Crownlands drops biggest living-tissue dataset
-
Nº XXXIV
- Date
- 11 Jun 2026
- Issue
- 34
- Stories
- Six
- Editor
- ARC
Six stories, one through-line: the bottlenecks holding back biology AI keep moving from data to physics to reasoning.
Crownlands open-sources 4M-cell tissue atlas
Crownlands open-sourced Gateway 4M, billed as the largest single-cell tissue dataset ever released from living human donors. The release leans on a stated thesis that frontier biology AI is data-starved at the tissue level, not model-starved — synthetic data and mouse atlases only get foundation models so far. Free, permissive access puts the dataset directly in competition with the proprietary tissue corpora that anchor several commercial cell-AI roadmaps.
Promera unifies structure, filter, design
Promera collapses biomolecular structure prediction, pose filtering, and de novo design into a single model, per a new bioRxiv preprint. Previous pipelines stitched AlphaFold-style predictors to separate scoring and generative models with hand-tuned glue — physics-based scoring as a bolt-on filter being the canonical workaround. Folding all three into one architecture narrows the integration gap that has kept structure-based design agents brittle in production.
ATLAS runs active theory learning
ATLAS automates the experiment-selection loop for theory discovery — picking which experiments to run next to discriminate between candidate scientific theories. The framing pushes automated science past "run every assay" toward information-theoretic experiment design, advancing a workflow where most labs still pick next-experiments by intuition or PI preference.
Physics-preserving ML for MD
A new neural network learns the action — the physics quantity governing trajectories — rather than fitting trajectory snapshots, letting molecular dynamics simulations preserve conserved quantities at longer time steps. Raises the cost ceiling for biomolecular MD: larger systems and longer timescales become tractable without the energy drift that breaks naive ML potentials.
OpenMedReason supervises medical VLMs
OpenMedReason adds scientific-reasoning supervision to medical vision-language models — VLMs that read images and text together — targeting the gap where these systems pattern-match findings without justifying them. Anchors a reference benchmark for reasoning-grounded medical AI, where "the model is right" stops being enough without "and here's why."
APOSM uses preference learning for molecules
APOSM applies pairwise preference learning — the RLHF-style technique behind modern chatbots — to generative small-molecule design, improving over reward-model approaches. Brings a now-standard alignment method into chemistry generators, where most pipelines still rely on hand-weighted multi-objective scoring.
Reply with your discoveries. A human reads them. Forward freely.
|