← Back to Home

Inferring Tumor Protein Maps from Pathology Slides: Hist2Prot Searches for Clues to Immunotherapy Response

Across seven independent cohorts comprising 1,774 patients, AI-inferred protein information improved performance on several clinical prediction tasks; the next step is to determine whether these clues can actually help guide treatment decisions.

By SURL BioNews

Beyond showing cell shapes and tissue organization, how much molecular information can a routine pathology slide reveal? A study published on November 13 in *Nature Cancer* introduced Hist2Prot, which attempts to infer protein distributions in individual tumor cells from routinely stained images, offering another approach to finding clues to treatment response.

The method targets barriers to using spatial protein analysis. Directly measuring where proteins are located in tissue helps researchers understand the tumor microenvironment, but is limited by cost, technical complexity, and tissue coverage. Hist2Prot uses data from approximately 51 million paired cells to learn the relationship between hematoxylin and eosin (H&E) images and protein signals. The resulting “virtual proteins” are model estimates and cannot be regarded as direct measurements.

In addition to internal testing, the study included independent data from the same platform and across platforms, totaling approximately 72 million additional cells. According to the paper’s abstract, the inferred results were consistent with measured protein distributions and retained some characteristics of immune, stromal, and tumor cells. In analyses of seven independent patient cohorts comprising 1,774 people, this information also improved predictions of molecular biomarkers, prognosis, and immunotherapy response compared with methods using H&E morphology alone.

However, a critical step remains between “better predictions” and “better treatment.” These results support further clinical evaluation, but do not yet demonstrate that selecting therapies based on the model can prolong survival or reduce ineffective treatment. The public abstract also does not provide complete performance figures for each task, so the total patient count alone cannot establish the reliability of each application.

The research team’s publicly released code repository makes the implementation requirements more concrete. In addition to whole-slide images, the documentation lists inputs including cell feature tables and segmentation masks identifying cell nuclei and expanded cell regions. It also provides code for preprocessing, training, inference, and evaluation. This means reproducing the study requires a data preparation workflow; the findings should not be interpreted as allowing users to upload any slide and obtain clinical answers.

By default, the repository splits training and test data at the patient level in a seven-to-three ratio, while also allowing users to specify group assignments. During evaluation, it processes patients who were not included in training one at a time. Patient-level splitting helps reduce bias caused by similar images from the same person appearing in both groups, but the default settings in the documentation cannot replace a full examination of each validation cohort in the paper.

Moving toward practical use will also require determining how differences in staining, scanning, and patients across hospitals affect the results, and testing its value for decision-making through clinical studies. Publicly available code does not mean unrestricted commercial use: the project specifies a CC BY-NC-ND 4.0 license, which includes restrictions on commercial use and the distribution of modified versions. For now, Hist2Prot offers a research approach that can be validated; both its clinical benefits and deployment requirements still need more evidence.

References

  1. Nature Cancer
  2. GitHub — Jsf826/Hist2Prot