A Deep Learning Framework to Predict Tumor Tissue-of-Origin Based on Copy Number Alteration

Area of Science:

• Computational biology
• Genomics
• Oncology

Background:

• Cancer of unknown primary site (CUPS) presents a diagnostic challenge, hindering effective treatment strategies.
• Accurate determination of the tissue-of-origin (TOO) for CUPS is critical for patient prognosis and treatment planning.
• Existing methods for TOO prediction often rely on various biomarkers, with limited exploration of copy number alterations (CNAs).

Purpose of the Study:

• To introduce CNA_origin, a novel two-step computational framework for predicting the tissue-of-origin of CUPS.
• To leverage gene copy number alteration (CNA) levels for enhanced accuracy in TOO prediction.
• To develop and validate a deep learning model for robust CUPS origin identification.

Main Methods:

• Development of CNA_origin, a computational framework employing a deep learning network.
• The network integrates an autoencoder for characteristic feature extraction from CNA data and a convolution neural network (CNN) for classification.
• Validation using real-world datasets with 10-fold cross-validation and independent dataset testing.

Main Results:

• CNA_origin achieved an overall accuracy of 83.81% in 10-fold cross-validation and 79% on independent datasets.
• The framework demonstrated significant accuracy improvements of 7.75% and 9.72% over a previously published method.
• The autoencoder effectively extracted key CNA features, and the CNN classifier provided robust and effective tumor origin prediction.

Conclusions:

• Gene copy number alterations are valuable indicators for predicting the tissue-of-origin in cancers of unknown primary site.
• The CNA_origin framework, utilizing deep learning, offers a powerful and accurate tool for CUPS diagnosis.
• This approach has the potential to improve treatment strategies and patient outcomes for CUPS.