CLinNET: An Interpretable and Uncertainty-Aware Deep Learning Framework for Multi-Modal Clinical Genomics.

Area of Science:

• Genomics
• Artificial Intelligence
• Neuroscience

Background:

• Identifying molecular drivers for neurocognitive disorders (NDs) is challenging due to variants of uncertain significance (VUS) and diagnostic platform limitations.
• Existing artificial intelligence (AI) models often lack interpretability and fail to address uncertainty, hindering clinical application.

Purpose of the Study:

• To introduce CLinNET, a multi-modal deep neural network designed to enhance gene curation and VUS interpretation for NDs.
• To improve the accuracy and biological relevance of gene prediction using a biologically informed architecture and confidence-based uncertainty quantification.

Main Methods:

• CLinNET utilizes a dual-branch deep neural network integrating sequencing data, gene expression, biological pathways, and gene ontology (GO).
• Employs layer-wise SHapley Additive exPlanations (SHAP) for robust interpretability and sparse networks enriched with pathway/GO data.
• Prioritizes tissue-expressed genes to enhance prediction accuracy and biological relevance.

Main Results:

• CLinNET achieved an F1-score of 76.4% and accuracy of 77.2% on ND datasets, outperforming existing methods.
• Uncertainty filtering improved precision to 87% while maintaining 73% high-confidence predictions.
• Identified significantly more ND-associated genes than random permutations, with 78 genes in the top decile linked to NDs (p-value = 1.2e-11).

Conclusions:

• CLinNET is a robust and interpretable AI tool for gene curation and VUS interpretation in neurocognitive disorders.
• The model demonstrates significant potential for identifying novel diagnostic genes and advancing individualized medicine.
• CLinNET's adaptability is underscored by its successful validation in prostate cancer datasets.