PathMethy: an interpretable AI framework for cancer origin tracing based on DNA methylation.

Area of Science:

• Epigenetics and Computational Oncology
• Cancer Genomics and Bioinformatics

Background:

• Cancer of unknown primary (CUP) accounts for 3%-5% of cancer diagnoses, posing a significant clinical challenge.
• DNA methylation patterns are crucial epigenetic markers for identifying the origin of metastatic tumors.

Purpose of the Study:

• To develop and validate PathMethy, a novel Transformer model for accurately tracing tumor origins in CUP cases using DNA methylation data.
• To improve diagnostic accuracy and provide biological insights into cancer metastasis.

Main Methods:

• Developed PathMethy, a Transformer model integrating DNA methylation data with functional pathway categories and crosstalk.
• Evaluated PathMethy's performance against seven competing methods across nine cancer datasets.
• Assessed the model's ability to predict molecular subtypes and agree with clinical diagnoses in CUP cases.

Main Results:

• PathMethy significantly outperformed existing methods in F1-score across nine cancer datasets.
• The model accurately predicted molecular subtypes within nine primary tumor types.
• PathMethy demonstrated high agreement with previously diagnosed sites in CUP cases, validating its clinical utility.

Conclusions:

• PathMethy offers a powerful, accurate, and biologically insightful approach for determining tumor origins in cancer of unknown primary.
• The model's integration of pathway information provides a global understanding of cancer biology.
• A user-friendly web server is available at https://cup.pathmethy.com for broader research and clinical application.