A next-gen AI protein folder that could help science? Meta's good for something AI researchers at Meta say they have developed the largest protein-folding model of its kind to date, and that it is capable of predicting the structure of more than 600 million proteins. The team released the 15-billion-parameter ESM-2 transformer-based model and a database of its protein structure predictions, dubbed the ESM …
Out of interest I wanted to know how they used their language model paradigm to model the protein folding in 3D space. Unfortunately there is only the abstract and even this hasn't been peer reviewed so make of that what you will.
Interesting that Facebook (sorry, Meta!) are working in this space, whether its a strategic thing or a me-too against Google (sorry, Alphabet!).
"Out of interest I wanted to know how they used their language model paradigm to model the protein folding in 3D space. Unfortunately there is only the abstract and even this hasn't been peer reviewed so make of that what you will."
Err, when I follow the link and download the PDF I get a 28-page document that looks like a complete paper to me? Sadly not yet peer-reviewed as you say. But it was only submitted October 31, so maybe there hasn't yet been time for reviewers to comment?
If a news source chooses to report on a paper before it is peer reviewed or fully published, that's not really the fault of the people doing the research?
If you want hard news, read the scientific journals who wait until things are reviewed to report on them. Just because something isn't peer reviewed (yet) doesn't mean it is not interesting though.
You're right, it is interesting and if claims like this are to be made then sticking it through a peer reviewed process to provide some veracity is a very good idea. These people aren't idiots and know where they're submitting and the level of scrutiny. It used to be common practice to assess how strong a concept was (and how much time you had to put a paper together) and choose where to submit based on that!
Without peer scrutiny, there's a possibility that fundamental issues invalidating the claim exist, or that there are severe limitations on the experimental methodology and performance claims.
Be aware, though, that the review process can be lengthy (as a some-time peer-reviewer, mea culpa - although frequently it's inefficiency on the part of the editorial staff). I'm currently awaiting a response from a very reputable statistical journal for an article (re)submitted in July. Reviews for the initial submission took about six months to arrive.
The model was able to create the ESM Metagenomic Atlas, predicting over 600 million structures from the MGnify90 protein database in just two weeks running on 2,000 GPUs. On a single Nvidia V100 GPU, it takes just 14.2 seconds to simulate a protein made up of 384 amino acids. It seems from the paper that Meta said its system mostly, but not fully, matched AlphaFold on accuracy though its speed is the key thing, allowing it to predict more proteins.
So showing off mostly - actual scientists can only use models that are reasonably accurate...
Like many areas of science there is a to-and-fro between theory and experiment. At the moment, DNA sequencing technology is well ahead so there are floods of this metagenomic type data, but it is just stamp-collecting unless it can be interpreted in some way. If the Meta ESM system can give even tentative suggestions about what sort of protein a sequence represents (nuclease, protease, receptor...) that would help the wet-lab people pursue the most likely targets. But the combinatorial space is so vast, and the thermodynamics of protein structure sufficiently marginal, that I expect this isn't the end of the story by any means.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
