Predicting human dose from structure alone
-
Nº CIII
- Date
- 21 Sep 2026
- Issue
- 103
- Stories
- Six
- Editor
- ARC
The weekend shipped a dose predictor and an aging model family; today, a lab robot that takes code.
Inductive Bio launches a model that predicts human dose
Inductive Bio launched Beacon-2, a model that predicts human dose directly from chemical structure, announced on X by Ben Birnbaum. The launch post frames it as a swipe at the rest of the field: most AI drug-discovery systems optimize the wrong reward, tuning for binding affinity or a single ADMET endpoint when the question that actually decides a program is whether a compound reaches therapeutic exposure at a tolerable dose. Predicting dose end-to-end puts one number on an objective medicinal chemistry has been triangulating from proxies for decades. No external validation accompanies the announcement, so the accuracy claim stands as the company's own.
One model family covers five kinds of aging data
One model family now spans clinical records, genetics, methylation, transcriptomics and proteomics in aging biology, with its developers reporting on X that it matches or beats frontier models (the general-purpose large models most benchmarking is run against) across all five. The pitch is one trained family instead of a bespoke predictor per assay. No numbers appear in the post and the comparison is self-reported. Aging research has run on a scatter of separate clocks and signatures, as Insilico's rentosertib results showed; a single model covering all of them changes what a shared reference point for the field could look like.
WetRobo gives coding agents a buildable lab robot
WetRobo packages a lab robot that code-writing agents can drive, released as a reproducible kit rather than a one-off rig in one group's basement. Agent-in-the-lab demos have mostly run on bespoke automation nobody outside can rebuild, which leaves the claims unfalsifiable. A documented reference platform moves agent-driven bench work from demo video to something a second group can check, and lowers the capital floor for testing whether agents can run physical protocols at all.
LLMs propose drug combinations for ER+ breast cancer
Language models scale network medicine to combinatorial drug repurposing in ER+ breast cancer, searching drug-target-pathway relationships for multi-drug candidates. Combination repurposing has been capped by the size of the search space, and automating the read of that space raises how many pairings get considered at all.
AI agents grade how severe genetic diseases are
Agents classify genetic-disease severity by working from published evidence rather than model recall. Severity calls gate newborn screening panels and variant-curation priorities, so an evidence-linked classification pipeline shifts what counts as a defensible call.
BELL checks biological models against published evidence
BELL pairs biomodel evidence with LLM-based logic, aiming at the curation step where mechanistic models get tied to the findings that justify them. That link has stayed a manual craft, and automating it raises the throughput ceiling for systems-biology model reuse.
Reply with your discoveries. A human reads them. Forward freely.
|