Imperagen raises £5M to make enzyme engineering less dependent on lab guesswork

Imperagen has raised a £5 million seed round to develop its computational enzyme engineering platform, which combines quantum physics-based simulation, custom AI models, robotics, and automated lab feedback.

PXN Ventures led the round, with IQ Capital and Northern Gritstone participating. The Manchester biotech spinout has now raised £8.5 million in total. It was founded in 2021 by Manchester Institute of Biotechnology scientists Dr. Andrew Currin, Dr. Tim Eyes, and Dr. Andy Almond.

The company has also named Guy Levy-Yurista as CEO. He brings experience across AI, life sciences, and enterprise technology. His job is straightforward: turn university-grade science into a platform industrial customers can use.

That matters because enzyme engineering has attracted plenty of AI optimism, and biology keeps punishing shallow software claims.

Why enzyme engineering is still hard

Enzymes are biological catalysts. They speed up chemical reactions, often with high specificity, which makes them valuable in drug manufacturing, food production, agriculture, biofuels, and materials processing. A better enzyme can mean cleaner chemistry, lower temperatures, fewer toxic reagents, higher yields, or cheaper production.

The problem is search space.

A protein’s function depends on its amino acid sequence and 3D structure. Even modest enzymes can have hundreds of amino acids. Mutating one site may help. Mutating three may break folding. Changing a residue near the active site may improve binding while reducing stability at industrial temperature or pH. The number of possible variants blows up fast.

Traditional enzyme engineering often uses directed evolution: create mutations, test variants, keep the useful ones, repeat. It works. Frances Arnold won a Nobel Prize for it. But it can be slow, expensive, and limited by what the lab can physically screen.

AI tools can narrow the search, but they run into a hard constraint. Models trained on weak or poorly matched biological data tend to produce tidy predictions that fail in wet-lab reality. A variant may look promising under assay conditions and then fall apart when moved into industrial production.

Imperagen is trying to close that gap.

The technical bet: physics, AI, and lab feedback

Imperagen says its platform has three main parts:

1. Quantum physics-based simulation for enzyme variant behavior
2. Custom AI models trained around specific enzyme engineering problems
3. Robotics and automation to generate experimental feedback in a closed loop

The order is worth noting.

Many AI-for-biology companies start with sequence data. Feed enough protein sequences into a model, learn patterns, predict variants. That approach has produced useful tools, especially for structure prediction and protein design. Enzymes, though, are chemistry machines. Their value depends on transition states, active-site geometry, electrostatics, solvent effects, binding dynamics, and reaction energetics.

Imperagen’s pitch is that quantum physics modeling can explore enzyme mutations computationally before the company spends lab time on them. In practice, “quantum physics-based simulation” likely means computational chemistry methods that estimate molecular behavior at the electronic level, especially around the active site where the reaction happens.

Full quantum treatment of entire proteins is usually too expensive, so serious systems tend to use hybrid methods, approximations, or focused modeling around chemically relevant regions. That’s normal computational chemistry. The trade-off is accuracy versus cost.

If the simulation is too simplified, it may miss protein-wide effects such as folding, long-range dynamics, aggregation risk, expression problems, and stability under process conditions. If it’s too detailed, throughput suffers and the method becomes impractical for screening millions of variants. The useful middle is hard to find.

Imperagen says it can explore millions of mutations computationally, then feed those outputs into custom AI models. That could be valuable if the physics engine produces informative features rather than another pile of noisy labels. A model trained on enzyme-specific simulation outputs plus targeted lab results may generalize better than one trained only on public protein databases.

The closed-loop lab work is the part developers and data scientists should watch most closely.

Closed-loop biology is an MLOps problem with pipettes

Closed-loop enzyme engineering means prediction doesn’t end the process. The system designs variants, tests them in the lab, captures assay data, updates the model, and proposes the next batch. Robotics and automation matter because biological feedback is slow and expensive compared with software telemetry.

This is active learning applied to wet-lab science.

Instead of randomly testing thousands of enzyme variants, the system can choose experiments that are likely to improve the model or identify useful candidates. Sometimes that means chasing likely winners. Sometimes it means testing uncertain variants because they expose where the model is wrong.

For technical teams, the useful questions are operational:

• How does the platform represent enzyme variants, structures, assay conditions, and simulation outputs in one data model?
• How does it track provenance from predicted mutation to physical sample to assay result?
• How does it distinguish failed biology from failed measurement?
• How does it handle batch effects across robotic runs?
• How often does the model retrain, and what triggers a new design cycle?
• How are uncertainty estimates used when selecting the next variants?

This is where many AI-biology systems get messy. Lab data is not clickstream data. Experiments drift. Reagents vary. Instruments need calibration. A failed assay may reflect a bad construct, a pipetting issue, a folding failure, or a real loss of activity. If the data pipeline treats all failures the same, the model learns the wrong lesson.

A strong closed-loop platform needs boring infrastructure: sample tracking, metadata discipline, versioned models, assay normalization, quality control, and audit trails. It also needs scientists who know when a model is confidently wrong.

Industrial scale is the credibility test

Levy-Yurista told TechCrunch that current enzyme engineering methods often fall short, including AI-powered approaches that pass trial-and-error stages but fail in practice at industrial scale.

That criticism is fair. A variant that works in a small assay may not survive the conditions customers care about. Industrial enzymes may need to tolerate heat, solvents, pressure, impurities, long shelf life, or unusual feedstocks. They may also need to express efficiently in microbial hosts and remain cheap to produce.

The commercial bar is unforgiving. A pharmaceutical manufacturer doesn’t adopt a new biocatalyst because the model architecture is elegant. It adopts one because the enzyme improves yield, purity, cost, process safety, regulatory posture, or supply-chain resilience.

For food, agriculture, and biofuels, price sensitivity can be even harsher. A better enzyme still has to fit existing production constraints. If it requires exotic conditions or expensive downstream handling, the technical win may not matter.

This is why Imperagen’s go-to-market plan matters as much as the science. The company says the new funding will go toward hiring AI specialists, research and development, expanding lab capabilities, and building a commercial function over the next two years. That suggests Imperagen is still moving from platform validation into broader customer work.

It’s early.

A crowded field

Imperagen sits in a market that includes companies such as Biomatter, Cradle Bio, and Absci, each attacking protein or enzyme design from different angles. The broader category has been energized by advances in protein language models, diffusion-based design, structure prediction, and high-throughput screening.

The technical split is becoming clearer.

Some companies lean heavily on generative AI. Some focus on experimental throughput. Some build around structure-based modeling. Imperagen is emphasizing physics-informed simulation, custom AI, and automated feedback. That mix makes sense for enzymes because catalytic activity depends on chemistry in ways broad sequence models may not capture.

But “physics-informed AI” can turn into a vague label without results. The useful evidence would be concrete:

• Hit rates compared with conventional directed evolution
• Number of design-build-test cycles needed to reach target performance
• Improvements in catalytic efficiency, selectivity, stability, or expression
• Transfer from computational prediction to industrial conditions
• Cost and time reductions against real customer projects

Benchmarks in enzyme engineering are difficult because each enzyme and substrate pair is different. Serious customers will still ask for comparative data, not platform language.

Why developers and AI engineers should care

Imperagen is a useful example of where applied AI is going in scientific domains: domain-specific systems that combine simulation, machine learning, automation, and feedback.

For AI engineers, the approach reflects patterns showing up across materials science, chemistry, drug discovery, and synthetic biology.

First, data generation becomes part of the product. If public data is sparse or poorly aligned with the target task, the company needs its own experimental engine. That changes the economics. The model is one component. The lab is part of the training system.

Second, simulation still matters. It can provide structure, priors, and synthetic signal. In enzyme engineering, physics can constrain the search space and help models avoid biologically silly suggestions. The hard part is knowing when simulation artifacts are contaminating the learning process.

Third, deployment means changing physical workflows. A recommender system can ship daily. A biocatalyst has to be manufactured, validated, integrated into a process, and often scrutinized under regulatory or quality systems. The feedback loop is slower, and mistakes cost more.

For technical decision-makers evaluating these platforms, the procurement question should be practical: does the vendor deliver improved enzyme candidates, or does it deliver software that still leaves your team doing most of the scientific heavy lifting?

A good platform should reduce cycles. It should also explain failures well enough that teams can make better next decisions.

The funding is modest, but the problem is worth solving

£5 million is not a giant round by AI infrastructure standards, but for an early UK biotech platform company, it’s meaningful. It gives Imperagen room to hire, automate more of the lab, and prove whether its physics-plus-AI workflow can produce commercially useful enzymes faster than conventional methods.

The promise is attractive. Better enzymes could make chemical manufacturing cleaner, improve drug production, reduce waste, and open routes to bio-based products that currently don’t make economic sense.

The risk is plain too. Biology has a long record of humbling computational certainty. Models that look strong in narrow experimental settings can fail when substrate, host organism, process conditions, or scale changes.

Imperagen’s funding gives it a chance to show whether its platform can handle those realities. The evidence will be in enzymes that work outside the demo.