Crinetics Pharmaceuticals

A few reflections from the recent roundtable discussion on AI/ML at DDC in San Diego (see my previous post) One of the clearest themes was the challenge of working with public data. These datasets are valuable, but they also come with real limitations — inconsistent annotation, aggregation from different sources, variable protocols, different error profiles, and uneven overall quality. There was broad agreement that better curation matters, but also that we need to be more explicit about uncertainty and more thoughtful about identifying and managing outliers before model training. Another recurring theme was the disconnect we so often see between in vitro and in vivo data. That remains such a central challenge. One idea that came up was whether ML could help by moving us toward more in vivo–centric models, rather than simply building better models of in vitro outputs. There was also discussion around whether we could identify early signs of in vitro/in vivo disconnects sooner, especially when they are target-, modality-, or project-specific. We also touched on foundation models and whether they could become the next “superheroes” in this space. That is an exciting idea, but it also highlights how difficult drug discovery really is. Getting a compound to clinic means balancing multiple interdependent properties, especially in ADMET, where trade-offs are everywhere. I also liked the “Santa wishlist” part of the discussion because it made clear what people would really love to predict: metabolites, dose–response in higher species, BBB penetration, and CYP inhibition. Interestingly, modeling complex organ toxicity was not really high on the wishlist. To me, that was telling. My interpretation was that these are often seen as too far downstream from where chemistry teams can act directly. The conversation stayed much closer to properties that can help guide compound design, triage, and optimization earlier in discovery. There was also clear awareness of the standard proactive toxicity screens, especially hERG and Ames. But beyond that, I did not come away with much insight into what smaller pharma companies are doing proactively for toxicology — or whether those efforts are systematic at all. For me, one of the biggest takeaways was the gap between what we would ideally like to predict and what is actually actionable in discovery. That is probably where AI has the greatest opportunity: not just to build impressive models, but to support better decisions early enough to matter. I really enjoyed charging this roundtable together with Mridula Bontha and hearing such a range of perspectives. I would be very interested to hear what others think — especially where you see the biggest unmet need for AI in discovery ADMET and early toxicology. Please feel free to add your thoughts in the comments. #DrugDiscovery #AIinDrugDiscovery #MachineLearning #ADMET #Toxicology #DrugSafety #PredictiveToxicology #MedicinalChemistry #Pharma