This capability could unlock new possibilities in medicine:
Artificial intelligence has altered the practise of science by enabling researchers to examine the vast volumes of data generated by current scientific instruments. Using deep learning, it can learn from the data itself and can locate a needle in a million haystacks of information. AI is advancing the development of gene searching, medicine, medication design, and chemical compound synthesis.
To extract information from fresh data, deep learning employs algorithms, often neural networks trained on massive volumes of data. With its step-by-step instructions, it is considerably different from traditional computing. It instead learns from data. Deep learning is far less transparent than conventional computer programming, leaving vital concerns unanswered: what has the system learnt and what does it know?
[...] For fifty years, computer scientists have unsuccessfully attempted to solve the protein-folding issue. Then in 2016, DeepMind, an AI subsidiary of Alphabet, the parent company of Google, launched its AlphaFold programme. It utilised the protein databank, which contains the empirically determined structures of over 150,000 proteins, as its training set.
In fewer than five years, AlphaFold had solved the protein-folding issue, or at least the most important aspect of it: identifying the protein structure from its amino acid sequence. AlphaFold can not explain how proteins may fold so rapidly and precisely. It was a tremendous victory for AI since not only did it earn a great deal of scientific reputation, but it was also a major scientific breakthrough that may touch everyone's life.
[...] AlphaFold2 was not meant to anticipate how proteins would interact with one another, but it can model how individual proteins assemble to build enormous complex units made of several proteins. We posed a difficult challenge to AlphaFold: Did its structural training set teach it chemistry? Was it able to predict whether or not amino acids will react with one another, an uncommon but crucial occurrence?
The protein databank contains 578 fluorescent proteins, of which 10 are "broken" and do not glow. [...]
Only a chemist with extensive understanding of fluorescent proteins would be able to utilise the amino acid sequence to identify fluorescent proteins with the correct amino acid sequence to undergo the necessary chemical changes to become fluorescent. AlphaFold2 folded the fixed fluorescent proteins differently than the broken fluorescent proteins when supplied with the sequences of 44 fluorescent proteins not found in the protein databank.
The outcome astounded us: AlphaFold2 had acquired knowledge of chemistry. It determined which amino acids in fluorescent proteins are responsible for the chemistry that causes them to shine. We hypothesise that the protein databank training set and numerous sequence alignments allow AlphaFold2 to "think" like a chemist and search for the amino acids necessary to react with one another to make the protein bright.
[Ed. note (hubie): an interesting podcast on Deepmind and all things AI]
(Score: 1, Troll) by looorg on Thursday June 23 2022, @11:16PM
"AI" used to play chess vs the meatbags. Now its relegated to cooking meth (i'm sure "chemistry" is some kind of breaking bad code) to pay the electricity bill.
(Score: 3, Interesting) by anubi on Thursday June 23 2022, @11:27PM (4 children)
(Copypasta)
In fewer than five years, AlphaFold had solved the protein-folding issue, or at least the most important aspect of it: identifying the protein structure from its amino acid sequence. AlphaFold can #not# explain how proteins may fold so rapidly and precisely.
(/Copypasta)
/S/not/now ?
"Prove all things; hold fast that which is good." [KJV: I Thessalonians 5:21]
(Score: 3, Informative) by Anonymous Coward on Friday June 24 2022, @12:31AM (3 children)
No, it is not a typo, but a rare shining example of truth in journalism. Neural networks are NOT capable of explaining their output.
(Score: 0) by Anonymous Coward on Friday June 24 2022, @10:52AM
Thanks! I was stumped at that.
(Score: 0) by Anonymous Coward on Friday June 24 2022, @01:19PM
Maybe they can, but they choose not to.
(Score: 0) by Anonymous Coward on Sunday June 26 2022, @09:35PM
Neural networks are still used in "AI"? I thought it was all statistical filters now.
(Score: 2, Informative) by Anonymous Coward on Friday June 24 2022, @12:29AM
https://www.technologyreview.com/2011/02/22/88992/physicists-discover-quantum-law-of-protein-folding/ [technologyreview.com]
https://arxiv.org/abs/1102.3748 [arxiv.org]
https://towardsdatascience.com/quantum-landscape-for-protein-discovery-62c0c86fc27e?gi=d45ce4cbb41b [towardsdatascience.com]
Protein folding is a quantum process. To calculate it properly AND quickly, you need a quantum computer, or at least some half-assed prototype of it.
What the neural networks, maliciously mislabeled "AI", can do, is GUESS at a possible folding configuration, guided by similarities to the proteins in the training set. Nothing more, nothing less.
(Score: 0) by Anonymous Coward on Friday June 24 2022, @12:29AM (7 children)
what is the cost now to train an AI or even update with a new data set each time?
Can we get it to do these tricks for free?
(Score: 0) by Anonymous Coward on Friday June 24 2022, @12:44AM (3 children)
Ai is becoming so powerful that its next experiment will be to try to get the Democrats re-elected in November even with gas at six bucks a gallon. Now that would be an accomplishment.
(Score: 0) by Anonymous Coward on Friday June 24 2022, @12:50AM (1 child)
Maybe you can get AI to explain to you how gas prices work, because clearly you don't understand it.
(Score: 0) by Anonymous Coward on Friday June 24 2022, @01:31PM
Cancelled pipelines, rising taxes, etc... I think Biden has some things to answer for when it comes to US gas prices.
(Score: 0) by Anonymous Coward on Friday June 24 2022, @03:00PM
Remember the observable levels of enthusiasm for Trump and Biden in the 2020 campaign?
(Score: 0) by Anonymous Coward on Friday June 24 2022, @12:47AM (2 children)
https://venturebeat.com/2021/10/15/ai-weekly-ai-model-training-costs-on-the-rise-highlighting-need-for-new-solutions/ [venturebeat.com]
So there's your answer, the units are "compute budgets of most governments":
example: Training costs for the SN Runaway Turing AI were 5.6 cbmg's.
(Score: 0) by Anonymous Coward on Friday June 24 2022, @01:03AM (1 child)
and we are supposed to react how to this?
(Score: 0) by Anonymous Coward on Friday June 24 2022, @11:45AM
By adding yet another check mark to the,
"fuck the planet, we have AI^H^H technology to feed".
Don't give a fuck?
Good, cause the planet doesn't either.
Enjoy your march to the dytopian future.
(Score: 1, Informative) by Anonymous Coward on Friday June 24 2022, @01:02AM
the system has learnt nothing and knows nothing.
(Score: 0) by Anonymous Coward on Friday June 24 2022, @02:29AM
Chemistry is you bang stuff together and see what happens. It's just like modern "AI" - you let it flow and see if anything interesting comes about.
(Score: 0) by Anonymous Coward on Friday June 24 2022, @03:33AM (1 child)
If it can't explain it, it doesn't know it.
What we have here is a pattern-matching system that matched patterns. It didn't establish any kind of mental schema, it just learned from a training set and responded by its patterns learned.
Well, duh.
(Score: 0) by Anonymous Coward on Friday June 24 2022, @05:14AM
that second paragraph wrote itself same way a NN works.
(Score: 2) by AnonTechie on Friday June 24 2022, @10:38AM
There are many worthwhile projects that could use additional help.
Choosing BOINC projects [berkeley.edu]
Albert Einstein - "Only two things are infinite, the universe and human stupidity, and I'm not sure about the former."
(Score: 0) by Anonymous Coward on Saturday June 25 2022, @04:00PM
and put it on a Windows computer and hook it up to a chemical plant!