← THING3 Prototype_001
The Loop
A study in learning from reality.
The idea
Engineering a better enzyme means searching an astronomically large space with a small testing budget. Change just three of the 317 building blocks in the enzyme shown, and there are 36,071,686,770 possible versions. A lab tray holds 96. So how do you choose what to test?
Why make it playable
The answer is easy to agree with and easy to forget when you just read it. So the piece asks you to commit to a prediction first, then run the strategies yourself on the same budget and watch the turn happen. We won’t spoil it here.
What’s in it
- 01 Real world
- A real enzyme (lipase B), and what swapping one of its parts can do.
- 02 The tray
- Why you can’t test everything.
- 03 The map
- Every version on one picture.
- 04 Your call
- Predict, then run three strategies.
- 05 The check
- One question to see whether the idea carried over, and a sample of the report a real audience could produce.
What’s real, and what’s illustrative
Real
- The enzyme’s structure (PDB 1TCA)
- The size of the search, and its arithmetic
- The yellow colour test used for lipases
- The fact that swapping serine 105 stops it working
- The published finding that learning from each round of results can reach better variants with fewer tests
Illustrative
- The landscape and the scores
- The “AI”: simple stand-ins, not any company’s technology
- The audience figures in the report
Simplified
- A flat map stands in for a space with many dimensions
- “Better” is a single number
If this were yours
In a piece made for you, the subject, the audience and the brand are yours, and the claims are reviewed with your experts. Where it’s useful, the questions and the report can run on your real audience, with consent and measures agreed up front. For a few investors, we’d test it in conversation instead.
Sources
- Directed evolution: Frances H. Arnold, Nobel Prize in Chemistry 2018.
- Machine-learning-guided evolution: Wu, Kan, Lewis, Wittmann & Arnold (2019), “Machine learning-assisted directed protein evolution with combinatorial libraries”, PNAS 116(18).
- Yang, Wu & Arnold (2019), “Machine-learning-guided directed evolution for protein engineering”, Nature Methods 16.
- Structure: PDB 1TCA, lipase B from Candida antarctica. Uppenberg, Hansen, Patkar & Jones (1994), Structure 2(4).
Independent THING3 prototype, built from public research. The simulation and the report’s audience figures are illustrative; your own answers are real and stay in your browser.