Machine learning-assisted directed protein evolution with combinatorial libraries

Zachary Wu, S. B. Jennifer Kan, Russell D. Lewis, Bruce J. Wittmann, Frances H. ArnoldProceedings of the National Academy of Sciences 2019 · DOI

TL;DR

This paper introduces a machine learning-assisted approach to directed protein evolution that significantly speeds up the process and identifies more effective protein variants. By using ML models to computationally screen vast libraries of protein sequences, researchers can explore sequence space more efficiently, leading to improved enzyme performance and the discovery of novel catalytic activities.

What Problem Does This Paper Address?

Directed evolution is a powerful technique for engineering proteins with novel or enhanced functions. However, it's often limited by the sheer scale of experimental screening required. Exploring the vast "sequence space" of possible protein variants, especially when multiple mutations are involved, can be prohibitively expensive and time-consuming. This paper tackles the challenge of efficiently navigating this complex landscape to find improved protein variants, particularly when aiming for specific catalytic activities or stereoselectivity.

Methods

The core innovation is the integration of machine learning (ML) into the directed evolution workflow. Instead of solely relying on experimental screening, the authors use ML models to predict the fitness of protein variants in silico. Here's a breakdown:

  • Combinatorial Libraries: They start by creating large libraries of protein variants with multiple mutations at specific positions.
  • In Silico Screening: A subset of these variants is experimentally tested to generate sequence-function data. This data is then used to train various ML regression models (e.g., linear models, decision trees, random forests, neural networks).
  • Prediction and Selection: The trained ML models predict the fitness of the entire theoretical library. A smaller, "predicted library" enriched in high-fitness variants is then selected for experimental testing in the next round.
  • Iterative Evolution: This process is repeated, with the best variants from each round serving as the starting point for the next, allowing for rapid exploration of sequence space and identification of beneficial mutation combinations.
  • Novelty: Unlike some previous ML approaches that focused on single mutations or linear models, this method explicitly models nonlinear, epistatic interactions between mutations by sampling combinatorial libraries and using diverse ML models.

Datasets

  • Human GB1 Binding Protein Fitness Landscape: The authors first validate their approach on a large, publicly available empirical fitness landscape for human GB1 binding protein. This dataset, collected by Wu et al., comprises measurements of 149,361 variants from NNK/NNS saturation mutagenesis at four positions known to interact epistatically.
  • Rma NOD Enzyme Evolution: For their example application, they generated their own sequence-function data by evolving the Rhodothermus marinus nitric oxide dioxygenase (Rma NOD) enzyme. This involved creating initial libraries and then iteratively testing predicted variants.

Key Insights and Ideas

  • ML as a Computational Microscope: Machine learning acts as a powerful computational tool to "see" into vast sequence spaces that would be experimentally intractable. It allows researchers to focus expensive experimental resources on the most promising variants.
  • Beyond Single Mutations: The approach effectively captures epistatic interactions (where the effect of one mutation depends on the presence of others), which are crucial for understanding complex protein landscapes and designing highly functional proteins. This is a significant advantage over methods that focus on individual mutations.
  • Accelerated Discovery: By enabling simultaneous exploration of multiple mutations, ML-assisted evolution allows for larger "steps" through sequence space in each round, dramatically accelerating the discovery of improved variants.
  • Diverse Solutions: The ML approach can identify a diverse set of sequences that achieve a desired outcome, offering flexibility for other engineering goals (e.g., increased activity, altered substrate tolerance).
  • Accuracy vs. Utility: Surprisingly, even models with moderate predictive accuracy (Pearson's r = 0.41) can be sufficient to guide directed evolution effectively, suggesting that the direction of prediction is more critical than perfect accuracy for this application.

Key Results

  • GB1 Landscape Validation: On the GB1 fitness landscape, ML-assisted directed evolution reached the global optimum fitness value in 8.2% of simulations, compared to 4.9% for single-mutant walks and 4.0% for recombination. It also required about 30% fewer variants to achieve similar results.
  • Enzyme Engineering for Stereodivergence: The authors successfully engineered the Rma NOD enzyme to produce both (S)- and (R)-enantiomers of a new-to-nature carbon-silicon bond formation reaction.
  • In two rounds of evolution, they identified variants with enantioselectivities of 93% ee for the (S)-enantiomer and 79% ee for the (R)-enantiomer.
  • This was achieved by exploring seven mutations across two sets of positions.
  • Increased Throughput: The ML approach significantly increases experimental throughput, enabling the exploration of larger and more complex sequence spaces.

Limitations and Future Work

  • Model Accuracy: While moderate accuracy is sufficient, the authors note that perfect accuracy isn't always achieved. They suggest that improving model accuracy, perhaps by collecting more data or using more computationally expensive simulations, could further enhance the approach.
  • Codon Usage Bias: The authors acknowledge that biases in sequence transformation into the host organism (due to cloning biases) can affect the observed ratios of sequences. They address this by visualizing encoded amino acid distributions.
  • Future Directions: The paper suggests that further development of ML algorithms for protein engineering, especially those that can handle more complex datasets and nonlinear relationships, will continue to advance the field. They also highlight the value of deep mutational scanning for generating comprehensive datasets to validate and improve these ML models. The ability to explore sequence space more efficiently opens doors for engineering proteins with increasingly complex and novel functions.

Summary updated

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% <i>ee</i> (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.
BibTeX
@inproceedings{wu2019machine,
  author = {Zachary Wu and S. B. Jennifer Kan and Russell D. Lewis and Bruce J. Wittmann and Frances H. Arnold},
  title = {{Machine learning-assisted directed protein evolution with combinatorial libraries}},
  booktitle = {Proceedings of the National Academy of Sciences},
  year = {2019},
  doi = {10.1073/pnas.1901979116},
}
Figure