Path2Omics Enhances Transcriptomic and Methylation Prediction Accuracy from Tumor Histopathology

Area of Science:

• Computational biology
• Bioinformatics
• Artificial intelligence in oncology

Background:

• Precision oncology relies on molecular data for improved treatment outcomes.
• Acquiring molecular data from patient samples is often expensive and time-consuming.

Purpose of the Study:

• To develop Path2Omics, a deep learning framework for predicting gene expression and methylation from histopathology images.
• To assess the performance of Path2Omics across various cancer types and tissue preservation methods.

Main Methods:

• Developed Path2Omics with two components: a formalin-fixed, paraffin-embedded (FFPE) model and a fresh-frozen (FF) model.
• Trained models on histopathology slides from The Cancer Genome Atlas (TCGA) across 30 cancer types.
• Validated performance on seven external datasets, including FFPE and FF samples.

Main Results:

• The FF model outperformed the FFPE model, even on FFPE-only datasets.
• An integrated model combining FF and FFPE predictions showed a 30% improvement over the FFPE model alone.
• Path2Omics accurately predicted gene expression for approximately 4,400 genes and demonstrated clinical relevance in predicting survival and treatment response.

Conclusions:

• Path2Omics effectively infers gene expression and methylation from routine histopathology slides.
• The framework holds significant potential for advancing precision oncology by leveraging accessible tissue data.
• Inferred molecular data closely matched actual values, supporting clinical utility in predicting patient outcomes.