Path2Omics Enhances Transcriptomic and Methylation Prediction Accuracy from Tumor Histopathology

Abstract

Precision oncology is becoming increasingly integral to clinical practice, demonstrating notable improvements in treatment outcomes. While molecular data provide comprehensive insights, obtaining such data remains costly and time-consuming. Here, we developed Path2Omics, a deep learning framework that independently predicts gene expression and methylation from histopathology across 30 TCGA cancer types. Path2Omics comprised two components: an "FFPE model" trained on FFPE slides and an "FF model" trained on fresh frozen (FF) slides. When evaluated on seven external datasets, the "FF model" outperformed the "FFPE model", even though six of the datasets consisted exclusively of FFPE slides. The "Integrated model" combined predictions from both, achieving a 30% improvement over the FFPE model alone and robustly predicting approximately 4,400 genes (out of 18,000). Importantly, the inferred gene expression closely matched actual values in predicting patient survival and treatment response. Overall, this study demonstrated the potential of Path2Omics to advance precision oncology using routine histopathology slides.