TL;DR
This paper introduces a novel machine learning pipeline using Variational Autoencoders (VAEs) and microfluidics to engineer dehalogenase enzymes. By leveraging VAEs to explore evolutionary relationships in a latent space, the researchers successfully designed novel enzyme variants with significantly improved solubility, stability, and activity, demonstrating a powerful approach for accelerating protein engineering.
What Problem Does This Paper Address?
Optimizing enzymes for industrial applications is a significant challenge due to the complex interplay of amino acid residues. While computational methods often rely on evolutionary information from homologous sequences, deciphering the intricate dependencies among residues remains difficult. Existing approaches struggle to efficiently navigate the vast sequence space to find functional and stable protein variants. This paper aims to bridge this gap by developing a machine learning-driven strategy that leverages the learned evolutionary landscape to design novel, improved enzyme sequences.
Methods
The core of this work is a pipeline that combines Variational Autoencoders (VAEs) with experimental validation.
- Variational Autoencoders (VAEs): The researchers trained VAEs on multiple sequence alignments (MSAs) of haloalkane dehalogenases. VAEs learn a compressed, low-dimensional "latent space" that represents the evolutionary relationships and dependencies within the protein family. The key innovation here is using the geometry of this latent space to guide the design of new sequences.
- Evolutionary Latent Space Strategy: They developed a "straight-line evolutionary strategy" where they sample points along a line connecting a query sequence's embedding in the latent space to the origin. This strategy is inspired by the observation that more stable or ancestral sequences tend to reside closer to the origin in the latent space.
- Microfluidics and High-Throughput Screening: Experimental validation was crucial. They employed the MicroPEX microfluidic platform for high-throughput screening of enzyme variants, enabling efficient characterization of solubility, thermostability, temperature profiles, and substrate specificity.
- Design-Build-Test Cycles: The pipeline was iterated through three design-build-test cycles. Each cycle involved designing new sequences using the VAE, experimentally producing and testing them, and then using the experimental results to refine the VAE models and the MSA preprocessing for the next round.
- Manual Curation: In later rounds, manual curation of VAE-proposed mutations was incorporated, guided by structural predictions from AlphaFold and stability assessments from MutCompute, to further refine designs.
- Conditional VAEs: To specifically address solubility issues, they also explored conditional VAEs (CVAEs) trained with solubility scores.
Datasets
- Haloalkane Dehalogenase (HLD) Sequence Datasets: Two primary datasets were used, derived from the EnzymeMiner tool:
- HLDI-IV: A larger dataset (12,053 sequences, 299 positions) used in the first two experimental rounds.
- HLDI-II: A smaller, more refined dataset (4,053 sequences, 293 positions) focusing on more soluble HLD subfamilies, used in the third round.
- Experimental Stability Data: Datasets of previously characterized ancestral sequences and engineered DhaA variants were used to validate the latent space's ability to capture stability information.
- No new benchmark datasets were explicitly introduced, but the collected experimental data from 20 variants is highlighted as a valuable dataset for future validation.
Key Insights and Ideas
- Latent Space Geometry as an Evolutionary Map: The VAE's latent space isn't just a compressed representation; its geometric structure directly reflects phylogenetic relationships and, importantly, protein stability. This allows for a more intuitive and guided approach to designing improved sequences.
- The "Origin" as a Hotspot for Stability: The observation that ancestral or more stable variants map closer to the latent space origin is a powerful guiding principle. The "straight-line evolutionary strategy" leverages this by moving towards the origin to find potentially more stable designs.
- Iterative Refinement is Key: The three design-build-test cycles were critical. Each iteration provided valuable feedback, allowing the researchers to improve MSA preprocessing, refine VAE models, and address specific challenges like low solubility.
- Indels are a Major Hurdle: The frequency and location of insertions/deletions (indels) proved to be a significant factor influencing solubility and overall success, highlighting the need for careful handling of these in sequence generation.
- AlphaFold Limitations for Novel Sequences: While AlphaFold is powerful for predicting structures of known proteins, its predictions for novel, VAE-generated sequences showed discrepancies with experimental secondary structure data, underscoring the continued need for experimental validation.
Key Results
- Improved Solubility: The success rate of obtaining soluble variants increased dramatically across the three rounds: from 11% in round 1 to 67% in round 2, and finally to 75% in round 3.
- Enhanced Stability: Nine of the 20 engineered variants showed a melting temperature increase of up to 9 °C, with an average improvement of 3 °C across soluble variants.
- Increased Activity: The most stable variant demonstrated a 3.5-fold increase in activity compared to the template.
- Sequence Similarity: Some highly improved variants shared as little as 67% sequence similarity with the original template, demonstrating significant engineering.
- 20 Multiple-Point Variants Characterized: A substantial number of novel variants were experimentally validated, providing a rich dataset.
Limitations and Future Work
- Generalizability: The study focused on a single enzyme family (haloalkane dehalogenases). The authors acknowledge that validating this approach on other protein families is necessary to assess its generalizability.
- Predicting Misfolding/Insolubility: While progress was made, accurately predicting and filtering out misfolded or poorly soluble designs in silico remains a challenge. Future work could focus on developing better scoring methods for this.
- Ensemble Learning: The authors noted that different VAE initializations led to scatter in trajectories. They suggest that ensemble learning could improve the robustness of the strategy.
- Transformer Architectures: The paper points to the potential of transformer-based architectures for improved sequence generation and learning from amino acid sequences.
- Manual Curation Insights: The manual curation step highlighted the limitations of current tools in predicting epistasis effects, suggesting that integrating more sophisticated models could be beneficial.
Summary updated