← Back to Home

Helping AI Read Tumor Cell Neighborhoods: EPFL Releases a Spatial Proteomics Atlas

spora brings together tissue images from more than 5,000 patients, allowing researchers to compare how different cells neighbor and interact with one another in tumors while establishing a common testing foundation for cancer AI. Data integration has taken a step forward, but reliable clinical predictions still require validation.

By SURL BioNews

A piece of tumor tissue can reveal more than what cancer cells look like. Where immune cells gather and which cells sit next to one another may also hold clues to disease progression and treatment response. On October 28, the Swiss Federal Institute of Technology Lausanne (EPFL) announced spora, a spatial proteomics resource that organizes tissue images scattered across different studies into a common format, giving artificial intelligence an opportunity to learn from more diverse cellular “neighborhoods.”

spora contains 13,546 multiplexed images from 35 study cohorts comprising 5,477 patients. According to EPFL’s announcement and the platform documentation, it covers 14 tissue sources and five imaging technologies: IF, Orion, IMC, CODEX, and MIBI. These technologies measure multiple proteins while preserving information about their locations in tissue, helping researchers identify cells and map the spatial relationships between tumors and surrounding immune cells.

Previously, these data were often difficult to analyze together: laboratories differed in their instruments, protein marker panels, file formats, and clinical annotations. The spora team therefore standardized images and protein names, organized patient data, and provided segmentation information delineating cell and tissue boundaries. However, standardizing formats does not eliminate experimental differences. The atlas retains the sources and annotations of the original studies so users can judge which data are suitable for answering their questions.

The atlas also seeks to connect molecular measurements with routine pathology interpretation. Of the 944,680 multiplexed image patches listed in the platform documentation, 30.7% have paired hematoxylin and eosin (H&E) staining data. This provides training material for “virtual staining”: researchers can test whether models can estimate the distribution of specific proteins from routine pathology images. However, paired data are available for only some patches, so the entire atlas cannot be treated as having the same degree of correspondence between molecular and pathology data.

Released alongside the atlas, the spora-io library helps researchers load and split images and sample across cohorts, while spora-bench provides a common evaluation workflow. A paper posted on arXiv on October 15 and marked as accepted at NeurIPS 2026 describes benchmarks spanning 12 study cohorts. These include cell type identification, diagnosis and treatment response prediction, and virtual staining, allowing existing models to be compared with task-specific models under consistent conditions.

The central question addressed by this evaluation framework is whether models remain reliable when patients, laboratories, or measurement technologies change. Strong performance within a single study may partly reflect familiarity with particular data characteristics; testing across studies can further examine whether the patterns a model has learned generalize. The atlas expands opportunities to investigate this question, but the amount of data, available protein markers, and clinical information remain uneven across tissues, requiring attention to the scope of applicability when using it.

For now, spora provides data and a testing foundation for research; these results do not yet establish that AI can improve patient diagnosis or treatment outcomes. Using models to interpret pathology or select therapies still requires validation for the intended use with independent patient data and within clinical workflows. Open data also come with conditions of use: the platform documentation notes that licensing varies across cohorts, and both subsequent research and commercial applications require checking each license individually.

References

  1. EPFL
  2. arXiv
  3. EPFL — Bunne Lab
  4. BioImage Archive