Noncoding RNAs and Deep Learning Neural Network Discriminate Multi-Cancer Types.

Area of Science:

• Biomarkers and Diagnostics
• Artificial Intelligence in Medicine
• Oncology

Background:

• Early cancer detection significantly reduces mortality rates, necessitating practical population-level screening solutions.
• Current screening methods may lack the comprehensive accuracy required for diverse cancer types.
• Noncoding RNA biomarkers offer potential for sensitive and specific cancer detection.

Purpose of the Study:

• To develop and validate an integrated artificial intelligence (AI) deep learning system for multi-cancer detection.
• To utilize noncoding RNA biomarkers for accurate classification of cancer versus healthy states.
• To establish a practical, affordable, and user-friendly framework for population-level cancer screening.

Main Methods:

• Integration of a deep learning neural network with selected noncoding RNA biomarkers.
• Validation of the system using massive datasets, real-world raw data, and exosome data from blood.
• Comparison of the AI system's performance against conventional machine learning models like random forest.

Main Results:

• The system achieved an Area Under Curve (AUC) of 96.3% for cancer vs. healthy detection.
• Validation on independent real-world data yielded an AUC of 78.77%, and 72% with blood exosome data.
• Individual cancer type discrimination reached 99-100% AUC using six biomarkers, outperforming traditional models.
• Simultaneous multi-cancer classification achieved 82.15% accuracy for heterogeneous tissues.

Conclusions:

• The developed AI-powered noncoding RNA biomarker system offers a simple, accurate, and affordable solution for population-level cancer screening.
• The system demonstrates high efficacy in binary cancer detection, individual cancer type identification, and simultaneous multi-cancer classification.
• This approach provides a robust and practical framework for advancing early cancer detection and improving patient outcomes.