BACKGROUND

Amyotrophic lateral sclerosis (ALS) is a rare progressive neurodegenerative disease that affects upper and lower motor neurons. As the molecular basis of the disease is still elusive, the development of high-throughput sequencing technologies, combined with data mining techniques and machine learning methods, could provide remarkable results in identifying pathogenetic mechanisms. High dimensionality is a major problem when applying machine learning techniques in biomedical data analysis, since a huge number of features is available for a limited number of samples. The aim of this study was to develop a methodology for training interpretable machine learning models in the classification of ALS and ALS-subtypes samples, using gene expression datasets.

METHODS

We performed dimensionality reduction in gene expression data using a semi-automated preprocessing systematic gene selection procedure using Statistically Equivalent Signature (SES), a causality-based feature selection algorithm, followed by Boosted Regression Trees (XGBoost) and Random Forest to train the machine learning classifiers. The SHapley Additive exPlanations (SHAP values) were used for interpretation of the machine learning classifiers. The methodology was developed and tested using two distinct publicly available ALS RNA-seq datasets. We evaluated the performance of SES as a dimensionality reduction method against: (a) Least Absolute Shrinkage and Selection Operator (LASSO), and (b) Local Outlier Factor (LOF).

RESULTS

The proposed methodology achieved 85.18% accuracy for the classification of cerebellum or frontal cortex samples as C9orf72-related familial ALS, sporadic ALS or healthy samples. Importantly, the genes identified as the most determinative have also been reported as disease-associated in ALS literature. When tested in the evaluation dataset, the methodology achieved 88.89% accuracy for the classification of sporadic ALS motor neuron samples. When LASSO was used as feature selection method instead of SES, the accuracy of the machine learning classifiers ranged from 74.07 to 96.30%, depending on tissue assessed, while LOF underperformed significantly (77.78% accuracy for the classification of pooled cerebellum and frontal cortex samples).

CONCLUSIONS

Using SES, we addressed the challenge of high dimensionality in gene expression data analysis, and we trained accurate machine learning ALS classifiers, specific for the gene expression patterns of different disease subtypes and tissue samples, while identifying disease-associated genes.