Introduction
The ability to expand the genetic code allows for the site-specific incorporation of noncanonical amino acids (ncAAs) into proteins, enabling the creation of "designer" proteins with novel chemical, physical, and biological properties. The pyrrolysyl-tRNA synthetase (PylRS)/tRNA<sub>CUA</sub> pair is the most versatile tool for this purpose. However, a significant hurdle remains: many evolved PylRS variants exhibit low catalytic activity, resulting in poor protein yields that limit their industrial and medical utility. Traditional directed evolution often struggles to overcome these efficiency bottlenecks due to the complex, non-additive interactions between mutations, known as epistasis.
ML-Driven Engineering Strategy
To address these challenges, researchers developed a hybrid machine learning (ML) framework to engineer the tRNA-binding domain (TBD) of PylRS. Unlike traditional approaches that focus on the catalytic active site, targeting the TBD offers a more generalizable route to improving enzyme efficiency across various substrates.
The workflow integrated supervised learning with zero-shot deep learning models:
- FFT-PLSR (Fast Fourier Transform-Partial Least Square Regression): This supervised model was used to navigate the epistatic landscape of the TBD. By converting protein sequences into numerical "spectra," the model could predict the performance of combinatorial mutations from a limited experimental dataset.
- Zero-Shot Deep Learning: To identify novel mutation sites beyond the initial library, the team employed three distinct models: ESM-1v (a protein language model), MutCompute (a structure-based model identifying suboptimal micro-environments), and ProRefiner (an inverse protein folding model).
This multi-stage approach allowed the researchers to explore a massive sequence space of over 11,000 possible mutation patterns, ultimately identifying highly synergistic variants that would likely be missed by random screening.
Results: Superior Incorporation Efficiency
The engineering process yielded two primary optimized variants: Com1-IFRS and Com2-IFRS. Com1-IFRS, derived from the initial supervised learning phase, showed an 11-fold increase in activity. The further optimized Com2-IFRS achieved a 30-fold increase in general activity compared to the baseline IFRS variant.
The performance gains were even more dramatic for specific substrates. For the ncAA 3-fluoro-L-phenylalanine (3FF), Com2-IFRS demonstrated a staggering 3,944.8-fold improvement in efficiency. Furthermore, the "modular" nature of these TBD mutations allowed them to be transplanted into seven other PylRS-derived synthetases. This transplantation successfully boosted the yields of proteins containing six different classes of ncAAs—including derivatives of phenylalanine, tyrosine, tryptophan, cysteine, histidine, and lysine—by up to 1,149.7-fold. In a practical application, the system produced 28.3 mg/L of engineered myoglobin, representing a 6.3-fold increase over standard methods.
Mechanistic Insights via Molecular Dynamics
To understand the physical basis of these improvements, the team utilized AlphaFold3 for structural prediction and conducted 50-ns molecular dynamics (MD) simulations. The analysis revealed that the mutations in the TBD did more than just increase binding affinity. They effectively reshaped the hydrogen bond network between the protein and the tRNA. This structural reorganization physically shortened the reaction distance between the tRNA and the ncAA, facilitating faster catalysis. Additionally, the mutations enhanced the dynamics correlation network within the enzyme, suggesting a more coordinated and efficient catalytic process.
Conclusion and Future Directions
This study demonstrates that ML-guided engineering of the tRNA-binding domain is a powerful, "plug-and-play" strategy for overcoming the efficiency limitations of the PylRS system. By navigating complex epistatic landscapes, the researchers produced variants that significantly enhance protein yields across a wide range of ncAAs. While challenges such as misincorporation of natural amino acids remain, this framework provides a robust template for optimizing other orthogonal translation systems, potentially paving the way for a universal toolkit in synthetic biology and protein engineering.
Summary updated