This sounds great:
“It usually takes an average of five to 10 years [to discover] one drug. And maybe we could accelerate that 10 times, which would be an incredible revolution in human health" https://t.co/Vl2vdygUPo
— James Pethokoukis ⏩️⤴️ (@JimPethokoukis) January 21, 2025
But you know what will be even better? Sometime in the medium future AI will have such a detailed understanding of human physiology—including all its variations—that it will be able to run virtual clinical trials. Anything that passes goes straight to Phase 3 human trials just to be 100% sure, and then gets approved. The entire pipeline will be reduced to a couple of years and will only go down from there.
How long until we get there? I don't know. Four or five years? This stuff is all closer than we think.
We do not have an understanding as to why some people have adverse reactions and others do not. AI would have to prove that it can predict those reactions in the tens of thousands of drugs we already have before anyone is going to even begin to trust it on never used drugs.
Much more than five years away. Further away than self driving cars. Probably further away than fusion power.
While AI might have a detailed understanding of human physiology, it would only have an understanding of average physiology. People vary. There would be no way of predicting how actual people would react unless it had a complete database of everyone's individual physiology. Not gonna happen. It might replace Phase 1 trials, but we can just about do that now. Later phases would still be needed.
FWIW, Phase 3 studies can't really either, no matter how hard we try, we do the best we can with an imperfect process (within the precautionary principle).
If anything, AI especially with quantum computing will likely be vastly superior in this respect, though Phase 3 will stick around for a very long time regardless.
Is there any reason to believe quantum computing will ever work?
AI can definitely make interesting and useful predictions based on pattern matching.
AI cannot simulare the physics, chemistry, and biology that actually determines how and whether a drug will work. It will especially fail on truly novel compounds or on ones in which phenomena which were not significant in other similar compounds have a major effect.
I think AI is going to be great at suggesting drug candidates, at identifying already rejected drug candidates that can be useful for subsets of patients (ie. drugs with bad side effects for patients with some genes that can be useful for patients who have different genes), and at suggesting novel drug combinations.
I think there is a danger that it will falsely reject many compounds with unusually or novel properties as drug candidates. That will become an issue if it is used to reject drug candidates before testing for efficacy.
I don't know why you credulous keep taking the word of an enormous collection of proven hucksters and liars.
This.
This is so scary it seems like it might be satire. But let me be clear:
* Using AI to accelerate the design and development of new drugs? Check. This is a great idea and likely to speed things up by a lot.
* Using AI to completely replace clinical tries with virtual trials? No. This is a very, very bad idea. No matter how good a model is, it's still a model. Many models are useful. Some are very useful. All models are wrong. All of them.
‘AI’ can process vast amounts of data to build very sophisticated models - if the data exists. Amassing a dataset of human physiology that could support models sufficient to replace in vivo testing requires a lot of research that AI can’t accelerate much if at all. KD keeps making this category error; lots of problems are not limited by computation, but by data. Someone else here stated it as, human data has to be collected in meat space and meat time, before we can analyze it in cyber time.
Bingo.
Drug candidate can work fine in the lab, but most wash out in clinical trials. Side effect, metabolism and metabolites, targets desired organ in the body but affects others too, hits the target--but their is no clinical effect, etc.
It is totally a category error. Some people seem to believe that life is like a video game: there exists an "intelligence" stat and when the value of your "intelligence" stat gets high enough, you get the "cure cancer" dialog option.
> But you know what will be even better? Sometime in the medium future AI will have such a detailed understanding of human physiology—including all its variations—that it will be able to run virtual clinical trials.
I was -sure- that this was snark on Kevin's part. Was I wrong? Maybe so. Sigh. Poe's Law is a harsh mistress.
Not at all. Kevin predicted AI would basically solve all the worlds problems last year and he'll never back down from that. Kind of like his years of completely wrong predictions on the economy since 2021. "Just a bit longer..." will be his mantra.
Sure man. Let's all drink the Kool-aide our corporate overlords are serving us. There is no possible downside.
it's very easy: we attribute to "AI" certain qualities and abilities (AI will be able to design drugs! AI will be able to model drug efficacy!) and then sit back and marvel at the prospect of AI doing those things.
It's a bit like musing (or getting investers jazzed) about the implications of LaPlace's demon. Yes, an entity that could do all of those things could do all of those things!, but it's thats different from explaining why or how the tech we are developing fits that category.
Good Lord. Calling it now:
"AI will be creating and testing our drugs in the medium future" is the new "self-driving cars are just around the corner" for Kevin's blog, since he never seems to account for complexity in handing things entirely over to computers. A computer-designed drug (or computer-driven car) that is good enough for 99.999% of situations will still cause thousands of deaths in a nation of 330 million people. And although human-designed drugs (and human-driven cars) also kill thousands of people, human-controlled anything is the status quo and the deaths caused by the human-controlled anything are thus less scary and (more importantly) less able to file corporation-bankruptable lawsuits over.
So this is where we need tort law reform.
When we have self driving cars that are safer than human drivers we need to recognize that people will still die and that by making the makers of these products liable for those deaths we are preventing their roll out and thereby killing people.
That means we set up something similar to the vaccine injury compensation program and make human drivers contribute more to it than drivers who use self driving. We eliminate torts except for clearly negligent actions (ie. drunk driving).
Doesn't Waymo have hundreds of self-driving cars operating in the U.S. right now? (Number of people killed = 0.)
This has to rank as the most pollyannish, uninformed post I've ever seen KD make. It's like he's totally forgotten the concept of GIGO. AI models know only what they have been trained to know, which means they know only what humans know, as far as observations are concerned. There is simply no way that the full complexity of a biological system like the human body can be successfully modeled in silico given our current knowledge base. I'm sure AI models can hallucinate whatever we might want them to, and a whole lot that we wouldn't want them to, and I'm also sure that they can make more sense of at least some of the complex systems that make up the whole than we mere watersacs can. But they cannot magically divine answers in the absence of data.
I can't imagine we will phase out Phase 3 trials for at least a few more decades, if this technology is even ready by then (color me skeptical, but not dismissing this tech could be legit within next decade).
How long until we get there? I don't know. Four or five years? This stuff is all closer than we think.
We'll get there in a fully autonomous EV. That's how long it will take.
Once AI is doing all the heavy lifting for drug development the prices will come way down, right?
AI doesn't understand human physiology. It understands, in a very limited way, papers about human physiology. That's a very limited understanding. Right now, despite all the hype, AI is maybe ever so slightly useful for candidate discovery but not much else. Finding candidates isn't the hard part. It's maybe 2-3% of drug development. We aren't even close to considering using AI to assess basic safety as would be done for a Phase 1 trial.
The problem with AI is that it is largely being done by computer science people, and computer science people know little about other fields and generally view their experts with contempt. They think those fields are as easy as software, but most software is at best mediocre and biology and chemistry are much harder. Biology is about reverse engineering and patching 3.5 billion year old kludge code and chemistry is about exploring a space that makes the set of all possible programs look tiny.
I don't think AI is useless in these fields, but if you said 20-30 years I might believe you.
" It understands, in a very limited way, papers about human physiology." No it doesn't. AI is fundamentally incapable of understanding anything. It's capable of making predictions, not understanding. It's amazing that Kevin so loves regression and doesn't understand that AI is simply non-linear regression.
Here's a hot take: human physiology boils down to DNA, right? And we can sequence that, and DNA codes for proteins, and we can infer which proteins are coded-for, and how those proteins fold, so haha, we should be able to predict all of human physiology, amrite? I mean, all that biology shit is just stamp collecting!
Sadly no. There's an entire field, systems biology, which continues to try to map out the gene and protein signaling networks that exist inside of cells. It's very, very complicated. Sure, maybe someday AI will map it all out. But I wouldn't place a bet on how long it takes.
Sheesh.
And you haven't even touched the epigene yet.
Good point. I thought about it but .... why make things even more complicated when just DNA-based gene signaling is already a massive snarled rat's nest? But yes, there's epigenetics, and then there's all the ways in which people are affected by the environment, and (haha) from what I understand, the percentage of genetic chimeras among humans is a lot higher than we might think: some hypothesize as much as 10%. And there's probably lots more.
I remember reading once that there's, like, two chemicals (vitamins) that get studied in carrots, but there are 4000 organic chemicals in carrots that could be affecting us. Same thing with oranges: I remember reading about interesting studies that showed that taking vitamin C didn't have anywhere near the good effect that actually taking orange juice did.
It's way, way more complicated than AI folks pretend.
"and how those proteins fold," That's not doable yet.
Half of the Nobel Prize in Chemistry this year went to guys who applied AI to predict protein structures. This is an area I know something about, and I can tell you: the performance of AI in predicting protein structures blows the mind of everyone who understands the nature of the problem.
You've been saying it's closer than we think for a while now and yet it never arrives. It was . . what? 10 years ago you predicted that self driving cars were only 2 or 3 years away.
These things have a history of not being nearly as close as you think. In this particular instance, I don't see the FDA approving any of this technology unless and until someone can explain how the black box of AI reaches its conclusions. How do you know it isn't just hallucinating treatments?
It won't take 5 years, it'll take 20 years. It will happen, in some capacity, but the FDA will test the process in parallel for a decade or more before allowing trials to skip from pre-clinical to Phase IIb or III. The first time that someone like Elmo pitches a hissy-fit and tries to skip to Phase III or even skip human testing in someplace like Antigua, Angola, or The Philippines, basically a country looking to earn some easy Dogecoin, someone will get hurt badly because the computer didn't model as precisely as human biology can get, the model didn't expect epigenetics to be as varied, or something random like lactose-intolerant left-handed red-heads have severe allergic reactions.
When that happens, everyone will recoil and the FDA will be vindicated and then they'll have their pipeline of testing the testing models completed before another Elmo hissy persuades a corrupt leader to let him try again.
PS: My industry will suffer from this but no one worth anything will shed a tear over quicker and safer drugs without using human guinea pigs. But that's what the first Elmo hissy fit will do: turn the Phase IIb or Phase III participants into human guinea pigs.
"Sometime in the medium future AI will have such a detailed understanding of human physiology Understanding and AI are mutually exclusive. AI ight be able to give right answers some/most of the time, but it will never ever provide understanding and will always flat out lie some of the time.
AI, at least in theory, and in some implementations, can "understand" things if the programming is designed and structured to do so, though that "understanding" so far is limited to fairly small domains. OTOH, the current generation of chatbots and wizards labeled "AI" do not make any attempt to "understand" at all.
Maybe AI will be that good and help expedite the process.
But I wonder if individual differences in people will make that more difficult. My understanding is that medicine is so difficult since we're so different. To make any one-size-fits-all drug so to speak, isn't easy at all.
AI also needs to know all the variables in order to work correctly. But it can't test for unknown factors – only the ones it knows about.
Kevin, Kevin, you believe just about anything any AI propagandists tell you! They are drunk on their own success, such as it is. Plus they are fundraising for new start ups (it's lucrative for founders even if the startups fail. The only ones that suffer are the employees with their worthless stock options). None of this stuff should be taken literally a nd not even seriously at this point in time.
Understanding physiology in such depth is much more difficult than this idiot thinks. It requires real experimentation--on human subjects which puts you under strict regulatory and ethical rules on top of all other difficulties.