iSOM-GSN

iSOM-GSN transforms high-dimensional multi-omic data into two-dimensional gene similarity grids using Kohonen's self-organizing map to enable application of graph convolutional neural networks (CNNs) for disease-state prediction and representation learning.

Key Features:

• Dimensionality Reduction: Uses Kohonen's self-organizing map to map heterogeneous multi-omic data such as gene expression, DNA methylation, and copy number alterations into a 2D grid representation.
• Gene Similarity Networks: Constructs gene similarity networks from the 2D grid to capture relationships between genes across different omic layers.
• Enhanced Representation Learning: Provides a structured 2D framework that enhances representation learning and supports downstream neural-network analyses.
• Disease Prediction: Applies convolutional neural networks on transformed 2D grids to predict disease states, reporting prediction accuracies of 94%–98% for breast tumor stages and prostate Gleason scores using only 14 input genes.
• Visualization and Interpretation: Produces grid-based visualizations that facilitate interpretation of multi-omic interactions and gene relationships.

Scientific Applications:

• Oncology: Prediction of breast cancer tumor stages and prostate cancer Gleason scores from integrated multi-omic inputs.
• Multi-omic Integration for Complex Diseases: Integration and joint analysis of heterogeneous omic types (gene expression, DNA methylation, copy number alterations) for disease-associated pattern discovery.

Methodology:

Transform multi-omic data into 2D grids using Kohonen's self-organizing map; generate gene similarity networks from the grids; apply convolutional neural networks (CNNs) to the resulting network/grid representations for disease-state prediction.