Designing Enzymes on Demand: How AI Is Rewiring Industrial Biotech

acib_open_innovation-pn0gwbmnm7mep6gvfoep1yj3d1ipxdg4q1xmb83d78
Author: Martin Trinker
Enzymes are nature’s master chemists: they can make (or break) molecules with impressive selectivity, under mild conditions, and often with far less waste than many traditional chemical routes.

Enzymes are nature’s master chemists: they can make (or break) molecules with impressive selectivity, under mild conditions, and often with far less waste than many traditional chemical routes.

For most of industrial history, enzymes were something you discovered in nature, not something you “made to order.” If you needed a catalyst for a reaction, you searched organisms, screened thousands of candidates, then improved the best hits through rounds of mutation and testing. This has delivered many success stories, but it also created a stubborn bottleneck: when the reaction is new-to-nature, hard to screen, or needs unusually high selectivity, the discovery-and-evolution route can become slow, uncertain, and expensive.

Computers started entering enzyme engineering decades ago, first as helpful calculators: modelling structures, estimating stability, predicting which mutations might fit, or simulating substrate binding. These tools were powerful, but they mostly accelerated incremental improvement of enzymes that already existed. The big dream was always bigger: to design a catalyst from scratch for a target reaction, like an engineer designs a machine.

Over the last years, that dream became real in flashes. We learned how to “sketch” catalytic ideas in silico, how to position key residues, and how to build new protein scaffolds. But there was a repeated catch: many designed enzymes showed low initial activities, forcing developers to compensate with large screening campaigns and extensive optimization, which limits how quickly and broadly enzyme design can translate into real industrial value.
Now, a new generation of methods is closing the gap. The most exciting development is not simply “better prediction,” but a new ability to generate enzyme structures that place catalytic atoms with near-atomic precision, and to do so in a way that produces high success rates in experiments. That is the central story of “Riff-Diff” (rotamer inverted fragment finder–diffusion), which combines machine learning with atomistic modelling to scaffold catalytic arrays into entirely new proteins.

A quick historical arc: how we moved from “assist” to “design”

A useful way to understand today’s leap is to separate three phases.
  • Phase 1 was computational assistance i.e. using modelling to help pick mutations, improve stability, and rationalize what experiments show. Valuable, but still fundamentally dependent on natural starting points.
  • Phase 2 was de novo enzyme design: trying to build catalytic function into a brand-new scaffold by placing a catalytic “idea” (the essential reactive residues) into a designed protein. This proved the concept, but the field repeatedly ran into the same industrial pain point: “It works” often meant “it works a little,” and getting to process-relevant performance typically required large experimental iteration.
  • Phase 3 is where we are now: generative AI plus atomistic refinement, focused specifically on what makes enzymes so valuable …

The core problem: enzymes don’t just need an active site – they need the right active site in the right pocket

To understand why this is so hard to achieve in de novo design, it helps to zoom in on what makes an enzyme an enzyme.

An enzyme is not “just a protein that binds something.” The magic is in the precise geometry and chemical environment around the reaction’s transition state. The catalytic residues must be positioned with very tight tolerances, and the substrate must be held in a pocket that encourages the right reaction pathway rather than side reactions.

Earlier computational strategies often struggled with two connected issues.

First, even if you know which residues you want (a catalytic “array”), embedding them into a brand-new protein so that every key atom is positioned correctly is difficult. Previous attempts at grafting catalytic arrays showed activity depends on how precisely the array is reproduced in the final active site.

Second, enzymes need binding pockets that bury and orient the substrate in a way that resembles natural enzymes. Using RFdiffusion’s substrate potential can reduce clashes, but it may fail to promote well-defined binding pockets – creating a trade-off between avoiding clashes and achieving strong substrate interactions.

Riff-Diff (rotamer inverted fragment finder–diffusion) is essentially a set of innovations aimed at solving those two pain points at once: high-precision catalytic placement plus realistic binding pockets.

How Riff-Diff works

Riff-Diff starts from what truly matters for catalysis: the catalytic “motif.” Instead of designing an entire enzyme at once, begin by defining the critical catalytic residues (the functional groups that must sit in exactly the right geometry for the reaction). Then build a protein around that motif while preserving precision. The method scaffolds these catalytic motifs into new protein backbones using diffusion-based structure generation, combined with atomistic refinement. The activity of the enzyme depends strongly on how precisely catalytic arrays are reproduced in the final active site, so the workflow is engineered to protect that geometry. Force the design to create a real pocket, not just “no clashes.” The authors note that avoiding clashes alone is not enough; many methods struggle to generate well-defined binding pockets that hold substrates in a productive orientation. Riff-Diff adds pocket enforcement so the enzyme has a realistic cavity and channel architecture rather than a vague surface groove. Refine structure and sequence iteratively with multiple tools, then rank in the ligand-bound state. The pipeline uses ligand-aware sequence design and structural refinement (described in the project offer as a hybrid diffusion/atomistic workflow using tools such as LigandMPNN, Rosetta FastRelax, ESM/AlphaFold, and ligand-state ranking). The goal is a compact set of candidates that are already “decision-ready,” rather than a massive library that still needs brute-force wet-lab screening.

The results: “one-shot” designs that start exceptionally strong

In a recent Nature Paper (https://www.nature.com/articles/s41586-025-09747-9) a group of scientists around Gustav Oberdorfer tested the method on two mechanistically distinct reactions, which is important because it suggests a general principle rather than a one-off success.
Case 1: retro-aldolases with high activity, stability, and stereoselectivity

For retro-aldol enzymes, the authors produced 35 designs and used SAXS measurements to confirm the expected fold for 29 of them, while CD and gel filtration confirmed all designs expressed as soluble, α-helical monomers.

That alone is a practical step forward: “made it and it folded” is not a given in de novo protein design.

Even more striking is catalytic performance. The top designs (RAD35 and RAD29) achieved kcat values around 3 × 10⁻² s⁻¹ and an estimated ~5 × 10⁶-fold rate acceleration over the uncatalysed reaction. The authors emphasized that these enzymes were orders of magnitude faster than previous computationally designed retro-aldolases.

The designs also show strong stability and selectivity signatures. CD melting curves indicated high thermodynamic stability up to 95 °C for most designs, and two designs achieved ~1,000 turnovers, with one showing 99% enantiomeric excess.

Finally, the paper reports that high-resolution structures of multiple designs reveal near-atomic active site precision – exactly the kind of evidence that turns “it worked” into “we understand why it worked,” which is what industry needs if it wants predictable design.

Case 2: Morita–Baylis–Hillmanases with unusually high “hit rates”
For the Morita–Baylis–Hillman reaction (a different chemistry problem entirely), the success rate is remarkable: in endpoint assays, 94% of designs in one set and 93.3% in another showed product formation above background and outperformed small-molecule nucleophile catalysts like imidazole and DMAP. In industrial terms, this is not a small detail. A high “hit rate” means fewer dead ends, less screening burden, and a faster path to the variants that are worth process development.

Why this is a big deal right now, not “someday”

Industrial biotechnology is entering a period where competitiveness increasingly depends on speed: speed to prototype, speed to a robust catalyst, speed to a cleaner route, speed to scale. What this new generation of computational enzyme design offers is not merely a better tool for scientists – it is a shift in project feasibility.

When “find a biocatalyst” becomes “design a compact set of plausible catalysts with high hit rate,” more companies can consider biocatalysis for steps they previously ruled out. And when catalysts start closer to performance targets, the remaining optimization becomes more focused, more predictable, and more economical.

That is the real significance of reaching “one-shot” capability: it changes what industrial teams can rationally attempt within standard R&D timelines.

While Riff-Diff’s message is not that directed evolution will disappear in the future, but that it enables a future where the starting point is far closer to what industry needs – sometimes even comparable to variants that previously required intensive in vitro evolution.

Riff-Diff shifts the whole economic picture in several ways

First, it reduces the “screening tax.” If you can produce more designs that are already functional, you spend less time on “finding anything that works” and more time on “optimizing the best candidates for your real-world process.” Second, it expands the design space to reactions that are hard to screen at scale. The classic paradigm – low initial efficiency compensated by high-throughput screening and directed evolution – is not always suited when the reaction is not found in nature or is difficult to access through screening. Better one-shot designs are especially valuable in exactly those cases. Third, it supports a move toward more sustainable synthesis. Enzymes can reduce hazardous reagents, enable milder conditions, and improve selectivity (reducing downstream separations). When computational design becomes more reliable, it becomes easier to imagine bespoke enzymes for cleaner routes to intermediates, fine chemicals, and specialty ingredients – especially where stereoselectivity matters.

What “industrial opportunities” looks like in this new era

At acib, we see these advances as more than a scientific milestone. They point to a practical partnership opportunity: turning state-of-the-art computational enzyme creation into deployable industrial solutions. Many companies are now asking variations of the same strategic question. Can we replace (or shorten) multi-year enzyme discovery and optimization cycles with a faster workflow that starts from a defined reaction need and ends with robust candidates ready for process development? Methods like Riff-Diff suggest the answer is increasingly “yes,” especially if industry and applied R&D partners define the target transformation, performance specs, and manufacturability constraints from day one. In real industrial settings, the sweet spot is often one of these scenarios. You have a valuable chemical step with an undesirable reagent, solvent, or waste stream and want a biocatalytic alternative. You need more control over stereochemistry or regioselectivity than traditional chemistry is giving you. You have an existing biocatalyst but want a step-change in activity, temperature tolerance, solvent tolerance, or substrate scope. You want a catalyst for a reaction that is not naturally accessible and would otherwise require long exploratory screening.

We’re entering the “designable enzyme” era – and now is the time to translate it into products

Computational enzyme design has been on a long arc: from early modelling and incremental mutation guidance, to first proof-of-concept de novo catalysts, to deep-learning-driven structure prediction, and now to diffusion-enabled generation of new enzyme backbones with high-precision active sites and realistic pockets. The Riff-Diff work is exciting because it attacks the historic bottleneck head-on: low initial activity and high experimental burden. It demonstrates a pipeline that can produce catalysts with strong activity, stability, stereoselectivity, and near-atomic agreement between design and structure across different reactions. acib is actively looking for industrial translation opportunities built on this new generation of computational enzyme design, especially where partners have a clear target transformation and a business case for faster, greener, and more selective catalysis.
Picture by acib