Science · Science brief · Research

Spora dataset posted on arXiv is a spatial proteomics corpus of multiplex images

The 15 October 2026 arXiv version counts 13,546 multiplex images from 5,477 patients.

Adrian Quill · Published Research paper: arXiv · 2 more · Explore evidence and context

In this story

  • von Querfurth and colleagues assemble spora from 13,546 multiplex images and 5,477 patients.
  • The paper's count is 35 cohorts, 13 tissue origins, and five imaging technologies, with paired histology where studies supply it.
  • The authors' benchmark covers cell type, diagnosis, treatment response, and virtual staining, and they say the scores are not saturated.
  • The arXiv record notes acceptance at NeurIPS 2026, and the paper does not show a change in care for a patient.

Spatial proteomics keeps proteins tied to their places in tissue, so models can learn from both identity and layout at once. Benedikt von Querfurth and colleagues posted spora, A Unified Multimodal Dataset and Platform for Spatial Proteomics, on arXiv on 15 October 2026, version 1, DOI 10.48550/arxiv.2610.24002.

The work is a dataset and benchmark study posted as an arXiv preprint. The arXiv record for that version carries the comment Accepted to NeurIPS 2026.

In the 2026 arXiv paper, the image count for spora is 13,546 standardized multiplex images, the patient count is 5,477, and the cohort count is 35, across five spatial proteomics technologies. In that same paper, the tissue-origin count is 13, the technologies are IMC, CODEX, Orion, MIBI, and IF, and the marker count is 295. The paper says images are resampled to 1.0 micrometers per pixel, samples are de-identified, and paired H&E images are included where the source studies have them. The authors write that assembling the collection took two years with biomedical experts, and that to their knowledge it is the largest publicly available spatial proteomics corpus assembled to date. They also write that the benchmark-cohort count is 12 and that performance on that benchmark is far from saturated.

Imaging mass cytometry of a lung tumor core from a tissue microarray
2026 EPFL/AIMM. Imaging mass cytometry of a lung tumor core from a tissue microarray. CC-BY-SA 4.0.

Phys.org published an EPFL account by Tanya Petersen on 28 October 2026. That account says the work comes from Charlotte Bunne's Artificial Intelligence in Molecular Medicine group at EPFL, with scientists from EPFL, University Hospital Zurich, the University of Zurich, Geneva University Hospitals, and the University of Geneva. The account calls Benedikt von Querfurth a doctoral student at EPFL and the first author, and calls Lukas Klein a postdoctoral researcher at EPFL and the University Hospital Zurich.

This is a single dataset and benchmark paper, and it does not establish that the atlas changes treatment for an individual patient.