Perturb-MARS: Reading mouse experiments through a human lens

· ai science · Source ↗

TLDR

  • Noetik combines multiplexed in-vivo mouse perturbations (Perturb-Map) with a human-only-trained foundation model (TARIO-2) to produce human-space tumor readouts from animal data.

Key Takeaways

  • Perturb-Map runs hundreds of CRISPR knockouts simultaneously in one mouse using Pro-Code barcodes, cutting cost and inter-animal variability vs. single-knockout cohort designs.
  • TARIO-2, trained exclusively on human cancer H&E, generalizes zero-shot to mouse tumor histology, predicting human spatial gene expression without retraining.
  • The combined system (Perturb-MARS) outputs human-gene-space readouts, including genes like CEACAM6 that have no mouse ortholog, grounding animal causal data in human biology.
  • Spatial resolution is preserved unlike Perturb-Seq, which requires tissue dissociation and destroys tumor microenvironment structure.
  • Claimed use case: answering combination therapy questions (e.g., next PD-1 bispecific target after VEGF) that standard preclinical apparatus cannot address.

Hacker News Comment Review

  • No substantive HN discussion yet.

Original | Discuss on HN