The pathway interaction method, originally proposed by Kelder et al. (31), analyzes biological networks, constructed by previously identified relationships between genes and examines their collective behaviors. In this paper, we demonstrated that this method could detect the disease expression subnetwork (in this study, clinical expression of autism spectrum disorder and/or psychosis) that was likely to have been affected by the deletion in the 22q11.2 region. The summary of our workflow is as follows. First, we created two large directed graphs, one for patients with autism spectrum disorder/psychosis and the other without, consisting of protein-protein interactions and the WikiPathways repository. Next, using two transcriptomics datasets of 22q11DS, we identified sets of pathways and genes that closely ‘interact’ with the 22q11DS pathway (22q11.2 copy number variation syndrome, WP4657). We then refined the interaction networks and identified a gene subnetwork that was exclusively associated with 22q11DS patients experiencing autism spectrum disorder/psychosis. The results were compared with that of pathway ORA.
In the subnetwork shown in Fig. 3B, a notable relationship is a propagation of a signal starting from CRKL, a commonly deleted gene in the 22q11.2 region, leading to GRB2, and then PDGFRB, which is a major hub within the subnetwork. Previous studies suggest that CKRL, GRB2, IL2, ITGB3, and FYN are involved with natural killer (NK) cell activation (38–42). Indeed, previous research observed decreased NK cell function among 22q11DS patients (43). In addition, there have been several studies supporting that reduced NK cell cytotoxicity is associated with neurodevelopmental disorders, especially autism spectrum disorder (44–46). Consequently, we suspect that NK cell function might be one of the important biological processes disrupted among the 22q11DS patients who experienced neuropsychiatric disorders. The upregulation of PDGFRB might give another piece of evidence for this mechanism. PDGFR-DD binds to NKp44 to activate NK cells, and also interacts with PDGFRB, to stimulate further downstream pathways (47–49). Therefore, the upregulation of PDGFRB might be the result of increased PDGFR-DD activity, via a possible feedback loop, to compensate for the reduced NK cell functions initially caused by the silencing of CRKL.
Besides the association with NK cell functions via PDGFRB, a number of studies reported that the downregulation of the PI3K/Akt pathway is associated with autism spectrum disorder and schizophrenia (50–53). PIK3CA and PIK3CB, two class I subunits of PI3K, as well as AKT1 are thus likely to have influenced the phenotypic expression of autism spectrum disorder and/or psychosis among the 22q11DS patients. Since GRB2 plays an important role in PI3K activation pathways (54, 55), GRAP2 (GRB2 related adaptor protein 2) can be considered to take part in the same pathway. There have been a few studies which suggested the association between GRB2/GRAP2 and schizophrenia (56–58), but we suspect GRAP2 might be actively involved in integrating and transmitting the effects of gene deletion leading to the expression of neuropsychiatric diseases. The role of other genes, HLA-A, HRAS, and PAK2, whose connection was not clearly demonstrated by the disease subnetwork, may be illuminated in conjunction with GRAP2 or its close interaction partners.
We can summarize our findings as follows. The differentiating feature of the Psychiatric Group over the Nonpsychiatric Group, i.e. the molecular signature of the 22q11DS patients who have experienced autism spectrum disorder and/or psychosis, compared to those who have not experienced such diseases, implies the involvement of NK cell functions and PI3K/Akt signalling pathway, as well as several genes such as CRKL, PDGFRB, and AKT1.
Compared to the pathway interaction method demonstrated here, pathway ORA detects individual genes that are significantly differentially expressed and are contained in previously defined pathways. Among the identified significantly expressed neuropsychiatric genes, current literature suggests that PRF1, GZMB, ADORA2A, ADRB2, and IL1B play important roles regarding the NK cell function (59–63). On the other hand, PRDX6, HSPA1B, and FKBP5 are known as biomarkers involved in regulation of the PI3K/Akt signalling pathway (64–66). Therefore, we can conclude, at the level of biological processes both the pathway interaction and the ORA methods imply certain biological pathways that are commonly involved, which may warrant closer look at such mechanisms in future study. Considering the most significantly differentially expressed genes were a bit far from the deleted genes of the 22q11.2 region, we suspect some negative feedback mechanisms that decrease the expression level of the genes that directly interact with both the deleted genes and the most significant genes. Still, we found that three of the genes (GZMB, PRF1, IL1B) from pathway ORA had direct protein-protein connections with the genes in Fig. 3B. Using this, we further extended our network, which is displayed in Fig. 5.
It would be beneficial to briefly compare the two methods we used. Considering our initial goal was to identify molecular interactions associated with the expression of autism spectrum disorder/psychosis among the 22q11DS patients, the pathway method seemed to offer a convincing way to identify a molecular subnetwork of the disease. One significant advantage of our method over the conventional pathway ORA is that the pathway interaction method can collectively detect genes of interest, where the connections between such gene sets indicate plausible biological relationships powered by prior knowledge instead of “guilt by association”. For example, GRB2, IL2, PIK3CA, and PIK3CB were not differentially expressed enough under the individual gene significance criteria, thus would not have been identified had we applied only pathway ORA, even though they have direct protein-protein interactions with other significant genes. On the other hand, pathway ORA seems to be better at identifying individual genes that are strongly expressed, regardless of the expression level of nearby genes. Another advantage of ORA is that gene significance criteria can be easily adjusted and modified, and can easily include any combination of variables, while the pathway interaction method depends only on gene weights, which first needs appropriate transformations. Our understanding is that the two methods complement each other and together reveal a more complete picture of the molecular interaction of 22q11DS, as shown in Fig. 5. Therefore, future research projects would benefit from exploiting both methods. Table 2 summarizes the characteristics of the two methods.
For our study, we have applied several adjustments to the original methods by Kelder et al. (31). First, the “interaction” between all possible pairs of two pathways was redefined as a source-target relationship, where the source was fixed as the 22q11DS pathway, and the focus was to choose the appropriate target pathways that interact with the source, rather than describing interaction signatures between all pairs. Next, the path threshold was initially set at 0.9 when selecting significant pathways, but then flexible threshold criteria were applied to collect actual paths, backed by our sensitivity analysis on threshold-score relationships. Flexible criteria were used to detect additional pathways that involve genes that did not have a large change, but still contributed to the paths by interacting with other significant genes. Calculating empirical p-values was also simplified; 1,000 samples rather than 10,000 samples, and each sample matches the number of genes having weight less than or equal to 0.8 and those having weight greater than 0.8. In addition, our focus was to offer a plausible interpretation of molecular mechanisms starting from an actually deleted gene on 22q11.2 (in our dataset, CRKL), by comparing results obtained from two parallel analyses of the Psychiatric and the Nonpsychiatric Groups. As for the final step, we compared our results with that of pathway ORA and demonstrated that the two methods could complement each other to obtain a more complete picture of the molecular perturbation from the gene deletion.
The value of pathways in this context is that pathways offer relatively safe start and end points. Pathways (at least those in the WikiPathways repository) are created and supported by previous publications, so that analysis revolves around plausibly grouped genes, making the results easy to interpret. Computationally speaking, pathway nodes are excellent source and target nodes, since it is much more efficient to find significant paths between pathways than to examine paths between all candidate gene pairs. In our study, we were able to use an already existing pathway (22q11DS pathway) as the source node, as it incorporated the deleted genes and their close interaction partners that were previously known. However, we would like to emphasize the flexibility of our method as well, since even if no published pathway existed, one would be able to directly create an appropriate gene-pathway relationship and use it as a starting point. The pathway interaction method can also include transcription-factor targets or other types of interaction databases, in addition to or instead of protein-protein interactions to further extend the base network.
As a final note, we would like to briefly discuss the data used. The dataset was obtained from whole blood rather than from brain cells, implying the tissue-specific gene expressions might have affected the key genes and pathways identified; for example, we found pathways involving innate immune systems and cell metabolism to be significantly over-expressed in both data groups. Considering our goal was to study the phenotypic expression of autism spectrum disorder and psychosis, actual brain cell expression data may provide clearer pictures on the dynamics between the deleted genes and related biological processes, for example, whether the perturbations in the cellular processes we observed was due to peripheral pathology or a secondary effect, as result of brain dysfunction. However, our dataset was still useful to examine the neuropsychiatric genes that might reveal crucial interactions via the immune-related responses. The research community also has discovered that neuropsychiatric functions and immune-related pathways tend to be closely associated (67, 68). We hope future experimental work would confirm and verify our findings.