Biomedical Artificial Intelligence · asia
Teaching AI to Read Slides with Gene Expression: PathLUPI Explores Cancer Molecular Prediction
Using transcriptomic data during training and pathology images alone for prediction, PathLUPI attempts to identify molecular clues to cancer from tissue morphology. Results across cohorts show potential while revealing gaps between hospitals and markers.
Tumors that look similar under a microscope may harbor different molecular alterations, affecting treatment choices and prognostic assessments. Molecular testing can provide this information but requires additional time and resources. PathLUPI, developed by a team at the Hong Kong University of Science and Technology, attempts to enable artificial intelligence to identify these differences from routine pathology slides. According to the research team's November 2 announcement, the findings have been published in *Nature Communications*.
The key to this approach is giving the model an additional source of data during learning: the transcriptome, which reflects gene activity. The researchers paired tissue images with gene expression data from the same patients to guide the model in learning associations between morphology and molecular pathways, then transferred these clues to a model branch that receives only images. Once training is complete, analyzing a new patient's whole-slide images does not require that patient's transcriptomic data.
According to the formal study figures in the announcement, the evaluation covered 20 cohorts, 11,361 patients, and 49 molecular oncology tasks, with 18 biomarker or molecular subtyping tasks achieving an area under the curve (AUC) of 0.80 or higher. AUC measures a model's ability to distinguish between different states. It cannot be directly interpreted as correctly classifying 80% of patients, nor does it establish that the model is ready to guide medication use.
Cross-cohort testing showed in particular that performance depends on the prediction target. For predicting BRAF mutations in colorectal cancer, AUC fell from 0.840 in internal testing to 0.655 in external testing; for predicting IDH1 mutations in glioma, the corresponding values were 0.992 and 0.920. These gaps are a reminder that strong performance on one task does not guarantee continued reliability when moving to another cancer, marker, or specimen source.
The data versions also need to be distinguished. The arXiv preprint released by Cheng Jin and colleagues in June 2025 likewise evaluated 49 tasks and 20 cohorts, but reported 11,257 cases and 14 tasks achieving an AUC of 0.80 or higher. This article uses the figures from the research team's latest announcement. The earlier preprint can support the account of the method and research context, but the statistics from the two versions should not be mixed.
The official code repository provides model weights and inference instructions, and clearly distinguishes the requirements for training and use: training requires paired pathology features and transcriptomic data, while inference requires neither transcriptomic data nor pathway data files. This gives other researchers an opportunity to reproduce and examine the results, but making code and models publicly available does not amount to completed clinical performance validation.
The practical possibility proposed by PathLUPI is to find more molecular information in existing pathology images, potentially helping to arrange subsequent testing in the future. However, it has not yet been shown to improve decisions or patient outcomes in prospective clinical workflows. Before it can enter clinical practice, acceptable levels of false negatives and false positives for each intended use still need to be clarified, along with whether the model can maintain sufficiently stable performance across different hospitals.