CNA_origin

CNA_origin predicts tumor tissue-of-origin from copy number alteration (CNA) profiles using a two-step deep learning framework to aid identification of cancers of unknown primary site (CUPS).

Key Features:

• Input data: Uses gene-level copy number alteration (CNA) profiles as the primary molecular input.
• Architecture: Employs a two-step deep learning architecture combining an autoencoder and a convolutional neural network (CNN).
• Representation learning: The autoencoder extracts efficient low-dimensional representations from raw CNA data while preserving salient features.
• Classification: A CNN classifier uses the autoencoder-derived features to predict tissue-of-origin.
• Performance: Achieved 83.81% accuracy in 10-fold cross-validation and 79% accuracy on independent datasets, with reported improvements of 7.75% and 9.72% over previous methods.
• Implementation: Implemented in Python.

Scientific Applications:

• Tissue-of-origin prediction: Assigns likely primary tumor origin for metastatic cancers of unknown primary site using CNA data.
• CNA feature extraction: Provides learned low-dimensional representations of CNA profiles for downstream analyses.
• Classifier benchmarking: Enables evaluation of predictive performance via 10-fold cross-validation and independent dataset testing.

Methodology:

Two-step pipeline comprising an autoencoder for representation learning and dimensionality reduction followed by a CNN classifier, evaluated by 10-fold cross-validation and independent dataset validation, implemented in Python.