Using Copy Number Variation Data and Neural Networks to Predict Cancer Metastasis Origin Achieves High Area under the...

Area of Science:

• Genomics
• Computational Biology
• Oncology

Background:

• Accurate primary tumor identification in metastatic cancer is vital for effective treatment and improved patient outcomes.
• Copy number alterations (CNAs) and copy number variation (CNV) are promising genomic markers for predicting metastasis origin.
• Existing models for cancer type prediction using CNV/CNA data exhibit suboptimal performance (low AUC values).

Purpose of the Study:

• To develop and evaluate advanced neural network models for predicting the origin of metastatic cancers using CNA profiles.
• To improve the accuracy and reliability of cancer type prediction from genomic data.
• To explore the relationship between CNV characteristics and cancer type.

Main Methods:

• Utilized a dataset of CNA profiles from twenty distinct cancer types.
• Developed and assessed two deep neural network architectures: a ReLU-based network and a 2D convolutional neural network.
• Implemented a second workflow involving stratification of cancer types by anatomical and physiological classifications, followed by shallow neural network construction for intra-cluster differentiation.

Main Results:

• Both deep and shallow neural network approaches achieved high Area Under the Curve (AUC) values.
• Deep neural networks demonstrated a precision of 60% in predicting cancer origin.
• Findings suggest a quantifiable mathematical relationship between CNV type, genomic location, and cancer type.

Conclusions:

• Advanced neural network models, particularly deep learning approaches, significantly enhance the accuracy of predicting cancer origin from CNA data.
• The study validates the potential of CNAs/CNVs as reliable biomarkers for metastasis origin identification.
• These findings offer a promising avenue for aiding pathologists in diagnosing cancer origins using accessible clinical genomic tests.