Delivering a 17x Specificity Boost in 6 Months

The Challenge: Hitting a Wall in Enzyme Optimization

A leading biotech company had invested significant resources into improving an enzyme’s activity and specificity. After five intensive rounds of directed evolution, they had achieved an 8x increase in activity and a 12x increase in specificity. However, they had reached a plateau.

Further progress was blocked by a critical lack of information. The enzyme had no known crystal structure (bear in mind that this was long before the release of Alphafold 2), very few similar sequences were available to serve as references, and there were no appropriate templates to build a high-quality homology model. This made standard, structure-based computational design methods impossible.

Our Solution

To bypass the need for structural data, our team deployed a novel sequence-based machine-learning algorithm.

• The algorithm first learned from a massive, general sequence space.
  
• It then focused on the enzyme’s evolutionary context to identify key functional regions.
  
• Finally, we used part of the client’s experimental data from their earlier efforts to fine-tune the model, teaching it to discriminate between active and non-active variants.

This approach allowed us to pinpoint the most promising mutations without ever seeing the enzyme’s 3D structure. To compare our results from the experimental ones, we started from the wild-type sequence, achieving better results with less lab effort. We performed two additional enzyme engineering rounds with this methodology to guide the search.

Results & Impact

• Specificity Boost
  Delivered a 17x increase in specificity, significantly outperforming the client’s previous best results of a 12x increase.
• Maximum Efficiency
  Traditional directed evolution efforts had screened over 32,000 variants. Our approach identified only 67 highly-specific, prioritized candidates for wet-lab validation, representing a remarkable 99.8% reduction in screened variants. This dramatic increase in efficiency de-risked the R&D process, saving significant time and resources.
• Validating a Novel Approach
  This case study successfully validated our sequence-based machine-learning algorithm for enzyme engineering, which laid the foundation for  ZYMEVOLVER 2.0, our next-generation platform for enzyme optimization.

Highlights

• 17x Improvement in specificity
  Outperformed client’s best results.
• 67 Sequences sent to lab
  A 99.8% reduction in the number of variants screened compared to the client’s original efforts (over 32,000 variants).
• 6 Months total project timeline
  From initial proposal to confirmed lab results.
• Validated a novel algorithm
  We validated a new sequence-based machine-learning approach for designing enzymes without structural data.