PathLUPI Uses Transcriptomic Training to Predict Molecular Features from Histology Images

Researchers have developed PathLUPI, a deep learning framework designed to extract genome-anchored histological representations by incorporating transcriptomic privileged information during training. Evaluated across 49 molecular oncology tasks using 11,361 cases drawn from 20 distinct cohorts, the model demonstrated superior performance over conventional methods that rely exclusively on whole-slide images for training.

In precision oncology, acquiring accurate molecular insights directly through genomic sequencing remains costly and time-consuming for widespread clinical implementation. At the same time, reliably predicting complex molecular characteristics and patient clinical prognosis directly from routine whole-slide histology images has presented a persistent computational challenge for existing deep learning approaches.

To address this limitation, PathLUPI integrates transcriptomic data as privileged information during its training phase. This training strategy allows the framework to learn genome-anchored histological embeddings from tissue architecture, enabling the model to perform molecular predictions using standard whole-slide images alone once deployed at inference.

Across extensive testing, PathLUPI attained an area under the receiver operating characteristic curve of at least 0.80 across 18 separate biomarker prediction and molecular subtyping tasks. The investigators noted that this performance surpassed standard baseline methods that were trained solely on routine histological whole-slide images.

Beyond biomarker identification, the model demonstrated utility in forecasting patient outcomes. PathLUPI achieved a concordance index of at least 0.70 across survival cohorts spanning five major cancer types, indicating its capacity to capture clinically relevant prognostic signals directly from routine diagnostic slides.

The embeddings generated by PathLUPI also identified distinct cellular morphological signatures tied to specific genotypes. The spatial fidelity of these learned morphological patterns was further substantiated through analysis of representative high-resolution spatial transcriptomics samples, providing a framework to connect molecular context with routine digital pathology workflows.

Subject of Research: Deep learning for molecular prediction from histology images

Article Title: Genome-anchored foundation model embeddings improve molecular prediction from histology images

Article References: Jin, C., Zhou, F., Yu, Y., Ma, J., Wang, Y., Xu, Y., Zhou, H., Jiang, H., Luo, L., Mao, L., He, Z., Zhang, X., Zhang, J., Chan, R. C. K., Yao, H., & Chen, H. (2026). Genome-anchored foundation model embeddings improve molecular prediction from histology images. Nature Communications. https://doi.org/10.1038/s41467-026-78947-2

Image Credits: AI Generated

DOI: 10.1038/s41467-026-78947-2

Keywords: PathLUPI, precision oncology, molecular prediction, whole-slide images, histopathology, transcriptomic privileged information, genome-anchored embeddings, biomarker prediction, molecular subtyping, cancer survival prediction, deep learning, spatial transcriptomics, cellular morphology, concordance index, receiver operating characteristic, computational pathology, cancer cohorts