Complex traits such as height and intelligence are always a fascinating research topic for geneticists. Two central questions on complex traits are how many genes are exactly behind a complex trait and whether they equally contribute to such a trait. In the era of genome project and next-generation sequencing, we seem to be approaching to the answers for these two questions. Nevertheless, the big data of this era only bring us confounding results instead of definitive answers. The study of complex traits with common SNPs shows that all autosomal SNPs contribute to the 45 % variance in height and the variance explained by each chromosome is proportional to its length^[1]. In another word, each gene in an individual’s genome could probably make its tiny contribution to his or her height. For complex diseases, such answer is obviously far beyond satisfaction, because we are trying to seek possible treatment, not probable statement.

In a recently published seminal paper, Boyle et al. proposed a new omnigenic model for understanding complex diseases^[2]. The omnigenic model proposes that all genes expressed in diseaserelevant cells have the influence on the functions of core disease-related genes and thus the genetic propensity for complex diseases can lay outside core gene pathways. The hypothesis behind the omnigenic model is that all genes expressed in a cell are tightly interconnected by an omnibus gene regulatory network and thus each of them can affect the development of disease. We found this model both troubling and enlightening and tested it with Alzheimer’s disease and Parkinson’s disease microarray data. Both of them are neurodegenerative diseases, whose causes are still poorly understood^{[3, 4]}.

First, we downloaded the expression data of Alzheimer’s disease and Parkinson’s disease from GEO database and then identified the differentially expressed genes (DEGs) between cases and controls in each dataset. The DEGs are classified into two categories --- down-regulated genes and up-regulated genes (Table 1). Our result shows the number of down-regulated genes and upregulated genes varies from dataset to dataset, which is ranging from a few of thousands to a couple of hundreds. Accordingly, the gene expression spectra are widely different among different disease datasets.

Second, we tried to find the common DEGs in down-regulated gene category and upregulated gene category for Alzheimer’s disease and Parkinson’s disease, respectively. Among four Alzheimer’s disease datasets, just dozens of common DEGs were found in both down-regulated genes and up-regulated genes (Fig. 1a and 1b). Among five Parkinson’s disease datasets, only three common DEGs were found in down-regulated genes whereas no common DEG was found in upregulated genes. Certainly, there exists a possibility that these disease cases share some prominent common core gene pathways, although they share very few common genes. To for or against such possibility, we performed the gene-GO term enrichment analyses for 1300 down-regulated genes and 1651 up-regulated genes in Alzheimer’s disease and 2042 down-regulated genes and 2469 upregulated genes in Parkinson’s disease. The geneGO term enrichment results are shown in Fig. 2. Although a couple of thousands of genes were used for each gene-GO term enrichment analysis, the most significant biological processes only contain a dozens of genes in each GO analysis. For the down-regulated genes in Alzheimer’s disease and Parkinson’s disease, their significant biological processes mainly involve in metabolic processes but not conspicuous cell degenerative pathways such as apoptotic process as we expected. Furthermore, the down-regulation of cell metabolism is a result of cell death but not a cause for it^[5]. Apparently, there are very few shared genes among different Alzheimer’s disease datasets or Parkinson’s disease datasets and our GO analyses failed to detect any common initiation for the neuron degeneration process in Alzheimer’s disease and Parkinson’s disease.

Third, we checked the expression levels of DEGs in normal organs, since the omnigenic model states that all genes expressed in disease-relevant cells can have the influence on the disease development process. And vice versa, the disease must have the impact on their expression level in diseaserelevant organs. Fig. 3 shows the expression levels of down-regulated genes and up-regulated genes in four Alzheimer’s disease dataset among six normal human organs. The down-regulated genes in four Alzheimer’s disease datasets exhibit a statistically higher expression level in brain and cerebellum than heart, liver, kidney, liver and testis while the up-regulated genes don’t (Fig. 3a and 3b). In five Parkinson’s disease datasets, the down-regulated genes show a similarly higher expression pattern in brain and cerebellum, although such pattern is less obvious in dataset GSE20295 (Fig. 4a). Moreover, the up-regulated genes in five Parkinson’s disease datasets show a more diversified expression pattern (Fig. 4b), which suggests that the pathogenic process might be more complex in Parkinson’s disease than in Alzheimer’s disease. Fig. 3 and 4 actually show that these down-regulated genes are specifically expressed in normal central nervous system (CNS), because they have a higher expression level in brain and/or cerebellum compared to the other four organs. That the CNS-specifically-expressed genes are down-regulated in neurodegenerative diseases is consistent with the assumption of the omnigenic model.

Finally, we examined the network properties of the down-regulated genes in Alzheimer’s disease and Parkinson’s disease using both protein-proteininteraction information and gene co-expression information. The key point in omnigenic model is that all genes expressed in disease-relevant cells are sufficiently interconnected by gene regulatory networks. Because our result shows that the downregulated genes in neurodegenerative diseases are also CNS-specifically-expressed genes, we infer that the down-regulated genes in neurodegenerative diseases must constitute a biological network which can be characterized by the power-law degree distribution and hierarchical structure^[6]. Fig. 5a shows that the degree distribution of the downregulated genes in Alzheimer’s disease to the probability P(k) fits the power law in protein-proteininteraction networks (the power-law fit is shown as a red line, R2 = 0.94, P-value < 2×10−16). Fig. 5b shows that the scaling of the clustering coefficient of the down-regulated genes in Alzheimer’s disease follows C(K)~K−1 in protein-protein-interaction networks (which is shown in a red straight line with negative slope, R2 = 0.37, P-values < 2×10−16). Our network property analyses also show that the downregulated genes in Parkinson’s disease have the same network properties as the ones in Alzheimer’s disease in protein-protein-interaction network (Fig. 5c and 5d). The network analysis based on gene coexpression information shows a similar result for the down-regulated genes in Alzheimer’s disease and Parkinson’s disease (Fig. 6a, 6b, 6c and 6d). That a gene set’s degree distribution fits the power law indicates that they constitute a scale-free network (there exist a small number of highly connected nodes, known as hubs, within the network). And the scaling of the clustering coefficients of two sets of down-regulated genes show that they have a hierarchical network structure as a protein-protein interaction network (sparsely connected nodes are part of highly clustered areas and different clustered areas are communicated by a few hubs). The hierarchical network structure exhibits the smallworld property of disease-related genes, which is predicted by Boyle et al’s paper^[2]. Thus, the results of our network analyses further support the omnigenic model.

There is a major difference between two network analysis results. The coefficient of determinations (R2) are smaller in gene co-expression network analyses than in protein-protein-interaction ones. We find that gene’s average degree and clustering coefficient is much higher in gene co-expression networks than in protein-protein-interaction networks and the genes with high degree and clustering coefficient in gene co-expression networks also outnumber their counterparts in protein-proteininteraction networks. For example, for the downregulated genes in Alzheimer’s disease, the proteinprotein-interaction network analysis shows that they have the highest degree of 50 while the gene coexpression one shows that they have the highest degree of 250 (Fig. 5a and Fig. 6a), which suggests that the protein-protein-interaction information might be limited. This limitation could lead to bias in our network analyses, although the bias is favorable towards our conclusion. In order to rule out such bias in our study, we performed the network property analysis for all genes with protein-protein interaction information and used them as the background (total 17644 genes). The result shows that all genes with protein-protein interaction information also constitute a genuine biological network (scale-free and highly modular, Supplementary Fig. 1). Consequently, our network analysis in Figure 5 is less likely to be biased since it is based on a small protein-proteininteraction network from a much bigger one. In addition, the large average degree and clustering coefficient from gene co-expression network analysis propose that the down-regulated genes in disease samples are often expressed in a modular fashion in normal brain regions.

In this comment, we used the expression data from Alzheimer’s disease and Parkinson’s disease to verify the omnigenic model in neurodegenerative diseases. We originally felt some qualm about this model, because it provides more challenge for those like us who work in the field of complex diseases. To our surprise, our results fully support it, although we have to say that our evidence is circumstantial. However, acquiring direct evidence for this model would not be an easy task. First, we should know how many genes are exactly expressed in a disease-relevant cell. Then, we have to thoroughly understand the cellular network in order to separate peripheral genes from core genes, since they play different roles in disease pathogenesis. In our view, the most important information that the omnigenic model conveys is that new methods, both experimental and bioinformatic ones, are needed for studying the complex diseases such as Alzheimer’s disease and Parkinson’s disease, because there are more genes than we expected participating in the process of neurodegenerative disease development.