Protein engineering has always had a bit of casino energy. You make a bunch of mutations, pull the lever, and hope your enzyme comes out faster, stronger, or at least not completely broken. This paper by Li and colleagues takes that whole routine and says: what if the slot machine had a map? I think that is the core idea here, anyway. I have read the abstract twice and the methods summary with the tense body language of a person trying to assemble IKEA furniture from a photocopy.

The enzyme with a day job

The enzyme in question is CYP2C9, a member of the cytochrome P450 family. In plain English: it is one of your liver's chemical workhorses, helping process drugs and other molecules. That matters because tiny sequence changes in CYP2C9 can change how well someone metabolizes medications such as warfarin and phenytoin. So this is not just "can we make a protein do a neat trick?" territory. This is also "please do not accidentally wreck someone's dose response" territory.

The old-school way to improve enzymes is directed evolution: introduce mutations, test them, keep the winners, repeat. It works, but it is slow and expensive. It is natural selection with lab equipment and a caffeine budget. The problem gets worse when function depends on lots of subtle internal interactions, which is where giant protein language models can start looking a little too confident for their own good.

VERnet: less vibes, more function

Li et al. built a model called VERnet to predict how sequence variants affect CYP2C9 function, using deep mutational scanning data as training signal rather than asking a general protein language model to just intuit the answer from sequence alone [1]. If I am reading this right, that is the paper's big philosophical move: stop treating "protein understanding" as one generic superpower and train for the thing you actually care about.

The training data came from earlier large-scale experiments measuring the activity of thousands of CYP2C9 variants [2]. According to the paper, VERnet reached 93.5% accuracy for interpreting CYP2C9 variants and outperformed more general-purpose variant effect predictors, including protein language model-based approaches [1]. That is the part where the anxious overachiever in me stops squinting and goes, OK, that is not nothing.

They also report that a fine-tuned version using AlphaFold3-optimized structural information did better on double-amino-acid substitutions, which is useful because biology loves making the "one weird mutation" story more complicated the second you look away [1].

Why this is more interesting than "AI, but for proteins"

What makes this paper fun is that the model was not just used to score existing variants. The authors used generative AI to guide virtual evolution at conserved positions and then validated selected CYP2C9 variants experimentally [1]. In other words, the model was not merely grading homework. It was suggesting what homework to assign next.

That fits with a broader trend in the field. Recent reviews argue that generative AI is becoming useful not just for making protein-shaped objects, but for linking sequence, structure, and function in a way enzyme engineers can actually use [3-5]. A 2024 Nature Genetics study also showed that protein language models can be strong variant effect predictors at scale, though papers like this new one make the case that task-specific, experimentally grounded models still have a real edge when the biology gets fussy [6].

And enzymes are very fussy. They are not just folded blobs. They are tiny molecular machines where one residue nudged the wrong way can turn "works beautifully" into "chemically unemployed."

The exciting part, with one eyebrow raised

If this kind of approach keeps working, you can imagine a much faster loop for enzyme engineering: screen in silico first, send only the best candidates to the lab, and spend less time mutating proteins like a raccoon rummaging through a silverware drawer. That could matter for drug metabolism research, industrial biocatalysis, greener chemistry, and maybe custom enzymes for therapeutic use.

The real-world mood around AI-designed proteins has also shifted lately. Reviews in Nature Biotechnology and Nature Reviews Bioengineering describe a field moving from "cool demo" to "show me the wet-lab validation" [4,5]. That is exactly why this paper is interesting. It does not just generate protein ideas and wander off like an overconfident intern. It brings receipts.

Still, a few limits are worth keeping on the table. This model is protein-specific, which is also why it works. That means the recipe may not transfer cleanly to every enzyme family. The whole system also depends on having high-quality functional data, and those datasets do not grow on trees. They grow in expensive assays, with graduate students nearby, looking tired.

The bottom line

I think this paper's honest contribution is not "AI has solved enzyme design." Absolutely not. It is more like: general protein AI is useful, but if you want an enzyme to do a particular job, you should train on that job and then test it in the real world. Sensible. Slightly unglamorous. Probably correct.

And honestly, that is refreshing. In a field with plenty of big claims and shiny model names, VERnet's best trick may be that it acts less like a magician and more like a very good mechanic.

References

1. Li C, Xu W, Yang H, et al. Generative Artificial Intelligence-Empowered Virtual Evolution of Enzyme with the VERnet Model. ACS Catalysis. 2026;16(8):7669-7682. DOI: 10.1021/acscatal.6c00759. PubMed: 42022779. PMCID: PMC13097142

2. Amorosi CJ, Chiasson MA, McDonald MG, et al. Massively parallel characterization of CYP2C9 variant enzyme activity and abundance. American Journal of Human Genetics. 2021;108(9):1735-1751. DOI: 10.1016/j.ajhg.2021.07.001

3. Xie WJ, Warshel A. Harnessing generative AI to decode enzyme catalysis and evolution for enhanced engineering. National Science Review. 2023;10(12):nwad331. DOI: 10.1093/nsr/nwad331

4. Yang KK, Wu Z, Arnold FH. Machine learning for functional protein design. Nature Biotechnology. 2024;42:216-228. DOI: 10.1038/s41587-024-02127-0

5. Kwok HF, Zhang Y, Yang M, et al. AI-driven protein design. Nature Reviews Bioengineering. 2025;3:1034-1056. DOI: 10.1038/s44222-025-00349-8

6. Hie B, Shanker V, Xu D, et al. Advancing variant effect prediction using protein language models. Nature Genetics. 2023;55:1812-1822. DOI: 10.1038/s41588-023-01470-3

7. Winnifrith A, Outeiral C, Hie B. Generative artificial intelligence for de novo protein design. Current Opinion in Structural Biology. 2024;86:102794. DOI: 10.1016/j.sbi.2024.102794

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.