69 Billion Molecules, Screened at 1,000x Lower Cost. The Paper Behind It Took 3 Years to Clear Peer Review.

69 billion drug-like molecules, screened at a claimed 1,000x lower computational cost than existing pipelines, in days. Getting that claim through peer review took longer: the platform first appeared as a bioRxiv preprint in April 2023, under the name VirtualFlow 2.0, and only cleared Nature Biotechnology's review on September 1, 2026 — more than three years later, published as AdaptiveFlow. In that time, the platform itself scaled to 5.6 million virtual CPUs and validated hits against two cancer targets. The screening got 1,000x faster. The verification didn't move at all. This is the only new bio-AI result that cleared our bar this week — below, what it shows, what it doesn't, and why a one-item digest beats stretching it into six.

Cheminformatics & drug discovery: the biggest library gets a cheaper search

"AdaptiveFlow" (Nature Biotechnology, s41587-026-03217-x, published September 1, 2026) is an open-source cloud platform for ultra-large virtual screening, led by co-corresponding authors Christoph Gorgulla (St. Jude Children's Research Hospital), Andrea Mattevi (University of Pavia), and Haribabu Arthanari (Dana-Farber Cancer Institute/Harvard Medical School), per St. Jude's announcement. It ships a screening-ready version of the Enamine REAL Space — 69 billion compounds, described as the largest ready-to-dock library assembled to date — and uses an 18-dimensional grid of molecular properties, with optional active learning, to prioritize promising chemical subspace before running expensive docking. That prioritization is what buys the claimed 1,000x cost reduction. The framework demonstrated linear scaling to 5.6 million virtual CPUs, described as a new benchmark for cloud-based drug discovery, and supports more than 1,500 docking protocols. As proof of concept, the team reports structurally validated hits against PARP1, an established cancer target with approved drugs, and FSP1, an emerging ferroptosis-suppressor target with few known inhibitors. What the announcement doesn't include is a binding-affinity number for either hit — no Kd, no IC50. That data may sit in the paper's full text or supplement, which is behind Nature's login wall as of this writing. Until we can read it directly, the validated-hit claim is sourced; the potency behind it isn't.

What this means for reproducible, local-first science

The interesting number this week isn't 69 billion or 1,000x. It's the gap between them: a screening platform that got fast enough to search a 69-billion-compound library in days took over three years to get its own claims through peer review. That isn't a knock on the authors — ultra-large screening platforms are exactly the kind of complex, many-moving-parts software that should take a while to verify properly, and the bioRxiv preprint has been citable and usable the whole time. It's a reminder of where the actual bottleneck in AI-for-science sits right now. It was never generating candidates at scale; every week's digest is proof that part is solved. It's confirming that what got generated is real, at a pace that doesn't take three years. That's also why this is a one-item digest instead of a six-item one: we'd rather publish one number we can stand behind than five we can't.