"In terms of antibiotic discovery, this is absolutely a first"
February 21, 2020 10:36 AM   Subscribe

Artificial intelligence yields new antibiotic (MIT): Using a machine-learning algorithm, MIT researchers have identified a powerful new antibiotic compound. In laboratory tests, the drug killed many of the world's most problematic disease-causing bacteria, including some strains that are resistant to all known antibiotics. It also cleared infections in two different mouse models.

Powerful antibiotic discovered using machine learning for first time (The Guardian):
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.

Powerful antibiotics discovered using AI (Nature):
The study is remarkable, says Jacob Durrant, a computational biologist at the University of Pittsburgh, Pennsylvania. The team didn't just identify candidates, but also validated promising molecules in animal tests, he says. What's more, the approach could also be applied to other types of drug, such as those used to treat cancer or neurodegenerative diseases, says Durrant.

Bacterial resistance to antibiotics is rising dramatically worldwide, and researchers predict that unless new drugs are developed urgently, resistant infections could kill ten million people per year by 2050. But over the past few decades, the discovery and regulatory approval of new antibiotics has slowed. "People keep finding the same molecules over and over," says Collins. "We need novel chemistries with novel mechanisms of action."
Machine Learning for Antibiotics (Science Translational Medicine):
An important feature of this work is that it's a close collaboration between virtual screening ML methods and actual assays, run specifically for this project. For example, the team started out by taking a list of FDA-approved drugs and a somewhat shorter list of bioactive natural products (2,335 compounds total) and running a growth-inhibition screen with them against E. coli bacteria. Machine-learning models are exquisitely sensitive to the quality of the data used to train them, and it's a very good idea to generate that data yourself under controlled conditions if you can. There are surely antibacterial numbers available in the literature for many of the compounds on that list, but they're going to be from assays run by different labs under different conditions, against different strains and at different concentrations, making those numbers close to useless for reliable machine learning fodder. So collecting fresh numbers under tighter conditions was an excellent start.
A Deep Learning Approach to Antibiotic Discovery (Cell, full text)
posted by not_the_water (34 comments total) 32 users marked this as a favorite
 
If AI has to stay busy inventing new antibiotics as each last one becomes no longer effective, that might postpone the machine uprising for a while.
posted by Greg_Ace at 11:02 AM on February 21 [15 favorites]


From the MIT article:
Preliminary studies suggest that halicin kills bacteria by disrupting their ability to maintain an electrochemical gradient across their cell membranes. This gradient is necessary, among other functions, to produce ATP (molecules that cells use to store energy), so if the gradient breaks down, the cells die. This type of killing mechanism could be difficult for bacteria to develop resistance to, the researchers say.
...
In this study, the researchers found that E. coli did not develop any resistance to halicin during a 30-day treatment period. In contrast, the bacteria started to develop resistance to the antibiotic ciprofloxacin within one to three days, and after 30 days, the bacteria were about 200 times more resistant to ciprofloxacin than they were at the beginning of the experiment.
posted by gwint at 11:02 AM on February 21 [16 favorites]


Two days after the grant ended the sneaky bacteria looked up and whispered, shhh give it another week and we'll hookup with corona.
posted by sammyo at 11:09 AM on February 21 [6 favorites]


This molecule, which the researchers decided to call halicin, after the fictional artificial intelligence system from “2001: A Space Odyssey,”

Wasn't that the same AI that kept coming up with ways to kill the humans due to conflicting priorities? I mean come on guys, think about your cool names for JUST a second

wait I'm going to develop a machine learning model to cross-analyze new things that need names against characters that don't have ironically inverted personal goals and make a mint in the product-branding game
posted by FatherDagon at 11:13 AM on February 21 [21 favorites]


> ...that might postpone the machine uprising for a while.

Or just make it even easier. Atlas Skynet Shrugged (and all of humanity dies from Super Polio).
posted by Godspeed.You!Black.Emperor.Penguin at 11:14 AM on February 21 [1 favorite]


Super Polio

Worst video game ever.
posted by Greg_Ace at 11:16 AM on February 21 [24 favorites]


lemme guess: it's radioactive / it kills everything / it's made of bleach — so technically it's a great antibiotic …
posted by scruss at 11:45 AM on February 21


It is good that we have another last-line of defense available to us. It is also interesting that combinatorial chemistry would have likely never found this compound in our lifetime, which speaks to the power of this approach. I might have overlooked this in what I've read, but I'd still be curious to know about the toxicity of halicin. There are classes of antibiotics that still work very well against certain bacterial infections (fluoroquinolones against UTIs, for instance), but which can cause serious damage to the mitochondria in our cells. Some effects can include deafness, nerve and joint damage, and even psychiatric disorders.
posted by They sucked his brains out! at 11:50 AM on February 21 [11 favorites]


Once upon a time I worked in pharma research, and the project I worked on was to throw a hundred thousand novel compounds at a couple dozen simple tests.

One of the things we looked for was specificity. If a drug was active in all of our tests, y'know why? Because it's poisonous, which means it will kill the healthy cells too. So we looked for drugs that were active in just one test.

So when I hear this killed so many kinds of bacteria, I think it is likely a powerful toxin.
posted by elizilla at 11:50 AM on February 21 [12 favorites]


While a cool piece of work, this compound is unfortunately likely to be incredibly toxic. It kills bacteria by making their cell membranes more leaky, so that they can't produce energy as efficiently and thus end up shutting down. This mechanism of action is exactly the same as the infamous 2,4-dinitrophenol, which, coincidentally, also has antibacterial activity through this same mechanism of action. Unfortunately, this mechanism isn't limited to bacterial cells—it'll kill your cells too.

When it comes to discovering new drugs, generating new chemical leads isn't the hard part—we have plenty of methods available to do that—but making sure these leads are safe is. Unfortunately, there isn't really a good way to place a constraint on the search space that reliably accounts for potential toxicity, so toxicity profiles usually have to be painstakingly validated in wet-lab experiments. The authors did try to predict potential toxicity in silico, but that only takes you so far—living organisms are just so complicated, and there's so much that we don't know about them, that you just can't reliably predict toxicity without actually exposing them to the compounds of interest.
posted by un petit cadeau at 11:51 AM on February 21 [16 favorites]


This looks interesting, and any lazy cynicism about it is tedious. It's a good post about an interesting thing, and there's lots to digest. Just reiterating that my favourite link in the post is the one to Derek Lowe's In The Pipeline blog.

Obviously, this paper is being published after it was shown to work successfully in mice:

From the In The Pipeline writeup: Halicin itself is shown in the paper to be effective in mouse models of drug-resistant bacterial infection, which is quite interesting. Topical infection with A. baumannii strain 288, which is resistant to all the usual antibiotics, was effectively treated with halicin ointment.
posted by ambrosen at 11:54 AM on February 21 [17 favorites]


I had to read a bit into the Cell paper to find out what's actually novel here, since it isn't, as the Nature article claims, the fact that they're training their model based on drug activity -- I was working on ML models for drug discovery almost ten years ago that did this, although the features were manually generated, consisting of a count vector with an index for every connected subgraph of the compound's graph representation (constructing these took a while) along with Tanimoto similarities to known active compounds, so it was a hybrid discrete-continuous representation. It seems that the method they're using for automated feature extraction is doing the heavy lifting here, which I guess the MIT article mentions, but in a way that made me unsure of whether they meant to say what it sounds like they were saying. Luckily, the paper describing their feature extraction methodology is available on PubMed!
posted by invitapriore at 12:01 PM on February 21 [8 favorites]



this is published in Cell, so lower likelihood of utter bullshit than a lot of ML-drug papers.



Once upon a time I worked in pharma research, and the project I worked on was to throw a hundred thousand novel compounds at a couple dozen simple tests.

One of the things we looked for was specificity. If a drug was active in all of our tests, y'know why? Because it's poisonous, which means it will kill the healthy cells too. So we looked for drugs that were active in just one test.

So when I hear this killed so many kinds of bacteria, I think it is likely a powerful toxin.


RTFA please. This is a trivial, sophomoric objection and one addressed in the paper (and indeed in antibiotic research in general). Many broad spectrum antibiotics are not toxic to mammals!

---

un petit cadeau's objection, with regard to mechanism & toxicity is important. the question is: how good a membrane disruptor is it, compared to other things - especially things toxic to mammalian cells? it didn't kill the mice, so it isn't quite 2,4 DNP, and if you're dying of a multidrug resistant infection, we already treat with some pretty shockingly toxic compounds already - we deal with the consequences of that later when the patient isnt dead from their infection........

also germane: is membrane disruption the primary mechanism? there are some indirect evidences that genes clustering to do with pH control and membrane integrity are found in resistant organisms, and the pH dependence of the drug itself is another indirect evidence, but more work is needed......

--

As usual, read the in the pipeline (derek lowe) piece. this is one of the few ML papers that I've read with both wetlab and in vivo ....this is thus work that I found compelling.

BTW: thankyou to not_the_water for linking to mostly good summaries, rather than wider media gibberish.
posted by lalochezia at 12:22 PM on February 21 [30 favorites]


While a cool piece of work, this compound is unfortunately likely to be incredibly toxic.

From the MIT press release: "Using a different machine-learning model, the researchers also showed that this molecule would likely have low toxicity to human cells."
posted by not_the_water at 12:45 PM on February 21 [4 favorites]


The real trick will be, if it works and it proves reasonably non-toxic to healthy cells, to keep the cost per dose under $4,491,887.39.
posted by delfin at 12:57 PM on February 21 [12 favorites]


Machine-learning models are exquisitely sensitive to the quality of the data used to train them, and it's a very good idea to generate that data yourself under controlled conditions if you can.
That is, this wasn't the new cool being applied to a field without understanding the field, it was the new cool rigorously built on a deep understanding of the problem. Nice.
posted by clew at 12:58 PM on February 21 [4 favorites]


That is pretty freaking awesome.
posted by captain afab at 1:08 PM on February 21 [1 favorite]


Clearing infections in two mouse models says that it is not super toxic to mammals.

I am skeptical about resistance proof. It is binding to something specific to the bacteria and that something can change to alter the affinity.
posted by dances_with_sneetches at 1:42 PM on February 21


delfin, that's also a consideration in pharma research. But unfortunately the other way around. If big pharma goes into their archive and finds a good candidate drug, but that candidate was synthesized 20 years ago, they put it back on the shelf and focus on making something that functions similarly but is new enough to patent, so they can charge $4,491,887.39 per dose.
posted by elizilla at 1:54 PM on February 21 [2 favorites]


If big pharma goes into their archive and finds a good candidate drug, but that candidate was synthesized 20 years ago, they put it back on the shelf and focus on making something that functions similarly but is new enough to patent, so they can charge $4,491,887.39 per dose.

It depends -- as I understand it, application is a factor in patent filings alongside the chemical structure itself, so repurposed drugs can be granted new patents if a sufficiently novel application is demonstrated.
posted by invitapriore at 2:46 PM on February 21 [1 favorite]


This boosted the chances that the drugs would work in radical new ways that bugs had yet to develop resistance to.
Can I just complain about the continued use of the word "bugs" in scientific reporting to describe microorganisms ? in 20freaking20 no less?

I still remember one time as a child, in the late 20th century, the doctor was describing some medication I'd be taking to prevent "bugs" from getting into my heart. "Amoxacillin?" I said, "Isn't that an antibiotic? what kind of insects are vulnerable to antibiotics?" and he said "oh no i mean micro organisms but I say bugs cuz it sounds folksy and people are too dumb to know what a microorganism is I guess?" or something along those lines.

It feels patronizing and a little insulting as well is what I'm saying.
posted by some loser at 4:00 PM on February 21 [2 favorites]



delfin, that's also a consideration in pharma research. But unfortunately the other way around. If big pharma goes into their archive and finds a good candidate drug, but that candidate was synthesized 20 years ago, they put it back on the shelf and focus on making something that functions similarly but is new enough to patent, so they can charge $4,491,887.39 per dose.


do you have evidence this happens beyond Ben Elton novels?
posted by lalochezia at 6:04 PM on February 21 [1 favorite]


It's certainly true with analgesics, where new, pattentable, drugs are given priority over existing off-patent drugs that are no worse but less profitable. This is a complete waste of research resources, but it is built into the system.
posted by sjswitzer at 6:10 PM on February 21


Can I just complain about the continued use of the word "bugs" in scientific reporting to describe microorganisms?

That really bugs you, eh?
posted by Greg_Ace at 6:28 PM on February 21 [8 favorites]


For the pedantically minded: 2,4-DNP is not a membrane disruptor. It works by shuttling protons back and forth through existing channels, not by punching new holes in cells that anything can pass through. "Uncoupler" is the better term.

Halicin may not work the same way, even if the net effect is still to disrupt the pH gradient inside and outside a compartment. It seems like it's suggested it hits iron homeostasis instead? That's a different mechanism.

Even if it were to work the same way, note DNP itself has a therapeutic window. (That's part of what makes it dangerous--people want to take it because it works and isn't guaranteed to kill you.) There's certainly room for a large window if you specifically want something that hits only bacterial cells and not mammalian mitochondria.

I was surprised to not see cytotoxicity experiments when I scanned the paper today, but the fact that this was an existing drug candidate means that this is probably well worked out already and one of the references point to it. There could still be issues but it definitely would not be some promiscuously toxic screening hit that a few posters speculate it was.

When it comes to discovering new drugs, generating new chemical leads isn't the hard part—we have plenty of methods available to do that

This is not correct. There are be specific projects at specific phases where it's easier to come up with leads than to test them, but lots of times it's not. "Find a new class of antibiotics that works via disrupting iron homeostasis" is absolutely a time when the lead itself is important.
posted by mark k at 9:13 PM on February 21 [8 favorites]


Doesn't this ultimately just train the bacteria to be more virulent?
posted by macross city flaneur at 9:21 PM on February 21


There's no reason to think that.

Generally the only thing you'd predict is it would cause them to evolve resistance towards this class of drug. Of course if it did that, even perfectly, we'd be back where we started, not worse off. They wouldn't get more infectious or anything.

With viruses resistance often comes at a price to the virus itself--they are less fit in other ways. I assume this is partially true of bacteria, but bacteria are a lot more complex than a virus and usually have more options to defeat a drug.
posted by mark k at 9:40 PM on February 21 [6 favorites]


The real trick will be, if it works and it proves reasonably non-toxic to healthy cells, to keep the cost per dose under $4,491,887.39.

And it'll probably work well until Big Agribusiness decides to feed it to all their animals to keep costs slightly lower and greatly accelerating the process of the bacteria developing resistance.
posted by JHarris at 10:15 PM on February 21 [1 favorite]


Antibiotic resistance does typically make bacteria less fit in the wild as well. The lay explanation is that bacteria tend to shed genes they’re not using because making gene products has an energy/nutrient cost, and when resistance arises not from a new gene but from a mutation in some core bacterial gene, those mutations tend to revert because they usually make that gene product work less efficiently. It’s rarely the case that the mutant grows better in the absence of antibiotics, because antibiotics tend to target very conserved parts of microbial biology, and if they’re that conserved over such a long evolutionary timescale there’s usually a really good reason.
posted by en forme de poire at 9:08 AM on February 22 [8 favorites]


Well said, en forme de poire. That's not necessarily true for wholly synthetic antibiotics, like those in this paper though.
posted by Orange Pamplemousse at 8:56 AM on February 23


Targeting conserved bacterial structures or proteins is absolutely a principle of rational antibiotic design so I’m not sure what you mean by that. While this antibiotic was not “rationally designed” in the typical way, the authors did also do the empirical experiment where they deliberately selected for resistance mutants with a lower dose of the drug. They didn’t get any over 30 days (which ironically actually made it harder to see what the drug’s mechanism of action was). Combined with halicin’s effectiveness across very diverged groups of bacteria (phylum level — Firmicutes, Actinobacteria, and Proteobacteria are separated by around 3 billion years of evolution), this suggests it’s targeting some aspect of bacterial physiology that’s pretty highly conserved. In that case, you would expect mutants to be less fit in the wild.
posted by en forme de poire at 9:56 AM on February 23 [2 favorites]


If something is strongly conserved across bacterial families, is it likely to also be present in animals?
posted by Joe in Australia at 4:14 PM on February 23


Bacteria are prokaryotes, we (and fungi) are eukaryotes. Antifungals are typically much more toxic to us than antibiotics.

"We need novel chemistries with novel mechanisms of action."

This is mostly legit, if you pare back the journalistic upsell, but this method by definition cannot rationally come up with novel mechanisms of action. They can figure out how to optimize an existing molecule, or use the criteria to screen a new library of molecules, but the input is a particular mechanism of action so you can't find new ones this way.
posted by porpoise at 5:25 PM on February 23 [1 favorite]


porpoise, the thing is that they didn’t just train it on antibiotics with known mechanisms of action. The input was growth rates of E. coli subjected to a couple thousand FDA approved drugs and a bunch of natural products; known antibiotics almost certainly were a very small minority of those. So it’s not just a matter of extrapolating from antibiotics where the mechanism is known.
posted by en forme de poire at 11:07 PM on February 23 [1 favorite]


« Older Nothing more British than fish and chips, brought...   |   Jackie Robinson vs. Malcolm X Newer »


This thread has been archived and is closed to new comments