The TOP10 AI x bio stories in April

April was a big month in AI × biology. Frontier AI labs made billion-dollar moves into drug discovery. A Nobel laureate's AI designed enzymes for chemistry that nature never invented. A $500 million initiative set out to build the data foundation for virtual cells. And a general-purpose AI matched the best human scientists on a real biological design challenge.

Here are the ten biggest stories from the past four weeks - in no particular order. Each one links back to the original issue if you want more of the context.

Let’s dive in.

The Chan Zuckerberg Biohub launched a five-year, $500 million initiative to build the open data foundation for predictive models of the human cell. The partner list reads like a who's who - Allen Institute, Arc Institute, Broad Institute, Wellcome Sanger, NVIDIA - and all data will be freely available. Biology's version of the Human Genome Project, if it works. (Issue 21)

Anthropic acquired Coefficient Bio - a stealth startup just eight months old, nearly all former Genentech drug discovery researchers - for over $400 million in stock. Coefficient had no public product, no disclosed revenue, and had never made an official announcement. The deal signals how seriously frontier AI labs are taking biology: Anthropic paid $400 million to skip the years it would take to build a biology team from scratch. (Issue 14)

Eli Lilly signed a multi-program deal with Profluent worth up to $2.25 billion for AI-designed recombinases - enzymes that can insert entire genes at specific genomic sites. A bet on AI designing the machinery that rewrites biology. It's Lilly's third major AI deal tracked by BAIO this year, after the NVIDIA co-innovation lab and the $2.75 billion Insilico partnership. (Issue 21)

OpenAI released GPT-Rosalind, its first model built specifically for biology and drug discovery, and signed Novo Nordisk for integration across drug discovery, manufacturing, and operations. On Dyno's ML hiring challenge, Rosalind's best submissions ranked above the 95th percentile of human experts on prediction. Just nine months earlier, OpenAI “didn't really have that much biology” in its training data, according to research lead Joy Jiao. (Issue 18)

DISCO, from Nobel laureate Frances Arnold's lab at Caltech, designed working enzymes for reactions no living organism has ever performed. The active sites have no close match among the 200 million-plus structures in the AlphaFold database. On one of organic chemistry's hardest selective-modification problems, a single DISCO design matched what previously took 14 rounds of laboratory evolution. (Issue 15)

Anthropic's Claude Mythos Preview took Dyno Therapeutics' ML hiring challenge - the same open-ended test used to evaluate human scientists - and performed at a level Dyno called “indistinguishable from the best-performing human candidate.” This is a general-purpose AI, not one trained on biology specifically, performing at expert level on real biological design. (Issue 16)

APOLLO, from Faisal Mahmood's lab at Harvard, was trained on 25 billion clinical records from 7.2 million patients spanning 33 years. Tested on 322 clinical tasks - predicting disease onset, treatment response, adverse events - it significantly outperformed baselines across the majority. The AI × bio field has been building virtual molecules, then virtual cells. APOLLO pushes the abstraction up, to virtual patients. (Issue 19)

About two million genetic variants in NIH's ClinVar database are stuck in diagnostic limbo - found in patients, but uninterpretable. Goodfire and Mayo Clinic found a way to extract clinical predictions from Arc Institute's Evo 2 by reading how the model's internal response changes when a mutation is introduced. The approach outperformed every existing method tested and, intriguingly, explains itself in plain language. (Issue 17)

Medra opened a 38,000-square-foot autonomous lab in San Francisco's SoMa district - general-purpose robots executing experiments end-to-end while scientists direct workflows in plain English. AI predictions are scaling exponentially; physical experiments aren't (just yet). Medra, backed by $52 million from Human Capital, Lux Capital, and Menlo Ventures, is trying to close that gap. (Issue 16)

Amazon launched Bio Discovery - a managed platform offering more than 40 biological AI models with lab partners including Ginkgo, Twist Bioscience, and A-Alpha Bio built in. Memorial Sloan Kettering used it to design nearly 288,000 novel antibody molecules for a pediatric cancer target. A major cloud provider packaging AI models and physical lab access into a single service is a new layer of infrastructure for the field. (Issue 17)

Runners-up: VCHarness builds virtual cell models autonomously (Issue 17). Tripso refuses to compress biology into a single number, finds what standard analysis missed (Issue 13). Genentech's CompBioBench tests whether AI agents can actually do computational biology (Issue 16). OpenAI Foundation commits $100M+ to Alzheimer's (Issue 15).

Until next time,
Peter at BAIO