Enzyme engineering projects fail for a predictable reason: the team treats “make this enzyme better” as one problem when it is really several, with very different success rates. Improving thermostability on a known scaffold is close to routine. Designing catalysis into a protein that never had it is a research programme.
This article separates those cases and describes what computation can realistically contribute to each.
Three problems, three risk profiles
1. Stabilising an enzyme you already have
The most tractable case. You have a working enzyme that denatures at process temperature, aggregates on storage, or dies in organic solvent. The fold is known, the mechanism is intact, and you need it to survive longer.
Computation is genuinely good here. Stability prediction across every possible point mutation, consensus analysis against the family alignment, disulfide engineering, and rigidification of flexible loops identified from simulation all produce candidate mutations with reasonable hit rates. Combining several stabilising mutations is usually additive, though not always.
The main trap is the stability–activity trade-off: the flexible regions you rigidify are frequently the ones catalysis depends on. Mutations far from the active site are safer, and any campaign should screen for retained activity, not just melting temperature.
2. Changing specificity or activity on a known scaffold
Harder, but well within reach. You want an enzyme to accept a different substrate, prefer one enantiomer, or turn over faster.
The computational approach is to model the substrate in the active site, identify the residues controlling the binding pocket’s shape and electrostatics, and generate variants that reshape it. Molecular dynamics shows whether the substrate stays in a productive orientation or drifts, which static docking cannot tell you.
Expect to screen. Computation narrows a combinatorial space to dozens or hundreds of variants; it rarely identifies the winner outright, because activity depends on transition-state stabilisation that current scoring functions model crudely.
3. Designing a new catalytic function
This is the ambitious case, and honesty about it matters.
The approach is to build a theozyme — a quantum-mechanical model of the ideal arrangement of catalytic groups around the transition state — then find or generate a protein scaffold that can present that geometry, and design a sequence around it.
It works, in the sense that de novo enzymes exist. It typically does not work well: designed enzymes usually show catalytic efficiencies several orders of magnitude below natural counterparts. Nearly every notable success was followed by directed evolution, which supplied most of the final activity.
The realistic framing is that computational design produces a starting point for evolution that would otherwise not exist. That is valuable. It is not a finished biocatalyst.
Why designing catalysis is so much harder than designing binding
Modern methods design binding sites well. Catalysis is different in kind.
A binder needs to be complementary to one stable structure. An enzyme must bind the substrate, stabilise a fleeting transition state more than it stabilises the substrate, and release the product. Bind the substrate too well and you have an inhibitor-like dead end rather than a catalyst.
Transition states are also poorly served by the tooling. Force fields are parameterised for stable molecules, not for bond-breaking geometries, so the very step that defines catalysis sits outside what standard molecular mechanics can model. That is why theozyme construction needs quantum calculations, and why the resulting geometric criteria are demanding — sub-Ångström positioning of catalytic groups, which design methods cannot yet reliably deliver.
Where deep learning has changed things
Sequence design has improved markedly. ProteinMPNN and similar methods produce sequences that fold as intended far more often than physics-based design did, which removes one major failure mode: designs that were geometrically correct but never folded.
Structure prediction now lets you check designs cheaply before committing to synthesis, and RFdiffusion can generate backbones conditioned on a fixed catalytic motif, so the scaffold search is no longer limited to structures that already exist in the PDB.
What has not been solved is predicting catalytic rate. Nothing available reliably tells you whether a designed active site will turn over quickly, which is why experimental screening remains unavoidable.
A workflow that reflects the actual risks
- Define the reaction and the mechanism precisely, including which step limits the rate. Projects that skip this optimise the wrong thing.
- Choose the honest starting point. A natural enzyme catalysing a related reaction is almost always a better base than a blank scaffold.
- Model the substrate and transition state in the site, with quantum treatment where the chemistry demands it.
- Generate variants: mutational scanning for stability and activity, or backbone generation with the catalytic motif constrained.
- Validate structurally — predict each designed sequence’s fold and confirm it matches the intent.
- Simulate. Molecular dynamics reveals whether the catalytic arrangement persists or relaxes away within nanoseconds. Active sites that collapse in simulation do not work at the bench.
- Filter for expression and solubility in the intended host before ordering anything.
- Screen a panel, and plan for directed evolution as a follow-on rather than a fallback.
Applications worth the effort
- Industrial biocatalysis — replacing steps that currently need heavy metals, high temperatures or hazardous solvents.
- Biosynthetic pathways — engineering enzymes to accept non-natural intermediates in a designed route.
- Plastic and pollutant degradation, where improving stability and activity of known hydrolases is a well-posed engineering problem.
- Therapeutic enzymes, where stability, immunogenicity and half-life matter as much as catalysis.
- Enzymes for non-natural conditions — solvent, pH or temperature regimes no organism ever needed to survive.
In every one of these, the value of computation is the same: it turns an intractable search space into a screening campaign of manageable size. It does not remove the screen.
