The question
How can a large protein sequence space be narrowed to a small, informative set of variants for experimental consideration?
Context
Protein engineering presents a combinatorial search problem. Labels are expensive, biological constraints matter, and the most valuable next experiment may be the one that reduces uncertainty rather than the one with the highest predicted score.
Why the problem was difficult
- Sequence space is too large for exhaustive experimental screening.
- Available measurements are limited compared with the number of possible variants.
- Candidate selection must balance predicted performance, uncertainty, and experimental diversity.
Approach
- Represented protein sequences using protein-language-model embeddings.
- Used active-learning logic to balance exploitation of promising regions with exploration of uncertain regions.
- Framed each round around which variants would be most informative to test next.
- Kept the computational ranking connected to practical validation planning.
Outcome or resulting capability
The resulting workflow provides a structured way to move from sequence representations to a prioritized experimental set. The work was published in AI Chemistry in 2026.
James' role
James developed the scientific-machine-learning framing and candidate-prioritization workflow.