Unsupervised machine-learning identifies clinically distinct subtypes of ALS that reflect different genetic architectures and biological mechanisms

BackgroundAmyotrophic lateral sclerosis (ALS) is a fatal neurodegenerative disease characterised by a highly variable clinical presentation and multifaceted genetic and biological bases that translate into great patient heterogeneity. The identification of homogeneous subgroups of patients in terms of both clinical presentation and biological causes, could favour the development of effective treatments, healthcare, and clinical trials. We aimed to identify and characterise homogenous clinical subgroups of ALS, examining whether they represent underlying biological trends. MethodsLatent class clustering analysis, an unsupervised machine-learning method, was used to identify homogenous subpopulations in 6,523 people with ALS from Project MinE, using widely collected ALS-related clinical variables. The clusters were validated using 7,829 independent patients from STRENGTH. We tested whether the identified subgroups were associated with biological trends in genetic variation across genes previously linked to ALS, polygenic risk scores of ALS and related neuropsychiatric traits, and in gene expression data from post-mortem motor cortex samples. ResultsWe identified five ALS subgroups based on patterns in clinical data which were general across international datasets. Distinct genetic trends were observed for rare variants in the SOD1 and C9orf72 genes, and across genes implicated in biological processes relevant to ALS. Polygenic risk scores of ALS, schizophrenia and Parkinsons disease were also higher in distinct clusters with respect to controls. Gene expression analysis identified different altered biological processes across clusters reflecting the genetic differences. We developed a machine learning classifier based on our model to assign subgroup membership using clinical data available at first visit, and made it available on a public webserver at http://latentclusterals.er.kcl.ac.uk. ConclusionALS subgroups characterised by highly distinct clinical presentations were discovered and validated in two large independent international datasets. Such groups were also characterised by different underlying genetic architectures and biology. Our results showed that data-driven patient stratification into more clinically and biologically homogeneous subtypes of ALS is possible and could help develop more effective and targeted approaches to the biomedical and clinical study of ALS.