It's Easter Sunday, and I hope you're spending it with people you like. BAIO will keep publishing on holidays, but some holiday issues may run lighter when the news cycle slows down.

Speaking of BAIO, nearly 600 of you have signed up since I launched BAIO six weeks ago. Thank you!

This issue, by the way, has a clear longevity theme. One of the reasons I started BAIO is a conviction I share with several of the people running the biggest AI companies: that AI will accelerate progress in aging biology research.

Dario Amodei at Anthropic has written about this explicitly. And this week, Anthropic put $400 million behind it - acquiring a team of former Genentech drug discovery researchers who launched their own startup barely eight months ago

Keeping with this longevity theme we also have a new temporal AI model from Gladstone and NVIDIA. It predicts how cells age across a full human lifetime. And Insilico Medicine argues that a single language model can replace the dozens of specialized aging clocks the field has relied on for a decade.

Let's dive in.

Anthropic acquired Coefficient Bio - a stealth startup just eight months old - for just over $400 million in stock. The team is fewer than 10 people, nearly all from Prescient Design, the unit inside Genentech (Roche's biotech subsidiary) that uses AI to find and design new drugs.

Co-founders Samuel Stanton and Nathan Frey built AI tools for designing drug molecules at Genentech before launching Coefficient Bio last year. Dimension, the venture firm co-founded by Zavain Dar, owned roughly half the company. The deal, first reported by The Information, is Anthropic's third acquisition in six months.

Coefficient Bio had no publicly known product, no disclosed revenue, and had never made any official public announcement. The team joins the healthcare and life sciences group that launched Claude for Life Sciences last October. Its head, Eric Kauderer-Abrams, told CNBC at the time: “We want a meaningful percentage of all of the life science work in the world to run on Claude.” In issue 12 a week ago, BAIO flagged signs of an unreleased biology workspace called Operon spotted in the Claude app. Buying a team from one of the world's top drug companies suggests this is moving fast.

Why it matters: Google has Isomorphic Labs (we covered IsoDDE in Issue 1). OpenAI pledged $1 billion for life sciences (Issue 11). Now Anthropic is paying $400 million for a biology AI team. The frontier AI companies are racing to become the default platform for drug discovery - and, in this case at least, they're paying $400 million to skip the time it would take to build a biology team from scratch.

Did you know? Anthropic's AI for Science program offers free API credits to academic biology researchers.

Most AI models for biology look at cells one at a time - a frozen snapshot. MaxToki, from Christina Theodoris at the Gladstone Institutes and co-authors at NVIDIA, learns how cells change across the full human lifespan. The name comes from a Japanese bullet train that's also a homonym for “time” in Japanese - fitting for a model built around speed and temporal modeling.

Training required processing nearly a trillion pieces of genetic data - comparable in scale to training a large language model, according to Gladstone.

Trained first on 175 million individual cells, then on 100 million aging trajectories built from about 3,800 donors spanning birth to age 90+, it predicts how aging unfolds in different tissues over decades.

The standout result: MaxToki predicted which genes drive aging in the heart - targets the researchers say they would not have tested otherwise. When the team activated those genes in human heart cells, the cells showed hallmarks of aging: increased inflammation, disrupted energy production, slowed calcium cycling, and irregular beating. In mice, the top two pro-aging genes caused significant cardiac decline within six weeks.

“Seeing them cause functional decline in heart cells, with exactly the kinds of changes we associate with cardiac aging in humans, confirms the model is capturing something real about the biology”, says Christina Theodoris, in an article on Gladstone’s website.

The model also detected accelerated aging in diseases it never saw during training. It inferred roughly 15 years of extra aging in pulmonary fibrosis patients and about 3 years in Alzheimer's. Notably, patients who had the same Alzheimer's brain pathology but no cognitive symptoms did not show the acceleration. That suggests the model appears to distinguish between having the disease and actually declining.

Theodoris previously created Geneformer, a widely used AI model for gene networks. Shinya Yamanaka - who won the 2012 Nobel Prize for showing that adult cells can be reprogrammed into stem cells - is a co-author. The findings have not been peer-reviewed.

Why it matters: Aging research has always been bottlenecked by time itself - you're studying a process that takes decades, which makes testing interventions difficult, slow and expensive. MaxToki lets researchers screen thousands of potential aging targets computationally, before committing to long-term studies. The team is already testing whether blocking the pro-aging genes it identified could protect against age-related heart decline.

Did you know? MaxToki is open source - pretrained models on Hugging Face, training and inference code on GitHub. The 175-million-cell training dataset will also be publicly released.

For over a decade, aging researchers have built specialized tools - called aging clocks - that estimate how old your body really is, as opposed to your birthday age. One clock reads chemical tags on DNA. Another reads protein levels in blood. A third reads gene activity.

Insilico Medicine CEO Alex Zhavoronkov - whose $2.75 billion Lilly deal led our Issue 12 - now argues the whole paradigm is obsolete, or at least ready to be absorbed. A preprint provocatively titled “The End of Aging Clocks” - from a man who helped build some of them - introduces Longevity-LLM. It’s a single language model with 14 billion parameters that handles DNA methylation, protein data, clinical blood markers, and gene activity data in one system. One model, four data types. The Insilico team built it in a ten-day sprint.

On estimating age from DNA methylation - one of the most established forms of biological age testing - the model reports a mean error of 4.34 years, beating the Horvath multi-tissue clock (4.61 years) on the same held-out test set. That comparison is fair in the narrow sense: both are predicting chronological age from the same type of data. But the Horvath clock dates to 2013. Newer clocks - like DunedinPACE, which measures the speed of aging, or GrimAge, which predicts mortality - are not compared against. On protein-based age prediction, Longevity-LLM performed within range of specialist clocks but did not beat the best of them.

The title is tongue-in-cheek - Zhavoronkov is part of the aging biomarker community and meets many of these researchers frequently. But the underlying claim is real enough: they believe a single AI model soon can do what many specialist tools used to do separately, and do it better.

Caveats: Every author on this preprint is an Insilico employee. The results have not been independently validated.

Why it matters: As the MaxToki story above illustrates, aging research is bottlenecked by time - you need reliable biomarkers that can stand in for decades of real aging when testing whether an intervention actually works. A single model that handles multiple data types and explains its reasoning? Well, it might, hopefully, be a more useful surrogate than the specialist clocks it aims to replace. That remains to be seen. But if it took ten days to build, the next version could come fast. This is part of the same MMAI Gym platform behind the drug discovery model we covered in Issue 6.

Did you know? Insilico runs LongevityBench - a leaderboard ranking how well commercial AI models handle aging biology tasks, from cancer survival prediction to proteome-based age prediction. Insilico is hiring.

📅 UNLOCK 2026 runs April 22 in San Francisco - a one-day, invite-only summit on AI in drug discovery and life sciences, taglined “Cutting Through the Hype.” The speaker list reads like a BAIO recurring cast. David Baker (2024 Nobel laureate in protein design) and Aviv Regev (head of Genentech Research) headline. Bo Wang and Ci Chu present X-Cell for Xaira (we covered X-Cell in Issue 9). Andrew White speaks for FutureHouse (Issue 3). Reshma Shetty represents Ginkgo (Issue 5). And Samuel Stanton - co-founder of Coefficient Bio, whose Anthropic acquisition leads this issue - is on the bill. Anthropic's own Jonah Cool from the life sciences team also speaks. Other names to watch: Jacob Kimmel (NewLimit, longevity), Martin Jensen (Gordian Bio, longevity) Shahin Farshchi (Lux Capital), and Alex Morgan (Khosla Ventures).

Genentech open-sourced GYDE (Guide Your Design and Engineering) - a web-based platform that puts AI protein design tools in one place for scientists who don't code. It wraps more than a dozen tools - including AlphaFold, Boltz, Chai-1, ProteinMPNN, BindCraft, and RFDiffusion - into a visual interface where you can view a protein's sequence, 3D structure, and experimental data side by side. Design a new protein, predict its structure, check its stability, and share the session with a colleague - all without touching a command line. Freely available for academic and non-commercial use. Preprint can be found here.