Please create an account to participate in the Slashdot moderation system

 



Forgot your password?
typodupeerror
AI Biotech Medicine

Google DeepMind's Spinoff Company 'Very Close' to Human Trials for Its AI-Designed Drugs (fortune.com) 40

Google DeepMind's chief business officer says Alphabet's drug-discovery company Isomorphic Labs "is preparing to launch human trials of AI-designed drugs," according to a report in Fortune, "pairing cutting-edge AI with pharma veterans to design medicines faster, cheaper, and more accurately." "There are people sitting in our office in King's Cross, London, working, and collaborating with AI to design drugs for cancer," said Colin Murdoch [DeepMind's chief business officer and president of Isomorphic Labs]. "That's happening right now."

After years in development, Murdoch says human clinical trials for Isomorphic's AI-assisted drugs are finally in sight. "The next big milestone is actually going out to clinical trials, starting to put these things into human beings," he said. "We're staffing up now. We're getting very close."

The company, which was spun out of DeepMind in 2021, was born from one of DeepMind's most celebrated breakthroughs, AlphaFold, an AI system capable of predicting protein structures with a high level of accuracy. Interactions of AlphaFold progressed from being able to accurately predict individual protein structures to modeling how proteins interact with other molecules like DNA and drugs. These leaps made it far more useful for drug discovery, helping researchers design medicines faster and more precisely, turning the tool into a launchpad for a much larger ambition... In 2024, the same year it released AlphaFold 3, Isomorphic signed major research collaborations with pharma companies Novartis and Eli Lilly. A year later, in April 2025, Isomorphic Labs raised $600 million in its first-ever external funding round, led by Thrive Capital. The deals are part of Isomorphic's plan to build a "world-class drug design engine..."

Today, pharma companies often spend millions attempting to bring a single drug to market, sometimes with just a 10% chance of success once trials begin. Murdoch believes Isomorphic's tech could radically improve those odds. "We're trying to do all these things: speed them up, reduce the cost, but also really improve the chance that we can be successful," he says. He wants to harness AlphaFold's technology to get to a point where researchers have 100% conviction that the drugs they are developing are going to work in human trials. "One day we hope to be able to say — well, here's a disease, and then click a button and out pops the design for a drug to address that disease," Murdoch said. "All powered by these amazing AI tools."

Google DeepMind's Spinoff Company 'Very Close' to Human Trials for Its AI-Designed Drugs

Comments Filter:
  • so good luck with pressing your button.

  • by simlox ( 6576120 ) on Monday July 07, 2025 @01:22AM (#65502302)
    Traditional calculations on proteins are very, very hard as the amount of approximations needed to get from the fundamental physics equations to something you can actually get into a super computer are so many, that the results are totally unreliable. Sometimes there are some correlation with experiments, sometimes it is just a noise generator - and it is hard to tell why. Machine learning models can do similar if not better predictions for less.
    • And then there's the toxicity screening part. You have to find something that adequately binds to your target in an inhibitory manner AND make sure it won't bind or accumulate anywhere else in a human body. And if it's a treatment for cancer you have to make sure that it will be infeasible for the tumor to use any of the mechanisms of resistance to evade it.

      • by dvice ( 6309704 ) on Monday July 07, 2025 @03:00AM (#65502378)

        Toxicity screening part became easy because AlphaFold2 created 3D model of all human proteins, so you just need to run the binding test against all of those to see if you have unwanted side effects or not. If the drug doesn't bind, it shouldn't in theory have any side effect.

        • But binding sites can change shape with different, otherwise uninteresting, chemicals present. E.g. metal allergies happen even though the metals tend to be found in bits too small to crosslink antibodies themselves, but can twist binding sites to cause reactions to otherwise benign proteins in the body. Shortcutting basic research could be a useful application but we won't be at a level to skip vivo studies until around the same time as we can make "Real" Human equivalent AI. There are too many interact
          • by ceoyoyo ( 59147 )

            we won't be at a level to skip vivo studies until around the same time as we can make "Real" Human equivalent AI.

            The latter is probably considerably easier than the former. We know that "real human equivalent [A]I" can be created. There are ten billion examples walking around. We don't have an example of an oracle machine that can tell you what a drug is going to do when you put it into one of them.

            We also have reasonable bounds on the amount of computation that goes into making one of those "real human" in

          • And it's not enough to screen the drug against every protein in your body. You also need to understand how it gets broken down, and test every product it gets broken down into to make sure none of them is toxic. And it can't get broken down too quickly, or it won't have a chance to work.

            And even if the drug doesn't bind to anything else on its own, it might still bind in the presence of some third molecule, so you need to consider all possible combinations with other molecules found in your body.

            Computati

        • by ceoyoyo ( 59147 )

          That's a nice story, but it's not true. Humans all make variations on the "standard" set of human proteins. Sometimes not important variations, sometimes very important. Proteins aren't one and done either; they're reused all over the place, in different ways. You can target the such and such a receptor on such and such a cell and your drug is going to have off target effects because that same receptor is used in a dozen other cell types to do two dozen other things.

          Not to mention drugs don't stay the same

    • Won't the real test be whether it's profitable?
      • Calculations are done to predict the experiments you want to do: You can't try everything. But that also put an upper limit to the cost of calculations: It has to be (much) cheaper than doing the actual experiments.
  • by backslashdot ( 95548 ) on Monday July 07, 2025 @01:26AM (#65502304)

    If they can figure out how to make an LNP that can efficiently deliver a large mRNA payload (30kb) into every cell type, EVERY disease suddenly becomes curable -- today.

    How? Because if you can put 30kb of RNA into a cell ... you can make all kinds of stuff. You can basically create a science lab within the cell to detect and fix what's broken (or destroy the cell if it is cancerous and has unwanted mutations).

    • In theory. If you can get it into the right cells. HIV for instance lays dormant in reservoirs. That's one reason it has only been treated and not cured yet, outside of very dangerous bone marrow transplants that replace the entire immune system.

      • If you can deliver 30 kb of RNA into every cell, you can excise that HIV from any such reservoir cells that contain latent provirus copies. Specifically, there's easily enough within 30kb to include gene excision tools such as CRISPR (which is only around 4kb max) and any other genes you may need to ensure you've fished it out. And btw, 30 kb is not an unreasonable size .. the coronavirus is about that size. Herpes is 150 kb.

        • You have 25-35 trillion cells. Some of them are in nearly inaccessible areas like bone marrow. Your body naturally has niches to protect things like stem cells from the systemic signals that would cause their demise (like hair coloring units). This poses challenges. The parent post touched on this dormancy that makes eliminating latent viruses difficult.

          The problem with CRISPR is it changes the DNA, which triggers cells DNA damage response mechanisms like P53. The body treats the CRISPR edit like a virus

          • Uh, that stuff in that Scientific American article from 2018 is pretty mitigated by now. Reference: https://www.oligotherapeutics.... [oligotherapeutics.org] For one, the cancer risk was exaggerated (cells with broken P53 are there anyway regardless of CRISPR), and second we have better CRISPR tech now -- and it continues to improve. More importantly we have hundreds of humans and plenty of animals edited with CRISPR and they haven't died or shown signs of cancer. https://www.technologyreview.c... [technologyreview.com]

            So far nobody has shown that concl

            • by ceoyoyo ( 59147 )

              More importantly we have hundreds of humans and plenty of animals edited with CRISPR and they haven't died or shown signs of cancer.

              The vast majority (all?) of those have been editing cells that are removed, then reinjected. One of the benefits of that is you can discard any suspicious ones.

              and they haven't died

              Oh yes, people have died.

          • Cells need blood to survive. Therefore they are all reachable. Besides, if a cell is so isolated it probably doesn't need to get treated. If it is virus infected, then how did the virus reach it? A virus is about the size of an LNP, if not bigger.

        • You'd run into 'software bugs' doing it too wide scale, like injecting CRISPR into rapidly reproducing cancer cells seems like a good way to infest yourself with mutated CRISPR sequences in what would likely end up feeling a lot like dying of acute radiation poisoning.
          • CRISPR has been extensively tested in cancer and no such thing has happened in any instance. Don't make up stuff based on speculation. Reference: https://pmc.ncbi.nlm.nih.gov/a... [nih.gov] Besides, for cancer there are better tech than CRISPR to identify and destroy the cell from within. And btw, most cancers are not as rapidly haphazardly mutating like you think .. if a cell mutated too fast it would not be functional and die. Anyway, I was talking about CRISPR for virus infected cells, which is also something peop

            • Ah thanks for the reference I thought you were proposing to insert a self perpetuating CRISPR hot fix into all the cells at once, but that sounds more normal. I do more software dev than biology these days, so have had more kneejerk responses to AI use cases of "oh no, you don't want to use it to do that" Suitable for all cell lines, but only ones with viruses, would be pretty tricky.
  • Before they have completed at the very least Stage 1 human trials, this is all without significance.

  • In the 1940's to late 1960's scientists and researchers worked together and shared. After https://en.wikipedia.org/wiki/... [wikipedia.org] the drug companies went profit driven and non-cooperative in case they got sued. Besides blockbusters were mostly discovered then, and clever dishonest shits played patent evergreening by teeny molecule variations, followed by slow release and combo's. Even the covid non-vaccines hid the negatives, and no real action has followed. Drug AI engine is NOT an AI engine but like a chess pro
  • The way Microsoft is calling everything Copilot.
  • It will be interesting to see if this company stays "good". For instance, they supported this effort: https://www.ebi.ac.uk/about/ne... [ebi.ac.uk] . Will this continue, will everything they learn going forward that is not patentable become trade secrets?

What the world *really* needs is a good Automatic Bicycle Sharpener.

Working...