Deep GONet

Deep GONet predicts phenotypes from gene expression data by integrating Gene Ontology (GO) annotations into a hierarchical, fully-connected deep neural network to produce interpretable associations between genes, GO terms, and clinical outcomes such as diagnosis, prognosis, and drug response.

Key Features:

• Integration of Gene Ontology (GO): Embeds GO annotations within a hierarchical neural network so that neurons represent specific GO biological functions.
• Self-Explainable Architecture: Identifies the most influential neurons for predictions and links them to biological functions to provide interpretability.
• GO-constrained Fully-Connected Architecture: Uses a fully-connected neural network architecture constrained by GO annotations to preserve gene-to-function relationships during learning.
• Phenotype Prediction from Gene Expression: Predicts clinical phenotypes such as diagnosis, prognosis, and drug response from gene expression profiles.
• Validation on Cancer Datasets: Demonstrated discrimination between cancerous and non-cancerous samples through experiments on cancer diagnosis datasets.

Scientific Applications:

• Precision Medicine: Supports prediction of clinical outcomes (diagnosis, prognosis, drug response) from patient gene expression data to inform clinical decision-making.
• Cancer Diagnostics: Enables discrimination between tumor and normal samples using gene expression-based phenotype prediction validated on cancer datasets.
• Mechanistic Interpretation: Links gene expression patterns to GO biological functions to aid interpretation of underlying biological mechanisms driving phenotypes.

Methodology:

Constructs a hierarchical, GO-constrained fully-connected neural network by embedding GO annotations so that each neuron corresponds to a specific biological function, identifies significant neurons associated with predictions, and validates performance on relevant datasets.