AI Insight
CellOS is a new 12-billion-parameter foundation model that learns from 390.5 million single-cell transcriptomes by integrating two complementary views of cellular state: gene expression patterns and functional perceptions. Unlike existing models that focus solely on reconstructing gene expression data, CellOS uses a three-stage training approach that aligns these multiple views through language modeling and a specialized prediction objective. The model demonstrates superior performance compared to current state-of-the-art systems across multiple tasks including cell-state classification, batch integration, and predicting how cells respond to perturbations.
Why it matters
This approach represents a significant step toward creating AI-powered "virtual cells" that can more accurately model and predict cellular behavior. Such models could accelerate drug discovery, disease understanding, and personalized medicine by enabling researchers to simulate cellular responses without conducting every experiment in the laboratory.
Understand the Science
⚠️ Preprint – Noch nicht peer-reviewed
Dieser Artikel wurde noch nicht von unabhängigen Experten begutachtet. Die Ergebnisse sind vorläufig und sollten mit Vorsicht interpretiert werden.
Foundation models learned from single-cell transcriptomes are central to the prospect of AI virtual cell that can represent, query and predict cellular state. However, most current single-cell foundation models learn from a single view of gene expression and are optimized primarily through reconstruction or next-token prediction. As a result, they capture expression abundance but can-not explicitly reconcile complementary views of cellular state. Here we present CellOS, a multi-view foundation model that learns cellular representations from paired expression and perception views. CellOS integrates complementary views through a scalable three-stage training strategy that combines causal cell-sentence language modelling, function-preserving dense-to-mixture-of-experts expansion and latent-space alignment via an LLM-JEPA objective. Using this framework, we trained a 12-billion-parameter model on 390.5 million single-cell transcriptomes. Across diverse benchmarks spanning cell-state annotation, batch integration and perturbation-response prediction, CellOS consistently outperformed state-of-the-art single-cell foundation models in cell-state annotation and perturbation-response prediction while preserving robust batch integration. Together, these results suggest that predictive alignment between complementary cellular views provides a scalable path toward representation-centric cellular world models and transferable AI virtual cells.
Source: CellOS: Learning a World Model of Cellular State through Joint Embedding Prediction