Anthropic Opens a Wet Biology Lab in the Bay Area 🔗

Anthropic disclosed that it has established a wet biology lab in the San Francisco Bay Area, where it can now run its own experiments alongside its models. Eric Kauderer-Abrams, head of life sciences at Anthropic, describes the lab as a way to close the gap between computational predictions and physical data. The company says the lab is not solely a drug discovery unit, and expects to work on preclinical research into diseases where commercial pharma incentives are weakest. Access to some advanced models remains restricted to limit dual-use biological risk. Alongside the Novo Nordisk deal, its life sciences programme now runs beside partnerships with Genentech and Adaptyv, and its published Model Hardware Standard for wiring agents into instruments.

OpenAI Connects ChatGPT to Epic EHRs and to Public Clinical Databases 🔗

OpenAI enabled ChatGPT to work inside Epic Electronic Health Records, so clinicians can review a patient's history, synthesise information across records and track changes over time from within the EHR. UCSF Health is piloting the integration. In parallel, OpenAI released a Healthcare Public Data plugin that connects ChatGPT to ClinicalTrials.gov, PubMed and other public sources for research and drug-safety lookup. Two moves landed in the same week. The dominant clinical AI competitor to Anthropic put its consumer product inside the dominant US EHR, and the same product gained a governed connection to the public sources most trial and pharmacovigilance work draws on.

ARPA-H Puts Federal Money Behind Autonomous Cardiology AI 🔗

Between 9 and 10 September, ARPA-H named three lead performers under its ADVOCATE programme, a 39-month initiative to build autonomous AI for cardiovascular care. Tempus AI received up to $9.5 million to build the first autonomous AI agent for heart failure, with a multi-centre prospective validation study. UpDoc received up to $9.2 million to build a physician-grade autonomous clinical system, with Microsoft, OpenAI and NVIDIA joining as technical partners. Atman Health received a further award for an agentic AI cardiologist and expects to file for FDA authorisation within 24 months. Roughly 1,454 of 3,143 US counties have no practising cardiologist. ADVOCATE is federal capital committed to closing that gap with autonomous software, with FDA validation built into the programme design.

Google DeepMind Releases AlphaGenome Atlas 🔗
Google DeepMind released AlphaGenome Atlas, a one-petabyte resource with AI-generated molecular-effect predictions for more than 9 billion possible single-letter DNA changes across the human genome. It was developed with scientific input from the Stowers Institute for Medical Research, the Broad Institute, the University of Exeter, Memorial Sloan Kettering Cancer Center and Stanford. Researchers can now rank variants across the whole genome and inspect the biological processes each is predicted to disrupt through a browser.

The signal: Sequencing has been affordable for a decade. Interpreting what a specific variant actually does is where the field still gets stuck. A comprehensive, browsable prediction map at variant scale is now openly available from the same lab that gave the field AlphaFold. It becomes an obvious base layer for target identification, rare disease diagnosis, and any AI tool that reasons over genomic variation, and it raises a competitive question for the private variant interpretation platforms that have built their business on this exact bottleneck.

Tempus Sets Out to Build the Largest Multimodal Whole-Genome Dataset 🔗

Tempus AI announced an initiative to build a research platform of 100,000 whole genomes paired with longitudinal clinical outcomes, with a stated long-term goal of one million. Existing population-scale genome programmes are drawn largely from general populations. Tempus is building this one around specific disease populations and outcomes and structuring it for AI model training. It sits inside Tempus's existing multimodal environment, where researchers can access genomic data alongside clinical histories, imaging and pathology through Tempus Lens.

The signal: If AlphaGenome Atlas is the reference layer, this is the therapeutic layer. Foundation models for biology need genomes at scale and genomes linked to what happened to those patients over time. Tempus already sells that linkage in oncology to pharma. It is now extending the same asset into whole-genome sequencing while, in parallel, taking federal money to build the autonomous cardiology AI that will sit on top of that data. The data layer and the AI product layer of chronic disease care are being assembled inside one company.

Lilly's TuneLab Federation Adds Twist and Ginkgo 🔗

Two companies joined Lilly TuneLab, the collaborative AI/ML drug discovery platform through which biotechs can use models trained on Lilly's proprietary data. Twist Bioscience signed on to supply antibody characterisation data services under standard protocols, feeding back into the shared antibody developability model AbLab. Ginkgo Datapoints signed on to supply ADME and antibody developability screening at scale, turning results around in ML-ready formats that plug directly into TuneLab's predictive models.

The signal: TuneLab is the clearest working example in life sciences of the federated data play in AI. Lilly opens up its models, participating biotechs contribute standardised experimental data back, and every party's model improves with the network. Adding Twist and Ginkgo means the physical means to generate that data at scale, and consistently, are now inside the network. This is the same structural bet Anthropic and Adaptyv have been making from a different direction: the next drug discovery advantage is standardised data generation at loop speed, and model quality alone will not carry a programme.