Deep GONet: self-explainable deep neural network based on Gene Ontology for phenotype prediction from gene expression...

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Genomic sequencing generates vast gene expression data, driving precision medicine development.
• Deep learning models offer accurate phenotype prediction but often lack interpretability.
• Integrating biological knowledge into deep learning is crucial for medical applications.

Purpose of the Study:

• To develop a self-explainable deep learning model for interpretable phenotype prediction.
• To integrate prior biological knowledge into a deep learning framework for medical applications.

Main Methods:

• Proposed Deep GONet, a self-explainable deep learning model.
• Integrated Gene Ontology into the neural network's hierarchical architecture.
• Constrained a fully-connected architecture using Gene Ontology annotations, mapping neurons to biological functions.

Main Results:

• Deep GONet demonstrated high performance in discriminating cancer and non-cancer samples.
• The model proved to be easily interpretable.
• Experimental validation on cancer diagnosis datasets confirmed model efficacy.

Conclusions:

• Deep GONet provides explanations for predictions by identifying key neurons and their biological functions.
• The model enhances understanding for biologists and physicians in precision medicine.
• Achieved both accuracy and interpretability in deep learning for medical diagnosis.