Enhanced Sequence-Activity Mapping and Evolution of Artificial Metalloenzymes by Active Learning

Tobias Vornholt, Mojmír Mutný, Gregor W. Schmidt, Christian Schellhaas, Ryo Tachibana, Sven Panke, Thomas R. Ward, Andreas Krause, Markus Jeschek

TL;DR

This paper introduces a novel active learning pipeline that significantly boosts the efficiency of engineering artificial metalloenzymes (ArMs). By integrating automated screening, next-generation sequencing, and a sophisticated machine learning model (Gaussian Process Regression), they achieved an order-of-magnitude increase in the success rate of identifying highly active ArM variants for a challenging hydroamination reaction.

What Problem Does This Paper Address?

Engineering enzymes for specific applications is a crucial but often serendipity-driven and laborious process. Machine learning (ML) offers a way to accelerate this by building models of the "sequence-activity landscape" – how changes in a protein's amino acid sequence affect its function. However, accurately modeling this landscape for large protein spaces with limited experimental data is difficult. Existing methods often struggle with experimental noise, inefficient data acquisition, and suboptimal strategies for exploring the vast sequence space. This paper tackles these limitations by developing a more efficient and effective ML-assisted directed evolution (MLDE) approach.

Methods

The core of this work is an integrated pipeline combining several key elements:

  • Informed Library Design: Instead of random mutagenesis, they strategically selected positions for randomization based on initial experiments, aiming for high impact.
  • Large-Scale Screening with Automation and NGS: They employed a lab automation platform for high-throughput activity assays and coupled it with Next-Generation Sequencing (NGS) for cost-effective, large-scale variant sequencing. This allowed them to generate thousands of sequence-activity data points.
  • Active Machine Learning (Gaussian Process Regression): They used Gaussian Process (GP) regression, a probabilistic ML model, to predict ArM activity. GPs are valuable because they provide not only predictions but also uncertainty estimates, which are crucial for guiding the next experimental steps.
  • Exploratory Round for Model Refinement: A key innovation is the inclusion of an "exploration" phase. Instead of solely focusing on finding the most active variants (exploitation), they specifically selected variants predicted to have high uncertainty. This helped the model learn more about previously underexplored regions of the sequence space, significantly improving its overall accuracy and ability to generalize.
  • Handling Experimental Noise: The model explicitly accounts for experimental noise by analyzing replicate measurements and incorporating this into the probabilistic framework, preventing overfitting to random fluctuations.
  • Diversified Sampling (DPPs): For the final "exploitation" phase, they used Determinantal Point Processes (DPPs) to select a diverse set of high-activity variants, ensuring they didn't just find very similar, highly active mutants.

Datasets

The primary dataset used was generated in situ for this study. It consists of sequence-activity data for thousands of artificial metalloenzyme (ArM) variants. Specifically:

  • Initial Library: Data from over 2,000 distinct ArM variants, generated through a combinatorial library of five selected positions. This was augmented with ~400 previously generated double mutant data points.
  • Exploration Round: Data from an additional 465 variants, selected based on the model's uncertainty estimates.
  • Exploitation Round: Data from 349 variants, selected based on high predicted activity and diversity.

The paper does not introduce a new public benchmark dataset but rather generates a rich, internally used dataset for their specific ArM engineering task.

Key Insights and Ideas

  • Active Learning is More Than Just Finding Hits: The "exploration" phase, where the model actively seeks out regions of high uncertainty, is critical for improving the model's accuracy and ability to generalize. This is a significant departure from simply focusing on the top predicted variants.
  • Experimental Noise is a Major Hurdle: Ignoring or poorly handling experimental noise can lead to models that overfit and perform poorly. Explicitly modeling noise is essential for robust MLDE.
  • Combinatorial Libraries are Key for Epistasis: To capture complex interactions between mutations (epistasis), combinatorial libraries that randomize multiple positions simultaneously are necessary, even if they initially yield more inactive variants.
  • Smart Sampling Maximizes Experimental Budget: By intelligently selecting which variants to test next based on model predictions and uncertainty, researchers can achieve much higher hit rates with fewer experiments.
  • Counter-Intuitive Mutations Can Be Found: The active learning approach helped identify highly active variants with seemingly counter-intuitive mutations (e.g., cysteine at position 111), highlighting the power of exploring beyond conventional wisdom.

Key Results

  • 12-fold Increase in Hit Rate: The active learning strategy, particularly with the exploration round, led to a 12-fold increase in the hit rate (variants with higher activity than the reference) in the final exploitation phase compared to the initial library.
  • 18-fold Higher Activity: The most active variant identified, Sav CTNLV, exhibited an 18-fold higher cell-specific hydroamination activity compared to the wild-type and a 3-fold improvement over the previously best variant.
  • Cost-Effective Data Generation: The combination of lab automation and NGS enabled the generation of sequence-activity data for thousands of variants at a significantly reduced cost per variant.
  • Robust Model Performance: The GP model, incorporating noise estimation and optimized descriptors, achieved a median R² of 0.54 and a Spearman correlation of 0.68, demonstrating its utility for guiding directed evolution.

Limitations and Future Work

  • ArM Specificity: The current work focuses on a specific ArM for gold-catalyzed hydroamination. While the methodology is generalizable, applying it to other enzymes or reactions may require re-optimization of descriptors and model parameters.
  • Scaffold Limitations: The authors note that the relatively exposed cofactor site in Sav might limit the achievable improvements. They suggest that applying this strategy to alternative scaffolds with more shielded active sites could lead to even larger gains.
  • Further Optimization: While the achieved correlation is sufficient for guiding evolution, further improvements in predictive performance could be sought by exploring advanced descriptors (e.g., from protein language models) or structure prediction algorithms.
  • Broader Applications: The authors envision this active learning approach being broadly applicable to tailor natural enzymes for industrial applications, engineer other protein classes like antibodies and biosensors, and design transporters.

Summary updated

Abstract

Abstract Tailored enzymes hold great potential to accelerate the transition to a sustainable bioeconomy. Yet, enzyme engineering remains challenging as it relies largely on serendipity and is, therefore, highly laborious and prone to failure. The efficiency and success rates of engineering campaigns may be improved substantially by applying machine learning to construct a comprehensive representation of the sequence-activity landscape from small sets of experimental data. However, it often proves challenging to reliably model a large protein sequence space while keeping the experimental effort tractable. To address this challenge, we present an integrated pipeline combining large-scale screening with active machine learning and model-guided library design. We applied this strategy to efficiently engineer an artificial metalloenzyme (ArM) catalysing a new-to-nature hydroamination reaction. By combining lab automation and next-generation sequencing, we acquired sequence-activity data for several thousand ArM variants. We then used Gaussian process regression to model the activity landscape and guide further screening rounds according to user-defined objectives. Crucial characteristics of our enhanced enzyme engineering pipeline include i) the cost-effective generation of information-rich experimental data sets, ii) the integration of an explorative round to improve the performance of the model, as well as iii) the consideration of experimental noise during modelling. Our approach led to an order-of-magnitude boost in the hit rate of screening while making efficient use of experimental resources. Smart search strategies like this should find broad utility in enzyme engineering and accelerate the development of novel biocatalysts.
BibTeX
@inproceedings{vornholt2024enhanced,
  author = {Tobias Vornholt and Mojmír Mutný and Gregor W. Schmidt and Christian Schellhaas and Ryo Tachibana and Sven Panke and Thomas R. Ward and Andreas Krause and Markus Jeschek},
  title = {{Enhanced Sequence-Activity Mapping and Evolution of Artificial Metalloenzymes by Active Learning}},
  booktitle = {ACS},
  year = {2024},
  doi = {10.1021/acscentsci.4c00258},
}
Figure