Biomedical Artificial Intelligence · global
Reading Molecular Clues from Pathology Slides: PathLUPI Tested in More Than 10,000 Cancer Cases
Using gene expression data during training and only pathology images for prediction, PathLUPI seeks to bridge tissue morphology and the molecular features of cancer. Results across cohorts show potential, but whether it can improve clinical decisions requires prospective research.
A cancer pathology slide can reveal the shapes and arrangement of cells, but may not directly reveal molecular features that influence treatment decisions. Identifying these clues earlier from routinely acquired images could help guide subsequent testing. The PathLUPI study, published on November 2 in Nature Communications, proposes an approach in this direction: exposing the model to transcriptomic information during learning, while using only pathology slide images for prediction.
The transcriptome records the state of gene expression. During training, PathLUPI uses paired tissue images and transcriptomic data to guide the model in learning image representations associated with molecular features. According to the official code documentation, once training is complete, the model receives features extracted from whole-slide images, without requiring transcriptomic information for each case being evaluated. This is central to its effort to fit into routine pathology workflows.
The evaluation in the formally published version covered 20 cohorts and 11,361 cases across 49 molecular oncology tasks. AUROC reached or exceeded 0.80 in 18 biomarker and molecular subtyping tasks. AUROC measures a model's ability to distinguish between classes; it cannot be directly interpreted as “correctly classifying 80% of cases,” nor does it mean that all evaluated tasks achieved the same level of performance. Prediction targets listed in the publicly available model resources include EGFR mutations in lung adenocarcinoma, breast cancer receptor status, and IDH1 mutations in glioma, illustrating the specific molecular clues the approach seeks to identify from morphology.
These figures also need to be distinguished from those in the earlier version. The preprint of the same study, released in June 2025, reported 11,257 cases and 14 tasks achieving an AUC of at least 0.80; the formally published version reported 11,361 cases and 18 tasks, respectively. Although both versions listed 49 tasks and 20 cohorts, their results should not be used interchangeably, and differences between the abstracts alone cannot establish why the figures changed.
The team also provides code, training instructions, curated labels, and five-fold validation splits on GitHub. On Hugging Face, it has released 185 model weight files covering five-fold models for 24 biomarker and molecular prediction tasks and 13 survival prognosis tasks. These resources give other researchers an opportunity to reproduce and examine the method. However, the weight files themselves are not accompanied by validation metrics or training metadata, and downloading a model does not amount to independent validation of its performance.
Several questions remain before this research tool can move into clinical practice: How will differences in slide preparation and scanning across hospitals affect performance? What interpretation thresholds should be set in actual patient populations to appropriately manage missed and incorrect classifications? The current results have not established prospective clinical benefit and are insufficient to support replacing molecular testing with image-based predictions. PathLUPI offers a way to bring molecular information into image learning. Whether it can help medical teams arrange testing more efficiently and thereby improve care still requires validation through studies in actual workflows.