Background
The pathophysiology behind atypical brain development in autism spectrum disorders (ASD) is highly complex. Elegant biological studies are continually unveiling an ever more diverse array of etiological factors and neurodevelopmental processes associated with ASD (e.g., [
1‐
17]). With such diversity, key questions arise as to what are the consistent and robust mechanisms involved and whether such mechanisms point to many independent disrupted pathways or some convergence on a few common pathways affecting large-scale biological systems and/or interactions between such systems [
18,
19]. One way to test this question is to examine pathophysiology at a level above genetics and non-genetic perturbations, such as the transcriptome, and examine whether the diversity of disrupted transcriptomic signals converge onto many independent or interacting systems.
Seminal work examining cortical transcriptome dysregulation in ASD has highlighted the dysregulation of multiple transcriptional programs. These programs include cell cycle/DNA damage, WNT signaling, cortical patterning and differentiation, and immune/inflammation at young ASD ages [
5] and apoptosis, repair and remodeling, synaptic and immune/inflammation functions at older ages [
8,
20,
21]. However, it remains unclear if such pathways are independently dysregulated or whether there is synergy between multiple dysregulated pathways. For example, prior work has shown downregulated synaptic and upregulated immune/inflammation signals in ASD cortical tissue [
8,
20,
21]. Pointing towards the idea that such dysregulated signals may not be independent, strong correlations are found between these dysregulated modules when collapsing data across both groups [
8,
20]. While this observation is important for generally suggesting statistical dependency between modules, more evidence is needed to suggest that such potentially synergistic effects among interacting modules are dysregulated in ASD. To test the hypothesis that transcriptome dysregulation in ASD extends beyond the level of single-dysregulated co-expression modules and involves dysregulation spanning interactions between larger systems-level processes, differences in between-module correlations need to be investigated. Furthermore, tests should also go beyond observations of statistical dependency in co-expression and test whether there are enhanced direct physical interactions between the protein products of such dysregulated modules compared to unaffected modules. An enhancement of direct physical protein interactions among dysregulated versus non-dysregulated modules would further suggest plausibility for synergistic interactions across disparate biological processes conferred by each individual co-expression module.
We examined this topic via re-analysis of two existing datasets that investigated multiple cortical regions. We tested the hypothesis that diverse molecular mechanisms are hierarchically disrupted in the cortical transcriptome of ASD and reflect interacting systems-level pathology rather than multiple independent types of molecular pathology. By “hierarchical disruption,” we refer specifically to evidence of dysregulation of individual co-expression modules as well as higher level disruptions in how such modules interact. Our approach is substantially different from previous work in leveraging the identification of consensus modules that robustly exist across datasets. We also account for known methodological differences intrinsic to the existing studies (e.g., age, gender, brain areas). Our approach also utilizes identical parameters to identify consensus networks across datasets, as these parameters vary across different studies in the literature (e.g., network type—signed versus unsigned, soft-power thresholds, deepSplit parameter for cutting dendrograms). These parameters can affect how modules are identified (i.e., clustering and cutting dendrograms) as well as affect the content (i.e., genes, module eigengene variability) composing different discovered modules, thus making direct cross-study comparisons across the literature somewhat difficult. Such a viewpoint of integrating information from multiple datasets currently does not exist in the literature, and we specifically implement this type of analysis in order to be best positioned to make inferences that are applicable across datasets.
Here, we also directly address “replication” as it pertains to ASD gene expression studies. While some existing studies we have re-analyzed here [
20,
21] constitute “conceptual” replications (i.e., studies that differ in numerous ways yet show roughly similar findings, such as similar gene ontology enrichment results) and are definitely important in their own right, more exact attempts at replication holding constant a variety of analysis issues may also prove insightful. For example, several independent studies may not detect certain dysregulated signals due to analytic or other methodological variance across studies. Repeated detection of such dysregulated signals across multiple studies under more uniform conditions of analysis may likely pull out such prominent signals that otherwise go undetected. Also of importance is how to formally quantify evidence for or against replication. Such formal quantification is missing in the ASD gene expression literature. In this study, we tackle these issues head on by analyzing multiple datasets under uniform analysis conditions (e.g., extracting consensus co-expression networks that exist in multiple datasets) and utilize new Bayesian methods developed directly from replication debates ongoing in other fields like psychology that more formally quantify the strength of evidence for or against replication.
This work also represents the first study to specifically aim at examining hierarchical disruption of the cortical transcriptome in ASD. That is, we go beyond examination of dysregulation at the level of single-gene co-expression modules and also examine whether dysregulation is present in higher level interactions between modules. We provide the first look at the full organization of correlations between gene modules across the transcriptome (i.e., eigengene networks) and examine how such connections manifest differently both at the level of inter-modular connectivity (i.e., connections between specific modules) as well as connectivity relevant to organization of clusters of highly correlated modules (i.e., meta-modules) [
22‐
24]. Localized subtle/specific changes in eigengene network organization or larger global patterns of network reorganization are both plausible predictions regarding how eigengene networks are organized differently in ASD. Both scenarios would lead to the prediction that the composition of meta-modules as well as connectivity within and outside of normative meta-module boundaries would differ in ASD.
Discussion
Here we provide the first detailed characterization of how the ASD cortical transcriptome is hierarchically disorganized both at the level of specific co-expression modules and at higher levels of eigengene network organization (i.e., connectivity between modules and meta-modules). We have pinpointed several novel co-expression signals that show strong evidence for replicable dysregulation across datasets [
20,
21]. Rather than pinpointing a single synaptic or immune-related module, we have identified several dysregulated synaptic and immune modules. These modules are differentiated in terms of cell type/compartment enrichment and/or show different biological process enrichment within the broader class of synaptic and immune-related processes. For example, while both M3 and M9 modules are downregulated in ASD and enriched in similar synaptic processes, their cell type/compartment enrichments differ. M3 is primarily enriched in neuronal markers, whereas M9 is specifically enriched in synaptic and postsynaptic density markers. Synaptic M3 and M9 modules also differentiate in how they interact with other modules (see Fig.
7c, d for example). These results provide an example of how subtle distinctions may be present within the class of downregulated synaptic signals.
We have also identified multiple types of ASD-upregulated immune/inflammation modules that are novel distinctions from past work. Although prior work has implicated interferon signaling, particularly with respect to M2 microglia markers [
20], here, we find evidence for two upregulated interferon signaling modules (M24, M27). These modules differentiate by M1 and M2 microglia activation states, with M27 enriched in M1 microglia markers while M24 is enriched in M2 microglia markers. Between-module connectivity evidence also suggests that these two interferon signaling modules are disrupted in different ways. M27 is abnormally connected to an important ASD-upregulated translation initiation (M25) and ASD-downregulated synaptic module (M9). Given the enrichment in M27 for M1 microglia activation markers, this evidence suggests that cytotoxic M1 microglia processes may be affecting synaptic proteins in ASD. On the other hand, M24 shows intact connectivity between M25 and M9 but aberrant connectivity between other modules (M2, M22). These results suggest that while upregulated interferon signaling can be linked to both M1 and M2 microglia phenotypes, such aberrant processes may have differing impact on ASD brain function and structure.
In addition to the multiple dysregulated interferon signaling modules, we have also uncovered novel evidence for ASD-upregulation of an immune/inflammation module (M12) enriched in the complement system and phagocytosis processes and M1 microglia markers. In conjunction with effects from interferon signaling modules, the addition of the complement system may be of particular importance given the known links between the complement system and synaptic pruning [
35,
36] and remodeling as well as enhancing pro-inflammatory states of microglia activation in ASD [
37‐
39]. Recently, the complement system has been noted as a prominent player in the pathophysiology of schizophrenia, particularly for its role in synaptic pruning [
40]. In the larger context of eigengene networks, it is interesting that all of these important immune/inflammation modules are members of the same meta-module in ASD and that such a meta-module also includes other prominent modules such as the ASD-upregulated M25 translation initiation module. The current data present a role for complement system signaling alongside interferon signaling and other immune processes working together and potentially in concert with other important modules relating to translation and also for their role in various types of microglia activation states.
New modules not highlighted at all by prior work were also identified. Two of these modules (M1 and M25) are heavily enriched in translation initiation and translation elongation-termination processes and are enriched in genes coding for proteins that make up the 40S and 60S ribosomal subunits (
RPL and
RPS genes). Translation has been an important topic in ASD primarily because of work on syndromic forms of autism related to mutations in
FMR1,
TSC1/2, and
PTEN [
6,
41], as well as the important cap-dependent translation gene
EIF4E [
42‐
45]. However, none of this work has specifically implicated ribosomal proteins themselves and no prior work on the cortical transcriptome in ASD has specifically implicated upregulation of translation initiation signals. These modules were dysregulated with respect to connectivity within and outside of normative meta-modular boundaries and showed specific abnormal interactions with each other as well as other ASD-upregulated modules (e.g., M27). Additionally, these translation modules were also a member of a meta-module in ASD that was composed of other upregulated immune/inflammation modules (M12, M24, M27), suggesting that they may play important roles integrating with upregulated immune/inflammation processes in ASD. Thus, not only have we discovered evidence for a novel and important upregulated signal in the ASD cortical transcriptome, but this finding also may have important implications with regards to its potential as a cross-cutting influence on other pathophysiological processes in ASD.
This novel finding of upregulated translation initiation and elongation-termination processes in ASD is important, as it agrees with other work on blood transcriptome markers. Our recent work on blood leukocyte gene expression has also uncovered upregulated translation initiation as a prominent signal in young toddlers with ASD and this signal is present alongside other upregulated immune/inflammation signals, particularly interferon signaling and phagocytosis [
46]. Further bolstering these inferences, a recent mega-analysis of seven different studies in the literature and also found ribosomal translation as one prominent upregulated process in blood [
47]. The presence of these dysregulated and highly connected translation initiation and immune/inflammation signals across brain and blood is potentially important because it may signal a unique opportunity to assay brain-relevant dysregulation in peripheral tissues and in vivo in living patients. This peripheral window into potentially brain-relevant dysfunction that can be assayed in living patients may be particularly important given the recent discovery of a direct linkage between the brain and lymphatic vessels of the immune system [
48]. Investigating this possible peripheral linkage to brain-relevant dysfunction in living patients using in vivo techniques like functional and structural neuroimaging [
49] will be an important next step in understanding whether peripherally dysregulated signals in blood play some role in linking directly to important macro-level neural systems dysfunction in living patients [
50]. We have also recently identified similar upregulation of translation initiation signals, particularly ribosomal proteins, in a rodent model of maternal immune activation [
51], indicating that sources of translation initiation upregulation in ASD may have pathophysiological impact in early fetal development and can be influenced by environmental factors. Furthermore, modeling the upregulated expression of a long non-coding RNA,
MSNP1AS, in neural progenitor cells also leads to differential expression of genes involved in translation, protein synthesis and which are localized to the ribosome [
52].
MSNP1AS was the first genome-wide association study (GWAS) hit in ASD [
53] and is known to be upregulated in expression in ASD cortex [
54]. Influence via this common variant may further indicate influence over this process of translation in the developing ASD brain. Future work using in vivo and in vitro models targeting these novel ribosomal protein genes from the M25 translation initiation module (e.g., hub genes shown in Additional file
3: Table S2) may be important for leading to further insights on the pathophysiology behind ASD.
In addition to implicating several new gene co-expression modules of significance to ASD, this work provides primary evidence supporting the idea that the cortical transcriptome is dysregulated at hierarchical levels and this hierarchical view of pathophysiology cannot be well understood from the vantage point of examining single co-expression modules in isolation. By identifying disruption in the interaction between-modules and in how eigengene networks are reconfigured into different meta-modular structures, this work presents a larger view on how multiple dysregulated signals may operate in conjunction with one another and potentially implicate important emergent interactions at the protein level. We show that a number of specific modules that are on-average up- or downregulated in ASD are also highly correlated and that this correlation can become stronger in ASD. This result is not apparent in prior work on this topic, with the closest result being the previous observation of a negative correlation when collapsing across both groups between single pair of modules enriched in synaptic and immune functions [
8,
20]. We have gone much further to show correlations between dysregulated modules including translation initiation modules and several other modules. We also demonstrated that beyond the statistical dependencies between co-expression modules, these dysregulated modules physically interact at the level of proteins. The disruption of these coordinated higher order interactions at a protein level suggests that systems-level phenomena are disrupted in ASD that coordinates disparate biological processes and which cannot be adequately characterized by viewing smaller elements (e.g., single genes, single co-expression modules) in isolation. Thus, a primary conceptual advance from this aspect our work suggests that we may need to move beyond arguments about single unitary processes, since the interactions between multiple dysregulated processes may underlie and better describe the pathology.
As a whole, the collection of ASD-downregulated modules appears to involve a number of processes occurring at the synapse. While synaptic processes are commonly discussed as important mechanisms [
21,
55], genes that are typically characterized as synaptic genes may have other pleiotropic roles in very early neural developmental processes. It is known that annotations in enrichment databases (e.g., GO, MetaCore) may be incomplete and an example of this can be seen in potential other interpretations of genes typically thought of as involved in synaptic processes. Casanova and colleagues recently showed that many high-risk ASD genes that have canonical roles in synapse development are also involved in very early stages of neural proliferation, growth, and maturation [
56]. As a specific example of this idea, Konopka and colleagues discovered that
NRXN3 plays a role in earlier neural progenitor biology that is different from its later function at the synapse [
57]. Early fetal brain developmental processes occurring as early as the end of the first trimester of gestation [
58] and are key signals of importance highlighted by prior studies on very early pathophysiology in ASD [
5,
12,
17,
49,
56,
59‐
62]. These specific neural developmental processes are developmentally prior to abnormalities in synaptic processes which emerge at later points in fetal development and continue to change throughout life as a result of postnatal experience and adaptation. Therefore, a nuanced interpretation of the role of synapse gene dysregulation in ASD could be that these genes have pleiotropic roles in both early stages of neural development (e.g., proliferation, growth, and maturation) and at later stages dealing with synaptic processes that continue throughout the lifespan.
The collections of modules upregulated in ASD showed evidence for several novel and emergent biological phenomena. To our knowledge, the novel signal of upregulated catabolism has not been implicated in any past work. Additionally, there are novel upregulated processes involved in protein targeting and localization that can be intertwined with translation processes (e.g., SRP-dependent co-translational protein targeting to membrane). Finally, we also found enrichment in several viral processes, responses to other organisms, cell proliferation, and vasculature developmental processes are highly prominent. These highly coordinated processes are associated with multiple cell types/compartments and the downregulation of synaptic processes—as evidenced by the strong negative correlations between upregulated and downregulated modules. This evidence is generally in agreement with past theoretical ideas [
60] that suggested that early manifestations of pathophysiology potentially emerging in fetal development could then trigger a later corrective phase of development characterized by downregulation of synaptic and neuronal processes and potential upregulation immune/inflammation (e.g., microglia activation) [
37‐
39], apoptotic, and other processes. A challenge for future research will be to unpack the relationships between known and novel upregulated processes with downregulated synaptic and neural developmental processes. However, it is important to underscore that these inferences emerge from the looking at the highly coordinated interactions between multiple dysregulated co-expression modules and are not obvious by simply targeting specific modules and looking at such elements in isolation. Thus, these new insights regarding systems-level phenomena in ASD can further guide future studies to unravel specific novel mechanisms (e.g., targeting hub genes for many of the dysregulated modules we have implicated and examining their impact on other connected systems-level processes; Additional file
3: Table S2).
Acknowledgements
We would like to thank Bhismadev Chakrabarti, Bonnie Auyeung, Meng-Chuan Lai, Simon Baron-Cohen, and Peter Langfelder for the helpful discussion on this work.