Machine learning compensates fold-change method and highlights oxidative phosphorylation in the brain transcriptome of Alzheimer’s disease

Abstract

Abstract Alzheimer’s disease (AD) is a neurodegenerative disorder causing 70% of dementia cases. However, the mechanism of disease development is still elusive. Despite the availability of a wide range of biological data, a comprehensive understanding of AD's mechanism from machine learning (ML) is so far unrealized, majorly due to the lack of needed data density. To harness the AD mechanism's knowledge from the expression profiles of postmortem prefrontal cortex samples of 310 AD and 157 controls, we used seven predictive operators or combinations of RapidMiner Studio operators to establish predictive models from the input matrix and to assign a weight to each attribute. Besides, conventional fold-change methods were also applied as controls. The identified genes were further submitted to enrichment analysis for KEGG pathways. The average accuracy of ML models ranges from 86.30% to 91.22%. The overlap ratio of the identified genes between ML and conventional methods ranges from 19.7% to 21.3%. ML exclusively identified oxidative phosphorylation genes in the AD pathway. Our results highlighted the deficiency of oxidative phosphorylation in AD and suggest that ML should be considered as complementary to the conventional fold-change methods in transcriptome studies.