Jyunmi

This sits in the very front of drug discovery pipeline.

Instead of screening millions of random molecules in the lab and hoping something sticks, you let AI generate a small, smarter list of candidates first.

The key question is, does any of this actually work in the real world?

Well, the team at MIT didn't stop at simulations.

They picked 26 different targets and ran eight experimental campaigns in real labs.

That means they actually made some of the AI's designs and tested them in experiments.

On several easier, well-studied targets, the AI designs worked about 80% of the time.

And so that's a big upgrade over throw a huge library at the wall and see what happens.

More interesting are the hard cases.

On nine new, very challenging targets, they decided to be strict.

They only allowed themselves 15 AI designs per target.

And even on that limit, they ended up with very tight binders for six of the nine targets.

you will sometimes see this described as nanomolar affinity.

And all that means is the binder still grabs the target very strongly, even when the binder is present in tiny concentration.

Tiny here means billionths of a mole per liter.

Now, this is not just AI juicing pretty molecules on the screen.

This is suggesting the models, labs built them, and quite a few worked well in the tests.

So let's take a look at Boltzgen under the hood.

Boltzgen is an L-atom model, which means it keeps track of where each atom sits in the 3D space, not just a rough cartoon of a protein.

It learns by doing several related tasks at once.