Nearly all protein structures known to science predicted by AlphaFold AI

The AI-powered protein-folding model AlphaFold has predicted more than 200 million proteins, nearly all such structures known to science, DeepMind said on Thursday. Proteins are complex biological molecules produced in living organisms from instructions stored in DNA. Made from as many as 20 types of amino acids, these nano- …

[Anonymous Coward]

AI

Although I still hesitate to call it Artificial Intelligence, I recognize that's a losing battle.

But this is what AI does superbly. While big pharma may gripe that DeepMind and AlphaFold AI didn't spit out a list of drugs that they could exploit, the value to researchers seems immense, on the order of the full sequencing of the human genome.

A pint to the boffins who worked on this project.

[sreynolds]

Re: AI

Did it get those weird bent ones that used to be found in British Beef at the turn of the century?

[NATTtrash]

Re: AI

Although the <sarc> tag was certainly noticed (and enjoyed!), it does go to the exact heart of the matter: designing proteins like this is all very good and nice, but this is just such a small step in the complete life of proteins and the processes they trigger and influence. And this is just for the part that we know about. So sure, let them have their minute of lime light with their marketing press release because it helps their bottom line. But please, let's hum Public Enemy while reading it. Fiddling behind your screen to simulate real life is nice, can help, but please do not forget to get out sometimes into the real world.

Did it get those weird bent ones that used to be found in British Beef at the turn of the century?

For example, nobody even knew that prions (those weird bent ones) could be transferred between animals "effectively" (not in the AI). We did find that out though when British farmers (and regulators) thought it was a good idea to mince dead cows (including their BSE brains) and feed them to other cows. Not in the Ai too. Cheap human behaviour however. And then it turned out these prions also worked pretty well in humans. I can remember watching the news on the Beeb at that time, noticing that the meat industry was working hard on their messaging through the royals, who stated: "There is nothing wrong with our excellent British beef! I had some this morning!"

...

...

Yes, indeed. A shiny endorsement.

But more to the point: the structure of a protein is just such a minute part of what a protein does, doesn't do, could do, or should. In addition it is subject to many interwoven processes, that might not even seem that important. Just one is what medicine/ pharmacology/ toxicology/ science calls "PK" or pharmacokinetics and pharmacodynamics, which tries to capture how that (in this case) protein is introduced, absorbed, and worked on on its way to the other end of the human tunnel. Going into it here would be excessive (and for some no doubt boring), but imagine this: we are going to make a protein tablet to treat that cancer right? OK... So how to we make sure your stomach, which is very efficient at digesting protein to generate energy, doesn't break it down together with the protein of that fry up you had this morning?

[Bebu]

Re: AI

> Did it get those weird bent ones that used to be found in British Beef at the turn of the century?

Good question.

Prions are basically two different conformations (shapes) of the same peptide (protein) - one naughty, one nice. Based purely on the sequence of amino acids in the peptide I don't see how this approach (Machine Learning) is able to detect or determine both conformations. Presumably picks up the normal (nice) one as I would assume it would fold like most other peptides in its training set.

Still calculating a likely structure first would likely expedite other techniques of structure determination - thinking x-ray crystallography, mass spec and nmr.

I would be interested in how accurately it predicts the structure of variants of the peptides with which it was trained

(ie modified with various amino acid substitutions.)

[VonDutch]

Re: AI

Traditional modelling methods (10 years or so ago) would take the model of the amino acid sequence, simulate heating to a high energy state so the chain moves around a lot in to random confirmations and then reduce the temperature and simulate the rearrangement in to a low energy conformation. Through thousands of rounds of this a number of classes of probable conformations would be produced with a likelihood of each one being "true".

The real complication comes where there are post translational modifications that either direct or inhibit the formation of different conformations. A lot of these PTMs can't be predicted from the genome (glycosylation) and they have a temporal factor (turning on or off of a protein by phosphorylation as an example).

The exponential amount of information moving from genomics to proteomics and beyond is a real challenge but the output of this project could help accelerate the hands on research a lot.

[Doctor Syntax]

Re: AI

"Traditional modelling methods (10 years or so ago) would take the model of the amino acid sequence, simulate heating to a high energy state so the chain moves around a lot in to random confirmations and then reduce the temperature and simulate the rearrangement in to a low energy conformation. "

I'd have thought a more productive approach would have been to simulate the synthesis, adding one amino acid at a time. As each emino acid is added there must be relatively few low energy configurations which wouldn't involve a reconfiguration of the existing molecule so substantial as to require significant energy input.

[Korev]

Re: AI

But this is what AI does superbly. While big pharma may gripe that DeepMind and AlphaFold AI didn't spit out a list of drugs that they could exploit, the value to researchers seems immense, on the order of the full sequencing of the human genome.

If these public models "spat out molecules" in public then we* couldn't patent them and therefore sell them! This in turn means that they wouldn't get developed as why would any company spend millions on trials and getting approvals for drugs that they couldn't sell.

* I work for a big pharma

[Doctor Syntax]

Re: AI

Maybe it should be the trials and approvals that carry the rights to be licenced, along with the methods of synthesis if these are novel.

A molecular structure is, if not a natural occurrence, at least a possibility of nature. It should be no more patentable than an algorithm. Proving that a substance is an effective and safe treatment and in what circumstances, on the other hand, is what makes it valuable.

[gandalfcn]

Re: AI

"big pharma may gripe that DeepMind and AlphaFold AI didn't spit out a list of drugs: If it had they would have no excuse to gouge.

[Steve Button]

Re: AI

The Register has asked DeepMind for further comment.

DeepMind replied "All your base are belong to us"

[mevets]

Re: AI

We have a rich history of redefining intelligence as something that machines can't do. Barely over a hundred years ago, arithmetic skills were an epitome of intelligence. Our definition will continue to adapt.

That said, it is arguable if a massively parallel successive approximation would signal intelligence. Had the machine autonomously decided to break into the various databases, steal the work, and claim it as its own, we could all agree on its brilliance.

[Lordrobot]

Re: AI

Couldn't agree with you more. This is a glorified CAD program and far far away from intelligence. But without Buck Rodgers, you get no bucks... Everything is a grift.

[Winkypop]

Bend me, shape me, any way you want me

Call it whatever you like.

In any case, it’s another big win for science!

[gandalfcn]

Re: Bend me, shape me, any way you want me

Amen to that corner of the debate

[EricB123]

Re: Bend me, shape me, any way you want me

Oh, thanks so bloody much. Now that silly song is stuck in my head.

[gandalfcn]

Re: Bend me, shape me, any way you want me

Wide eyed and legless again?

[Anonymous Coward]

DeepMind (aka Google) Hard At Work!

Link: https://www.theregister.com/2021/09/30/royal_free_deepmind_representative_action_uk/

No connection, of course, with the same DeepMind (aka Google) slurping 1.6 million medical records from the Royal Free Trust -- without any user consent, and likely in total violation of the law.

Protein analysis, millions of medical records......What else should UK citizens know about exactly what is going on here?

Quote (William Burroughs): "The paranoid is a person who knows a little about what is going on."

[Korev]

Re: DeepMind (aka Google) Hard At Work!

I'm 99.9999% certain that the records from the Royal Free Trust would not have been useful to this project. There's a huge difference between "patient X visited and we diagnosed Y" records from the RFT and the protein sequences and structures that you'd use to make models like these.

I'm not defending the slurping BTW just pointing out the difference.

[Cuddles]

Re: DeepMind (aka Google) Hard At Work!

The vast majority of protein structures are recorded in the Worldwide Protein Data Bank - http://www.wwpdb.org/

They're freely available for anyone to download and do any research they want. Even if the paper initially describing a protein structure is behind a paywall at a journal, the structure itself will almost always be deposited in the protein data bank. Other similar resources include the European (not EU, it's based in the UK) UniProt database, although there will be a lot of overlap between them. I don't know the details of where all AlphaFold's training data came from, but it was almost certainly from one or more of the publicly accessible databases which exist specifically to enable this kind of research.

[Tom 7]

It's not clear how fully accurate AlphaFold's predictions are.

Also the lack of knowledge how they change shape in the presence of others. Not saying what they have done is not bloody amazing because it is but I'm getting the impression around 200 million proteins need xray diffraction (with AI help) and then we need to mix those 200 million proteins with every other one to see how they interact. So there is a bit of work left to do.

Me - I'd then run it on the LUCA proteins to see if we can get back to LUCAs ancestors - I think it likely there were many proto-life forms and two or more merged to make LUCA that went and ate all the others or just left them no room to evolve.

Might be nice to have an AI on bio-security so we can ensure some of the stuff NATTtrash alludes to dont get out. It would obviously have to be a completely separate AI wtih no knowledge of Deep Blue etc to prevent it getting any ideas above its station which would need an AI to ensure complete information AI separation which obviously wouldnt work...

Let me drink about this some more..

[Little Mouse]

Re: It's not clear how fully accurate AlphaFold's predictions are.

"It's not clear how fully accurate AlphaFold's predictions are". This.

I'm going to assume that at least some of the results have been independently verified and that there is a measure of how accurate the 200M predictions are likely to be, but I see no mention of this in the article.

[Ellipsis]

Re: It's not clear how fully accurate AlphaFold's predictions are.

AlphaFold was already pretty good at CASP13, then blew everybody else out of the water at CASP14. CASP15 is in progress right now.

Part of the problem is that experimentally determining a protein’s folded structure is very laborious, which is why so few are actually known.

What I find impressive is that DeepMind has achieved in a few months in its own lab what Rosetta@home has got nowhere near to in fifteen years with a worldwide distributed network…

[Howard Sway]

improving machine-learning to help us discover new potential drugs for rare diseases

Also potential drugs fot other things too. What would you prefer today sir? Google Soma, Facebook Soma or Amazon Soma?

[Anonymous Coward]

Re: improving machine-learning to help us discover new potential drugs for rare diseases

"Also potential drugs fot other things too"

Non-addictive heroin? Too cheap to meter...

[gandalfcn]

Why can't Brit politicians be like their scientific counterparts, competent, innovative and world beating?

Shame this is now owned by the unspeakables.

[Lordrobot]

I see Prions made it into the discussion

The Theory of Prions is anything but accepted science. While the kuru ate the brains of their dead relatives or something else or nothing else and developed encephalopathy there is NO DIRECT link to spongiform encephalopathy and prions. The claim is that prions fold over on themselves and self-replicate. Of course, there is simply no instance in all of the biochemistry where this occurs. The claim is that it occurred in a petri dish.

I did my thesis work on rabies. The rabies virus has a gyrus that heads straight for the central nervous system. That is the same claim for the prior which is preposterous. So in one breath the PRION folds and self-replicates the process unknown... must be magic and the topper is this short peptide also has a gyrus for the central nervous system.

Expecting a CAD program which is all this "AI" program is, could not possibly come up with such a crazy set of notions because there is no mechanics known for the prior claims.

CJD has two variants. One is genetic the other is not. There are a lot of antidotal transmission claims even by eye doctors transplanting corneas. When a patient dies of CJD nobody wants to touch the body, or perform an autopsy but various brain banks will accept the brain. So add that to the virtues of the prion that it is contagious on contact with bodily fluids. Yet there is not a single case where spouses or caretakers have contracted CJD. I digress. My point is that this CAD program has no biochemical basis of any kind to find prions because the claims of the Prion are based largely on fantasy and claims of singulariy.

One of our posters referred to Prions and the good and the bad. That is indeed the claim that PrP, PrPD, types 1 and 2. PrPD they say causes the many encephalopathies.

I think its' great to have these CAD programs doing heavy lifting of protein structures. But I am sceptical of claims of pathological peptides especially those with such exaggerated claims as the prion... deadly, infectious, indestructible, a CNS gyrus, and self-replicating. This is the stuff of Stephen King...My theory is that it is an organism that can penetrate the blood-brain barrier that craps prions.

[MachDiamond]

This sounds like another good tool added to the toolbox. It's not the whole house built, but just a framing square that can prove useful in many situations.