Heat can help an industrial reaction run efficiently, yet it can also cause the enzyme driving it to lose its shape and activity. DeepEvo, an artificial-intelligence approach reported in 2024, searches for enzyme sequences that remain functional at higher temperatures.
In a study published in BioDesign Research, Huanyu Chu and colleagues combined computational design with laboratory testing. Their main demonstration used glyceraldehyde-3-phosphate dehydrogenase, or GAPDH, an enzyme involved in the breakdown of sugars. The team produced variants with improved high-temperature performance, while also showing which designs failed along the way. Original study.
How DeepEvo searches for heat-tolerant enzymes
The approach links two models. A Variant-generator learns patterns in sequences from an enzyme family and produces new sequences. A Thermo-selector then screens for features associated with high-temperature tolerance.
The selector was trained using proteins from organisms with different optimal growth temperatures. Those labels provide a large training resource, but describe the organism rather than directly measuring each protein’s stability. They are useful clues, with imperfect predictive power.
Selected sequences were fed back into the generator’s training process. The aim was to enrich its output for promising designs. The researchers then applied further filters, including conserved functional features and confidence in predicted structures, before choosing proteins to make and test.
Which designs worked in the laboratory?
From a larger computational collection, the researchers selected 30 sequences for experimental testing. Of those, 23 produced soluble proteins that could be purified, and 17 showed catalytic activity at 30°C. Eleven had detectable activity at 65°C; eight stood out for relatively strong high-temperature tolerance.
These stages separate proteins that could be produced from those that also worked in the heat. Because the candidates had already passed computational filters, the results describe a selected set rather than a success rate for arbitrary enzyme sequences.
The researchers also tested the closest natural sequence counterparts of the eight highlighted variants. Seven designed variants had a marked activity advantage at 65°C over counterparts with little or no detectable activity in that assay. This comparison helped show that the search could improve performance rather than simply recover sequences resembling naturally heat-tolerant enzymes.
Activity and stability tell different stories
An enzyme’s activity at a high temperature and its resistance to thermal unfolding are related but distinct properties. The researchers measured both. The eight highlighted variants also had higher melting temperatures than their corresponding natural homologues, supporting improved thermal stability.
Industrial usefulness involves further questions. An enzyme must perform for a useful length of time, at relevant substrate and product concentrations, and in the presence of other mixture components. Its production yield and cost matter too. The reported laboratory results do not establish those outcomes for a manufacturing process.
The same distinction applies to structural predictions. A confident predicted fold is helpful when selecting candidates, but cannot replace an activity or stability measurement. Our article on AlphaFold 2 and enzyme structure explores that complementary role.
A more focused route to enzyme engineering
DeepEvo’s appeal is the possibility of directing laboratory effort towards sequences more likely to have a desired property. Its generator was trained for an enzyme family, so the main experiment does not establish a universal design system for every protein.
The study nevertheless provides a concrete link between computational selection and measured enzyme behaviour. Applications in food production, chemical manufacturing or environmental processes will depend on testing the relevant enzymes under the conditions in which they would actually be used.
Update note, 27 September 2026: This archived report was revised to clarify the experimental counts, distinguish activity from thermal stability, and qualify proposed applications. Its original publication date and URL are retained.
References
- Chu, H., et al. (2024). High-Temperature Tolerance Protein Engineering through Deep Evolution. BioDesign Research, 6, 0031. DOI: 10.34133/bdr.0031. Open full text.
Image sources
- Structure: PDB 3KV3, the GAPDH reference used for functional-motif selection in the DeepEvo study. It is the C151S mutant from Staphylococcus aureus, not a DeepEvo-designed variant. Mukherjee, S., Dutta, D., Saha, B. and Das, A. K. (2010). Crystal structure of glyceraldehyde-3-phosphate dehydrogenase 1 from methicillin-resistant Staphylococcus aureus MRSA252 provides novel insights into substrate binding and catalytic mechanism. Journal of Molecular Biology, 401, 949–968. DOI: 10.1016/j.jmb.2010.07.002.
- Visualization: original EnzymeNews rendering with Mol*, using RCSB PDB data. Biological assembly 1 is shown as protein ribbons; substrate, cofactor and water are omitted.
