entee
Joined 13 years ago
Karma
2,979
Submissions
634
Member since
Jul 19
Recent submissions
Comment · 11 days ago
Parent comment reached out asking for examples so thought I'd post my answer here:
We thought clearing amyloid beta plaques would slow alzheimer’s disease. We now have antibodies that do clear the plaques quite effectively, however they barely show any clinical benefit. Some trials failed entirely, others cleared a statistical bar that clinicians agree would be hard to detect by someone just living with the person. Essentially all patients given this drug still decline at more or less the same rate they declined before getting the drug. Definitely not the “we took away the cause of the disease, so at least they’ll stay about the same as they are today” effect we were hoping for.
To find this out, many trials costing (hard to get the specific number) but hundreds of millions of dollars had to be run. And this was easily the consensus target for solving Alzheimer’s.
If AI magically comes up with a new idea (Eliminate Protein X and Alzheimer’s will stop in its tracks), it will take us that long and that much money to know if it was right.
And before that we’ll have to run a huge number of lab experiments, likely involving developing new model systems (engineered mice usually, but not always) to test the concept before a human gets the drug.
So even assuming an AI model today is a “Genius” level talent, and that the target we need to go after is really already in some pre-existing paper or dataset, it will take a huge amount of investment to know.
And again, likely we need more data about how Alzheimer's works, because our best ideas haven’t worked the way we expected. Maybe AI can help us design ways to get that data, or ideas for what to gather. But it’s not going to magically come up with an answer where we simply do not know.
There’s a huge cardiology trial using Lp(a) as a target that just failed because it turns out that lowering cholesterol using the drug doesn’t actually reduce cardiac events. This was the most genetically validated open target in cardiology, and it turns out to not work how we thought. Hundreds of millions of dollars lost. (https://www.cnbc.com/2026/09/08/novartis-cholesterol-setback...)
Take another unrelated example: If AI told me, “I’ve done a bunch of geology, and I know there’s a huge vat of oil under the ground over there”. You’re gonna need to drill to find out. And if you give the AI all the information you have about geology and it says, “Maybe there’s oil under the ground over there, but I can’t be sure unless you take these rock samples”. Then someone needs to go out and get the rock samples, grind them up, stick it in a machine, get the data.
All those things take time and money. Ideas are cheap, measurements cost time and money. In drug discovery, we’re often against the frontier of what we know. Past the frontier you need new data. In computational fields (coding, video games, maybe in some ways mathematics) you can convert compute into data. In biology and the physical world, you have to do measurements. Compute can’t help you at the edge of data in the physical world.
Comment · 11 days ago
Have a comment below on this, but generally things don't work out in the clinic because we didn't understand the biology adequately, and not because we didn't have a molecule.
A better binder only helps if you know what to bind to and that if you bind to it, you'll have the desired effect.
This is why the industry has crowded around a very limited set of "validated" targets. It's because we know if we do something with them, they will work. Obviously we're likely missing a lot of great ideas though.
AI might help find some of those ideas, but we won't know until someone tests them in a human. Essentially, without more extremely expensive and difficult data to acquire, it's not likely to be that much better at picking winners.
I'm in the sector and I believe strongly in AI helping (I've bet my entire career on it). It will get there eventually, but it will take much longer than people seem to think at the big labs.
On crowding in big pharma: https://www.lek.com/insights/life-sciences-pharma/biopharma-...
Comment · 11 days ago
Some trials fail because we didn't have a "better" drug for a given target, many fail because we don't know the right biology. Biology and knowing precisely how to modulate the right thing is more gating than "design a better compound".
Better for what is the question and while sometimes the answer is "solubility, metabolism, off target toxicity", more often it's "you misunderstood something about the biology of this disease".
Maybe you needed more potency than your assays said you needed.
Maybe you missed something else the target does.
Maybe you had the right effect on the target but the body found a way to compensate (in an undesirable way).
Maybe the target looks like it matters in all sorts of model systems, but in real life it doesn't matter at all!
Many such cases. AI might marginally help in these scenarios, but doesn't look like it will eliminate the problem. AI can't help us with what we fundamentally don't know and have zero data for, or for which data is extremely hard to get.
Comment · 1 year ago
A lot of this post relies on the recent open ai result they call GDPval (link below). They note some limitations (lack of iteration in the tasks and others) which are key complaints and possibly fundamental limitations of current models.
But more interesting is the 50% win rate stat that represents expert human performance in the paper.
That seems absurdly low, most employees don’t have a 50% success rate on self contained tasks that take ~1 day of work. That means at least one of a few things could be true:
1. The tasks aren’t defined in a way that makes real world sense
2. The tasks require iteration, which wasn’t tested, for real world success (as many tasks do)
I think while interesting and a very worthy research avenue, this paper is only the first in a still early area of understanding how AI will affect with the real world, and it’s hard to project well from this one paper.
https://cdn.openai.com/pdf/d5eb7428-c4e9-4a33-bd86-86dd4bcf1...
Comment · 3 years ago
I think what’s missing is what the software allows. It could be BMW/Merc etc are way more conservative on what the allow the system to do and when they force the driver to take over. In certain contexts Merc is actually willing to assert and stand by a higher level of autonomy than any other manufacturer: (https://www.motortrend.com/news/mercedes-benz-drive-pilot-le...). Taking that at face value it’s possible they can do it and choose not to because they don’t want the liability. Whatever systems are in regular cars are then either borked or deliberately have less hardware.
Tesla is uniquely risk tolerant for better or worse. You also don’t hear about people getting into accidents in a BMW on self driving because they don’t make the same claims and have tons of safeguards.
Comment · 4 years ago
Transcription factors often are partially disordered, just to name one. A bunch of others here:
Comment · 4 years ago
> It’s only marginally less useful to actual biology than full on X-ray structures anyway.
I'm not sure what you're implying here. Are you saying both types of structures are useful, but not as useful as the hype suggests, or that an X-Ray Crystal (XRC) and low confidence structures are both very useful with the XRC being marginally more so?
An XRC structure is great, but it's a very (very) long way from getting me to a drug. Observe the long history of fully crystalized proteins still lacking a good drug. Or this piece on the general failure of purely structure guided efforts in drug discovery for COVID (https://www.science.org/content/blog-post/virtual-screening-...). I think this tech will certainly be helpful, but for most problems I don't see it being better than a slightly-more-than-marginal gain in our ability to find medicines.
Edit: To clarify, if the current state of the field is "given a well understood structure, I often still can't find a good medicine without doing a ton of screening experiments" then it's hard to see how much this helps us. I can also see several ways in which a less than accurate structure could be very misleading.
FWIW I can see a few ways in which it could be very useful for hypothesis generation too, but we're still talking pretty early stage basic science work with lots of caveats.
Source: PhD Biochemist and CEO of a biotech.
Comment · 4 years ago
Such a database would be hugely helpful across chemistry. Right now it’s extremely expensive to access databases like Reaxys or Scifinder, and they’re not usually programmatically searchable at scale. Some databases do exist based on the patent literature (https://depth-first.com/articles/2019/01/28/the-nextmove-pat...) but they’re not as well curated or complete. A pubchem like database for reactions would be really awesome.
Comment · 4 years ago
As a fellow perfectionist who has started a company, one thing that has helped me is realizing that most decisions are a lot more reversible than they appear. Even in the legal and financial domain, most things that you might obsess over are fixable if you make a mistake, and decent lawyers will tell you which ones you really have to avoid. Sometimes it'll cost you money and time, but the biggest cost is avoiding making decisions.
Always remember, no decision is a decision. Usually that's the worst choice because almost any decision, even a wrong one, at least moves the ball in some direction, allowing you to gather more information. The only guarantee in this game is that stasis will kill you, so bias towards action. When in doubt, try to evaluate "most probable bad outcome" which is different from "worst possible outcome".
Good luck!
Comment · 4 years ago
There's a lot that can be learned with building-block based experiments. If you do a building block based experiment then train a model, then predict new compounds, the models do generalize meaningfully outside the original set of building blocks into other sets of building blocks (including variations on different ways of linking the building blocks). Granted that's not the "fully novel scaffold" test, however it suggests that there should be some positive predictive value on novel scaffolds.
We've done work in this area and will be publishing some results later in the year.
Comment · 4 years ago
This is true. Getting datasets with the necessary quality and scale for molecular ML is hard and uncommon. Experimental design is also a huge value add, especially given the enormous search space (estimates suggest there are more possible drug-like structures than there are stars in the universe). The challenge is figuring out how to do computational work in a tight marriage with the lab work to support and rapidly explore the hypotheses generated by the computational predictions. Getting compute and lab to mesh productively is hard. Teams and projects have to be designed to do so from the start to derive maximum benefit.
Also shameless plug: I started a company to do just that, anchored to generating custom million-to-billion point datasets and using ML to interpret and design new experiments at scale.
Comment · 5 years ago
Not a chiphead, but saw this in the article that might be a reason ARM is better for this kind of thing:
"The theory goes that arm64’s fixed instruction length and relatively simple instructions make implementing extremely wide decoding and execution far more practical for Apple, compared with what Intel and AMD have to do in order to decode x86-64’s variable length, often complex compound instructions."
Not sure it's true, not an expert. But it doesn't sound wrong!
Comment · 5 years ago
Most of the methane comes from cattle belching, so it's basically impossible to harvest at scale. The lagoons of manure also do, and that can be harvested though it's hard.
https://climate.nasa.gov/faq/33/which-is-a-bigger-methane-so...
Comment · 5 years ago
The catalyst is the part I'm most curious about. Carbon capture is a hard problem at scale. Once you have methanol, I'm not terribly surprised that you can convert it to other things enzymatically. Hard to evaluate without full text of paper though.
Comment · 5 years ago
I'd agree if it works as advertised. Level 5 or very close to it is the key. No system has shown that, much less Tesla's. In the meantime, doing a live experiment with 3,000+ pound machines moving at 30+ mph seems like a bad idea.
Currently: Founder, Anagenex.com