Discovering Alzheimer’s disease subtypes with imaging and genetic signatures via multi‐view weakly‐supervised deep clustering

Abstract

AbstractBackgroundThe heterogeneity of neurodegenerative diseases, including Alzheimer’s disease (AD), has hampered precision diagnosis and treatment. Machine learning methods enable the identification of genetically‐explained disease subtypes with distinct brain phenotypes.MethodGene‐SGAN is a novel, multi‐view, weakly‐supervised deep clustering method that jointly considers phenotypic and genetic data, thereby deriving subtypes of disease‐related brain changes with genetic associations. We apply the method to derive subtypes related to AD. Analyses used structural MRIs and whole‐genome sequencing data from the Alzheimer’s Disease Neuroimaging Initiative (1,533 participants, including controls and patients with Mild Cognitive Impairment (MCI) and AD).ResultGene‐SGAN identified four distinct subtypes within MCI/AD: A1, normal anatomy; A2, focal medial temporal lobe (MTL) atrophy; A3, severe atrophy in neocortex and MTL; A4, severe cortical but relatively modest MTL atrophy (Figure 1). The four subtypes reveal significant differences in seven known AD‐related genetic variants (Figure 2). The identified subtype‐associated SNPs not only resemble previous findings, but also suggest genetic protective effects contributing to the observed resilience (A1), potential inflammatory mechanisms underlying subtypes (A2&A3), and the effect of SNPs on atypical atrophy patterns of AD (A4). Clinically, A1 and A3 participants showed the best and the worst cognitive performances (Figure 3). A3 participants had the most abnormal CSF Aβ (p<0.001 vs. A1&A4 and p = 0.039 vs. A2), and exhibited the highest white matter hyperintensity (WMH) volumes (p<0.001 vs. A1&A4 and p = 0.017 vs. A2). However, A2 participants showed significantly higher CSF phospho‐tau than A3 participants (p = 0.019). A4 participants were the youngest group (p<0.001 vs. all other groups), suggesting the inclusion of early onset AD participants, who have a higher prevalence of hippocampal‐sparing variants such as posterior cortical atrophy. The four genetically‐associated subtypes also possess significant differences in a large set of plasma/CSF biomarkers, which are related to distinct biological mechanisms contributing to the heterogeneity of AD.ConclusionDeep learning revealed four morphologic/genetic MCI/AD subtypes. The derived subtypes display distinct neuroanatomical patterns and genetic determinants and also differ in many clinically interpretable biomarkers. These subtypes provide potential for drug discovery and repurposing, optimization of clinical trial recruitment, and personalized medicine based on an individual’s genetic profile.