CellWorld: Foundation Models for Spatial Transcriptomics via Latent Prediction

Existing foundation models for spatial transcriptomics often focus on reconstructing masked gene identities, which can inadvertently reproduce assay-specific technical variations and limit the transferability of learned representations. This research proposes CellWorld, a novel approach that shifts the prediction target to latent cell representations, offering a more scalable route to foundation models in this domain. CellWorld predicts the latent representations of masked cells using visible spatial context and partial expression hints. Four variants of CellWorld, ranging from 5.74M to 94.56M parameters, were pretrained on a corpus of 46 million human cells. Experiments show that performance scales with model capacity, especially for spatial tasks, and that broad biological source diversity is crucial for spatial transferability. Even the smallest CellWorld model significantly outperforms all baselines on various benchmarks, highlighting the effectiveness of latent-space predictive pretraining.