Herbert B Schiller

The principles for AI-enabling longevity data are: 1⃣ Maximize answers per unit time by starting the gating experiments now and avoiding sequential dependencies, since much biology can’t be sped up. 2⃣ Data has to span the layer where we want AI to provide answers (physiological or organismal for longevity) to be task-shaped. 3⃣ Richer data is better, since models can find structure we wouldn’t have known to look for. 4⃣ When we want treatments, the data needs to contain or link to outcomes generated by interventions or by time, to enable causal inference. And we will get the most progress if we focus on how to bridge slow and fast, to feed AIs the right data ASAP. Read the full piece here: norngroup.substack.com/p/ask-not-what…

Our latest paper is out @ScienceTM! Natural killer cell immunotherapy reverses lung fibrosis by eliminating senescent fibroblasts | Science Translational Medicine science.org/doi/10.1126/sc… #Immunotherapy #Fibrosis #Aging #Senescence #NKcell

Excited to share our RegVelo paper in Cell cell.com/cell/fulltext/… We unify RNA velocity + GRNs into one model → better OOD prediction of perturbations (e.g. gene KOs), with examples incl. neural crest KO predictions 🔬 Big thanks to W Wang, Z Hu & T Sauka-Spengler 🙏

TWEETORIAL | We mapped PPFE at single-cell resolution — and found an unexpected blueprint for lung fibrosis. A thread on our new paper in @ScienceAdvances science.org/doi/10.1126/sc… Data live at PPFEatlas.com

In 2005, Irina and Michael Conboy at Stanford asked whether young blood could rejuvenate old tissue. Years later, this fueled a wave of longevity-pilled “young blood” headlines, startups, and even Bryan Johnson’s experiments using his son’s plasma. But what does the science really show, and where did the young blood story go off the rails?

“Where does biology actually compute?” I had the privilege of unpacking this question today on a fantastic panel with 2 of the sharpest minds in the space: @iamjohnnyyu of @tahoe_ai & @glebkuz of @ManifoldBio, masterfully moderated by @FabioZB_I of Polyphron. 🧵

AI in pathology may be entering a new phase. A new @NatureMedicine Research Briefing * highlights SPARK, an “agentic AI” system that did more than detect patterns on pathology slides, it autonomously generated biological hypotheses from pathology and spatial omics data. The system linked tumor morphology with prognosis, metastasis, MSI, PD-L1 status, and even possible tumor evolution patterns across multiple cancer types. Still early and entirely retrospective, but the idea is fascinating: AI not only helping diagnose cancer, but potentially helping researchers discover new biology.

AI predictions ultimately need validation in the wet lab—but with limited resources, how do you decide what to test while controlling error rates and maximizing discovery? Introducing TxConformal: a new statistical framework for controlling false discoveries in AI-driven drug discovery. We prospectively validated this approach in an AI-guided A. baumannii molecule screen at @genentech Grateful for an incredible multi-year collaboration with @YingJin531, @EmmanuelCandes, @jure, @DamantNathaniel, @GabrieleScalia, and the broader team! Learn more:

We built an AI that will tell you what will get your grant rejected. It’s a tough critique, but it’s better to know when you can still do something about it 👇 @qedScience

This paper exists because teams of people started an unglamorous data collection project many years ago. The Interventions Testing Program (ITP) is a long-running project by the NIH-NIA with the goal to identify interventions that extend lifespan in mice. The ITP was not built around the analysis this paper eventually performed. Instead, the paper’s method emerged because a group of scientists had access to a sufficiently large dataset generated by the ITP and realized there were better questions to ask of it. Computational power is scaling rapidly. What isn't scaling is the generation of biological data that requires organisms to age, populations to evolve, or longitudinal measurements to accumulate across years. If someone starts a 10-year dataset now, by the time it matures the analytical tools available to interrogate it will be dramatically more powerful than anything we have today. But if nobody starts the dataset, those tools have nothing to work with. For those thinking about the overall progress of longevity, it’s now obvious that AI will be bottlenecked by the lack of high quality biological datasets. nature.com/articles/s4158…

What if you could test a drug on a human organ and simultaneously know what would have happened without it? That's what we built. Proud to share our work in @NatureBiotech: digital twins of human lungs 🫁🔥 (paper: nature.com/articles/s4158…) We created digital twins of ex vivo human lungs: multimodal AI models trained on 951 human lungs from the world's largest EVLP dataset at @UHN. Physics-informed ML across physiology, biochemistry, transcriptomics, proteomics, metabolomics, and imaging, all forecasting together. The key insight: the physical lung receives the treatment. The twin is the untreated control. Paired causal inference on the same organ. No separate cohort. No intersubject noise. Result: we detected drug efficacy with 6 lungs. Traditional methods would need 18. This is what precision preclinical evaluation looks like. From Virtual Cells → Virtual Organs → Virtual Patients. One step toward virtual organs replacing animal testing. Huge congratulations to Elly Zhou (a phd student I co-supervise) for leading this work with exceptional rigor, and to Andrew Sage and @SKeshavjee for building the foundation that made it possible.

On May 12, we're hosting a day-long AI x Bio symposium at @novonordiskfond HQ, with Mads Krogsgaard Thomsen, CEO of the NNF, delivering the keynote, alongside 25+ leading scientists, founders, and investors. 🧵

What if we could use a foundation model to simulate human biology from mouse data? Today, we're sharing Perturb-MARS, a platform for genetics and drug treatment in vivo at SCALE. ... and we HUMANIZE the read-outs using TARIO-2.

Wrapping up a massive month for PerturbAI with some #BTS moments from the @OpenAI Forum! We showcased our technology, unpacking how AI can help biology move from static maps toward predictive, causal models of living systems. We loved answering your thoughtful questions and were thrilled by the incredible interest our demo drummed up. Our AI agents and models are designed to move biology from data to insight to therapeutic action, helping accelerate drug discovery from the ground up. If you missed it, watch the replay of the Forum here: forum.openai.com/home/videos/ev… #AI #Biotech #DrugDiscovery #CRISPR #Genomics

Big news - we received IND clearance from CDE to start a clinical study of the inhalable rentosertib. I think that this is the first drug discovered using generative AI to reach human clinical stage in the inhalable formulation. Godspeed, my friends and thanks to our teams around the world for making this happen. I feel that this year is the year of massive acceleration on all fronts. insilico.com/news/pdmncgky5…

What biology can do with AI in the next few years depends on what gets measured, standardized, and shared today. We're setting up our 4th(!) Impetus Grants @impetusgrants round to fund open, AI-enabling biological datasets, and we're currently assembling the review team. If you want to learn more or contribute to this round, click here: norn.group/impetus-grants…

Health and wellness will be the biggest market in the world. It’s inevitable.

📣 We are excited and thrilled to announce APOLLO, a healthcare system-scale multimodal temporal foundation model for virtual patient representations. Trained on 25 billion clinical events from 7.2 million patients across 33 years and 28 modalities, APOLLO learns a unified atlas of medicine. Turning labs, notes, pathology images, medications, and diagnoses into coherent, computable longitudinal trajectories. APOLLO is disease-agnostic by design, a single model that learns the shared structure underlying human health and disease across every specialty, modality, and stage of care. The possibilities are enormous: earlier risk prediction, treatment response modeling, clinical trial matching, biomarker discovery, and a new generation of agentic systems built on rich patient representations. Read the pre-print: arxiv.org/pdf/2604.18570 Read our blog post about the work: linkedin.com/pulse/apollo-m… 👏 🎉Huge congratulations to Andrew Zhang , @TongDing99, Sophia J. Wagner, and the rest of the team.

🔓🧬First big unlock from vibe science-ing: rapid access to publicly available datasets. Sounds basic. It isn't. If you've ever tried to pull raw data from a paper you care about, you know: the metadata is a mess, the supplementary tables are unstructured, the file formats don't match, and by the time you've got it working you've lost half a day. For people without strong bioinformatics skills, it's often a dead end entirely. 1/🧵

Hot take: AI hypothesis generation gets the headlines. AI lab robots get the demos. Meanwhile, computational biology/bioinformatics is quietly doing more for actual scientific output than either. Aside from protein structure predictions, nobody set out to revolutionize computational biology. It probably happened as a side effect of teaching AI to write code. Whether you're a bench scientist who touches data occasionally or a bioinformatician who lives in it, the impact is enormous, and I think a lot of people who claim to build AI scientists haven't fully realized it yet. I've been doing bioinformatics for over 12 years across most data types biology produces, and I haven't been THIS excited in a long time: One, speed and scale of analyses I already know have gone through the roof and second, things I always wanted to do but didn't have the bandwidth for become possible. After a few months of vibe science-ing, I can tell that there's definitely a lot that can be done wrong and that many easy unlocks remain underused. Over the coming weeks, I'll share some of my observations and easy-but-impactful fixes: Dataset assembly at scale. Multi-agent reviewer patterns. Forcing the model to 'see' its own mistakes... and a questionable but very fun deep dive using genomics tools to dissect Kpop songs (this stuff is addictive!) If you're figuring this out too, I'd love to hear what's worked for you.