[nharada]

From the abstract[1]:

> After decades of effort, 17% of the total residues in human protein sequences are covered by an experimentally-determined structure. Here we dramatically expand structural coverage by applying the state-of-the-art machine learning method, AlphaFold2, at scale to almost the entire human proteome (98.5% of human proteins).

[1] https://www.nature.com/articles/s41586-021-03828-1

[vmception]

Basically they are saying that decades of distributed protein folding was useless and everyone would have had more utility mining cryptocurrency if it existed several years earlier

But at least it inspired someone to make and release this

[dekhn]

you're conflating two different disciplines: distributed protein folding studies the biophysical process of proteins folding over time, while protein structure prediction makes a static single predict of what is believed to be the final structure adopted by the protein in the folding process.

I think many people believe that given infinite computer time the protein folding simulations would produce the same output as the static prediction (modulo a number of complex details) but use far, far more computer time to get there.

The fundamental observation from the DM AF2 paper that I've been able to glean (which I kind of sort of already believed) is that careful multiple sequence alignments of 30-100 evolutionarily related proteins is enough to produce coarse distance constraints that can be used to guide a structure prediction to a good answer quickly. And that depended on new ML technology that didn't exist before.

[vmception]

thanks for that explanation!

[cing]

Just in case you're not joking, it's worth noting that the majority of distributed molecular simulation (past and present) is spent studying "folded proteins" to discover structures of proteins that are often hidden from methods like AlphaFold (currently). For example, https://www.nature.com/articles/s41557-021-00707-0

[ramraj07]

I don't know if you know, but doctors spent 1,300 YEARS using the wrong anatomy book. A few years and compute time isnt the end of the world. I'm sure oracle's DB2 test suite has burned more carbon than protein folding labs have.

[Jabbles]

A third way in which you are wrong is that AlphaFold derives a lot of its power by referring to previously-solved protein structures, or parts of them. It doesn't fold the proteins from scratch in an "alpha-zero" way.

[vmception]

so its more like protein folding was useless until an AI could make sense of the 17% solved variations and using that for the other 83% of proteins found in humans?

> After decades of effort, 17% of the total residues in human protein sequences are covered by an experimentally-determined structure. Here we dramatically expand structural coverage by applying the state-of-the-art machine learning method, AlphaFold2, at scale to almost the entire human proteome (98.5% of human proteins).

I just don't actually understand the quote from the article if it isn't comparing the same thing

[dmitryminkovsky]

> experimentally-determined structure

refers to structures determined by means of physical examination, with like crystallography, not to attempts at predictive computational analysis prior to AlphaFold, which were not accurate compared to AlphaFold.