Articles

How Machine Learning Guides the Next Round of Enzyme Screening

Experimental results from successful and unsuccessful variants can help researchers prioritize candidates and explore combinations of mutations.
Written byAimee Cichocki
Presented byFrances Arnold
Laboratory scene with machine learning visualization on tablet

Google Flow 2026

Register for free to listen to this article
Listen with Speechify
0:00
2:00

An enzyme variant that fails to improve performance may still contribute to the next discovery. Its screening result can help a model learn which sequence changes influence activity—and guide researchers toward more promising candidates.

During her Pittcon 2026 keynote presentation, Nobel Laureate, Frances H. Arnold, described how her laboratory incorporates machine learning into directed evolution. The approach changes how researchers use experimental data and decide which variants to test next.

For analytical scientists, it gives screening measurements a second role. Beyond identifying the strongest performers in one round, the results become evidence for planning subsequent experiments.

Learning From More Than the Best Performers

Directed evolution typically advances through repeated cycles of mutation, expression, and screening. Researchers identify improved variants and use them as starting points for further modification.

Arnold explained that this process can require hundreds of experiments. Advancing through small sequence changes can also constrain the search: a useful combination of mutations may be difficult to reach by selecting only improvements at each individual step.

Working in analytical science?

Register for a FREE Separation Science account to subscribe to the Separation Science Newsletter.

Subscribe for free

Machine learning offers a way to investigate more combinations computationally. Researchers train models on measured variants, use predictions to prioritize additional sequences, and then test those candidates experimentally.

A study from Arnold’s group demonstrated this strategy with combinatorial libraries, including an application to enzymes that selectively produce different product enantiomers. The work showed how models trained on tested variants could help identify productive sequences without experimentally screening every combination.

The model extends the search that a laboratory can consider. Measurements still establish whether the proposed variants deliver the desired performance.

Making Screening Data Informative

Arnold emphasized that results from poorer performers can contribute information about sequence–function relationships. Their value, however, depends on what they reveal.

A weak but measurable response may help distinguish how different mutations affect activity. A dataset dominated by variants with indistinguishable, near-zero performance may offer less information for predicting useful candidates.

Research on training-set design from Arnold and colleagues examined this issue. The study found that reducing the inclusion of minimally informative, very low-fitness variants improved the machine-learning-assisted optimization strategy tested.

For laboratories, this makes experimental design part of model development. The selection of variants, the measured endpoint, and the ability to resolve differences all influence what a model can learn.

Choosing a suitable data analysis approach therefore requires attention to the information available in the measurements, alongside the complexity of the prediction task.

Updating Predictions With New Experiments

Arnold described an iterative process in which each experimental generation supplies additional data for model updates. Researchers can then revise their predictions and choose the next set of variants.

She also highlighted uncertainty quantification as a tool for deciding where to investigate next. This creates opportunities to balance testing candidates expected to perform well with experiments that could improve understanding of less-explored sequences.

Generative AI could extend the process further by proposing starting sequences when researchers lack a suitable enzyme to modify. Arnold presented this as a developing opportunity, with evolutionary optimization providing a route to improve initial designs.

The distinction matters: predicting a sequence, identifying detectable activity, and developing a useful catalyst remain separate stages. Each requires experimental evidence.

Machine learning changes how laboratories allocate screening effort and extract value from their results. Analytical measurements remain the link between a proposed sequence and demonstrated function—and the basis for deciding which experiment follows.

This article was produced under Separation Science's AI Editorial Policy.

Add Separation Science as a preferred source on Google

Add Separation Science as a preferred Google source to see more of our trusted coverage

Meet the Author(s):

  • Aimee Cichocki is the Editorial Director at Separation Science and Chromatography Forum. Aimee brings a broad range of experience in creating, editing, and formatting scientific content. With a degree in medicinal chemistry, a 10-year background in formulation chemistry, an MBA, and a diverse background in publishing, Aimee guides editorial initiatives at Separation Science and Chromatography Forum. Aimee is dedicated to ensuring the delivery of informative, reliable, and practical content to our audience of analytical scientists.

    View Full Profile

Speaker

Here are some related topics that may interest you:

Loading Next Article...
Loading Next Article...