The genetic basis of sporadic amyotrophic lateral sclerosis (ALS) is not well understood. Using large genome-wide association studies and validated tools to quantify genetic overlap, we systematically identified single nucleotide polymorphisms (SNPs) associated with ALS conditional on genetic data from 65 different traits and diseases from >3 million people. We found strong genetic enrichment between ALS and a number of disparate traits including frontotemporal dementia, coronary artery disease, C-reactive protein, celiac disease and memory function. Beyond C9ORF72, we detected novel genetic signal within numerous loci including GIPC1, ELMO1 and COL16A and confirmed previously reported variants, such as ATXN2, KIF5A, UNC13A and MOBP. We found that ALS variants form a small-world co-expression network characterized by highly inter-connected 'hub' genes. This network clustered into smaller sub-networks, each associated with a unique function. Altered gene expression of several subnetworks and hubs was over-represented in neuropathological samples from ALS patients and SOD1 G93A mice. Our collective findings indicate that the genetic architecture of ALS can be partitioned into distinct components where some genes are highly important for developing disease. These findings have implications for stratification and enrichment strategies for ALS clinical trials.
Properties of the ALS biological networksWe assessed the network structure of the physical protein-protein interaction network, coexpression network, and genetic interactions network. Specifically, we asked whether some genes play a more influential role than others. Most complex networks have a small-world property characterized by relatively short paths between any pair of nodes (genes) 15 . In smallworld networks, perturbing any given node is thought to also perturb neighboring nodes and the entire network in general. Quantitatively, a network is considered small-world if its "smallworldness" index is higher than one (a stricter rule is small-worldness >=3) 16 . Further, the clustering coefficient for the target small-world network should be higher than the clustering coefficient of a comparable random network. Also, the average shortest path length of the target network should be similar or higher (but not substantially higher) than a comparable random network.7 First, we evaluated the degree to which each network assumed a small-work network structure. The co-expression interaction network consisted of 95 nodes and 132 edges, had a small-world index of 6.06, a diameter of 13, and average shortest path length of 5.12. The clustering coefficient was 0.102, which is higher than the clustering coefficient of a random graph with the same number of indices (0.031). The physical protein-protein interaction network consisted of 85 nodes and 41 edges, had a small-world index of 4.43, a diameter of 5, and average shortest path length of 82.03. The clustering coefficient was 0.326, which is also higher than the clustering coefficient of a random network with the same ...