Cox-PASNet

Cox-PASNet applies a pathway-based sparse deep neural network to integrate high-dimensional gene expression and clinical data for cancer survival prediction and interpretation of prognostic genes and pathways.

Key Features:

• Pathway-based sparse deep neural network: Nodes correspond directly to genes and pathways to capture nonlinear and hierarchical biological effects.
• Integration of genomic and clinical data: Combines high-dimensional gene expression data with clinical information within a single model.
• Use of pathway databases: Leverages pathway databases to model biological mechanisms and impose biologically informed structure on the network.
• Sparse coding techniques: Employs sparse coding to manage complexity in high-dimensional, low-sample size (HDLSS) datasets.
• Heuristic optimization for HDLSS: Uses a heuristic optimization solution to improve training robustness with limited sample sizes.
• Implementation: Implemented in PyTorch.
• Empirical performance: Demonstrated superior predictive performance and identification of significant prognostic genes and pathways in evaluations.

Scientific Applications:

• Cancer survival prediction: Predicts patient survival outcomes by integrating genomic and clinical variables.
• Identification of prognostic genes and pathways: Facilitates biological interpretation to identify significant prognostic factors.
• Modeling nonlinear and hierarchical associations: Captures complex nonlinear and hierarchical relationships between genes, pathways, and clinical features.
• Application to specific cancers: Applied to glioblastoma multiforme (GBM) and ovarian serous cystadenocarcinoma (OV) datasets.
• Informing treatment development: Provides interpretable prognostic insights that can inform treatment development and personalized medicine.

Methodology:

Integrates high-dimensional gene expression and clinical data into a pathway-structured sparse deep neural network with nodes mapped to genes and pathways, uses pathway databases and sparse coding, applies a heuristic optimization strategy for HDLSS training, and is implemented in PyTorch; evaluated on GBM and OV datasets.