Towards a diagnostic test for sporadic ALS utilising deep learning and SNP microarrays

A variety of common and rare genetic factors have been implicated in the development of amyotrophic lateral sclerosis (ALS), and the evidence is that a genetic component is present in most affected individuals. However, our current understanding of ALS genetics causally explains only a small proportion of sporadic cases which represent over 90% of all people with ALS. This limits the utility of genetic testing in screening, diagnosis and management to the 15-20% of people with ALS who carry a known pathogenic variant. Capsule Networks (CapsNets) constitute a deep learning method that has demonstrated strong performance in using genotyping data to predict individuals at risk for ALS. However, their use is constrained by a lack of generalised, flexible, and validated implementations across comprehensive datasets that account for the technical, biological, and clinical heterogeneity found in real-world disease scenarios. In this study, we build upon this method to address existing limitations, to develop a new model that is validated across diverse ALS populations, can handle discrepancies between genotyping technologies, and is applicable to individual external samples. Using large-scale datasets from over 47,000 individuals from 13 countries, genotyped with nine different genotyping platforms, our model achieved high precision and sensitivity in distinguishing between individuals with ALS and non-affected controls. Moreover, in simulations of population screening for ALS, its performance was comparable to that of conventional genetic screening for known ALS gene mutations, such as FUS and C9orf72. Our results demonstrate that this flexible and validated method could support the development of a genetic screening test for identifying individuals at risk and expediting ALS diagnosis. This would be applicable to all individuals, regardless of their family history or presence of known ALS mutations.