DeePathology: Deep Multi-Task Learning for Inferring Molecular Pathology from Cancer Transcriptome.

Area of Science:

• Computational Biology
• Bioinformatics
• Artificial Intelligence in Oncology

Background:

• Molecular cancer pathology often relies on limited biomarkers due to computational challenges with whole transcriptome analysis.
• Existing methods struggle to efficiently process and interpret complex genomic data for cancer subtyping.

Purpose of the Study:

• To develop a novel Deep Neural Network (DNN) architecture for comprehensive molecular cancer pathology.
• To enable simultaneous inference of biological sample properties, including tissue and disease type, from whole transcriptome data.
• To overcome computational limitations in analyzing large-scale transcriptomic datasets.

Main Methods:

• Implemented a multi-task and transfer learning DNN architecture.
• Encoded whole transcription profiles into an 8-dimensional latent vector.
• Utilized mRNA transcription profiles from 10,750 clinical samples across 34 classes and 27 tissues.

Main Results:

• The DNN achieved superior performance in predicting tissue-of-origin, normal vs. disease state, and cancer type.
• Achieved 99.4% accuracy in identifying cancer subtypes within specific tissues.
• Demonstrated robustness against noise and missing data in transcriptomic profiles.
• The low-dimensional latent space significantly improved sample discrimination compared to original gene expression profiles.

Conclusions:

• The developed AI system, DeePathology, offers a powerful tool for molecular cancer pathology and oncological research.
• Highlights the potential of artificial intelligence to address computational challenges in analyzing complex biological data.
• Provides a robust and accurate method for cancer classification and subtype identification.