Will Your Next Drug Owe Google Rent?

Google can now leave its signature, i.e. a “watermark” inside a protein. Yes… a protein sequence, and not the text on your research paper, the actual protein you make in the lab.

Just yesterday, Google DeepMind introduced SynthID Bio and here is how it works. When an AI designs a protein, it often has several acceptable amino acids to choose from. Google nudges some of those choices to create a hidden statistical pattern. Make the protein, test it, and the pattern can still be detected.

The team tried this with binders against three different targets and reported comparable binding performance to the designs of the original proteins (the ones without the watermarks).

Google says this could help DNA synthesis companies identify where unfamiliar sequences came from and stop AI-generated material being mistaken for natural biology in public databases. All of which is sensible and probably useful too, but once again, I’m going to put on my tinfoil hat here...

Just a few weeks ago, Anthropic announced plans to watermark Claude’s writing, and a lot of people weren’t thrilled. They worried their work would be branded as AI-generated when they had merely used Claude as an editor or translator. Anthropic says the mark identifies model involvement, not the person using it.

The crux of it is that many didn’t want to lose ownership or credit of their own writing, but this gets much more interesting when we’re talking about a billion-dollar drug.

So imagine a biotech that spends years developing a therapeutic protein. It generates its own experimental data, does the lab work, patents the molecule and eventually gets it through clinical trials. Somewhere in the discovery process, it used a commercial AI platform. And suppose its finished protein still carries that platform’s watermark.

Five years later, the model provider comes knocking. Its contract included a royalty on drugs developed using the platform. That watermark is effectively the unique ID that proves this company used their platform. Suddenly, the company supplying the AI gets a cut of a drug it never manufactured, tested in patients or brought to market.

Now, this arrangement isn’t as exotic as it sounds. Iambic’s February deal with Takeda includes potential royalties on medicines arising from their AI drug-discovery collaboration. That’s a negotiated research partnership, not a watermark-based tollbooth, but it tells us the royalty business model already exists.

Google offers another precedent. Its publicly released AlphaFold 3 model weights and their outputs are restricted to certain non-commercial uses, while its terms explicitly say Google doesn’t claim ownership of original outputs. Now, AlphaFold 3 is a different tool, and those aren’t royalty terms. Still, owning the result and controlling access to the tool are already two different things.

To be clear, Google hasn’t announced any royalty plan for SynthID Bio. It's sharing research materials, and the current watermark can be weakened or removed through protein redesign. So today’s technology is hardly an automatic royalty collector.

But imagine future models that are substantially better than the alternatives, with stronger watermarks and commercial contracts attached. You could own your patent and still owe someone rent because of the tool you used along the way.

And here’s the irony, AlphaFold 3 itself was trained using experimental structures from the public Protein Data Bank, built through contributions from researchers around the world.

Big Tech has invested enormous resources in turning accumulated scientific knowledge into powerful tools. Yes, that work has incredible value, but we may eventually find ourselves paying to access tools built on the decades of science we collectively produced, and now we have to pay for again if these tools help us invent something worth selling.

Of course, I hope we never get there, but greed is a powerful incentive.

The NIH launched Linked Discoveries, an AI-powered tool that helps researchers navigate the growing complexity of scientific literature. The system connects more than 29 million PubMed papers through citations, related studies, retractions, genes, diseases and chemicals, giving researchers a broader view of how different discoveries connect. The goal is to help scientists find relevant evidence faster and improve reproducibility in research. [Link]

Estonia partnered with Owkin to build a “sovereign AI” platform for biology using national health data. The collaboration will apply Owkin’s K Pro AI scientist to secure Estonian datasets, with an initial focus on areas including oncology. The idea is to develop AI capabilities around a country’s own biomedical data while maintaining control over sensitive health information. [Link]

Anthropic’s Claude helped identify a CRISPR-like system, but scientists are debating what counts as an AI discovery. The company used hundreds of AI agents to search genomic databases and reported finding an unusual genetic system, although researchers noted that parts of the finding had appeared previously and still require experimental validation. [Link]