TERRA: A Tissue World Model Pretrained on 112M Cells
mo_lotfollahi · x · 2026-08-06
A research team has introduced TERRA, a world model for human tissues. The model is pretrained on a massive corpus of 112 million cells, spanning 20 tissues and 26 disease conditions, with the majority of data newly generated via Xenium 5000-plex spatial transcriptomics.
Much like ViT for vision or ESM for proteins, TERRA aims to provide a universal foundational representation for human tissue, capturing its multi-scale logic from genes to cells and their microenvironments. Borrowing from the JEPA architecture championed by Yann LeCun, TERRA learns by making predictions in embedding space to understand and model complex human tissue states.
Related event: TERRA: Tissue World Model Trained on 112 Million Cells(4 posts)→
More from Research
- Open Models Beat GPT-5.6 Sol on Retrieval at 1/100th the Cost — girishkumama · 2026-08-06
- Meta Researcher's Advice: Treat Published AI Results with a Hierarchy of Evidence — fredahshi · 2026-08-06
- AI Search of Neanderthal Proteins Resurrects 'Extinct' Antibiotics — ArtificialOther · 2026-08-06
- ContinualSkillBench: Explicit Skill Libraries Offer Little Advantage for AI Agents — dair_ai · 2026-08-06
- Zalando's Presentation: Forecasting Metrics Problems in Practice — mariofilhoml · 2026-08-06
- Tackling LLM Overthinking: DASH Algorithm Boosts Math Reasoning Accuracy — Chia-Hsuan Lee · 2026-08-06