Until now, finding these molecular workhorses has been like speed-dating with a blindfold on - expensive, slow, and mostly disappointing. But a team of researchers just taught AI to play matchmaker, and the results are turning heads in biotech labs worldwide.

The Enzyme Discovery Problem (It's Worse Than You Think)

Enzymes are biology's catalysts - proteins that make chemical reactions happen faster, cleaner, and at room temperature instead of requiring industrial furnaces. They're already everywhere: in your laundry detergent breaking down stains, in cheese production, in the synthesis of pharmaceuticals. The global enzyme market is worth tens of billions of dollars annually.

But here's the catch: when scientists want an enzyme for a new reaction, they're essentially searching through a database of millions of protein sequences hoping to find one that works. Traditional methods involve expressing hundreds of candidate proteins in the lab, purifying each one, and testing them individually. This process costs months of labor and substantial resources per successful hit.

The research team, led by scientists at a synthetic biology company called Tierra Biosciences, decided there had to be a smarter way.

Teaching AI to Think Like a Biochemist (Sort Of)

The approach uses something called contrastive learning - a technique where AI learns by comparing things rather than memorizing labels. It's the same general idea behind how you learned to distinguish cats from dogs as a kid: not by studying textbook definitions, but by seeing lots of examples and noticing what makes them different.

The researchers built a dual-encoder system. One encoder processes protein sequences (the amino acid chains that make up enzymes). The other encoder processes chemical reactions (what goes in and what comes out). The AI learns to match them up - essentially creating a shared "understanding" of which proteins catalyze which reactions.

They trained this system on hundreds of thousands of known enzyme-reaction pairs scraped from biochemistry databases. The model, which they call CLEP (Contrastive Learning of Enzyme and reaction Pairs), learns to embed proteins and reactions into the same mathematical space. Proteins that catalyze similar reactions cluster together, even if their sequences look nothing alike.

The Part Where They Actually Test It

Here's where this paper stands out from the usual AI hype cycle. The team didn't just publish benchmark numbers and call it a day. They ran actual wet-lab experiments.

They used CLEP to search for enzymes that could perform specific reactions of industrial interest. The model ranked candidates from a database of over 30 million protein sequences. Then they picked the top predictions and actually expressed and tested them in the lab.

The hit rate was striking. For some reaction types, over 70% of the AI's top picks showed genuine catalytic activity. Compare that to traditional homology-based searches, which often yield hit rates in the single digits for novel reactions. The model was finding functional enzymes that sequence-similarity methods would have completely missed.

One particularly satisfying result: CLEP successfully identified enzymes from organisms that had never been experimentally characterized before - proteins sitting in genomic databases that nobody had ever tested. The AI essentially mined dark biological matter for useful catalysts.

Why Dual Encoders Beat Single-Track Thinking

Previous approaches to computational enzyme discovery typically worked in one direction: you have a known enzyme family, and you search for related sequences. This works great when you want more of what you already have. It falls apart when you need something genuinely new.

The dual-encoder setup flips the script. Instead of asking "what proteins look like this one?", it asks "what proteins could catalyze this reaction?" The reaction encoder captures the chemistry you care about, independent of any particular protein family.

This matters because enzymes that evolved independently can perform identical reactions through completely different structures - a phenomenon called convergent evolution. Single-encoder methods miss these. Dual encoders catch them.

The Bigger Picture

This isn't just about faster enzyme hunting. It's about what becomes possible when discovery accelerates by an order of magnitude.

Enzyme engineering currently has a brutal economics problem: the upfront cost of finding a starting point often kills projects before they begin. If AI can reliably surface good candidates from sequence databases, the calculus changes. Reactions that were "too expensive to pursue" become tractable. Sustainable chemistry gets cheaper. Drug synthesis pathways that required rare or expensive catalysts might find alternatives hiding in microbial genomes.

The researchers note limitations, of course. CLEP learns from known enzyme-reaction pairs, so it's bounded by existing biochemical knowledge. It won't discover entirely unprecedented chemistry. And lab validation remains essential - the model predicts potential activity, not guaranteed function.

But as a first-pass filter to narrow millions of candidates down to dozens worth testing? That's exactly where AI shines.

What Comes Next

The team has made their model available, which means other labs can start stress-testing it on their own problems. Expect to see follow-up studies applying CLEP to everything from carbon capture enzymes to antibiotic biosynthesis.

The underlying approach - contrastive learning across different data modalities - is gaining traction across biological AI. Similar architectures now connect proteins to text descriptions, to 3D structures, to gene expression profiles [3]. Each new pairing teaches AI something different about how biology works.

For enzyme discovery specifically, the next frontier is probably integrating these learned representations with protein structure prediction tools like AlphaFold. If you can predict both what a protein does and what it looks like, rational enzyme design starts looking less like alchemy and more like engineering.

Until then, there's a whole lot of sequence database left to mine. And apparently, the AI is getting pretty good at finding the gems.

References

1. Rocks JW, Truong DP, Rappoport D, et al. Dual-encoder contrastive learning accelerates enzyme discovery. Proceedings of the National Academy of Sciences. 2025. DOI: 10.1073/pnas.2520070123. PMID: 41843673

2. Rives A, Meier J, Sercu T, et al. Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences. PNAS. 2021;118(15):e2016239118. DOI: 10.1073/pnas.2016239118

3. Lin Z, Akin H, Rao R, et al. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science. 2023;379(6637):1123-1130. DOI: 10.1126/science.ade2574

4. Yu T, Cui H, Li JC, et al. Enzyme function prediction using contrastive learning. Science. 2023;379(6639):1358-1363. DOI: 10.1126/science.adf2465

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.