theshape

Volume I

The First Folds

Biology

MMXXVI

— frontispiece —

Volume I

The First Folds

i

An opening

“Before the form, there is the wish to fold.”

opening epigraph · draft

ii

The Body's Short Words

1

Semaglutide took about forty years. From the first time someone looked at the incretin effect and figured GLP-1 was the molecule doing the work, to the week a tweaked version of it became the drug your relatives have heard of. Forty years of some of the best people the field has, spent on a single peptide.

We file that under success. I've never managed to read it as anything but slow.

2

And slow is still the rate. The newest is retatrutide, reta if you're close to it. It hits three receptors at once and the weight-loss numbers are the best anyone's posted, and after the better part of a decade and a small army, it's still in trials, not something a doctor can write you yet.

The body makes thousands of these, signaling peptides, and that's before you get to the ones cut out of bigger proteins, or those that mean nothing until the cell adds a modification the genome never wrote down. Nearly every one of them could be another GLP-1, and almost none has been started on. Our fastest proven route from sequence to medicine, with the incretin program standing in as the best case and not the average, runs to something like a couple of decades a molecule.

3

Do that arithmetic against the thousands the body actually runs on, and we never finish.

Peptides being good drugs was settled a hundred years ago, by insulin. The pace never changed: a century of knowing these are some of the best medicine a body will ever hand us, and we still pull them out one forty-year haul at a time.

We can sequence a peptide in an afternoon, have it synthesized by Friday, and still not read it.

4

Read, in the dull literal sense. You give me the sequence, I tell you what the thing does once it's in a body, but not whether it sticks to the receptor you already suspected. I mean the shapes it takes while it's drifting around in water, and 'takes' is already the wrong word, since it never holds still long enough to take one. It moves through a population of them, and the population is the behavior you're trying to predict. And that happens to be the one situation a tool designed to find a single native structure is worst at, however good it is at finding that structure.

5

That's why the system that cracked protein folding didn't just shrink to peptide size and crack this too. Folding was a search for one answer. Peptides don't give you one, they give you a fog, and you have to describe all of it. Almost nobody has, because proteins got solved and that's where everyone went.

6

If you could read them, a row of doors that have been shut for decades comes open. Take the body's own antibiotics, the host-defense peptides: a graveyard for thirty years. They look great in a dish, and then a real body clears them within minutes, so they never make it to a patient. But they kill bacteria physically, by tearing the membrane open, about the only mechanism a microbe has real trouble evolving around. Turning one into a usable drug, though, meant tuning a shape nobody could see, so it got done by hand, a few analogs a paper, year after year. Make the shape legible and that slog collapses into a search you can run.

7

Then there's the personalized cancer vaccine. It runs on a guess: which of a patient's own tumor peptides will end up displayed on the cell surface in a form some T cell will register. We are bad at the guess, and that is much of why the response rates keep underwhelming. Under the immunology, it's reading again.

A lot of things that look unrelated are the same problem underneath. Most of how your immune system tells you from a pathogen is this. The neuropeptide signaling nobody can drug, too. And the venom work, where people still pull one good hit at a time out of snakes and cone snails. The biology is all there, we just can't read what's in front of us.

8

And no, I did not solve peptide structure. But something I'd built for a problem with no peptides in it at all turned out to read them, what they do and where the chain folds back on itself, at roughly the level of the systems that lean on the big labeled corpora and the training budgets everyone assumes you need. Mine leaned on neither, and it was tiny. Someone with the proper training would most likely have known the approach was hopeless, which is more or less why it was left lying around for me.

9

And it reads in seconds. That part I didn't expect to matter as much as it does. The usual way to get near this kind of answer is to chain a structure predictor into a dynamics run and put a model on top to read what comes out, a job you submit and come back to, sometimes the next day. Mine returns the answer while you're still sitting there. When a read costs a day you ration it, spending it only on candidates you already trust. At a second you stop rationing. You can turn it loose on a whole library, or drop it inside a loop that redesigns against its own output, and you start asking things nobody asked before because asking had a price. You're not speeding up the old work, you're doing work that was never on the table.

10

Why it works takes the rest of the book. The thing's shape happened to fit the problem, and I'll stand behind that even when it costs me. The usual tools don't, and you can't buy your way past that with compute. I'm not putting a name to the shape yet. The later chapters are how I earn it.

This is Volume One, biology. The First Folds, because small folded things are where I started, and they turned out to be the easiest place to read. Nobody expects that, because the field keeps filing a peptide as a protein that came up short. There are more volumes after it, and an ending I'm keeping to myself. Inside this one the order runs peptides, then proteins, then tissue. You're at the smallest.

11

Peptides are just the first thing it read. I run a small operation that builds one structure, and intelligence emerges within it rather than getting wired in piece by piece. It doesn't tell domains apart, so point it at the next layer and it does the same thing, because to it a domain is just more to take in. It never branches into a product line, it just stays one thing and gets wider, one volume at a time.

Next chapter is one of the doors. What's past them all isn't a theory of everything either. Just something I still can't put a name to, and I'm afraid you and I are going to have to figure it out together.

12

— colophon —

The First Folds is one half of Volume I.

The other half is the model.

They open together.

Request the Volume →

13

A new kind of intelligence.

Hold what cannot be held.

Continuous reasoning

Generative compression

Cross-domain transfer

Separate surfaces of the same architecture.

Across thresholds, recognition.

An ongoing series.

The First Folds

Solve et Coagula

The Hidden Harmony

Apokalypsis

Older than Any Language

One touch. Endless rings.

For researchers

For organizations

The First Folds

The First Folds

Contents

The Body's Short Words