Powerful Antibiotic Discovered Using Machine Learning For First Time (theguardian.com) 54
A powerful antibiotic that kills some of the most dangerous drug-resistant bacteria in the world has been discovered using artificial intelligence. The Guardian reports: To find new antibiotics, the researchers first trained a "deep learning" algorithm to identify the sorts of molecules that kill bacteria. To do this, they fed the program information on the atomic and molecular features of nearly 2,500 drugs and natural compounds, and how well or not the substance blocked the growth of the bug E coli. Once the algorithm had learned what molecular features made for good antibiotics, the scientists set it working on a library of more than 6,000 compounds under investigation for treating various human diseases. Rather than looking for any potential antimicrobials, the algorithm focused on compounds that looked effective but unlike existing antibiotics. This boosted the chances that the drugs would work in radical new ways that bugs had yet to develop resistance to.
Jonathan Stokes, the first author of the study, said it took a matter of hours for the algorithm to assess the compounds and come up with some promising antibiotics. One, which the researchers named "halicin" after Hal, the astronaut-bothering AI in the film 2001: A Space Odyssey, looked particularly potent. Writing in the journal Cell, the researchers describe how they treated numerous drug-resistant infections with halicin, a compound that was originally developed to treat diabetes, but which fell by the wayside before it reached the clinic. Tests on bacteria collected from patients showed that halicin killed Mycobacterium tuberculosis, the bug that causes TB, and strains of Enterobacteriaceae that are resistant to carbapenems, a group of antibiotics that are considered the last resort for such infections. Halicin also cleared C difficile and multidrug-resistant Acinetobacter baumannii infections in mice. Three days after being set loose on a database of about 1.5 billion compounds, the algorithm returned a shortlist of 23 potential antibiotics, of which two appear to be particularly potent.
"[The senior researcher] now wants to use the algorithm to find antibiotics that are more selective in the bacteria they kill," adds The Guardian. "This would mean that taking the antibiotic kills only the bugs causing an infection, and not all the healthy bacteria that live in the gut. More ambitiously, the scientists aim to use the algorithm to design potent new antibiotics from scratch."
Jonathan Stokes, the first author of the study, said it took a matter of hours for the algorithm to assess the compounds and come up with some promising antibiotics. One, which the researchers named "halicin" after Hal, the astronaut-bothering AI in the film 2001: A Space Odyssey, looked particularly potent. Writing in the journal Cell, the researchers describe how they treated numerous drug-resistant infections with halicin, a compound that was originally developed to treat diabetes, but which fell by the wayside before it reached the clinic. Tests on bacteria collected from patients showed that halicin killed Mycobacterium tuberculosis, the bug that causes TB, and strains of Enterobacteriaceae that are resistant to carbapenems, a group of antibiotics that are considered the last resort for such infections. Halicin also cleared C difficile and multidrug-resistant Acinetobacter baumannii infections in mice. Three days after being set loose on a database of about 1.5 billion compounds, the algorithm returned a shortlist of 23 potential antibiotics, of which two appear to be particularly potent.
"[The senior researcher] now wants to use the algorithm to find antibiotics that are more selective in the bacteria they kill," adds The Guardian. "This would mean that taking the antibiotic kills only the bugs causing an infection, and not all the healthy bacteria that live in the gut. More ambitiously, the scientists aim to use the algorithm to design potent new antibiotics from scratch."
Cure for AIDS and cancer? (Score:1)
Re: (Score:2)
Not blindly. But with HIV at least its concievable that using the technique to find compounds that target a specific protein or whatever might be a productive effort.
Although largely keeping HIV surpressed is totally a thing we can already do. The problem is disentangling that nasty shit from peoples cellular reproduction machinery. After all, its just DNA, and lentiviruses in particular are quite clever at integrating themselves into people genetic makeup
Re: (Score:3)
"[The senior researcher] now wants to use the algorithm to find antibiotics that are more selective in the bacteria they kill,"
Finding a potent biocide is one thing, finding one that doesn't kill all sorts of other stuff that you really don't want killed is quite another. For example hard UV and gamma rays are very effective at killing all bacteria, but that doesn't make them practically useful against MRSA.
Re: Cure for AIDS and cancer? (Score:5, Insightful)
Re: (Score:2)
Arsenic cures Aids and all cancers. The problem is finding compounds that are a wee more specific.
Number one on the list was fire.
Re: (Score:3)
We already have drugs that kill HIV. The problem is that the virus "hides" in places where it's not accessible to the drugs that will kill it. Medication will keep a HIV+ person's blood clear of the virus so they won't get sick or infect anyone else, but if they stop it the virus will come back. An AI system could potentially be helpful in figuring out a way to crack the T-cells that HIV likes to hibernate in, but it's not the exact same problem as just targeting a viral protein.
Just imagine... (Score:4, Interesting)
How many products, do you think, could have great benefits to us but fell to the wayside because they were developed to tackle a different problem and weren't particularly effective against that?
Re: (Score:2)
How many products, do you think, could have great benefits to us but fell to the wayside because they were developed to tackle a different problem and weren't particularly effective against that?
Post-it notes?
Teflon coating?
Pacemakers?
Super Glue?
Re: (Score:2)
Those are all things that made it into widely used products. What's that got to do with my question?
Re:Just imagine... (Score:4, Informative)
Re: (Score:1)
I know, what does that have to do with my question?
Re:Just imagine... (Score:4, Insightful)
The relevance to your question is that while a drug may work for something for which it was not intended, you won't know that unless you go through extensive and exhaustive testing. The obvious reason is that the new use might be good in isolation, but if that condition usually comes associated with others for which the drug has bad effects, then it isn't viable.
More directly to your question, it is sort of unanswerable in the same sense that whether an algorithm halts is unanswerable except in restrictive circumstances. What is an effective (meant in the sense of a recipe of instructions) test for whether something has another use? Over what duration does one look? 1 year, 10 years, 100 years, etc. There is no recipe of instructions because how can we be sure the recipe is exhaustive of all necessary instructions in all the possible sequences.
Say a physicist conjures up Really Exciting Substance A. A can be used as film to make things slide easier. How do we know it won't work for cancers? Why would you even try? It might, but you cannot know ahead of time and maybe it will but only in combinations with Really Exciting Substance B that increases the shelf life of food. The possibilities are probably not infinite in the sense the universe might only have a finite number of particles. Include an endless time and the possibilities do become infinite, but repetitive until either the heat death of the universe or a Big Crunch.
Re: (Score:1)
Re: (Score:2)
The classic case of this is Viagra. I forget what its intended purpose was, but so many people in the clinical trials reported boners that the molecule was repurposed for just that use.
Re: Just imagine... (Score:1)
Re: (Score:1)
Re: (Score:2)
Re: Just imagine... (Score:2)
When it comes to medicine they test that crap all the time. Are used to think that they would just go back in the lab and shake a few files together and do a bunch of trials to see what itâ(TM)s good for. Wellbutrine is an antidepressant but they use it to treat smoking addiction marketed under the name Zyban. Imitrex did not start out as a migraine medication. Reglan is now not only an anti-emetic but used to kickstart more productive lactation in women.
Re: (Score:2)
How many products, do you think, could have great benefits to us but fell to the wayside because they were developed to tackle a different problem and weren't particularly effective against that?
That may be an interesting philosophical question, but from a practical standpoint it's just an invitation to pointless speculation. If a product or process is found to be effective for some use that wasn't intended, it doesn't fall by the wayside; see Post-it Notes. In modern pharmaceutical development, the only way it happens is if a beneficial side effect slaps the clinicians in the face during trials.
Re: (Score:2)
Re: (Score:3, Insightful)
I think you do not understand the mechanics of evolution...
Re: (Score:2)
With these machine learning approaches we have someone who supervises the process. That is why we can direct a progress here. We can make backups of useful results to which we can fall back if the process goes awry. The same applies to all trial and error approaches. For them to work you need some kind of memory where failures are recorded so you don't repeat them all the time.
In evolution ho
Re: (Score:2)
For them to work you need some kind of memory where failures are recorded so you don't repeat them all the time. In evolution however you have randomness and environmental circumstances. Evolution knows no backups. Evolution does not keep a record of failures either. It can repeat the same 'mistake' (defined by circumstance) ad infinitum.
Kinda, but also forgetfulness. A classic example are puzzle games where you can't go through a locked door, but maybe there's a key or a lever or whatever that'll open that door. It can't remember that the door is locked forever, the confidence has to fade over time. There's lots of other examples where challenges that were too hard and punished you earlier give you high rewards later. How often and how far it should go in trying to find novel solutions is an active research topic, it's easy to get stuck in
Re: (Score:2)
Throw enough monkeys and typewriters in a room and eventually you'll get Shakespeare.
Are you implying that Shakespeare evolved from an un-holy mating of monkeys and typewriters? Because that's impossible. Typewriters hadn't even been invented in his time.
Re: (Score:2)
That and monkeys don't do that sort of thing unless the money is right.
Re: (Score:1)
That and monkeys don't do that sort of thing unless the money is right.
That's why we get Fast and Furious 17 instead.
Re: (Score:3)
That's what they did until now!
Currently people are trying to come up with potential candidates for drugs based on molecules that have been proven to work. It's more or less a blind trial and error game.
What AI brings to the table is that it can sift through test results and find similarities between molecules that have a similar effect (in the tests) but come from a different 'familty'. In this way it is able to find different molecules than the traditional ones.
As a matter of fact the article explicitely
Re: (Score:1)
Re: (Score:2)
Um ... "bothering"? (Score:2)
One, which the researchers named "halicin" after Hal, the astronaut-bothering AI in the film 2001 ...
This was quoted from TFA, so, apparently, someone hasn't seen or doesn't remember the film 'cause (*spoilers*) things don't really work out that well for (a) the crew and (b) HAL. HAL kills all but one of the crew, then gets lobotomized and (basically) killed.
Re: (Score:2)
To be fair, HAL killing then did cause a spot of bother....
Re: (Score:2)
Re: (Score:2)
But we find out in the sequel that it wasn't actually HAL's fault; it was upper-level management who abused HAL's directives.
Similarly, if this method works, the drug company executives will probably only use it to find more ED treatments.
Never thought I'd see it in my lifetime (Score:5, Funny)
>A powerful antibiotic that kills some of the most dangerous drug-resistant bacteria in the world has been discovered using artificial intelligence.
An antibiotic that uses artificial intelligence?! Wow! Will the wonders never cease?
Re: (Score:2)
Reminds me of the time I shot an elephant in my pajamas!
Re: (Score:2)
How did he get into them?
Re: (Score:2)
I'll never know.
Re: (Score:2)
Through the pyjamas fan club. I used to be into them myself. Now I'm more into trunks.
Obligatory XKCD (Score:2)
Just a new algorithm but still not that useful (Score:5, Informative)
The mere use of a computer to predict drug candidates isn't novel.. There have been papers using ML to predict drug candidates for years now and there have been programs using things like simulated receptor fit or structural similarity to suggest candidates for decades now. The only novelty claimed by this paper is the use of a novel neural net representation that is claimed to offer better performance (not that clear it really does from paper but seems plausible).
The reason it's not particularly useful is that predicting a molecule might have some activity against a known target isn't really the hard part. Chemists working in the area can generally come up with potential candidates pretty easily by hand or with existing hand assisted computational simulations. The hard part isn't just finding a compound with activity at the target site (and certainly not the very easy task of fucking up a bacteria). The hard part is finding something that kills bacteria (or whatever) without fucking other stuff up.
Besides, automated screening methods are well developed so if all you want to know is which of the compounds in your library bind to some target or kill some bacteria you can brute force a ton of candidates so it's not that helpful to improve your guesses about which will have the strongest effect a little bit.
Now, if they come up with ML that can help screen out the candidates that will screw up something else in the body or otherwise fail in vivo better than chemists can by hand (or at least help them do so) I'll be much more impressed. But it's a lot harder to guess something won't screw anything else up than it will do a specific task.
Re: (Score:1)
Unclear... (Score:2)
So, have less powerful antibiotics been using machine learning for a while? Or was an antibiotic discovered using a machine, and is now learning for the first time?
don't repeat the same mistake (Score:2)
great, now that they have this new antibiotic, don't mess it up so that in 50 years we'll have the same problem again where no antibiotic works anymore.
Re: (Score:2)
We will be lucky if these new antibiotics last even that long.
Comment removed (Score:3)
Excellent work, but... (Score:2)
Cool! How soon can the beef and dairy industry start using this stuff with their cows? And how long will it take to come up with yet another antibiotic for the new-antibiotic-resistant bacteria? Hurry up, we're gonna need something in about 18 months. :-/
This is interesting news, but it gets infuriating to be constantly chasing such a rapidly moving target.