The engine
MABE is an experiment.
Not a product. Not a tool to ship. A test of one idea: that an engine given only physics can predict and design across every kind of molecular problem there is.
Take only what things are made of, and what they are in. See how far that goes.
Picture a chemistry model kit, the kind with balls and sticks. Now replace the sticks with elastic strings. Bonds stretch, bend and break, and their tension depends on everything else in the network. Push one atom and all the others move. Then put the whole assembly in something: water, salt, heat, a field. The surroundings pull on every string.
MABE is the quantum-mechanical version of that kit. If the idea is right, you should be able to hand it atoms and conditions and watch it build its way up to structures as intricate as proteins. Not because it learned them. Because the physics dictated them.
The rules of the experiment
The constraint is severe, and the constraint is the point.
The engine may use
- What things are made of: which atoms, how many electrons, the constants of nature.
- What things are in: the solvent, temperature, pressure, fields, neighbours.
The engine may not use
- Scores or parameters fitted to known answers.
- Templates or look-alikes from known structures.
- Folding rules learned from a database of solved proteins.
- Another program’s prediction, passed off as its own.
Known structures and measurements are still precious. They are how we mark the engine’s work. They are never how it gets its answers.
Either result is a result
If the engine works in a domain, a class of problems becomes something you can reason out from physics. If it fails, we learn exactly where building up from physics breaks, and what is missing. Failures here are not bugs to be patched over. They are data.
So the engine never says “this is the best.” It gives the leading options, how likely each is to work under your conditions, what could make it fail, and which experiment would tell us most. You can always see which parts are measured and which are predicted.
What it has been pointed at
Ice, proteins, colour, starlight, salmon, medicine, glass. They look like different sciences. They are the same physics, and each carries a plain label for how far along it is.
- In development
How water becomes ice
No molecule carries the plan for a crystal. We are asking whether the lattice, and the way it bends light, can emerge from physics alone.
- Planned
Folding a protein without being shown one
Machine learning proved the folded shape can be computed. We want to know whether it can be derived, from atoms and surroundings alone.
- In testing
From an atom to the colour you see
One unbroken chain of physics from electrons to perceived colour, that can also be run backwards.
- In development
Don't find the star. Design the experiment.
Astronomy can look but not touch. If the physics really is one unbroken chain, a measurement made on a lab bench can be carried all the way out to a star.
- Designed
What the rain carries to the salmon
A molecule shed by tyres is killing coho salmon in British Columbia's streams. A capture system for stormwater.
- In development
Does a binder have to be biological?
Therapeutic antibodies are grown in living cells, and priced accordingly. Binding is physics, so the binder could be made another way.
- In development
Colour without pigment
A butterfly wing is blue with no blue in it. Surfaces that make colour from structure alone.
- In development
A window that manages the sun
Ultraviolet, heat and light handled by the molecular structure of the glass surface. Passive, permanent, no moving parts.