ALS is a devastating neurodegenerative disease that kills the body's motor neurons, resulting in progressive paralysis and a survival of 2-4 years. However, some motor neurons degenerate, like fast-firing motor neurons, while others do not, like slow-firing motor neurons. Can we distinguish the neurons that are resistant or vulnerable to ALS using machine learning?
Based on 1200 fast-firing and slow-firing motor neurons, I trained a machine learning model (support vector machine–singular value decomposition) to classify new neuron populations as resistant or vulnerable to ALS. The model produced an average accuracy of 88.30% and sensitivities and specificities of over 95%. Using this model, I assessed a population of 8973 visceral and beta motor neurons, resolving new distinct populations of resistant and vulnerable cells. Differential expression between these two populations yielded 21 genes expressed at a p-value < e-100, and these genes were thresholded to 12 when a ±0.5 log2 fold change was applied. Among these 12 genes, 8 have already been connected to ALS in previous studies, affirming the potency of my approach and ML model. One of these 8 is Pard3b, the loss of which directly causes a form of ALS. Furthermore, genetic ablation of Sst is further corroborated as a therapeutic approach. The 21 genes identified are prime targets for future therapeutic methods in ALS, and the machine learning framework can be extended to other diseases too.