DeepFeature: feature selection in nonimage data using convolutional neural network.

Area of Science:

• Computational biology
• Bioinformatics
• Artificial intelligence in medicine

Background:

• Classical statistical tests struggle with complex patterns in high-dimensional omics data.
• Deep neural networks, particularly Convolutional Neural Networks (CNNs), show promise but face challenges in interpreting biomedical results.
• Meaningful interpretation of deep learning models for omics data remains an open problem.

Purpose of the Study:

• To present DeepFeature, a novel pipeline applying CNNs to non-image omics data for effective feature selection.
• To enable the interpretation of deep learning models in a biomedical context.
• To facilitate the discovery of biological mechanisms from high-dimensional data.

Main Methods:

• Developed DeepFeature pipeline to transform omics data for CNN fitting.
• Integrated Snowfall compression algorithm for enhanced feature representation.
• Utilized region accumulation and element decoder to identify key genes from class activation maps.

Main Results:

• DeepFeature achieved superior predictive performance in cancer type prediction tasks.
• The pipeline successfully identified key pathways and biological processes relevant to cancer.
• Demonstrated the ability to extract meaningful biological insights from complex omics data.

Conclusions:

• DeepFeature offers an effective approach for applying deep learning to high-dimensional biomedical data.
• The framework facilitates the discovery of causal biological mechanisms.
• Enables more effective use of advanced AI methods for biological insights.