Algorithmically-created medicine to be used on humans for first time
bbc.co.uk
bbc.co.uk
Giant hand-waving oversimplifications are certainly misleading-or-linkbait and usually both. So the rule means we should change it. It's just a question of what to change it to. Generally there's a subtitle or heading or photo caption or first sentence that says what the article is actually about—especially in major media, where the headlines are written by specialists who have nothing to do with the article. But I couldn't find any representative phrase in the article itself this time.
But, I guess if every algorithm is an AI, then this counts. But then so does every other instance of algorithms being used in science, which is everything these days.
Perhaps AI is just a buzzword used to gain attention???
More likely we follow the pattern we've had for the last 50-60 years; things we call "AI" stop being called "AI" after we understand how they work well enough.
Then we're also on observation studies with casuality since the 1980s between statistician and econometrician. We're using propensity modeling now for observational studies and quasi experiment. It's where a new wave of interesting stuff is happening all the while AI is taking up this weird hype in the programming world for healthcare.
Data science and ML really likes to make things black boxes while ignoring the data and not accounting for uncertainty. I'm going to flat out state, you're not doing this in healthcare. There is no way in hell this is going to work without statistic. Judea Pearl is gonna have to marry statistic with CS if he wants any real headway with causality and even then I'd take Rubin, Rosenbaum, and Angrist.
We're doing the same in other domains now.
More generally, the naysayers in this case are likely to be people who actually have some knowledge of drug development, and have seen countless companies that promised to revolutionize drug discovery by using computers, most of which are no longer in existence. If an AI-powered company can suddenly start passing Phase III clinical trials at an unprecedented rate, then I'll be genuinely impressed.
One of the first examples I remember reading about this was an antenna designed by an AI at NASA[1]. This was back in the days when "AI" wasn't the buzzword so much as "genetic algorithms" was. It ended up looking like this crazy zigzag thing.
1) https://www.nasa.gov/centers/ames/research/exploringtheunive...
1. Figure out what enzyme or receptor you have to mess with to cause the disease. You can skip this step if you decide just to blast away in a phenotypic screen.
2. Design a molecule that will actually mess with the enzyme/receptor to prevent it from malfunctioning in whatever way is causing the disease.
3. Figure out how to actually synthesize said molecule.
4. Now make sure this molecule can actually get to the sites where it needs to go to hit the target...
5. ... that it lasts there long enough and in enough potency to do its job ...
6. ... that it doesn't screw anything else up along the way (or get converted by your body into something that does)...
7. ... and that there's no compensatory mechanism that still causes the disease after your drug successfully gummed up what it was supposed to. Or maybe you failed at step 1 and you're barking up the wrong tree.
Roughly speaking, the last four steps will correspond to the different trials you have to run to get your drug approved. About 90% of drug candidates that make it to step 4 fail to make it all the way through step 7.
The difficult things that kill drugs are understanding their toxicity profiles and other off-target effects. We suck at this because we have so little data (and the effects are quite complicated), and modern AI techniques are mostly borne on shoveling piles of data and hoping for the best--not a good match.