AI Designed a More Stable Enzyme. Did It Preserve Activity?

If an AI-designed enzyme handles heat better, I'd still want to know how much useful chemistry it can do.

As I learn more about protein engineering, I'm interested in how teams make that choice. A stability result can help us choose candidates, but the enzyme still has a job to do: turn a substrate into a product.

Before moving a variant forward, I'd want evidence for both.

What the AI actually did

In EvoSeq-ML, published as an accepted manuscript on November 3, 2026, researchers trained generative models using datasets enriched with ancestral sequences: computational estimates of earlier evolutionary relatives.

The models proposed new ethylene-forming enzyme sequences. The authors report better measured thermostability than for designs generated from modern-sequence data. All tested ancestral and ML-generated enzymes produced ethylene, but the ML-generated variants had lower activity than wild type.

That does not establish a universal stability–activity trade-off. It shows why these properties need separate measurements.

Use the thermal screen to compare thermal behaviour

Intrinsic differential scanning fluorimetry, or DSF, follows changes in a protein's natural fluorescence as it heats up. Those changes can reveal unfolding transitions and an apparent melting temperature, usually written as Tm. This is a measurement of thermal behaviour, not catalytic activity. A review of DSF methods explains the fluorescence readouts and their interpretation.

For a larger set of suitable variants, I would consider a plate-based screen. The system measures intrinsic fluorescence in standard 384-well plates using 10 µL sample volumes.

That format leaves room to compare variants alongside controls and repeat measurements without using a large volume of each protein. I'd use it to identify reproducible differences worth following up, provided the samples give an interpretable fluorescence signal.

Measure activity under conditions that matter

Alongside the thermal measurements, I'd use a functional assay that tracks substrate consumption or product formation. I would also check how activity holds up over the intended operating time.

These are different questions. A heating ramp and a prolonged reaction do not test the same behaviour. A 2026 review of enzyme stability in biocatalysis makes this distinction between structural stability and the ability to retain activity under process conditions.

My decision table would look like this:

Hypothetically, one variant might start faster but lose activity sooner. Another might start slower and keep producing for longer. I'd choose between them using the product made over the time that matters for the application.

Use stopped-flow when the timing could explain the difference

If two promising variants behave differently in the activity assay, the next question may be why.

Stopped-flow rapidly mixes reactants and records a changing signal. It can help investigate early reaction behaviour before steady-state turnover is established, often called pre-steady-state kinetics.

The Applied Photophysics SX20 supports absorbance and fluorescence measurements. That gives a team options for following a suitable reaction, depending on which observable signal reports the step of interest.

A useful manufacturer example is chymotrypsin breaking down p-nitrophenyl acetate. Using SX20 absorbance measurements, the application note shows an initial burst of product formation followed by slower steady-state production. An endpoint measurement would lose that timing information.

For an engineered enzyme, that kind of time course could help test whether a particular reaction phase differs between variants.

I would reserve it for candidates where understanding the reaction could change the next design decision.

Give the next model more than a winner label

For teams building AI design tools, this is where the experimental results become especially useful.

I would keep thermal behaviour, catalytic activity and retained activity as separate measurements linked to each sequence. I'd also retain the assay conditions, controls and uncertainty. A missing activity measurement should stay missing; it should not become a zero or an automatic failure.

That preserves the information needed to decide what “better” means. A team can set a minimum activity requirement, then compare the stability of candidates that meet it. Another project might prioritize total product formed over a longer reaction.

Desmond Gilmour

Email: dgilmour@nicoyalife.com

Phone: +1 519-546-9876

I work at Nicoya Lifesciences. The views expressed here are my own and do not necessarily reflect those of Nicoya. This is a personal publication, not an official company communication.