Foldit

> Is this not a solved problem now?

This is an interesting question.

Can any contemporary “ai” solution be considered a solved problem?

A lot of these neural net architecture are designed first and foremost as approximation engines without any guarantees the provided approximation is even any good.

This can be useful, but especially in a medical sense, also insufficient or even dangerous.

> AlphaFold's protein structure prediction results at CASP were described as "transformational" and "astounding".[93][94] Some researchers noted that the accuracy is not high enough for a third of its predictions, and that it does not reveal the mechanism or rules of protein folding for the protein folding problem to be considered solved.[95] Nevertheless, it is considered a significant achievement in computational biology[92] and great progress towards a decades-old grand challenge of biology.[93]

https://en.m.wikipedia.org/wiki/Protein_folding

Protein folding is still very much an unsolved problem.

What AlphaFold is doing is essentially matching a sequence of amino acids to solved protein structures, albeit dynamically, and with some advanced constraints. In essence, it's not predicting how the protein will fold, but rather, how the protein will look once it's done folding. And in many many cases, there's no matching fold for a sequence.

Also, proteins are dynamic. They interact with other proteins. They sample multiple conformations, they frequently rearrangle, or become disordered, and then re-ordered. AlphaFold is just a single snapshot.

If we want to understand how proteins fold, we need more than just AlphaFold.

The goal of most/all machine learning is hunting down performance maxima in a super high dimensional space. In this case the space is the positions and angles etc of the atoms that comprise a protein and the "performance" is the stability of the protein.

For the number of atoms that comprise most proteins, iterating through all the possible positions would take an unimaginable amount of time, so you have to have some kind of search method to identify good position-space-areas to investigate more closely.

My guess as to where people help in is getting away from bad local maxima. In my experience playing foldit, sometimes you can see pretty clearly that the stability is not good and it's not going to get much better with small changes - the algorithm has found a bad local maxima of performance - so you can manually move big chunks of the protein around to explore a new part of the position-space. This kind of evaluation, knowing when to stop climbing a small hill and instead go looking for bigger hills, seems to be something that humans are pretty decent at. Of course there's also a million algorithms to do the same thing without humans.

Depends. If you manage to fold a prion it will spread to other users and corrupt their memory, introduce odd rendering artifacts and finally brick their computer.

;)

From Washington though.