George Church

So to give an example, AlphaFold...

Last time I checked, anyway, at least it's all changed.

If you substitute an alanine for a serine in a serine protease, it will have exactly the right fold.

It will be precise to a fraction of an angstrom overall average.

but it won't function.

It just won't function.

And that's where you need either extraordinary precision or just knowledge of what happens evolutionarily or happens in experiments to say that no, an alanine won't work.

And so I think

There's all kinds of combinations of AI tools that can give you deeper insight into that.

I mean, I think the way that it's working now, which will get us a long way, won't get us the whole way, is we have something that kind of works, and we make libraries inspired by that, make variations on it.

And then whichever of those variations work, we make variations on that, and we can just keep going.

It's kind of like the way evolution worked, except different.

Now we can do it at incredibly high speeds.

And in principle, evolution might incorporate a few base pair changes in a million years.

Now we can make billions of changes in an afternoon.

And it's all guided in such a way that you get rid of the wastefulness of having a bunch of neutral mutations and a bunch of lethal mutations.

You can have things that are quasi-neutral but likely to be game-changing, have more of a focus on those.

Another thing that's been missing and none of the

AI protein design tools that I know of are particularly good at it yet, but we're trying to, as we speak, trying to improve this, is non-standard amino acids.

Because a lot of these tools depend on having libraries of 3D structures which use 20 amino acids and large language models where you line up all the sequences of 20 amino acids.