PGVDA: a pathway-aggregated genetic dosage framework for interpretable disease classification using machine learning.

Area of Science:

• Genetics
• Neurology
• Computational Biology

Background:

• Neuromuscular junction disorders (NMDs) and inflammatory polyneuropathies (IPNs) are distinct but share biological pathways.
• Limited genetic comparisons exist, hindering understanding of underlying differences.
• Machine learning (ML) offers potential for distinguishing these diseases based on variant patterns.

Purpose of the Study:

• To develop an interpretable ML framework, Pathway-based Genetic Variant Dosage Average (PGVDA), for classifying NMDs and IPNs.
• To identify key genes and pathways that differentiate these two neuromuscular diseases.
• To leverage genetic variant data for improved disease classification and understanding.

Main Methods:

• Utilized nonsynonymous variants from 667 UK Biobank participants.
• Employed logistic regression for variant association and pathway enrichment analysis.
• Developed PGVDA by averaging log odds ratios of variant dosages within pathways.
• Applied dimensionality reduction and leave-one-out cross-validation for ML model evaluation.
• Interpreted results using SHAP values for pathway- and variant-level insights.

Main Results:

• The PGVDA-based ML framework accurately classified NMDs and IPNs.
• Identified five key PGVDAs and 10 genes within specific pathways that differentiate the diseases.
• The logistic regression model demonstrated the best performance.
• Pathway-level variant aggregation proved effective for classification.

Conclusions:

• Pathway-level genetic variant analysis using PGVDA provides an accurate and interpretable method for distinguishing NMDs and IPNs.
• This approach highlights specific genes and pathways crucial for differentiating these neuromuscular conditions.
• Further external validation is recommended to confirm generalizability.