A graph-embedded deep feedforward network for disease outcome classification and feature selection using gene...

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Gene expression data poses a significant challenge for predictive modeling due to a high dimensionality (p) relative to sample size (n), often described as 'n≪p'.
• This imbalance hinders the direct application of deep learning techniques for disease outcome classification.
• Existing sparse learning methods can be limited by the scale-free structure of gene networks, which is not conducive to standard convolutional neural network architectures.

Purpose of the Study:

• To develop a robust classification model for gene expression data that addresses the 'n≪p' challenge.
• To integrate external gene network information into deep neural network architectures to improve predictive accuracy and feature selection.
• To overcome limitations posed by the scale-free structure of gene networks in deep learning models.

Main Methods:

• Proposing Graph-Embedded Deep Feedforward Networks (GEDFN), a novel deep learning architecture.
• Integrating external relational information of genes into the network architecture.
• Achieving sparse connections between network layers to mitigate overfitting.

Main Results:

• GEDFN demonstrated high classification accuracy in both simulation experiments and real-world analyses.
• The method was validated using RNA-seq datasets for breast invasive carcinoma and kidney renal clear cell carcinoma from The Cancer Genome Atlas.
• GEDFN provided easily interpretable feature selection results, highlighting its utility.

Conclusions:

• Graph-Embedded Deep Feedforward Networks (GEDFN) offer a powerful approach for disease outcome classification using high-dimensional gene expression data.
• The integration of gene network information enhances the robustness and interpretability of predictive models.
• GEDFN represents a valuable advancement in graph-guided classification and feature selection methodologies.