CellWorld: Latent-Space Pretraining for Spatial Transcriptomics
A recent preprint on arXiv (2608.06659) presents CellWorld, a foundational model designed for spatial transcriptomics that transitions the focus from gene measurements to latent cell representations. This method seeks to eliminate assay-specific technical variations that hinder the adaptability of current models. CellWorld generates latent representations of masked cells by utilizing visible spatial context and partial expression clues. The authors trained four variants, containing between 5.74M and 94.56M parameters, on a dataset comprising 46 million human cells. Experiments conducted under controlled conditions indicate that performance enhances with increased model capacity, especially for spatial tasks, while spatial transfer is influenced by additional factors. This work was announced as a new submission by the researchers on arXiv.
Key facts
- CellWorld predicts latent representations of masked cells from visible spatial context and partial-expression hints.
- Pretrained on a corpus of 46 million human cells.
- Four variants with 5.74M to 94.56M trainable parameters.
- Performance improves with model capacity, especially on spatial tasks.
- Spatial transfer depends more on factors other than model capacity.
- Paper available at arXiv:2608.06659.
- Announcement type: new.
- Published on arXiv.
Entities
Institutions
- arXiv