Optimizing gene selection and module identification via ontology-based scoring and deep learning

Abstract

Motivation: Understanding gene interactions and their biological significance is a key challenge in computational biology. The complexity of biological systems, coupled with high-dimensional omics data, necessitates robust methods for gene selection and interaction analysis. Traditional statistical techniques often struggle with the hierarchical nature of gene ontology (GO) terms, leading to redundancy and limited interpretability. Meanwhile, deep learning models require biologically meaningful input to enhance their predictive power. Results: We present an integrated framework that enhances gene selection and uncovers gene interactions by combining a novel statistical algorithm with a deep neural network model. The statistical algorithm ranks differentially expressed genes by correlating their expression scores with the semantic similarity of their biological context, utilizing GO information to align genes with known pathways. The deep neural network then identifies interaction modules by integrating genes from different clusters based on regulatory pathway data. This model effectively navigates the hierarchical complexity of GO terms structured as directed acyclic graphs, employing a feed-forward architecture optimized via back-propagation. Our results demonstrate improved gene selection accuracy and enhanced discovery of biologically relevant interactions, providing valuable insights into complex disease mechanisms.