Non-linear gene sets for digital biomarkers of amyotrophic lateral sclerosis

Abstract

Amyotrophic lateral sclerosis (ALS) is a fatal neurodegenerative disease caused by the loss of motor neurons. Accurate and accessible blood-based diagnosis of neurodegenerative diseases, including ALS, is becoming increasingly critical. It would be of a major clinical advantage to be able to distinguish ALS based on gene expression profiles of blood cells, but this still remains to be established. We analyzed existing data of peripheral blood mononuclear cells (PBMCs) of ALS patients by machine learning, the Maximum Mean Discrepancy (MMD), minimizing the issue of multi-collinearity shown in a multiple regression model, and identified a non-linear combination of gene sets to classify healthy controls and ALS, since non-linear approaches can capture intricate gene-gene interactions and threshold effects, which are often critical for accurately distinguishing disease from healthy states. In the gene expression profiles of PBMCs, a combination of expression levels in 3 genes, PRKAR1A, QPCT, and TMEM71, enabled us to classify ALS with an area under the curve (AUC) accuracy of 0.83 from the public database, which were then confirmed by laboratory blood samples with an AUC accuracy of 0.85. Furthermore, we found the expression levels of PRKAR1A, QPCT, and TMEM71 in motor neurons derived from induced pluripotent stem cells (iPSCs), the cell type at the core of the pathology, classified ALS with an accuracy of AUC 0.79. This approach of discriminating ALS with non-linear gene combinations may be useful for identifying ALS molecular biomarkers for blood-based diagnosis as well as for arriving at a completely new perspective of ALS pathogenesis.