A deep learning model to classify neoplastic state and tissue origin from transcriptomic data.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Transcriptomic data analysis is crucial for understanding tissue states.
• Deep learning offers potential for improved accuracy in biological data interpretation.
• Existing methods may lack comprehensive classification across diverse tissue types.

Purpose of the Study:

• To develop a multitask deep learning model for classifying disease state, tissue origin, and neoplastic subclass.
• To leverage publicly available whole transcriptomic (RNA-seq) datasets for robust model training.
• To establish a framework for integrating large-scale transcriptomic data for clinical applications.

Main Methods:

• A multitask deep learning model was designed for tissue classification.
• Publicly available RNA-seq data from 10,116 patient samples were utilized.
• A uniform pipeline analysis was applied to training and validation datasets.

Main Results:

• The model achieved 99% accuracy for disease state classification (ROC-AUC 0.98).
• Tissue origin classification reached 97% accuracy (ROC-AUC 0.99).
• Neoplastic subclassification accuracy was 92% (ROC-AUC 0.95).

Conclusions:

• This study presents the first multitask deep learning algorithm for tissue classification using uniform transcriptomic data analysis.
• The developed model demonstrates high accuracy in classifying disease state, tissue origin, and neoplastic subclass.
• The framework facilitates clinical diagnosis and the development of cell-based treatment strategies.