Scientists discover new antibiotics using AI
euronews.com
euronews.com
When you invent new weight loss drug you can sell millions of pills to anyone.
If you invent new antibiotic it is immediately classified as "reserve antibiotic" and the sale is restricted.
Is that so medical professionals can break it out as a means of last resort to save a patient and ensure the course is finished?
If so that mightn't be a bad idea before we destroy the potential of new antibiotics too by having the general public misuse them render them useless over time.
Oh, and you still have to prove it to be safe and effective through years of clinical trials.
AI is basically still only helping us with the easiest parts, and this particular class is drugs is essentially a limited resource (to avoid force-evolving stains of bacteria that are resistant or immune).
At the end of the day, there's not as much research because there's a finite amount of resources to go around and there are easier / lower hanging fruit.
Anyway, how does holding antibiotics in reserve help anything? Why let the bugs evolve resistance to one antibiotic at a time, rather than hitting them them with multiple antibiotics and thus raising the evolutionary hurdle to resistance? Was this strategy designed by the bad guys in a Jacky Chan movie?
The government may have contributed to the discovery of nuclear power, but for the last 50 years it has been the largest impediment to nuclear power.
The private sector has never done anything like what NASA has accomplished, and NASA is far and away the most successful explorers and engineers in human history. I have no reason to believe anyone else could do it.
Has Space-X left Earth orbit?
> NASA spent a quarter of a TRILLION dollars… to send out a handful of probes
Look at NASA's budget (or any other source) and you'll find they do far, far more.
The all caps, btw, are against HN guidelines. Why have so many started using them this week?
Nonsense. NASA spent 50 years using disposable lift vehicles. Only the government could bring itself to create single-use articles that cost hundreds of millions of dollars. SpaceX has fixed this by building the Falcon 9, which is not only reusable, it has the highest success rate of any orbital rocket in human history. They're not only "doing something like what NASA has accomplished", they are objectively doing it better and far, far cheaper.
> and NASA is far and away the most successful explorers and engineers in human history. I have no reason to believe anyone else could do it.
No belief necessary. SpaceX has already done it.
> Has Space-X left Earth orbit?
Yes.
> Look at NASA's budget (or any other source) and you'll find they do far, far more.
My point wasn't that they didn't do a bunch of stuff, but rather that they didn't do enough to justify their $0.25 Trillion per-decade budget.
> The all caps, btw, are against HN guidelines.
Ah, I see that in the guidelines. I'll avoid them in the future, thanks!
What exploration and engineering has SpaceX done to match NASA (except in one narrow area). NASA has a telescope at L2 looking back over 13 billion years, a helicopter flying on Mars, it has left the Solar System, it has satellites and probes all over orbit and other planets.
If only the world were as simple. But it isn't, and everyone knows that those with power (in general, governments) will never be regulated effectively enough for everyone - but that doesn't make them never to be trusted with monopoly, nor does it mean that the government's monopoly is bad.
Sorry, I don't really want private militaries nor private tax collection. I'm happy the government runs things like IDs and drivers licenses. In fact, various department of motor vehicles (or whatever your local thing is called) is an excellent example of how a monopoly can be a good or bad experience. I'm originally from Indiana and the motor vehicle folks are great there and the service is easy to use - by design - but it isn't like that everywhere in the US. This has nothing really to do with how "effective" their regulation is.
I'm not sure actual monopolies are better - are you really satisfied with your electricity provider or ISP?
When we can safely give multiple therapies without harming the patient, we do it. Standard HIV treatment uses 4 different drugs to raise the evolutionary hurdle for the virus.
You can't have a government without a military or taxes, so it makes sense that government handles those things. They are requirements for the existence of government, and can't be delegated without existential risk. Medical research is not a core government function.
> I'm happy the government runs things like IDs and drivers licenses.
You'd probably be happier if you weren't required by law to get a license in the first place.
> In fact, various department of motor vehicles (or whatever your local thing is called) is an excellent example of how a monopoly can be a good or bad experience. I'm originally from Indiana and the motor vehicle folks are great there and the service is easy to use - by design - but it isn't like that everywhere in the US. This has nothing really to do with how "effective" their regulation is.
"Sometimes governments can handle basic services" doesn't strike me as a great argument for entrusting the government with complicated, vital services.
> I'm not sure actual monopolies are better - are you really satisfied with your electricity provider or ISP?
The alternative to government monopoly on the development of antibiotics isn't a private monopoly, it's no monopoly, and ideally without regulations which destroy the market incentives to develop antibiotics.
If you have some ideas, do share them.
> rather than hitting them them with multiple antibiotics and thus raising the evolutionary hurdle to resistance?
It was surrounded by some rudeness and is probably a poor idea (but I'm no expert), but it was an idea.
It was a joke. Please laugh.
Plus the normal gut flora would be disrupted by the multiple antibiotics as well, so there would be more work involved in the recovery of the symbiotes and commensals.
If they have novel mechanisms of action, quite a lot. If it's just another entrant in a class of antibiotics for which we have numerous drugs (and thus resistances) already, probably not as much.
You want to preserve your own bacterias, you won't survive without them. Combinations are effective, but they might kill too much of your friendly gut bacterias.
The underlying problem is the lack of political will and cooperation to implement these solutions.
Market-driven capitalism without any government incentives (or disincentives) has no answer here. You just proposed something that's not market-driven capitalism. That the government should get involved in some way to fund and incentivize what it deems to be good for society in the long run. The market on it's own can't do this.
When making brain-dead claims, at least be fair and compare it to the track record of other economic systems. To do otherwise is shameful.
Are there examples of this? I've heard people say this happens, but I've never heard a concrete example.
I googled 'reserve antibiotic' and found Ceftobiprole, and see it's for sale for $40,000/gram, so clearly it's possible to charge money for a 'reserve'.
They did a virtual screen of a large compound library.
QSAR / QSPR / Virtual Screening has been around since the late 1950s.
The secret sauce here was the large experimental dataset - on the order of 10^5 compounds - they generated to train the model.
Maybe the explainability part is kind of novel for a neural network based approach; but not clear that you couldn’t identify substructure classes better with a pure bioinformatic approach.
"The insight here was that we could see what was being learned by the models to make their predictions that certain molecules would make for good antibiotics," James Collins, professor of Medical Engineering and Science at the Massachusetts Institute of Technology (MIT) and one of the study’s authors, said in a statement.
...
"What we set out to do in this study was to open the black box. These models consist of very large numbers of calculations that mimic neural connections, and no one really knows what's going on underneath the hood," said Felix Wong, a postdoc at MIT and Harvard and one of the study’s lead authors.
> not clear that you couldn’t identify substructure classes better with a pure bioinformatic approach.
Lots of things aren't clear, but what supports your particular claim? Has it ever been done? And is it addressed in the article?
So, yes, it’s been done at scale for decades.
The article makes all sorts of incorrect claims. For example, it says it has been 60 years since the last antibiotic was discovered, and that this works enables computational screening of antibiotics for the first time.
Here’s work from 2022 that used conventional computational screening to discover a novel antibiotic:
https://phys.org/news/2022-10-discovery-antibiotic-resistant...
HN is a forum of intellectual curiosity. The near-universal critiques are the opposite of that.
Do you trust journalists that much? Have you ever read an article about something you are an expert on? You shouldn't trust articles that much. I trust HN discussions much more since there are plenty of people who will weigh in when things are wrong or misrepresented, unlike these articles.
Edit: Gell-Mann Amnesia effect, that was what this is called. Don't trust the media just because you aren't an expert.
> I trust HN discussions much more since there are plenty of people who will weigh in when things are wrong or misrepresented, unlike these articles.
Why are those people, with no responsibility, little expertise, etc., more trustworthy than other people? It seems like the common bias of social media - for some reason, people trust strangers with no credibility, which exposes them to the enormous amount of misinformation and just BS that now floods our world. If the article author or a professor wrote the same thing in a HN comment, it seems like you'd believe it.
I don't trust them, I trust the discussion in aggregate. The discussion includes people on all sides in a healthy forum. If you are on a subreddit you are in an echo chamber so many voices will be missing, but HN has good representation of experts and people from all sorts of fields so I trust these discussions to unearth better truths than what a typical journalist can.
> If the article author or a professor wrote the same thing in a HN comment, it seems like you'd believe it.
No, if they wrote the article in a comment with that headline it would get comments saying why it is wrong, just like here. The ability for people like you to object is why I trust it more.
Regarding things I already know - or even OPs I've read but which most commenters have not - HN is mostly misinformation.
I participate in HN, obviously - only because I've found nothing better (a complement to it, in a way, and maybe there's no known way to improve).
If you are saying that most discussions here on HN have zero people who are right, then I strongly disagree. It is very rare for there to be no truths in HN discussions. On reddit however it is very common for me to see not a single reasonable post to be found.
For greater context, in the drug discovery world, "explainability" of QSAR/QSPR has been a longstanding goal. It's (rightly) not considered sufficient to have "black box" models that make predictions -- it's too expensive and risky to carry drug candidates into the lab based on the output of an algorithm, so subject-matter experts want to know why the algorithms are making the predictions that they make.
Do you see the irony? What better describes social media? Personally, I see plenty of good, valuable, useful science journalism; it's got plenty of flaws, like any human endeavor - including social media comments.
But attacking journalism is normalized, and thus people know they can do it without being questioned and get some social media likes from it. It would be interesting to see the same responses turned on individual social media comments - what flaws could we find? :)
> For greater context, in the drug discovery world, "explainability" of QSAR/QSPR has been a longstanding goal. It's (rightly) not considered sufficient to have "black box" models that make predictions -- it's too expensive and risky to carry drug candidates into the lab based on the output of an algorithm, so subject-matter experts want to know why the algorithms are making the predictions that they make.
Ironically, that's what the OP is about; that was the point of their research.
Social media is not one thing. You've got multiple people in this thread who know the subject, in detail, telling you that the headline is wrong. These aren't youtube video comments.
> Ironically, that's what the OP is about; that was the point of their research.
Yes, I know. That's why I wrote that. The research is about a relatively obscure technical problem. The headline made it sound like a revolutionary breakthrough in antibiotic development.
Fair enough, but science journalism is not just one thing.
> You've got multiple people in this thread who know the subject, in detail, telling you that the headline is wrong.
That appears on every HN discussion of anything. It's not a signal of wrongness or rightness, any more than the color of the banner at the top - if it's orange, the OP must be obviously wrong!
Also, your comment is a baseless claim - like the strawperson journalists you rail against - that HN commenters are more informed than the authors. I expect that very few HN commenters are as informed.
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC1426172/
is a new class of antibiotics discovered around 2005 that also works against resistant bacteria such as MRSA and VRE.
So while it is great they are coming up with new antibiotics, it is by no means the first in 60 years.
Disclaimer: spelling and name of antibiotic may be incorrect. It is not my field, and the spelling is phonetic from myself. If pressed, I will ask the scientist for the specific name.
https://pubmed.ncbi.nlm.nih.gov/20334458/
Discovered about at the same time or later. Also works against MRSA.
While this speeds up the screning process quite a lot you might as well feel like someone might just have thrown something out that might actually turn out useful and you can't backcheck without resorting to testing everthing.
When I read the headline I was expected something more along "a network that knows what processes we have and therfor could imagine what compounds we could come up with" combined with that "inverse" network that predicts weaknesses of bacteria. Then those two would be matched and we would have to figure out the missing steps how to actually build the compound coming from that "black box" prediction.
The benefit would be that we would be presented with something that could exist but we hadn't thought of because of our processes and limited resources.
While I value the speedup and cost effect I'd think we could do much better here in exploiting the "creative" and powerful numerical inverse potential of ANNs especially in this area and outsmart those evolutionary skills of bacteria that way.
a lot of biology, as well as where its been exploited, relies on specific binding, along with this comes the hope that what you test won't inadvertently show off-target effects. Through looking at what's already available for information you can actually use the lack something existing within a human system as a probabilistic edge over the future success of their application. in a regard its like an artistic rendering that makes use of negative space, its an impressive direction.
Additionally, it can't be called an antibiotic until it goes through the necessary human clinical trials, where all kinds of problems can crop up - they don't seem to provide any evidence that their AI-enabled scan for human cell toxicity actually worked.
Finally, the real problem with antibiotic resistance is that microbes always evolve resistance over time if the antibiotics are overused (i.e prophylactic use in industrial animal agribusiness, and careless prescriptions of antibiotics from doctors).
LLMs are just the ones that are popular right now because even non technical people can interact with them.
In the mid 2010s it was image classification, for example.
Modern AI is the synthesis of new techniques and old ideas plus huge datasets and huge computing power.
this was essentially my point, that a /s might have made more clear, and made cost less internet points. ai and how the term is used commercially are very different things. the latter tending to be a myopic view of the former that is constrained by dollar-seeking.
ai is getting to be old as shit. ai hype is in its latest wave of commercialization. llms being the current so hot right now.
It's still not "intelligence" but it does use techniques inspired by brains and have some of the same advantages in "seeing" patterns.
Marketing.
AI is "artificial intelligence". That's a pretty intentionally vague term. It doesn't imply any specific technological implementation, but it does imply the device/software makes decisions autonomously that would otherwise need a human to intervene. There's a million holes in that definition, and that's why the adjective "AI" kind of sucks when used on its own. It's also worth noting it's been used in sci-fi for eons. It's also been used to describe non-playable characters in video games for a long time, which obviously didn't have the sort of "AI" we think of now.
LLMs are neural networks that work with "tokens" from text. So a big statistics-based predictive AI. There are more complex implementations that can consume other kinds of content as well, but "LLM" almost always refers to a neural network that consumes small chunks of text called tokens, and outputs more tokens based on that input. The magic we see with LLMs happens because given enough data, language itself builds up a surprisingly good model of "thinking".
There's some really good explanations in the sibling comments! I'm no expert, and I think someone can probably correct me here if I messed up something there.
I don't want to be too rigid and I also like to play with names, but the use of the word AI is getting out of hand and laypeople and professionals alike are believing stuff that these programs are not (yet) capable of doing.
This started for me some 10 or 15 years ago with the word 'cloud'. I was out of the field for some time and had a hard time understanding what the 'cloud' was. The word 'smartphone' also comes to mind, although this one is less elusive.