Summary
A bioRxiv preprint describes a physics-based framework that combines machine learning, molecular mechanics and microkinetic modelling to predict enzyme reaction rates and selectivity. In tests on an oxidative amidase, the authors report near-experimental accuracy across substrates, mutations and enzyme homologues.
Researchers have presented a computational framework intended to make enzyme engineering more physically grounded and scalable. The bioRxiv preprint, posted on September 17, 2026, combines field-aware machine learning with molecular mechanics and microkinetic modelling to connect molecular interactions with measurable catalytic rates and product selectivity.
The authors applied the framework to a newly engineered oxidative amidase called OxiAm. They report that it predicted catalytic rate constants with near-experimental accuracy, distinguished quantitatively between two competing reaction outcomes, and worked across different substrates, mutations and enzyme homologues. The study also used transition-state analysis to identify features of the reaction mechanism and guide the design of enzyme variants for pharmaceutical synthesis.
A model built around reaction energetics
Enzyme engineering changes an enzyme's amino-acid sequence or surrounding conditions to alter how efficiently and selectively it catalyses a chemical reaction. Conventional rational design often depends on structural information, experimental screening and expert judgement. A central challenge is that an enzyme's performance depends not only on its static three-dimensional shape but also on changing atomic interactions as a reaction proceeds.
The new framework focuses on the reaction's free-energy landscape: the set of energetic states and barriers that molecules pass through during catalysis. Lower barriers generally correspond to faster elementary reaction steps, while differences between competing pathways can influence which products form. A useful model therefore needs to represent both the molecular environment around the reacting atoms and the sequence of steps leading to each possible product.
According to the preprint, field-aware machine learning is combined with molecular mechanics to represent enzyme electrostatics at quantum-mechanical accuracy while allowing efficient atomistic exploration. Molecular mechanics provides a computationally tractable way to calculate forces and energies for large molecular systems. The machine-learning component is intended to improve how the model represents the electronic environment relevant to the reaction.
The framework then uses microkinetic modelling. This approach combines the rates of individual reaction steps into predictions for overall catalytic rates and selectivity, including cases with competing, multistep pathways. In the authors' design, the molecular calculations supply free-energy information and the microkinetic model translates that information into experimentally meaningful outputs.
Testing the approach on OxiAm
The researchers used the oxidative amidase OxiAm as a test case. An amidase catalyses reactions involving amide bonds; in this system, the model examined whether the reaction proceeded through hydrolysis or aminolysis. Hydrolysis uses water to break a chemical bond, whereas aminolysis involves an amine and can produce a different product. Predicting the balance between these pathways is important when an enzyme is being developed for selective synthesis.
The authors report that the framework predicted catalytic rate constants close to the measured experimental values. They also report that it quantitatively resolved selectivity between hydrolysis and aminolysis, rather than treating the enzyme as having a single undifferentiated reaction outcome. The reported generalisation across substrates, mutations and enzyme homologues suggests that the method was used beyond one fixed enzyme sequence and one reaction setting.
Transition-state ensemble analysis provided a mechanistic layer to the predictions. A transition state is the high-energy configuration through which reactants pass as chemical bonds are broken or formed. Examining the collection of configurations around this point can show which interactions influence the reaction barrier and can indicate where sequence changes may improve a desired pathway.
The work is a bioRxiv preprint rather than a peer-reviewed journal publication. The supplied record presents the results as a demonstration on OxiAm and describes the reported accuracy and generalisation, but does not provide numerical rate constants, error values or the size of the tested substrate and variant sets in its abstract. Its practical significance therefore lies in the proposed modelling strategy and its initial application, while broader use across industrial or therapeutic enzyme programmes will depend on further testing.
If validated across more reaction classes, the approach could help shift enzyme design from selecting variants mainly through static structural reasoning and experimental iteration toward calculations that explicitly account for reaction dynamics, energy barriers and competing pathways. That would be particularly useful for biocatalysts in which product selectivity matters as much as reaction speed.