Gene targeting in amyotrophic lateral sclerosis using causality-based feature selection and machine learning.

Area of Science:

• Genomics
• Computational Biology
• Neuroscience

Background:

• Amyotrophic lateral sclerosis (ALS) is a rare neurodegenerative disease affecting motor neurons, with an elusive molecular basis.
• High-throughput sequencing and machine learning offer potential for identifying ALS pathogenetic mechanisms.
• High dimensionality of gene expression data presents a challenge for machine learning in biomedical analysis.

Purpose of the Study:

• To develop a methodology for training interpretable machine learning models for ALS and its subtypes classification.
• To utilize gene expression datasets for identifying disease-specific patterns.
• To address the challenge of high dimensionality in analyzing biomedical data.

Main Methods:

• Applied a systematic gene selection procedure using Statistically Equivalent Signature (SES) for dimensionality reduction.
• Trained machine learning classifiers (XGBoost, Random Forest) on ALS RNA-seq datasets.
• Used SHapley Additive exPlanations (SHAP values) for model interpretability and compared SES with LASSO and LOF.

Main Results:

• Achieved 85.18% accuracy in classifying C9orf72-related familial ALS, sporadic ALS, and healthy samples from brain tissue.
• Identified key genes with high determinative power that are documented as disease-associated in ALS literature.
• Attained 88.89% accuracy in classifying sporadic ALS motor neuron samples, outperforming other feature selection methods like LASSO and LOF in certain contexts.

Conclusions:

• The SES methodology effectively addresses high dimensionality in gene expression data analysis for ALS.
• Developed accurate, interpretable machine learning classifiers specific to ALS subtypes and tissue samples.
• Successfully identified disease-associated genes, contributing to understanding ALS pathogenetic mechanisms.