Machine learning reshapes yeast enzyme to favor a stubborn aldehyde and survive heat

Enzymes rarely give engineers everything they want at once. Make a catalyst faster and it often becomes fragile; make it sturdier and it frequently slows down. A team reporting in Biotechnology for Biofuels and Bioproducts has now shown that a carefully assembled machine learning pipeline can push against that trade-off in yeast alcohol dehydrogenase I, an NAD+-dependent enzyme that normally prefers short aldehydes but was retargeted here toward trans-2-nonenal, a nine-carbon unsaturated aldehyde best known as the molecule behind age-related body odor. The study, published on 1 November 2026, combined docking simulation, saturation and random mutagenesis, and three predictive models to find variants that cut trans-2-nonenal down to trans-2-nonenol while refusing to fall apart at elevated temperature.

The target enzyme was not chosen at random. Yeast alcohol dehydrogenase I is commercially available, highly active with ethanol, and has been studied so thoroughly that its kinetic and structural properties are well mapped. It is a tetramer of four identical 347-residue subunits, and its activity declines as the carbon chain of the alcohol lengthens. That bias toward short-chain substrates is exactly what limits most alcohol dehydrogenases when industry wants them to process medium- and long-chain aldehydes, which feed into value-added products such as long-chain fatty alcohols used in pharmaceuticals and healthcare and intermediates relevant to hydrocarbon fuels. Shifting the specificity window toward longer substrates, the authors argued, could reduce competition with native cellular reactions and widen the enzyme’s industrial utility.

Trans-2-nonenal gave the project a second hook. The volatile unsaturated aldehyde arises from oxidative degradation of fatty acids in human skin and has drawn attention in the cosmetic and healthcare sectors as a biomarker of aging and oxidative stress. Its reduced form, trans-2-nonenol, is less volatile and less toxic and has potential as a fragrance ingredient. Converting one into the other with an engineered enzyme therefore offered both a demanding test case for expanding substrate range and a plausible industrial bioconversion.

The engineering began structurally rather than computationally. Using the crystal structure of YADH bound to a nicotinamide-8-iodo-adenine-dinucleotide cofactor analogue, the researchers docked trans-2-nonenal into the active site with AutoDock 4.2.6 and found ten residues within 4 Å of the ligand: T45, W54, L56, H66, W92, Y119, C153, M270, Y294, and V295. Two of these, H66 and C153, coordinate the catalytic zinc and were off limits, so the team built alanine substitutions at the remaining eight positions and measured how each change affected activity on trans-2-nonenal. Alanine at T45 or V295 hurt the enzyme, while swapping out the bulky aromatics W54, W92, Y119, M270, or Y294 left activity intact or better. Those five permissive positions became the mutagenesis canvas.

From that canvas the researchers constructed a library of 190 variants carrying single or multiple substitutions at the five sites, assayed dehydrogenase activity by following NADH consumption at 340 nm, and quantified thermostability as the change in melting temperature relative to the wild type in a fluorescent thermal shift assay spanning 37 to 98 °C. The screen found 19 variants with improved activity toward trans-2-nonenal and three with enhanced thermostability. Single substitutions told a nuanced story: ten W54 replacements raised activity at the cost of stability, W92F did the same, and among the M270 mutants only M270L gained stability while M270G, M270N, M270Q, M270S, and M270V gained activity. One variant, Y119F, modestly improved both properties at once, and Y119W raised stability while lowering activity. Substituting Y294 reduced both traits, and most multi-substitution combinations lost activity entirely or dropped below detection.

That lopsided dataset, roughly half the variants inactive, shaped the machine learning design. Rather than forcing one model to handle everything, the team split the problem. A support vector classifier trained on all variants estimated the probability that a given sequence retained any detectable activity, using the experimentally determined detection limit of 0.01 relative activity as the active-inactive boundary. A support vector regression model, trained only on active variants and using z-scale physicochemical descriptors of the five mutated positions, predicted the magnitude of activity on a log2-transformed scale, so that a twofold activity gain corresponds to roughly one unit. A second support vector regression predicted ΔTM, the melting-temperature shift relative to wild type. Hyperparameters were tuned with Optuna over 50 trials using five-fold cross-validation, with Spearman’s rank correlation as the objective for the regression models and ROC-AUC for the classifier.

The three predictions were then fused into a single ranking score: the probability of activity multiplied by predicted activity multiplied by an exponential stability weighting, exp(ΔTM/τ), with the scaling parameter τ set empirically to 5 °C. At that setting a predicted stability gain or loss of 5 °C shifts the weighting term by roughly a factor of e, and a sensitivity analysis across τ values from 1 to 15 °C showed the top 22 candidates stayed 95.5 to 100 percent identical. The score was applied to all 205 unexplored five-site combinations, and the 22 highest-ranked variants went to the bench. As a control on enrichment, 10 variants were drawn at random from the lower-ranked pool and characterized under identical conditions.

The validation was decisive. Of the 22 top-ranked variants, 21 retained catalytic activity; one, LWMMY, produced no soluble protein and was dropped. Four variants, FWMMY, FWYGY, LWFMY, and YWYGY, exceeded wild-type activity on trans-2-nonenal, with the best performer more than doubling specific activity, and LWFMY also showed improved thermal stability. Every one of the 10 randomly sampled low-ranked variants, by contrast, showed no detectable catalytic activity and substantial decreases in thermal stability. Kinetic characterization sharpened the picture: wild-type YADH has a KM of 1.5 ± 0.3 mM and a kcat of 1479 ± 231 min−1 for trans-2-nonenal, while LWFMY reached a kcat of 25,187 ± 6012 min−1 with a KM of 13.0 ± 3.3 mM, a seventeen-fold turnover gain paid for with reduced affinity that only matters at low substrate concentrations.

Substrate specificity profiling across aldehydes from C2 to C10 revealed that the engineering did more than speed up one reaction. Wild-type YADH favors short-chain aldehydes from C2 to C4. FWMMY, W54F, and W54L broadened activity toward C5 and C6 substrates, and Y119F and Y119M kept the wild-type profile while boosting short-chain activity. The striking cases were FWYGY and YWYGY, which lost activity toward the native C2 to C4 substrates entirely and became selective for longer chains, especially heptanal and trans-2-nonenal, a complete specificity shift from acetaldehyde to the non-native C9 target. The corresponding single mutant M270G showed activity only toward C7 and C9 aldehydes. Docking simulations offered a structural rationale: replacing the bulky methionine at position 270 with glycine enlarged the binding pocket, widened the entrance measured between residues 270 and 294 from 5.9 Å to 10.4 Å, and reoriented the hydrophobic chain of the substrate by roughly 5.6 Å at its distal end while keeping the aldehyde group near the catalytic zinc.

The stability results rounded out the dual optimization. After 10-minute incubations across 55 to 75 °C, wild-type activity collapsed sharply while Y119F and LWFMY held on, with half-inactivation temperatures of 59.8 ± 1.3 °C and 60.7 ± 0.4 °C respectively, about 3 °C above the wild type’s 57.0 ± 0.3 °C. At 60 °C the half-lives of Y119F and LWFMY were 29 and 19 minutes against 13 minutes for the wild type. The authors are candid about limits: predictive accuracy was constrained by the small, imbalanced training set in which high-performing variants are rare, and screening at 10 mM substrate likely enriched variants suited to high-substrate conditions, a bias that may suit industrial reactors but not low-substrate assays. Still, the framework cut the experimental search space dramatically, recovered beneficial mutations from limited data, and showed that machine learning can find mutational combinations that partially escape the activity-stability trade-off that usually constrains enzyme design.

Subject of Research: Machine learning-guided protein engineering of yeast alcohol dehydrogenase for dual optimization of substrate specificity and thermostability

Article Title: Machine learning-guided dual optimization of the substrate specificity and thermostability of yeast alcohol dehydrogenase

Article References: Ogawa, Y., Kondo, A., & Shirai, T. (2026). Machine learning-guided dual optimization of the substrate specificity and thermostability of yeast alcohol dehydrogenase. Biotechnology for Biofuels and Bioproducts, 19(1), Article 89. https://doi.org/10.1186/s13068-026-02839-2

Image Credits: AI Generated

DOI: 10.1186/s13068-026-02839-2

Keywords: yeast alcohol dehydrogenase, machine learning, protein engineering, substrate specificity, thermostability, trans-2-nonenal, directed evolution, biocatalysis, support vector regression, enzyme kinetics, Machine, learning-guided