Nobody tells a protein how to fold
Properties are not carried in. They emerge from the whole, and each whole becomes a part of the next.
A protein leaves the ribosome as a string. Nothing hands it a diagram. Within moments it has folded into a precise, working shape, and it did that by responding to everything around it: the water, the salts, the warmth, its own neighbouring atoms.
Change the surroundings and the shape changes. The fold was never written in the chain alone. It belongs to the chain and its world together.
Ice is something water does
A single water molecule is not cold, or slippery, or six-sided. Put enough of them together under the right conditions and a lattice appears that no one molecule contains.
And the molecule itself is changed by the company it keeps. A water molecule inside ice responds to light differently from the same molecule alone in vapour. Its properties are not fixed numbers it carries from place to place. They are answers to the question: in what?
What emergent means
Parts interact, and something new appears with properties of its own, properties that were in none of the parts. Then that new thing becomes a part, and interacts, and something new appears again.
Electrons and nuclei make an atom. Atoms make a molecule. Molecules make a droplet, a crystal, a membrane, a cell. At every step the level above is created by the level below, and is more than a list of it.
Nothing along the way is static. A bond is not a stick with a length. It is closer to an elastic string, whose tension depends on everything else attached to the network. Push on one atom and every string around it stretches.
So what should a design engine learn?
There are two ways to predict what a molecule will do.
One is to study a very large number of molecules whose behaviour is already known, and learn the patterns. This works remarkably well, as far as the examples reach. It proved that the answers can be computed.
The other is to learn what the molecule itself is responding to. The physics is the same for a protein, a pollutant, a pigment and a pane of glass. An engine that understood the physics would not need to have seen your molecule before, or anything like it.
That is the wager we are making. Don’t train on the molecules. Learn the physics, and let the molecules happen.
Known structures are the exam, not the textbook
We do use what is known: measured structures, spectra, binding strengths. But only to check the engine’s work, never to feed it answers.
The engine builds from what things are made of and what they are in. Then we hold the result up against reality. If it matches, that is evidence the physics is right. If it doesn’t, we have found exactly where something is missing, and that is worth as much.
We may be wrong about how far this goes. Finding out is the experiment.