Decision to improve
How can researchers choose a small, informative set of Rubisco variants from a protein sequence space too large to test exhaustively?
What shapes the outcome
Protein engineering offers far more possible sequences than a laboratory can test. Measurements are expensive, biological constraints matter, and the best next experiment may be the one that teaches the most rather than the one with the highest predicted performance.
Constraints and uncertainty
- The possible sequence space is too large for exhaustive laboratory screening.
- Available measurements cover only a small fraction of possible variants.
- Each experimental round must balance predicted performance, uncertainty, and a diverse set of candidates.
How the system works
- Mapped protein sequences into numerical representations using protein-language-model embeddings.
- Used active learning to balance testing promising regions with exploring uncertain ones.
- Optimized each round around which variants would be most useful to test next.
- Kept the computational ranking connected to practical experimental validation.
What this system enables
The workflow turns a vast sequence space into a prioritized, informative experimental set and creates a repeatable loop in which each result can improve the next selection. The work was published in AI Chemistry in 2026.
What I built
I developed the scientific decision framing, protein representation, active-learning strategy, and candidate-prioritization workflow.