machine learning into the directed evolution workflow. Combinatorial sequence space can be quite expensive to sample experimentally, but machine-learning models trained on tested variants provide a fast method for testing sequence space computationally. We validated this approach on a large published empirical fitness landscape for human GB1 binding protein, demonstrating that machine learning-guided directed evolution finds variants with higher fitness than those found by other directed evolution approaches. We then provide an example application in evolving an enzyme to produce each of the two possible product enantiomers (i.e., stereodivergence) of a new-to-nature carbene Si-H insertion reaction. The approach predicted libraries enriched in functional enzymes and fixed seven mutations in two rounds of evolution to identify variants for selective catalysis with 93% and 79% ee (enantiomeric excess). By greatly increasing throughput with in silico modeling, machine learning enhances the quality and diversity of sequence solutions for a protein engineering problem.

Keywords: catalysis; directed evolution; enzyme; machine learning; protein engineering.

Fig. 1.

(A) Directed evolution with single mutations. If limited to single mutations, the identification of optimal amino acids for N positions requires N rounds of evolution. (B) Directed evolution by recombining mutations found in best variants from a random combinatorial search. (C) Machine learning-assisted directed evolution. As a result of increased throughput provided by screening in silico, four positions can be explored simultaneously in a single round, enabling a broader search of sequence–function relationships and deeper exploration of epistatic interactions.

Fig. 2.

(A) Highest fitness values found by directed evolution and directed evolution assisted by machine learning. The distribution of fitness peaks found by iterative site-saturation mutagenesis from all labeled variants (149,361 of 20⁴ possible covering four residues) is shown in red. The distribution of fitness peaks found by 10,000 recombination runs with an average of 570 variants tested is shown in blue. The distribution of the highest fitnesses found from 600 runs of the machine learning-assisted approach is shown in green. A total of 570 variants are tested in all approaches. For reference, the distribution of all measured fitness values in the landscape is shown in gray. (B) The same evolutionary distributions are shown as empirical cumulative distribution functions, where the ordinate at any specified fitness value is the fraction of evolutionary runs that reach a fitness less than or equal to that specified value. Machine learning-assisted evolution walks are more likely to reach higher fitness levels compared with conventional directed evolution.

Fig. 3.

Carbon–silicon bond formation catalyzed by heme-containing Rma NOD to form individual product enantiomers with high selectivity.

Fig. 4.

(A) Structural homology model of Rma NOD and positions of mutated residues made by SWISS-MODEL (47). Set I positions 32, 46, 56, and 97 are shown in red, and set II positions 49, 51, and 53 are shown in blue. (B) Evolutionary lineage of the two rounds of evolution. (C) Summary statistics for each round, including the number of sequences obtained to train each model, the fraction of the total library represented in the input variants, each model’s leave-one-out cross-validation (CV) Pearson correlation, and the number of predicted sequences tested.

Fig. 5.

A library’s fitness values can be visualized as a 1D distribution, in this case as kernel density estimates over corresponding rug plots. This figure shows subplots for each library illustrating the changes between input (lighter) and predicted (darker) libraries for the (S)-enantiomers (cyan) and (R)-enantiomers (red). The initial input library for set I is shown in gray. The predicted (darker) libraries for each round are shifted toward the right and left of the distributions for the (S)- and (R)-enantiomers, respectively. For reference, dotted lines are shown for no enantiopreference (i.e., 0% ee).