Ligo Biosciences wants enzyme design to take days, not years

Ryan Bednar9 min read
Ligo Biosciences wants enzyme design to take days, not years

Ligo Biosciences wants enzyme design to take days, not years

In May 2024, Google DeepMind published AlphaFold3 in Nature. The paper described a model that went well beyond its predecessor, and it arrived without the code. Researchers got a web server with usage caps and restrictions on commercial work, plus a description of a method that nobody outside the company could run.

The reaction from the field was loud. Hundreds of researchers signed an open letter arguing that a method published in a journal should be reproducible by the people reading it, and Nature took public criticism for making the exception. DeepMind later said it would release the model for academic use.

A few people decided not to wait for that. Three researchers who had met in a synthetic biology lab at Oxford started rebuilding the model from the paper, in public, under a license that let anyone use the result commercially. The repository collected more than a thousand stars and a reputation that preceded the company attached to it.

That company is Ligo Biosciences, a Y Combinator Summer 2024 startup. The reimplementation was never the business. Ligo is building design models for enzymes, aimed at industrial chemistry, and the line the founders use for what they want is worth quoting directly: enzymes for each reaction, in days rather than years.

The rebuild was the resume

Reproducing a model like AlphaFold3 from a paper is not a weekend of copying pseudocode. The published description leaves gaps, and the Ligo team hit several of them. They have written up the places where they had to depart from the supplementary pseudocode to get training to work, including reordering part of the model so a later stage could still update its internal representation, and adopting a loss-scaling approach from the diffusion literature. They were candid about which of those choices they were least sure of.

The engineering was the harder half. They adapted an attention kernel to keep memory use manageable, wrote custom GPU kernels, and reused proven components from OpenFold rather than rewriting everything. Their own profiling found that over 60 percent of the model's operations were memory bound, which is the sort of thing you only learn by actually training the model rather than reading about it.

They put the whole thing on GitHub under Apache 2.0, training code included. The license did more work than the star count. When DeepMind released AlphaFold3's code later that year, the weights came with academic-only terms. Ligo's version is the one a company can legally build a product on, and that distinction is why their repository is still in use.

The project is also honest about its state. It covers single-chain protein prediction, with support for other molecule types still in training, and the authors describe it as research software rather than something production ready. None of that blunted the signal. For a startup whose business depends on training frontier-scale biology models, rebuilding DeepMind's flagship from a paper is a more persuasive credential than a deck.

Enzymes are the cheapest chemistry nobody can order

Enzymes are nature's catalysts, and they are remarkably efficient at it. Reactions that otherwise demand high heat, high pressure, or expensive metal catalysts can often run in water at mild temperatures when a suitable enzyme exists. Industry has been buying this advantage for decades. Enzymes are why laundry detergent works in cold water, and they appear throughout pharmaceutical and food manufacturing, usually because the enzymatic route uses less energy and produces less waste than the conventional one.

The problem has never been demand. It is supply. For most reactions a chemical company would like to run, no suitable enzyme exists, because evolution was not optimizing for anyone's factory. The ones industry already uses are largely the lucky overlaps, cases where something nature built happened to be close enough to something useful.

When the overlap isn't there, the standard answer is directed evolution: start from the closest natural enzyme available, generate variants, screen them in the lab, keep what works best, and go around again. The technique is good enough to have earned Frances Arnold a share of the 2018 Nobel Prize in Chemistry, and it underpins most industrial biocatalysis today.

It is also slow and expensive, and it inherits its starting point. Every cycle costs real laboratory time, a commercially useful result can take years of cycles, and the search never strays far from the protein you began with. Directed evolution improves what already exists. It does not produce a catalyst for a reaction that nothing in nature comes close to performing.

Why the folding models didn't already solve it

Ligo's founding argument is that the modeling boom of the last few years, AlphaFold included, did not fix this gap.

Structure prediction answers a different question. It tells you what shape a sequence will fold into, which is useful but not the same as telling you what that shape will do chemically. Generative protein design has gotten good at producing stable new folds and molecules that bind to a chosen target. Binding, though, is a weaker requirement than catalysis. A designed protein can hold onto a molecule and leave it chemically untouched, and from a manufacturer's point of view that result is worthless.

The company's framing is that no existing model has learned the principles of catalysis itself, which is why the tooling still mostly nudges enzymes that already exist instead of designing for a reaction a customer names. Their stated goal is a model that starts from the reaction a chemist wants to run and designs a candidate catalyst for it. Independent benchmarks of that claim are not public yet, which is the honest status of a company a couple of years old working on one of the harder problems in computational biology.

Three founders, one lab, one pub

The team is small and specific. Edward Harris is CEO, with a computer science background from Princeton followed by medicine at Oxford. Emily Egerton-Warburton, the chief scientific officer, is a biochemist. Arda Göreci, the CTO, studied cell and systems biology at Oxford and has described getting pulled into deep learning for biomolecular design by the original AlphaFold work. They met in an Oxford synbio lab, and by their own telling the company idea took shape in a pub.

That mix is rarer than it sounds. Plenty of teams have the machine learning to train a large model, and plenty of labs have the biochemistry to know which catalyst would be worth money. A team that can do the GPU kernel work and the wet-lab reasoning in the same room gets to iterate on both halves without a translation layer in between.

The decision to go public with the AlphaFold3 work before having a product looks, in hindsight, like the best distribution move available to a company with no marketing budget. Their posts drew attention from researchers and investors, which is how the YC application happened. For a technical startup selling to a skeptical industry, the open repository functions as proof that the team can do the work.

Open source as the opening move

Ligo's structure makes more sense once you notice how much of its early public surface is a gift to its own field.

Open weights and training code for a model like this lower the floor for everyone, academics included, and that is the stated motivation behind the reproducibility complaints that prompted the project. It also builds a constituency. Researchers who run your implementation become people who know your company exists, read your next release, and occasionally fix your bugs. The repository invites exactly that, and the authors note plainly that bugs and subtle errors are likely.

The commercial product does not have to be open for this to pay off. Ligo has been building out data infrastructure with partners, and in a field where the training data is the moat, that work is where a defensible position would come from. Giving away a reimplementation of someone else's architecture costs the company very little and buys it credibility, recruiting reach, and a seat in the conversation. Keeping the enzyme models and the data pipeline proprietary is where the business lives.

What has to be true

The case for Ligo rests on a few things that are not settled yet.

The models have to work on reactions that matter commercially, and that has to be demonstrable rather than asserted. The public record right now is an open-source release, named data partnerships, and a founding story with good coverage, not measured activity data on designed catalysts. The company is young enough that this is reasonable, and specific enough in its claims that the evidence will eventually be checkable.

The chemical industry also has to buy. Manufacturers are conservative for sound reasons, since a catalyst that fails at scale takes a production line with it. Selling a designed enzyme means clearing validation hurdles that have nothing to do with how good the model is. Companies in this space usually get there through partnerships and pilot programs rather than a product launch, which is slower than software but does build real switching costs once a process depends on you.

Both of those are the ordinary risks of selling computational biology to industry. What makes Ligo interesting is that the team picked the part of the problem the current generation of tools skipped. Structure prediction is largely solved and increasingly commoditized. Binder design is crowded. Catalysis is the step where the value is, and the step where the models still fall short.

If enzyme design turns into something a chemist can request for a specific reaction and receive in a week, a lot of industrial processes get cheaper and cleaner without anyone outside those plants noticing. Ligo is betting that getting there depends less on model size than on treating catalysis as its own problem, separate from folding, and starting on it early. The open-source work suggests a team willing to do the unglamorous version of that bet in public.

Related Posts