When Biology Became National Infrastructure

South Korea just built its own version of an AlphaFold-style protein model.

It is called K-Fold. The Korea Advanced Institute of Science & Technology (KAIST) says it can predict proteins, protein complexes, small-molecule interactions and other biological structures, and in some internal tests it apparently gets close to AlphaFold3 while running much faster.

Whether those benchmark claims hold up independently, who knows. But I actually think that is the least interesting part of the story.

The really interesting part is why South Korea thinks it needs one at all.

When AlphaFold first appeared, the reaction was basically, look at this incredible thing that AI can do. A problem scientists had wrestled with for decades suddenly seemed solvable.

Now the reaction is starting to look more like, our country should probably have one too.

KAIST actually describes K-Fold as part of something called “sovereign bio AI.” And that sentiment tells you quite a lot about where the world is going.

For most of the past few decades, science moved in the opposite direction. Researchers shared structures through the Protein Data Bank. Genomic databases crossed borders. Drug companies used software built somewhere else. If another country had a better tool, you used the better tool.

That logic is getting harder politically.

The US wants domestic semiconductors. Europe wants sovereign cloud infrastructure. Countries are building their own foundation models. Supply chains are being pulled home. Technology that used to be treated as a product is increasingly treated as strategic capacity.

And now, it seems that biology may be next.

From my view, this is all very strategic because Korea is not only building the model, it is also building a national biomedical dataset, combining genomic and clinical information to support precision medicine and drug development.

Put those two things together and you can clearly see a very focused direction.

A domestic biological model trained on deeply characterized domestic health data looks more like national infrastructure.

There is also a strange possibility here. This more nationalistic world could, in one narrow way, actually help personalized medicine.

A huge amount of biomedical research has historically depended on whichever populations happened to have the best datasets. That has consequences. Genetic variants differ in frequency between populations. Drug metabolism can differ. Disease risks differ. Environmental exposures differ. Models trained heavily on one ancestry do not always transfer cleanly to another.

So a Korean government asking why Korean medicine should depend mostly on biological data collected somewhere else is not an unreasonable question.

Japan can ask the same thing. So can India, Nigeria, Canada, etc.

If enough countries decide they need much better biological data on their own populations, we could end up with a far richer picture of human biology than we have today.

And the funny thing is that better national data may eventually make the national idea less important.

Once you collect enough information, you start seeing all the differences within a population too. Different genetic backgrounds, disease subtypes, exposures, and responses to the same drug.

You start by trying to understand the Korean patient better, and eventually you get closer to understanding the individual patient better.

That is where this gets interesting for personalized medicine.

A more nationalistic world could push countries to build biological datasets they probably should have built anyway.

The risk is what happens if they never share them.

If every country decides its biological data are a strategic asset and locks them away, we do not get better global medicine. We get very sophisticated national silos.

So K-Fold may be less interesting as a competitor to AlphaFold than as a sign of what comes next.

Biology is becoming something countries think they need to control.

Whether that ultimately gives us a richer picture of humanity, or simply a more fragmented one, may depend on what they do with all that data once they have it.

Anthropic is giving Claude the ability to operate laboratory equipment through a new standard for AI-controlled instruments. Genentech has already tested it with liquid handlers, robotic arms and plate readers, pushing AI one step closer to running experiments rather than just analysing them. [Link]

Moderna and Merck’s personalized cancer vaccine uses machine learning to decide which tumor mutations should go into each patient’s treatment. The selected neoantigens are encoded into a custom mRNA therapy, giving one of the clearest examples yet of software helping determine the actual composition of a medicine. [Link]

A veteran medicinal chemist used ordinary ChatGPT and Gemini to help design a new renin inhibitor. Sheo Pharmaceuticals used the models to study decades of failed compounds, then synthesized around 200 molecules before arriving at its lead candidate, without relying on a specialized AI drug discovery platform. [Link]