DeePathology

DeePathology applies a Deep Neural Network (DNN) with multi-task and transfer learning to encode the entire transcriptome and infer mRNA and miRNA expression profiles, tissue type, and disease state for molecular cancer pathology.

Key Features:

• Multi-Task Learning Architecture: Employs a Deep Neural Network (DNN) using multi-task and transfer learning to simultaneously infer mRNA and miRNA expression profiles, tissue type, and disease state.
• Dimensionality Reduction: Encodes the entire transcription profile into an exceptionally low-dimensional latent vector of size 8 to enhance sample discrimination.
• High Accuracy in Classification: Tested on mRNA transcription profiles from 10,750 clinical samples across 34 classes (one healthy and 33 cancer types) from 27 tissues, demonstrating superior performance over previous methods and classical machine learning approaches and achieving up to 99.4% accuracy in identifying correct cancer subtypes for tissues with multiple cancer types.
• Robustness: Designed to be robust against noise and missing values to maintain reliable performance under suboptimal data conditions.

Scientific Applications:

• Tissue and Cancer Type Identification: Predicts tissue-of-origin and distinguishes between normal and diseased states, including identification of specific cancer types.
• Transcriptome Analysis: Analyzes the whole transcriptome rather than relying on limited biomarkers to provide comprehensive molecular pathology insights.

Methodology:

Deep Neural Network employing multi-task and transfer learning to encode the entire transcriptome into an 8-dimensional latent vector and infer mRNA and miRNA expression profiles, tissue type, and disease state; trained and evaluated on mRNA transcription profiles from 10,750 clinical samples across 34 classes from 27 tissues.