Last issue we said the house of AI × bio is being built brick by brick. To continue that analogy: this week the walls started connecting to each other.

☑️ OpenAI shipped its first dedicated life sciences model and signed Novo Nordisk as a customer.

☑️ Formation Bio, valued at $1.8 billion, argues that discovery was never the bottleneck - development is - and is using AI to rescue stalled drug candidates.

☑️ Valence Labs released a framework that forces AI to explain its biological reasoning step by step.

☑️ And Helical raised $10 million to build the orchestration layer that lets pharma teams actually use AI models in practice. The era of standalone demos is, hopefully, giving way to systems.

Oh, and someone also invented a beanie that reads your mind. Food…I mean sensors for thought.

Let's dive in.

OpenAI released GPT-Rosalind this week - its first model built specifically for scientific research in biology, drug discovery, and translational medicine. Days earlier, Novo Nordisk announced a partnership to deploy OpenAI's capabilities across drug discovery, manufacturing, and commercial operations, with full integration planned by end of 2026.

The clearest signal of what Rosalind can do came from Dyno Therapeutics, whose ML hiring challenge has become a benchmark for AI in biology. Dyno tested GPT-Rosalind on unpublished RNA sequences the model couldn't have memorized. Its best-of-ten submissions ranked above the 95th percentile of human experts on prediction and hit the 84th percentile on sequence generation. That's the same test where Anthropic's Claude Mythos reached the 90th percentile on prediction (we covered it in Issue 16) - but Rosalind is a dedicated biology model, not a generalist.

How fast did this come together? In an OpenAI podcast, research lead Joy Jiao said that when the Ginkgo collaboration began in July 2025, “we didn't really have that much biology in our training data.” Nine months later, they've shipped a domain model and signed the one of the world's biggest pharma companies.

The team's internal tagline, Jiao said half-jokingly: “Scale test time compute to cure all disease.” She said the models could help with drug repurposing for rare diseases, and that predicting drug toxicity for a specific person is on the roadmap. Her own motivation is simpler: “I don't really ever want to touch a pipette again.”

Why it matters: Just a few days ago Anthropic put Novartis CEO Vas Narasimhan on its board. Now OpenAI has both a domain model and a pharma giant as a customer. Frontier AI labs are embedding in the pharma value chain.

Did you know? OpenAI also launched a free Life Sciences Codex plugin connecting to over 50 scientific tools and databases, available on GitHub - open to all developers, not just Rosalind users. OpenAI is hiring.

The number of drug candidates has nearly doubled in the past decade. The number of FDA approvals - roughly 50 per year - hasn't budged. Formation Bio CEO Ben Liu thinks that tells you where the real bottleneck is, and it's not discovery. “What we think the world has wrong is that drug discovery is the bottleneck, and it has not been the bottleneck for a long time,” Liu told Forbes.

The New York-based company, valued at $1.8 billion with $615 million raised from Sequoia, a16z, and OpenAI's Sam Altman, buys stalled early-stage drug candidates and uses AI to push them through clinical trials up to 50% faster. Its drug-picking team is led by Mikael Dolsten, who retired as Pfizer's president of worldwide R&D.

One example: Formation rescued a knee osteoarthritis drug originally developed by Merck KGaA, building an AI model trained on 23,000 patients and 48,000 MRIs to predict who would benefit most. "Without AI, I probably would have abandoned it," Dolsten said. The drug is now in late-stage trials.

Why it matters: In the story above, OpenAI's Joy Jiao said the models could suggest new uses for existing drugs. Formation Bio is applying the same logic - rescuing candidates that stalled before approval and using AI to find the patients they'll actually work for. “Drug discovery is going to be commoditized with China and AI,” Liu says to Forbes. “If we don't solve this drug development bottleneck, we're actually not going to end up with more medicines for patients.”

Did you know? Formation Bio has already licensed one drug to Sanofi for roughly $630 million. An earlier venture the co-founders helped set up, Aditum Bio, sold a lean muscle mass drug to Eli Lilly for up to $1.9 billion in 2023. Formation Bio is hiring.

AI models are getting better at predicting what happens when you apply a drug or genetic change to a cell. But ask them why it happens and they'll give you fluent, plausible text that may be wrong - and unlike in math, there's no simple way to check. A preprint from Valence Labs, Recursion's research arm headquartered at Mila in Montreal, attacks that gap directly.

VCR-Agent is a multi-agent system that generates structured chains of biological reasoning - not free-form text, but step-by-step graphs where each claim (“this drug binds this target”, “this gene gets upregulated”) can be independently verified against external databases and tools like Boltz-2, the structure prediction model. The system was applied to 18,950 drug-cell scenarios from the Tahoe-100M perturbation atlas. Its verifier pipeline caught 28.2% of faulty drug-target claims and refined 87.3% of gene expression predictions.

Why it matters: Virtual cells - computational models that simulate how cells respond to drugs or genetic changes - are one of the most ambitious goals in AI × biology. But a virtual cell that predicts the right answer without explaining why is hard to trust and hard to build on. VCR-Agent adds an explanation layer, and the explanation seems to help: models trained on its structured reasoning traces outperformed STATE, a model trained directly on perturbation data, at predicting gene expression changes.

Did you know? Valence Labs is hiring. The VCR-Agent code and the full VC-TRACES dataset are both open on GitHub.

The biological foundation model space has a gap between what models can do in a notebook and what pharma teams can do with them in practice. Biologists need to run in-silico experiments without writing code. ML engineers need production environments for model fine-tuning and alignment. Both need to work on the same data and trust the same results. London-based Helical, which just raised a $10 million seed round led by redalpine, is building the platform that sits between them.

The product has two sides. A Virtual Lab lets biologists design experiments and evaluate biological outcomes without code. A Model Factory gives ML engineers a production environment with VS Code integration and experiment tracking. Helical personalizes models to a specific disease area, runs experiments at scale, and validates results against real biological evidence.

The company claims teams see up to 3-5x higher hit rates versus traditional approaches and have compressed timelines from years to weeks. Large pharma R&D teams are already deploying it across target identification, biomarker discovery, and therapeutic design in areas including cardiometabolic disease, neurology, immunology, and oncology.

Why it matters: Foundation models for biology keep improving. But a model that can predict a drug's effect is useless to a pharma team if they can't reproduce the result, trace how it was generated, or hand it to a regulator. Helical is building the layer between the model and the decision.

Did you know? Helical was founded in early 2024 by three school friends who later went into tech, data science, and medicine. One of their pharma collaborations is public: a project with Pfizer on predictive blood-based safety biomarkers. Helical is hiring.

Many brain-computer interfaces require surgery. Sabi, a Silicon Valley startup profiled by Wired this week, is betting that a wearable can get close enough - if you pack it with enough sensors and train a large enough AI model on what they pick up.

The device reads brain activity using EEG - electrodes on the scalp that detect electrical signals from neurons. The fundamental limitation of non-invasive approaches is that the skull gets in the way, weakening the signal compared to implanted devices like Neuralink. Sabi's workaround is brute-force sensor density: 70,000 to 100,000 miniature electrodes in a single cap, orders of magnitude more than a typical EEG headset. CEO Rahul Chhabra told Wired the first product ships by end of year, initially decoding internal speech at roughly 30 words per minute.

Sabi is also building what it calls a brain foundation model. Trained on 100,000 hours of data from 100 volunteers, the model aims to generalize across brains rather than requiring calibration for each user. Vinod Khosla, an early OpenAI investor, is backing the company. “If you're going to have a billion people use BCI for access to their computers every day, it can't be invasive,” he told Wired.

Why it matters: If the brain foundation model works at scale, the applications extend beyond typing into neurological diagnostics and new ways of interacting with AI.

Did you know? Sabi says it trains its AI on encrypted neural data rather than raw brain signals, and is working with neurosecurity experts from Stanford to audit the full technology stack. "Neural data is the most private kind of data that a person could possibly have," Chhabra told Wired.

Biomni Lab - Phylo Bio's platform now lets any scientist create, fine-tune, pre-train, and optimize biological foundation models on their own datasets by describing what they want in natural language. The new capability, called GPU-as-a-tool, has AI agents spin up and orchestrate GPU sandboxes on your behalf. Demos show fine-tuning Borzoi, scGPT, and ESM2, pre-training protein language models from scratch on UniRef, and building multi-task ADMET models (absorption, distribution, metabolism, excretion, toxicity) across 22 endpoints. BAIO readers may remember Phylo Bio from Issue 1.

☑️ ”Protein language models: fluent, but clueless”

☑️ Stanford's Markus Covert proposes goals for a Human Cell Project

☑️ Calico's Oliver Hahn: computational biology is quietly outperforming headline-grabbing AI

☑️ How AI can beat cancer - and why the US is massively underinvesting in biopharma AI