AI-predicted protein structures could unlock vaccine for COVID-19 coronavirus... if correct... after clinical trials DeepMind has shared its AI software's homework detailing the structure of six proteins linked to SARS-CoV-2 aka the coronavirus that causes COVID-19. The machine-learning startup hopes the findings can help scientists develop a vaccine to halt and reverse the global spread of the bio-nasty. Knowing these viral protein …
... now there is money for AI, so it is done with AI. Usually, those algorithms are informed by (learning? Lets make that extra-deep learning to make sure the funding comes through!) and run against known structure to test how well they perform. So it's not like there weren't any yardstick they could use to compare the expected quality of their structure to those of other modeling efforts. Unfortunately, even 98% or 99% correct predictions can be completely worthless, imagine taking just one wrong turn when driving somewhere.
Even if they get the structure of some virus proteins right, that does not give a vaccine. I'd expect that they already plenty of lot of known proteins that are shared between the new virus and other known Coronavirus strains. So yes, it sounds very much like a PR stunt. Made it into the Register through, so it was a good one!
" I'd expect that they already plenty of lot of known proteins that are shared between the new virus and other known Coronavirus strains."
Ditto I was reading this paper yesterday (mainly for its MHC * content) :
https://www.mdpi.com/1999-4915/12/3/254/pdf.
The spike protein is thought to be the key to binding to cells via the angiotensin II receptor.
* Old research interest of mine. Part of the major mechanism the immune system uses to distinguish self from non-self
Agreed, the number of variables and conditions for folding are huge and makes this extremely challenging, we still have a long way to go in our understanding before we can predict with certainty. So it's back to the lab to do some real research on the bench rather than being sat around with fancy MacBooks thinking your going to arrive at the answer 42 without fully understanding the question first.
"Oddly enough, once the protein's been made it doesn't need AI. It just folds itself."
It's an analogue computer of great power. Seriously the most important property of a sequence is to fold quickly and cleanly. The efficiency of it active or whatever is a secondary requirement.
Not quite, Natural Selection does not measure methods, it measures outputs, usually at the organism level.
Sure correct folding is necessary for much protein function and we have prions and chaperone proteins to get it wrong and right.
The only way NS measures methods and mechanisms is if they are very energetically wasteful. But there are some very wasteful ones out there. Beta-Catenin at the end of point of Wnt signalling comes particularly to mind.
Let's be skeptical, but not cynical. Cynicism is a feedstock of the antivaxers.
So here's to all the efforts of the epistemologists, data modellers and clinicians who are doing their best at this time. Have a beer to keep your spirits up (but not too many just yet).
Er - sorry, I forgot to mention the politicians....
What about the epidemiologists? Don't they get a look in?
Epistemologists want to know, because that's what epistemologists do.
(There's still a dent in my skull from having the difference between episteme and noesis hammered in all those years ago.)
This does not mean anything. The whole months-long effort in vaccine development is the testing for effectiveness and side-effects. The invention of new vaccine candidates takes only hours nowadays. This AI only does something that has never been a problem in the first place.
"Determining the structure of the virus proteins might also help in developing a molecule that disrupts the operation of just those proteins, and not anything else in the human body."
Well it might, but predicting whether a 'drug' will NOT interact with any other of the 20000+ protein in complex organisms is well beyond current science. If we could do that we could predict/avoid toxicity and other non-mechanism related side-effects & mostly we can't.
There are 480 structures on PDBe resulting from a search on ‘coronavirus,’ the top hits from MERS and SARS. PR stunt or not, they did win the most recent CASP ‘competition’, so arguably it’s probably our best shot right now - and I am certainly not satisfied that they have been sufficiently open in explaining their algorithms though I have not checked in the last few months. No one is betting anyone’s health on this, and it is not like making one wrong turn in a series of car directions. Latest prediction algorithms incorporate contact map predictions, so it’s not like a wrong dihedral angle sends the chain off in the wrong direction. A decent model would give something to run docking algorithms against with a series of already approved drugs, then we take that shortlist into the lab. A confirmed hit could be an instantly available treatment, no two year wait as currently estimated.
Translation: "We don't know if this is remotely close to reality, or if they're of any relevance to SARS-CoV-2, but here's some protein structures to look at, they might be useful in the future some day. Or maybe not. Did we mention AI?".
Yup, thanks.
Naaaah. Can't possibly be a stupid marketing stunt.
Naaaah. Can't possibly be a stupid marketing stunt.
Well yes, a good possibility. But it can also be trying to build on the open-source model of putting it out there for others to build and improve upon. Essentially opening that "peer review" to a larger audience quicker.
We shall see.
What bothers me, besides the obvious PR stunt, is that they say this prediction is licensed. How can a prediction from software be protected by, I presume, patents? And if this can be protected without even verifying which predictions actually work, what's to stop someone spitting out millions of random, untested predictions just in case they can claim ownership later when one of them is proven to work?
I've never been really fascinated by A.I until I started thinking about the uses for it. Of course, the PR stunt is clear in this case, but I think the increased focus to develop and use A.I for anything and everywhere is needed. There are plenty of places where scientific funds are distributed poorly and inefficiently thus making less impact, but it would be great if countries started funding these things rather trying to increase their (for example) military effectiveness
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
