Machine learning predicts distinct biotypes of amyotrophic lateral sclerosis

Abstract

Amyotrophic lateral sclerosis (ALS) is a neurodegenerative disease that is universally fatal and has no cure. Heterogeneity of clinical presentation, disease onset, and proposed pathological mechanisms are key reasons why developing impactful therapies for ALS has been challenging. Here we analyzed data from two postmortem cohorts: one with bulk transcriptomes from 297 ALS patients and a separate cohort of single cell transcriptomes from 23 ALS patients. Using unsupervised machine learning, we found three groups of ALS patients characterized by synaptic dysfunction (34%), neuronal regeneration (47%), and neuronal degeneration (19%). Each of these ALS subtypes had unique patterns of transcriptional dysregulation that could represent novel therapeutic targets. We then developed a supervised machine learning model that was about 80% accurate at predicting ALS subtype based on patient demographic and clinical data. Together, we established three biologically distinct subtypes of ALS that can be predicted by clinical and demographic data.