Hacker News Reader

Comment by xzel | Hacker News Reader

People have been doing this exact thing for two decades at least but obviously with less computing power. There's literally nothing new about the idea. The real trick is being incredibly lucky and finding something that actually works in humans after multiple trials. I'm sure you know this based on your comment and this isn't really directed at you (truly wish you best of luck, I really hope the computing power and skill we have saves lives) but I absolutely hate these types of articles, which I've seen becoming more frequent the past year. AI + scientific challenge + possibility = puff article. This is the type of semi-hysterical reporting I expect from a local news station.

View on HN

I share some of your skepticism, but I think this particular paper is worth a little less cynicism. It’s true that statistical analysis has always helped inform medicinal chemistry efforts. They used to be called QSAR ;) What’s different here is the scale of information that was processed. As you point out the computational power and therefore the algorithms we can apply has enabled different experiments which I very much hope will yield better and more abundant medicines. Papers like this are a sign of these methods bearing fruit!

I totally agree, and I think my comment could be summarized as such from another angle. I think the recent Coronavirus + AI = cure spam has me a little jaded. I'll add this, this paper is really well done and like I said I really hope this works in humans!

I’m usually also skeptical of getting excited about successful animal trials for new medications, but with this I’m actually leaning more on the side of optimism.

With cancer treatments and antivirals in mice, we’re not so much targeting the pathogen as targeting the host immune system in the hopes it ends up nerfing the intended target (tumor/virus/whatever).

Given that the compound seems effective against C. Difficile (even if it’s in mice), I’d expect it to work elsewhere.

Of course, I’m not a doctor and have no idea what I’m talking about so grain of salt required.

Problem with many of these molecules that are found to work in mice, when put in human which is considerably more complex, say it goes thru the liver, the liver may break down most of it when circulating the blood stream. So now the drug needs to also resist that type of break down. Now multiply that by 10 to 100 of other types of reactions your body can generate from this.

Are humans more complex than mice or just different? I've always assumed the latter, but I'm in no way an expert.

The mouse genome is about 14% shorter than the human genome, although that could mean a lot or a little depending on certain other factors.

Well c diff is in the gut so it’ll work there probably.

> People have been doing this exact thing for two decades at least but obviously with less computing power.

Here's an analogy which shows how the cynicism here isn't actually useful wisdom:

Lee Sedol lost at go.

Kasparov lost at chess two decades before.

The Kasparov loss was 'exactly the same thing albeit with less computing power'

Correct. I don't understand why this contradicts the argument.

"Computers can learn complicated games with more computing power"

"Compute power goes up"

"Computer learns more complicated game"

I mean, if you really think of AlphaGo on an extremely high level, it's just a really elaborate way to create and learn a dictionary of moves to take under different circumstances. Of course that's going to be completely dependent on amount of memory and CPU power.

I think the criticism is that it's not obvious whether success here was a function of improved performance, expanded throughput, expanded testing, or sheer luck.

Chess engines have clearly improved in both design and computing power over the years; doubling an engine's resources or pitting a new engine against an old one produces straightforwardly better play. But the drug-discovery technique in use here may not be "playing better" in terms of producing higher-quality predictions.

To extend the chess metaphor:

- Deep Fritz is a stronger player Deep Blue even with 4% as much computing power. This story does not appear to be an algorithmic breakthrough of that source.

- Deep Blue lost to Kasparov in 1996, then beat him in 1997 with double the computing power. That's a clear improvement in play, but not an improvement in efficiency. This story might represent such a change, modelling more prospective drugs to test higher-confidence candidates.

- If an AI that can only win 2% of games against humans plays 10 games, it has an 18% chance of beating someone. But over 100 games, it has an 87% chance of a win. This result might be a team with a larger testing budget claiming the 'first win' without any AI-side improvement.

- If a dozen grandmaster-level chess AIs play GMs, one of them will have to get the first win against a human. Labeling this result a 'breakthrough' in AI terms might be outright publication bias among equivalent projects.

As far as the drug, none of that really matters, except that efficiency improvements would have more potential to increase drug discovery. The drug itself is still useful, and the discovery is a proof of concept; in 1980 no possible computer would have beaten Kasparov. But this is being hailed as a breakthrough in AI in seriously questionable ways. The BBC article, for example, managed to imply that this specific project was novel and important for using neutral nets to produce a significant result.

It's a bummer, from the headline you'd think the ML designed a molecule and tests with it showed results, instead it compared known drug interactions against available compounds, sort of a rainbow table matrix. You know, what researchers already do for a living.

There are differences over time, though. This drug repurposing, for instance, allows us to take drug candidates for other diseases and apply them to new areas. That is a fairly new approach when combined with ML in drug development, as in the past one would see brute force computation without repurposing. Repurposing gets better and better over time, as well, since there are more and more drugs on the lists that we've already run through a gamut of testing.

> There's literally nothing new about the idea. The real trick is being incredibly lucky and finding something that actually works in humans after multiple trials.

I guess I'm not sure where the dismissiveness is coming from here. Are claiming this could have been trivially done before? If so, why didn't you or someone else do it already?

Or are you claiming it's an uninteresting result that is not worthy of publication or attention?

I'm saying I'm tired of the puff piece articles and reddit style headlines. I was saying NN's and other "AI" style models has been used in this way for decades for these types of things. 0 disrespect to the actual science; the paper is great and I truly hope it works in humans. We need more drugs against the inevitable fight against drug resistance.

Okay, fine... what do you want to be different about the situation, then?

Do you not want any non-technical summary articles like this to be written? So that only those with the training to understand a Cell journal article would be able to learn anything about the result?

Or do you prefer that no journal articles be published that rely on 2020-era NN models, because older articles based on less state-of-the-art NNs have been published already?

The story could be about perseverance paying off, instead of a breakthrough portending big changes in near-future drug development.

I'd much prefer there be more detail in the article about AI research in drug discovery, how it still needs clinical trials in humans. I agree with the sub-comment as well. I think your last question doesn't make sense to me though as I haven't said anything similar to that.

But I think you raise an interesting point about non-technical summary articles: what do they do and who are they for? Does the non-technical public need to know about this research? Do they gain anything vs. reading the study's actual summary? I'm not sure. I do think there isn't much for non-technical people to get from this article that would be useful. I think the best reason would be for younger people to pique their interest in the field. Though, I honestly don't know the answer.

Yeah, good points. I guess such non-tech articles serve several purposes.

I do think this one gives non-technical people get something useful, though. More broadly, it'll increase science understanding among the non-scientist/non-technical public, at least on this topic, and good does tend to come out of that.

This is a novel and important result in antibiotics. It's also a proof-of-concept for using ML to produce vital drugs with novel mechanisms, rather than incidental alterations or discoveries in noncompetitive spaces. It might be an incremental speedup or computing-power advance in ML drug discovery also, but it could equally just be the result of a lucky break or a particularly large lab-test budget. (In which case, "why didn't someone do it already?" is closer to asking why nobody else bothered to win the lottery.)

It's not a major theoretical advance in ML drug-discovery techniques or the first big step in ML drug discovery. It's certainly not the invention of ML drug discovery or neural nets as an ML technique, both things I've seen implied in news stories on this work.

This is attention-worthy, absolutely. (I'll leave "publication-worthy methodology" to experts.) But it's newsworthy on actual merits, as a drug breakthrough and a demonstration of an increasingly-important technique. So I share the frustration when lazy or confused reporting implies this is the same style of ML-theory breakthrough as CNNs, Transformers, or even neural nets themselves.

Yeah, I used linear regression through Excel in my 5th grade science fair to draw a nice line over my data points that showed how plants grew with different fertilizers.

Linear regression is everywhere in science and that can be classified as machine learning artificial intelligence.

Still, it's cool that science is using more data and more intricate algorithms to produce fits.

I don't understand, they didn't find ONE they found NINE.

That doesn't sound like luck.

Finding nine is more like luck then finding one. Note that just one of the nine (so far) is validated. If I tell you to roll a six on a dice and you do that’s pretty lucky. If I tell you to roll an even number (I.e., give you wider scope) and you do ... less amazing. They could have lab tested every compound in the set but it would have taken too long. The ML for them to 9. I’m not saying that this isn’t good work. Just addressing the statistical point re 1 v 9.

That is sort of the thing about receptors, it is really had to know what will actually work. When you run these types of programs you'll get a number of possibilities that the simulator will spit out, that be from an AI or more traditional stat generated model. There were a few articles very similar to this with Coronavirus cures. They then have to go work in a well designed clinical trial, against a placebo, and then in humans, which is way more unlikely. It's most likely that all 9 don't work, unfortunately. I'm not saying the fact they found them was luck, the luck is if they actually work.

Reply on news.ycombinator.com