Revealing Alzheimer’s disease genes spectrum in the whole-genome by machine learning

There are many different supervised machine learning approaches. We implement a prediction algorithm by SVM (Support Vector Machines) using the e1071 package of R, published by David Meyer. SVM was chosen here, as it can allow the assignment of different weights to different classes. The kernel type used in training and predicting is radial. In addition, a 5-fold cross validation was performed, in order to assess the quality of the model.

After initially collecting 335 AD-associated genes [Additional file 2], we then classified these genes into four categories, based on the strength of supporting evidence (the number of positive evidence of family-based studies and case-control studies). These were labeled C1-AD: probable pathogenic genes. C2-AD: high confidence genes. C3-AD: related genes, and C4-AD: possibly associated genes. Three hundred thirty-five AD-non associated genes as C5 group [Additional file 3]. Our goal is, within the scope of whole genome analysis, to find a stable relationship between a pair of genes, so as to find the AD candidate genes closely linked to genes known to be associated to AD.

According to the pathogenic mechanism of AD, we recruited the brain-specific gene network data from GIANT. Using these five gene groups, along with their evidence classification, and integrating brain-specific gene network data, we trained an evidence-weighted, network-based classifier, using an SVM approach. We randomly subdivided the total genes into two parts (10-fold cross-validation), which were used as training dataset and testing dataset, respectively. The classifier first identified network patterns (The relationship between any gene and 670 genes (335 AD associated genes +335 AD non-associated genes) as the features of SVM model). Then, we used the testing dataset to test the accuracy of the classifier, and divided them into initial two categories (AD related and AD non-related). We found that the average correct rate was 80.59% and the highest correct rate reached 84.56% with the ROC curve (receiver operating characteristic curve) as shown in Fig. 2. The ROC curve was constructed by “ROCR” package in R with a threshold of 0.561. Finally, we applied this classifier integrating the classification of known AD-associated genes to identify new AD candidate genes which interact closely with the known AD associated genes in the brain-specific network and divided those candidate genes into different groups by comparing the probability of each group and choosing the largest one. This method resulted in a comprehensive, genome-wide, ranked list of AD candidate genes.

To screen a more credible candidate gene list, we first annotated all AD associated genes (known AD-associated genes and AD predicted candidates) using the Gene Ontology resource (GO: http://​www.​geneontology.​org/). We then performed a GO functional enrichment analysis of 335 known AD-associated genes and selected the top 10 GO items (Table 1) with P-values (adjusted by false discovery rate [19]) below a max value of 6.87e-11. Finally, we chose the AD predicted genes annotated on those GO items to further obtain high-confidence candidate genes. After this filter, we arrived at a total number of 832 AD predicted genes [Additional file 4].

In previous research, one prediction method has been based upon similarities in sequence-based features in disease genes [9‐11]. Herein, a set of features was chosen from the genes, in order to assess our predictions.

The feature set (described in Table 2) reflects the structure and content of each gene examined. Table 3 lists the differences among the features present among the different sets of genes. Using the Mann-Whitney U test, we discovered highly significant differences in gene length, exon count, transcript count, 3’UTR length and 5’UTR length between the gene sequences of the AD-associated gene set and the control set of genes. Besides, there were also highly significant differences in transmembrane domain, signal domain and paralog calculated using the chi squared test. We reached the same conclusion when comparing the AD candidate dataset and control dataset.

Meanwhile, there was no statistically significant difference between the sequence patterns of AD-associated genes and AD candidate genes (as described in Table 4). The size (in bp) of the genes in both the AD-associated dataset and the AD candidate dataset is significantly larger than among controls, and this is similar to previous findings on AD association [9], which also report generally that genes associated with disease tend to be larger than those involved in normal phenotypes. In much the same way as larger total gene length is associated with AD, the length of 3’ UTR and 5’UTR, transcript count and the number of exons per gene, is in the AD-associated dataset and AD candidate dataset all larger. Genes in the known AD-associated dataset and the AD candidate dataset had a median count of 10 exons and 8 transcripts, while genes in the control dataset had a median count of 5 exons and 3 transcripts. We also found that there were significant differences in the length of 3’ UTR and 5’UTR in both AD-associated genes and AD candidate genes, which both had a larger median length of 3’UTR and 5’UTR, while genes in the control set had a smaller median length (described in Table 3).

Furtherly, we added genes annotated to non-mental-health diseases for comparison [14]. Differences calculated as p-value by using the Mann-Whitney U test or chi squared test between any two sets of four gene sets are shown in Table 4. On the basis of above results, we guessed that the differences between the sequence patterns of AD-associated genes and non-mental-health genes (as Group 1) were greater than that of AD-associated genes and AD candidate genes (as Group 2), but less than that of AD-associated genes and control genes (as Group 3). That is to say, our guess was that the p-values of Group 1 were smaller than that of Group 2, but larger than that of Group 3. Finally, most of the results were up to our expection, including length of gene and 5’ UTR, transcript count, transmembrane domain, signal domain and paralog.

Graphs presenting the distributions of each feature in the four datasets are shown in Fig. 3. It’s clear that the peaks of feature values of the known AD-associated dataset and AD candidate sets shift rightward, when compared to those of the control set. To our knowledge, the number of exons is also correlated to total gene length. However, the differences in 5’ UTR and 3’ UTR length have not been explained in terms of correlations to other feature differences.

Since the genes of the known AD-associated dataset were selected from the literature published before 2015, we also checked the predicted AD candidate gene set by scanning papers published post-2014 to verify the accuracy of our prediction (described in Table 5). As a result, we found that the AD candidate genes which were also reported in the post-2014 research could be roughly divided into several types. First, our gene candidates were identified as being associated with AD genes [20‐22]. For example, DAB1, a novel candidate liability/protective gene, was identified by functional enrichment analysis of 3 AD Genome-Wide Association Studies (GWAS). Second, the genes were simply associated with risk of AD [23‐25]. ANXA1 and CDC25C were identified as potentially contributing to AD susceptibility, by applying ICSNPathway analysis to the AD GWAS meta-analysis data. Third, abnormal changes in gene modification level or expression level were shown to exist in AD cases, when compared to the controls [23, 25]. For instance, DNA methylation levels within the CRTC1 gene were decreased in human hippocampus tissue affected by AD, suggesting that CRTC1 methylation plays an important role in AD pathophysiology.