Reproduce “STimage-1K4M: A histopathology image-gene expression dataset for spatial transcriptomics” on OpenResearch

About this paper

Jiawen Chen, Muqing Zhou, Wenrong Wu, Jinwei Zhang, Yun Li, Didong Li

STimage-1K4M assembles 1,149 histopathology slides with 4,293,195 spatially aligned image regions and gene-expression profiles spanning 50 tissue types and 10 species. A contrastive-learning demonstration shows that transcriptomic supervision improves CLIP- and PLIP-based image representations for tissue classification and cortical-layer separation, although the resulting models remain below the performance of the pretrained UNI pathology model and are affected by cross-study batch effects.

Key insights

• Molecular supervision can improve generic image representations: fine-tuned CLIP and PLIP embeddings outperformed their zero-shot versions on linear-probe classification and showed clearer separation of annotated cortical layers.
• The dataset combines broad biological coverage with substantial imbalance, including concentrations in human and mouse samples, brain and breast tissue, and cancer-related studies.
• Cross-dataset training remains difficult because platform differences, gene-set variation, and batch effects can reduce performance, so scale alone does not resolve spatial transcriptomics integration challenges.

Summary by alphaXiv.