structural-biology/ai produced the result/Nature Chemical Engineering 2024 · v2
Robot lab and learning agent redesign an enzyme to survive higher temperatures
Researchers built a self-driving laboratory in which a statistical learning model chose which enzyme sequences to build, and robots assembled and tested them. Over twenty rounds, four independent agents each found designs more heat-stable than the natural starting proteins.
spectrum · one line per step, placed by what the step does · bright lines used AI
Self-driving laboratories to autonomously navigate the protein fitness landscape
Nature Chemical Engineering, 2024
doi:10.1038/s44286-023-00002-4 · record aix-00004 v2 · checked 2026-10-07
- AI was for
- Property prediction, Classification, Experimental design
- Model family
- Gaussian process
- Checked by
- Experimental4 tested, 4 worked
- Code
- available
The finding the paper is about came from the AI.
What this research was about
Enzymes are proteins that speed up chemical reactions, and each one is spelled out by a sequence of amino acid building blocks. Change a few of those building blocks and the enzyme may work better, work worse, or stop working altogether. Scientists picture this as a landscape: every possible sequence is a point, and its height is how well the protein performs. The trouble is that the landscape is astronomically large, the shape is not known in advance, and the only reliable way to learn the height at a point is to build that protein and measure it. Each measurement costs time and materials, so only a tiny fraction can ever be tested.
This work targeted heat tolerance in a sugar-cutting enzyme family known as glycoside hydrolase family 1. The researchers assembled a space of 1,352 possible sequences by mixing and matching 34 pre-made gene fragments drawn from natural relatives of the enzyme. They then asked whether a machine could run the whole search on its own: decide what to build, have robots build and measure it, learn from the answer and decide again. Four separate agents each began from the same six natural sequences and ran twenty rounds, picking three sequences per round.
Where AI came in
The AI was the part that decided what to make. Each agent used a Gaussian process, a statistical model that fits a smooth guess to scattered observations and reports how unsure it is at every untested point. Here it did two jobs at once: sorting sequences into likely working or likely broken, and estimating the heat tolerance of the ones expected to work. After every round of experiments the model was refitted to everything measured so far and used to predict a value, an uncertainty and a probability of being functional for each sequence not yet tried.
Those predictions fed a selection rule that favoured sequences both promising and still uncertain, weighted by the chance they would work at all. This stood in for the expert judgement a protein engineer would normally supply when choosing the next batch, and the predictions stood in for experiments that were never run on the rest of the space. The chosen designs went straight to the robots for gene assembly, protein production and a heat-stability assay, and the results returned to the model. Each agent tested less than 2% of the space; its top design was then rebuilt and measured by hand.
Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.
The work
Technical · from the record
The SAMPLE platform couples a Gaussian process agent that models protein sequence–function relationships to a fully automated robotic pipeline that assembles genes, expresses proteins in a cell-free system and measures enzyme thermostability. Four independent agents, each seeded with the same six natural glycoside hydrolase family 1 sequences, ran 20 design–test–learn rounds over a combinatorial space of 1,352 sequences, selecting three sequences per round. Each agent found sequences at least 12 °C more stable than the six initial natural sequences while testing less than 2% of the combinatorial landscape, and the top design from each agent was then rebuilt and measured by hand, where all four were more thermostable than the top natural sequence Bgl3.
How AI was used
A multi-output Gaussian process with a linear Hamming kernel was fitted to the study's own measurements, jointly classifying sequences as active or inactive and regressing thermostability for active sequences. The modelling and acquisition choices were first benchmarked on a compiled cytochrome P450 dataset using tenfold cross-validation and 10,000 simulated protein engineering trials comparing random selection, standard UCB, UCB positive and expected UCB, plus batched variants. In the live campaign, each agent refit its Gaussian process to all data it had collected, predicted thermostability, uncertainty and probability of activity for every untested member of a 1,352-sequence combinatorial space built from 34 pre-synthesised gene fragments, and selected sequences by the expected UCB criterion, which multiplies the upper confidence bound by the classifier's probability of activity. Selected designs were sent to the automated pipeline for Golden Gate assembly, PCR verification with EvaGreen, cell-free expression and a T50 thermostability assay, with rule-based quality checkpoints returning failed experiments to the queue; the resulting measurements were returned to the model to close the loop for 20 rounds per agent. A further Gaussian process trained on the pooled data from all agents was used as a reference landscape for analysing each agent's model predictions and decision-making.
The shape of the work
Structural · the record, drawn
AI
Benchmark model and acquisition functions on P450 data
Testing outputs against ground truth.
We tested the multi-output GP model by performing tenfold cross-validation, where a GP classifier was trained on binary active/inactive datawhere the paper describes this · verbatim
no AI
Design combinatorial GH1 sequence space
Producing candidate objects that did not previously exist.
We designed a combinatorial glycoside hydrolase family 1 (GH1) sequence space composed of sequence elements from natural GH1 family memberswhere the paper describes this · verbatim
AI
Fit agent's landscape model to observed data
Fitting model parameters, including fine-tuning an existing model. The AI stood in for expert judgement.
The SAMPLE agent uses a Gaussian process (GP) model to build an understanding of the fitness landscape from limited experimental observationswhere the paper describes this · verbatim
AI
Predict activity and thermostability across the space
Running a trained model over new data to predict, classify or score. The AI stood in for physical experiment.
We plotted each agent’s model predictions for all 1,352 combinatorial sequences over the course of the optimizationwhere the paper describes this · verbatim
no AI
Select sequences to test by expected UCB
Iterative search over a space.
The agents designed sequences according to the Expected UCB criterion, chose three sequences per round, and ran for a total of 20 roundswhere the paper describes this · verbatim
no AI
Robotic gene assembly, expression and thermostability measurement
Physical execution, by hand or by robot. Its result feeds back into an earlier step.
a fully automated robotic system that experimentally tests the designed proteins by synthesizing genes, expressing proteins and performing biochemical measurements of enzyme activitywhere the paper describes this · verbatim
no AI
Human characterization of top designs
Testing outputs against ground truth.
We expressed the enzymes in Escherichia coli and performed lysate-based thermostability assays (Methods).where the paper describes this · verbatim
AI
Unified landscape model for analysing agent behaviour
Extracting understanding from model behaviour.
we trained a GP model on all sequence–function data from all agents, which we refer to as the ‘unified landscape model’where the paper describes this · verbatim
What the record says
Technical · every part carries its own basis
+ in the paper~ our reading− not reported
How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.
The learned agent chose every sequence that was built and measured; the reported thermostable enzymes are the direct output of the model-driven design-test-learn loop
The SAMPLE agent uses a Gaussian process (GP) model to build an understanding of the fitness landscape from limited experimental observationswhere the paper describes this · verbatim
We found that all four machine-designed enzymes were substantially more thermostable than the top natural sequence (Bgl3)where the paper describes this · verbatim
A more complete set of data including the code to interpret the data is accessible atwhere the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- Trained model weightsWhether the trained model is available is not stated.
- Version of multi-output Gaussian process (GP classifier for active/inactive plus GP regression for thermostability)Which version of the model was used is not stated.
- What step 1 replacedThe paper gives no basis for what the AI stood in for.
- What step 8 replacedThe paper gives no basis for what the AI stood in for.
About this article
Record aix-00004, version 2, checked by a person on 2026-10-07. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY-4.0; quotations are at most 25 words. How we work · Report an error