Reproduce “CellWorld: From Gene-Level Reconstruction to Latent Cell Prediction in Spatial Transcriptomics Foundation Models” on OpenResearch

About this paper

Haiping Liu, Qian Zhao, Lijing Lin, Jingyuan Sun, Hongpeng Zhou

CellWorld introduces a foundation model for spatial transcriptomics that employs latent cell prediction as its pretraining objective, moving beyond gene-level reconstruction to learn more robust representations. The model achieves state-of-the-art performance across all 11 linear-probe and 7 fine-tuned spatial benchmarks on held-out datasets, with its smallest variant outperforming existing baselines.

Key insights

• Latent cell prediction, when augmented with partial-expression hints, is a more effective and scalable pretraining objective for spatial transcriptomics foundation models than direct gene-level reconstruction, yielding more robust and transferable representations.
• Effective pretraining for spatial transfer is not solely dependent on the sheer volume of cells but crucially benefits from broad biological source diversity and extensive optimization steps.
• Incorporating minimal partial-expression hints is critical for preventing predictor degeneration and stabilizing training during model scaling, enabling specific target prediction despite the inherent cell-level ambiguity in ST data.

Summary by alphaXiv.