DeepPathway: Predicting Pathway Expression from Histopathology Images

In plain English

The model uses stained tissue slides to estimate which biological pathways are active in different regions.

What they found

• The authors report that predictions distinguished selected pathway activities in normal and tumor regions in TCGA images.
• They also evaluated hypoxia-signature predictions against pimonidazole staining in brain-tumor samples.

The catch

• Predicted pathway activity is not a substitute for direct transcriptomic measurement.
• The publisher abstract does not report complete sample counts or quantitative performance metrics.

Methods at a glance

Study type

Contrastive model development and external image evaluation

Biological system

Cancer tissue microenvironments

Organism

Human

Dataset

Paired spatial transcriptomics and H&E images, TCGA tissue images, and brain-tumor samples with hypoxia staining

Sample size

Not reported

Model architecture

Contrastive learning model

Training objective

Predict pathway expression from histopathology images

Code

The publisher links code on GitHub and Zenodo.

Data

Dataset access is described in the accepted manuscript.

Tasks and metrics

Tasks: Separate pathway activity in normal and tumor regions; Predict hypoxia signatures

Metrics: Complete quantitative metrics are not reported in the publisher abstract

Baselines: Gene-level histopathology prediction approaches discussed by the authors

Principal findings

• The authors report differentiated pathway activities across normal and tumor regions and agreement with hypoxia staining in the described samples.

What remains unresolved

External validation: TCGA images and brain-tumor samples with pimonidazole hypoxia staining

• The abstract does not provide complete cohort sizes.
• Complete benchmark metrics require the accepted manuscript.