AI search of Neanderthal proteins resurrects ‘extinct’ antibiotics

https://www.nature.com/articles/d41586-023-02403-0

by mfiguiere • 3 years ago

79 52 3 years ago

c_crank3• 3 years ago

As far as 'AI risk' stuff goes, I expect this kind of shit to actually kill people.

astrange2• 3 years ago

We have medical trials for a reason. It doesn't matter what the drug discovery process is as long as you test them after discovering them.

ggm• 3 years ago

Agree. Personally I'm strong on "AI has real risks", which are mostly inflated beliefs in what it tells us. The protocols for drug testing are so high AI (if it works) is most likely to reduce prescreening delay, and if it had a risk it would be ignoring novel forms which don't conform to training, not high lethality: the drug test regimen deals with that already.

pixl97• 3 years ago

I'm assuming you mean 'We' as the US. Assuming everyone holds high standards may be a mistake.

jjk166• 3 years ago

One could imagine a sufficiently advanced AI coming up with drugs which are designed to appear harmless in drug trials despite still being dangerous, either because it legitimately wants to murder us or because its learning function will be incredibly biased towards passing drug trials. I mean, we already sometimes do that (see the Thalidomide Tradgedy), but presumably AI can do it faster and more efficiently in ways that we would struggle to anticipate.

shrimp_emoji• 3 years ago

Open source genomes of engineered super viruses combined with garage-grade nucleotide synthesizers will with more probability than anything else I can think of.

Or, heck, maybe the super virus escapes from a BSL4 lab on accident before the garage phase. :D (There's been precedent, and I'm not alluding to COVID.)

bulla• 3 years ago

What precedents?

XorNot• 3 years ago

I love how the Longtermism people won't shut up about this without having one single clue what they're actually talking about.

As though virus particles spontaneously pop into existence when you mix the right DNA strands together (a process that gets wildly expensive beyond a couple hundred basepairs and is destroyed if the proteins on your fingers get anywhere near it).

That's notwithstanding the fact that you can't just turn a dial labelled "lethality" to produce a gene sequence.

vouaobrasil• 3 years ago

Yes, if AI can do this, expect AI to also find diseases that are perfect bioweapons...such as a long-lasting common cold that merely weakens you for months, and is extremely contagious. Perfect for weakening an entire country before invasion, without even a hint that it actually is a bioweapon until it's too late.

pawelmurias2• 3 years ago

Weapons that attack the entire planet are not perfect.

Baeocystin• 3 years ago

How? Not snark, genuine question. What do you see about this form of active molecule search that makes it uniquely dangerous?

c_crank3• 3 years ago

It seems very close to resurrecting ancient diseases and protein strands 'just because we can'.

a_bonobo3• 3 years ago

This is the code used in the paper:

https://gitlab.com/machine-biology-group-public/pancleave

This package implements a scikit-learn-based random forest classifier to predict the location of proteoylytic cleavage sites in amino acid sequences. The panCleave model is trained and tested on all human protease substrates in the MEROPS Peptidase Database as of June 2020. This pan-protease approach is designed to facilitate protease-agnostic cleavage site recognition and proteome-scale searches. When presented with an 8-residue input, panCleave returns a binary classification indicating that the sequence is predicted to be a cleavage site or non-cleavage site. Additionally, panCleave returns the estimated probability of class membership. Through probability reporting, this classifier allows the user to filter by probability threshold, e.g. to bias toward predictions of high probability.

In the face of current hype around LLMs and 'fear of AI', calling a Random Forest Classifier 'AI' is a bit... far

beeforpork• 3 years ago

Every algorithm is now called AI, unfortunately. 'AI' is being used 'exponentially'.

wrp• 3 years ago

Under current usage of the term, I think the bimetallic switch in my office thermostat qualifies as an AI.

IanCal3• 3 years ago

Random forests are absolutely AI as the term has been used for a very long time in academia and in public.

bluenose69• 3 years ago

Agree completely, particularly with the smirk I can see in the final word! As an academic, "AI" means "fund me". It's the new "big data". And "exponential" is the new "rising" (or, sometimes, "falling").

jjcon2• 3 years ago

The AI term has been coopted by Hollywood notions of synthetic human-level intelligence but it really is just referring to the academic discipline which Random Forests definitely falls under and has for decades.

a_bonobo3• 3 years ago

Yes I agree, but OP's link is not aimed at academics, it's aimed at the general public. And you can't expect the general public to know that distinction: they see 'AI', they think 'ChatGPT', not Random Forests

nerdponx2• 3 years ago

We've also been using the term Machine Learning for a long time to try to avoid this. Using "AI" where "ML" would suffice is just trying to cash in on recent hype for clicks.

Still an interesting result of course.

tasogare• 3 years ago

In the face of current hype around LLMs and 'fear of AI', calling a Random Forest Classifier 'AI' is a bit... far

Just because the state of art evolved doesn't mean we have to erase history of the field. This is a ML algorithm so calling it AI is perfectly in line.

The fear you mention is built on the lack of understanding of what ML is. Showing that some AI has "dumb" yet useful implementations can help show the limits of this category of technology.

antegamisou• 3 years ago

Lol the wikipedia article of RFs is literally quoted to be "Part of a series on Machine learning and data mining" yet you have MLDevSecBullshitOPs-type experts here schooling you on what consists AI or not.

https://en.wikipedia.org/wiki/Random_forest

jeroenhd• 3 years ago

A random forrest classifier is a classic example of AI, quite popular before hardware caught up with the computational needs of neural networks.

Random forests of significant size also suffer some of the same inexplicability problems that neural networks suffer from, so it makes even more sense to make the comparison.

fodkodrasz• 3 years ago

At a job where the company specialized in mathematical optimizations (think operations research, logistic planning, timetable optimization, etc.) we jokingly called the solvers with heuristics "A.I. technology, short for Advanced If technology". Some customers got the joke, some didn't, but they were generally content with the results :)

selimthegrim• 3 years ago

Now the question is, could this be done for stuff like marine immune genomes?

lakomen• 3 years ago

Those are extinct for a reason. Why recreate those? For the same reason you don't recreate dinosaurs ( if you could ).

Hey let's recreate long extinct antibiotics, whatever could go wrong?

AnimalMuppet• 3 years ago

Sarcasm? Just in case it wasn't...

The antibiotics aren't going to break out of the lab, knock down people on the street, and inject itself into their arms. If the antibiotics work against today's bacteria, and are safe for today's humans, then what's the problem?

AI search of Neanderthal proteins resurrects ‘extinct’ antibiotics

Related Stories