Folding a protein without being shown one

Predicting a protein’s shape from its sequence was one of the great open problems, and machine learning has largely cracked it by learning from the enormous library of structures that experimentalists solved over decades. That work showed the answer is computable.

Our question is different. Can the fold be derived? Given only the atoms in the chain and the world the chain is sitting in, with no library of solved proteins to learn from, does the right shape emerge?

To the engine, this is the same operation as finding the shape of a water molecule, at a much larger scale, with one crucial addition: the surroundings. A protein in empty space is not the protein of interest. The fold belongs to the chain plus its water, its salts and its neighbours.

Where this stands

This one is planned, not done. We have a design for how the pieces of the engine fit together to attempt it, and an honest list of what stands in the way.

• Speed. Even a tiny protein of ten amino acids is far larger than anything the engine has computed so far. Getting there is an engineering problem, not a new-physics problem, but it is a large one.
• Surroundings. The engine needs to place layers of water around the chain and treat near water in detail and far water more simply.
• Search. A small protein has more possible shapes than could ever be tried one by one. The search has to be clever.

The first real test

Take one of the smallest proteins known to fold on its own. Give the engine its atoms and its water. Compare the shape that comes out with the shape measured in the lab.

If it lands close, the claim has teeth. If it doesn’t, the way it misses will tell us which piece of physics we left out.

A test we are especially curious about

Take the water away and see whether the protein comes undone. If the fold really belongs to the chain and its world together, it should.