Clustering the autisms using glutamate synapse protein interaction networks from cortical and hippocampal tissue of seven mouse models

The specific identity of all mouse strains used is shown in Table 1. Littermate mice were co-housed in groups of 2–5 under standard laboratory conditions. At 60 days of age, mice were deeply anesthetized with isofluorane, decapitated, and brains were removed. We chose day 60 because we wanted to focus on adult animals, since the majority of animal models have been behaviorally tested as adults (see Table 1). We used two males and two females for each genotype to focus our study on robust, non-sex-dependent effects, except FMR1^−/y mice and controls, which were all male since FMR1 is an X-linked gene. We used frontal cortex for all models and hippocampus for some models because these two brain regions have been frequently analyzed in electrophysiology and biochemical studies of ASD models (see Table 1). For frontal cortex, the rostral 3 mm of cortex was cut with a razor blade in a metal brain mold, making sure not to include any striatal tissue in the section, and the olfactory bulb was removed; for hippocampus, bilateral hippocampi were removed with curved forceps. Tissue was frozen in liquid nitrogen and stored at − 80 until homogenization. All work was performed under an approved animal protocol at Seattle Children’s Research Institute (#15580).

VPA mice were prepared at McMaster University in compliance with standards of the Canadian Council on Animal Care and with approval from the McMaster University Animal Research Ethics Board. CD-1 female mice were mated until a sperm plug was detected (E0). On day 12.5 after conception (E12.5), pregnant mice received a single intraperitoneal (i.p.) injection of 500 mg/kg sodium valproate (VPA; Sigma, Oakville, ON, Canada) dissolved in 0.9% NaCl solution, while controls were injected with only saline. E12.5 was chosen to match previous reports from our group and others (see [32]). Pups were weaned on postnatal day (PD) 21 and subjected to behavioral assays (three-chamber sociability, elevated plus maze, and marble-burying assays for social behavior, anxiety, and repetitive behavior, respectively) on PDs 29–34. Animals were killed by decapitation on PD 35, and brains were rapidly dissected and stored at − 80 °C.

Tissue was homogenized in 0.32 M Sucrose in HEPES buffer, pH 7.4 with Sigma protease (Cat # P8340) and phosphatase (Cat # P5726) inhibitors (Sigma Aldrich), using 12 strokes of a glass-Teflon homogenizer. Samples were spun at 1000×g for 5 min to pellet membranes, then spun at 10,000×g for 15 min to pellet P2 synaptosomes. Synaptosomes were solubilized on ice in 200 ul lysis buffer (150 mM NaCl, 50 mM Tris (pH 7.4), 1% NP-40, 10 mM NaF, 2 mM sodium orthovanadate + protease/phosphatase inhibitor cocktails [Sigma]) for 15 min, spun at 4 °C at 10,000×g for 15 min to remove insoluble material, and protein concentration was measured by BCA assay (Thermo-Fisher).

QMI beads (Luminex) were prepared as previously described [32], with each bead color-class coupled to a distinct immunoprecipitating antibody, as shown in Table 2. Equal amounts of protein from each matched pair of animals (transgenic vs. wild type littermate or VPA- vs. saline-treated control) were incubated with QMI beads overnight at 4 °C, with constant rotation. Beads from each sample were then distributed into 32 wells of a 96-well plate, approximately 250 beads of each class per well, and each of 16 probe antibodies was added, in duplicate, to individual wells. Beads were then washed with ice-cold Fly-P buffer [50 mM Tris (pH 7.4), 100 mM NaCl, 1% bovine serum albumin, and 0.01% sodium azide], incubated for 30 min with streptavidin-PE (1:200, BioLegend), washed again, and read on a custom refrigerated Bioplex 200 flow cytometer (BioRad), which recorded the bead classification (corresponding to IP’d protein, X) and PE fluorescence (corresponding to the amount of probe antibody target protein, Y) of each bead. An above-background reading for IP:X Probe:Y indicates the occurrence of a protein complex containing both X and Y [33].

Data were exported in .xml files containing all data on a bead-by-bead, well-by-well basis. A custom Javascript was written to generate histograms showing bead distributions for a given bead class in a given well and to extract the median fluorescent intensity of each bead class in each well for export to Excel and R (faculty.​washington.​edu/seps/program). A custom MatLab script, “Adaptive Non-parametric statistical test with an adjustable alpha Cutoff” (abbreviated ANC), previously described in detail [30], was used to identify interactions that changed significantly in > 70% of experiments; these interactions are referred to as “hits.” ANC first uses a K-S test to compare histogram distributions of technical replicates to both discard duplicate wells that are significantly different from each other (presumed manual error) and to adjust the alpha value based on technical error. K-S test results from comparisons between an experimental sample and a matched control are then corrected for multiple comparisons and technical errors to obtain a final p value. “Hits” were interactions with p < 0.05. Please see [30] for details. Prior QMI analysis in both T cells [30] and neural tissue [32] found that N of four biological replicates are sufficient to produce a consistent number of significant hits, so an N of at least four matched pairs was used. To eliminate batch effects due to both technical and biological variation, we limit comparisons to ASD model animals and co-housed, littermate controls euthanized on the same day and run on the same assay plate; ANC statistics are therefore based on consistent differences in paired comparisons for N = 4 experiments (each run with technical replicates). Workflow and examples of smoothed histograms are shown in Fig. 1.

Data matrices for each matched pair were exported from Java to Excel. For each matrix position, we divided the median fluorescence value (of the two technical replicates) of each ASD model animal by its wildtype littermate control and log₂-transformed the result. Then, log₂ fold change (log₂FC) matrices from N = 4 experiments were averaged to generate a single mean log₂FC matrix per genotype/tissue type, shown in Additional file 1: Table S1. ANC significant hits were identified and imported into Cytoscape for visualization. Significant interactions are represented by an edge connecting two protein nodes; the color and width of the edge corresponds to the direction (red = up, blue = down) and magnitude of the change (Fig. 1e, f). Changes in protein abundance (IP probe of same target protein) are represented by loops.

To cluster samples by shared ANC-significant hits, we used the hclust(dist()) and heatmap.2 functions in R. To cluster samples by average fold-change matrices shown in Additional file 1: Table S1, we first performed principal component analysis to reduce noise due to nonspecific background fluctuations using the “PCA” function in the “FactomineR” package for R; then we used the “HCPC” function in the same package to cluster genotypes/tissue types by principal components. To test the robustness of clustering, we used the “pvclust” function in R. All options were used in the default settings.

Western blots were run on cortical tissue using standard protocols. Briefly, cortical P2 fractions were lysed in lysis buffer, protein concentrations were normalized using BCA assays, equal amounts of protein were loaded into each well and run at 110 V. Protein was transferred onto PVDF membranes, blocked with 5% milk in TBS-T, primary antibodies were incubated overnight at 4C, followed by washes, species-specific secondary antibody incubation (anti-mouse or rabbit, 1:10,000, Jackson Immunoresearch), and luminol detection (Pierce Femto reagent). Antibodies used (all 1:1000 dilutions) the following: Ube3a clone E-4 (Santa Cruz), pAKTs473 clone D9E, pAKTt308 clone 244F9, panAKT clone 40D4, pMTOR polyclonal Cat #2971, and pS6 clone D57.2.2 (all from Cell Signaling).

Overall, we found 32 statistically significant differences across the 7 mouse models (Additional file 1: Table S1 Sheet 2, and Fig. 2). Of 240 total IP_Probe combinations, 9 proteins (IP probe for the same target) and 18 proteins in shared complexes (PiSCES—IP probe for different proteins) showed differences in abundance across the 7 models. Four PiSCES (Homer1_PSD95, Homer_NMDAR1, SynGAP_PSD95 and NL3_FYN) and three abundance measures (FYN, SynGAP and PSD95) were significantly different in multiple models, while the remaining differences were unique to a single model.

Shank3B^−/− animals (Fig. 2a) showed a reduction in Shank3 protein levels and an increased co-association of Homer with PSD95 and NMDAR1. This is counterintuitive, since Homer-PSD interactions are likely mediated by Shank proteins, and may be expected to be reduced in Shank3 animals. These results may reflect changes in Shank1/2 vs. 3 scaffolding, or an increase in these activity-labile interactions [32] may be downstream of reduced synaptic activity in mutant animals [27]. Interestingly, these same two interactions were changed, but in the opposite direction, in FMR^−/y mice (see below), which have increased basal activity, potentially supporting an activity-dependent mechanism (see discussion). We also observed a small increase in SynGAP_PSD95 (also activity-labile) and a small decrease in FYN_NL3, interactions that were both observed in the opposite direction in the TSC2^+/− model.

Shank3Δex4–9^+/− animals (Fig. 2b) were tested as heterozygotes because the heterozygotes showed abnormal behavior in the original publication [28] and more accurately represent the human condition, a heterozygous deletion/mutation. Consistent with only moderately reduced Shank3 levels, a modest reduction in Shank3 (log₂FC = − 0.227) was not significant by ANC. Remarkably, there was no overlap in significant hits with Shank3B knockouts; in fact, SynGAP_PSD95 was significant in the opposite direction in the two models. Similar to the TSC2^+/− animals, Shank3Δex4–9^+/− animals showed significant decreases in SynGAP_SynGAP and SynGAP_PSD95, in addition to reductions in SynGAP_Homer1A and SynGAP_NMDAR1 that were unique to this model. Finally, a moderate increase in FYN_Shank3 was observed. While published electrophysiology revealed reduced excitatory transmission in both the Shank3B^−/− and Shank3Δex4–9^+/− animals [27, 28], the experiments were performed in different brain areas (striatum and hippocampus, respectively) and showed small but important differences, such as reduced vs. increased miniature EPSP frequency, respectively. In summary, while QMI data from the Shank3Δex4–9^+/− animals highlight reduced SynGAP associations with NMDARs and scaffolds, the Shank3B animals show differences in Homer-PSD-NMDAR complexes but no changes in SynGAP. These data suggest that the molecular deficits in the two animal models may be quite different, consistent with the different isoforms that are affected in the two models [24].

Ube3a^2xTg mice (Fig. 2c) showed an expected increase in the amount of Ube3a and reduced co-association between GluR1_PSD95 and NL3_GluR2. Prior work in the cortex of Ube3a animals showed reduced glutamatergic transmission [34] and reduced scaffolding of GluRs is therefore consistent with prior observations.

E12.5 VPA mice (Fig. 2d) are the only non-genetic model analyzed here. Mice were generated by injection of VPA on E12.5, and the efficacy of the treatment was confirmed by behavioral testing (as in [32]) of the adult offspring before dissection and QMI analysis. We observed a large increase in the amount of co-associated Homer_PSD95. In all other models, the amount of Homer_NMDAR1 correlated with Homer_PSD95, and VPA mice were trending towards an increase in this interaction as well (log₂FC = 0.48, NS). In addition, levels of Fyn were increased, PSD95 were decreased, and interactions between SAP97_Homer1A and Shank3_NL3 were increased. Prior reports in VPA-treated rat cortex showed enhanced NMDAR-mediated synaptic currents and enhanced LTP [35], consistent with the observed increase in Homer-NMDAR scaffolding. Decreased levels of PSD95 have also been reported by Western blotting [52].

TSC2^+/− mice (Fig. 2e) showed large reductions in the abundance of SynGAP and SynGAP_PSD95. The mTOR activator Rheb (ras homolog enriched in brain), which is directly suppressed by the TSC1/2 complex, is activated by SynGAP following NMDAR stimulation [36], so the reduction of SynGAP may be a homeostatic response to chronically activated Rheb. Reduced levels of NMDAR2B were also observed, along with increased abundance of complexes containing SAP97_mGluR5, Homer1_SAP97, and Fyn_NL3. Taken together, these data indicate reduced NMDAR2B and SynGAP expression, abnormal scaffolding of mGluR5 to Sap97, and abnormal FYN signaling, which could contribute to the altered LTP phenotype reported in the hippocampus of TSC mice [37].

Fragile X mice (Fig. 2f) showed increased abundance of complexes containing NMDAR2A_NL3 and NMDAR2B_NL3. Both NMDA receptors [38] and Neuroligins [39] bind PSD95, which could mediate this observed interaction. FragileX mice also showed reduced complexes with Homer1_PSD95 and Homer1_NMDAR1, demonstrating disrupted Homer-Shank-PSD95-NMDAR complexes, consistent with previous reports [40‐42]. These activity-dependent interactions were also significant hits in the Shank3B and E12.5 VPA models, but in the opposite direction, possibly reflecting hyper- vs. hypo-activity of cortical neurons in these models. Finally, reduced levels of PI3K and increased Fyn were detected, consistent with disrupted kinase cascades downstream of mGluR5 in FMR1 mice [43‐45].

CNTNAP 2 KO mice (Fig. 2g) showed the greatest number of ANC hits (7) as well as many large but non-significant changes. The abundance of GluR2 was reduced, accompanied by reduced GluR2_Shank3, NMDAR1_Shank3, and NMDAR2B_Homer1A, consistent with reduced scaffolding and expression of glutamate receptors. In addition, the detected levels of Sap97 and SynGAP were reduced, while Fyn was increased, a change also observed in the Fragile X model. While the CNTNAP2 gene product CASPR is known to cluster at the nodes of Ranvier following myelination [46], acute CASPR knockdown acts cell-autonomously to reduce both AMPA and NMDA-mediated EPSPs [47], congruent with reduced NMDAR and AMPAR levels and scaffolding observed here.

In four models, we also isolated P2 fractions from the hippocampi of the same animals that supplied cortical tissue. We identified 45 statistically significant differences across the 4 mouse models, with the majority of differences, 21, found in the VPA hippocampus (Fig. 3). Nine proteins showed differences in protein abundance (IP probe for the same target), and 32 protein interactions showed differences (IP probe for different targets). However, only three complexes (Homer_mGluR5, Sap97_NMDAR1, and SynGAP_NMDAR2A) and 1 abundance measure (PSD95) were detected in multiple models, while the remainder was unique to a single model. Below, we describe the findings from each model, compared with prior data from the cortex of the same model.

Shank3B^−/− hippocampal tissue (Fig. 3a) showed an increase in PSD95 levels and a decrease in FYN levels, neither of which were observed in cortical tissue. A decrease in NMDAR1_mGluR5 likely reflects disrupted scaffolding linking the two receptor types via PSD95/Shank3/Homer linkages [24]. We observed an increase in Homer1_SYNGAP, PSD95_GluR1, and, counter-intuitively, PSD95_Shank3. The latter interaction may reflect elevated expression of an alternative isoform of Shank3 that lacks the PDZ domains in complex with PSD95, possibly mediated via another protein such as Homer. Besides this interaction, no major changes in Shank3 were detected, likely due to the fact that very little Shank3 was detected from Shank3 IPs or probes, consistent with low hippocampal Shank3 expression. We are unable to relate these changes to known electrophysiological abnormalities in these animals, since to our knowledge, hippocampal electrophysiology has not been reported in this model.

Shank3Δex4–9^+/− animals (Fig. 3b) showed reduced levels of SynGAP, consistent with cortical tissue from these animals. Interactions involving SynGAP_NMDAR2A were reduced, while SynGAP_GluR2 were increased. Complexes containing NL3_NMDAR2A and _FYN were reduced. Complexes containing PI3K_Sap97, _GluR2, and _SynGAP were all increased. Hippocampal electrophysiology in this model indicated reduced basal AMPA-mediated transmission, and a failure of hippocampal LTP that was correlated with failure to maintain spine expansion following a tetanizing stimulation [28]. Our results indicate that SynGAP, a critical mediator of signal transduction downstream of NMDARs [36, 48], is dysregulated in hippocampal tissue prior to any type of stimulation. Further, changes in FYN and PI3K suggest downstream disruption of signaling cascades.

Ube3a^2xTG hippocampus (Fig. 3c) showed the expected increase in Ube3a expression, the only change that was consistent between hippocampus and cortex. A reduction in mGluR5 levels, Homer_mGluR5, and Homer_GlurR2 suggest reduced Homer-mediated scaffolding. Ube3A_Homer interactions were strongly increased, although the significance of this increase is unclear since Ube3a has not been documented to bind directly to or ubiquinate/degrade Homer proteins. The amount of PSD95_Shank3 was increased, as was SAP97_NMDAR1 and SAP97_Shank3. Finally, the amount of Homer1A was increased. These data demonstrate complex changes in scaffolding of AMPA, NMDA, and metabotropic glutamate receptors mediated by both Homer and DLG scaffolds in the Ube3a^2xTG animal. Hippocampal electrophysiology has not been reported in these animals, although LTP disruptions due to lack of small conductance potassium channel 2 (SK2) channel regulation have been reported in the Ube3a knockout animal [49].

VPA hippocampus yielded 21 significant QMI hits, the most of any sample tested, all in the positive direction. PI3K was involved in six significant interactions, with _mGluR5, _NMDAR1, _NMDAR2A, _PSD95, _HOMER1A, and _PI3K. These disruptions in PI3K, which controls AKT/mTOR signaling, is consistent with several reports implicating dysregulated mTOR signaling in the VPA model [32]. The amount of Homer_PSD95, Homer_NMDAR2B, and Homer_NMDAR1 were each increased by almost twofold, reflecting increased NMDAR scaffolding and/or expression. Levels of detected NMDAR1 and NMDAR2B were also increased. These data support prior studies showing increased NMDAR expression in rats following VPA exposure in the cortex [35], although note that a separate study did not find differences in mRNA expression in the cortex or hippocampus [50]. Other notable hits included SynGAP_NMDAR2A and B, SAP97_NMDAR1 and 2B, and Homer_mGLUR5. Comparing these results with VPA cortex, only 2/5 QMI hits in the cortex were shared with the hippocampus, Homer_PSD95 and SAP97_HOMER1A. However, several other interactions that were significant in hippocampus were trending towards significance in cortex; for example, Homer_NMDAR1 and _NMDAR2B were increased by 1.29 and 1.37-fold in the cortex, respectively, but were not significant by ANC criteria (see Additional file 1: Table S1).

Several limitations of our study should be noted. The background strain was different among several of the models, the age of the VPA mice was different from the other six models, and mice of both sexes were used. Our analysis approach, in which each mutant animal was normalized to a matched littermate control, was designed to cancel out these effects, as well as assay-dependent batch effects, to identify differences caused by each mutation. However, this experimental design prevents us from making wildtype-to-wildtype comparisons (since batch effects cannot be normalized for mice run on different assay plates), so we are unable to unambiguously demonstrate that our clustering was not driven by uncorrected effects stemming from these differences in background, age, or sex. QMI is a candidate-based approach and shares limitations with all antibody-based assays, including potential antibody cross-reactivity and issues of binding epitope access in native protein complexes. The absence of a detected interaction cannot be interpreted as unambiguously indicating that the interaction does not exist in vivo, since occlusion of binding sites could lead to false negative results. We carefully selected and screened all antibodies used in the QMI panel [32], but antibody caveats are unavoidable. We used NP40 detergent after pilot data showed that it produced higher mean matrix MFIs than TritonX100, Digitonin, or Deoxycholate [32]. However, NP40 does not fully solubilize the core postsynaptic density, where several of our protein targets are enriched (discussed in [32]). Since detergents that solubilize the PSD also disrupt protein interactions, detergent selection is necessarily a trade-off, and further studies could more thoroughly quantify differences in synaptic QMI networks due to different detergent conditions. Finally, many of our interactions vary with neuronal activity or by brain region. Small variations in microdissection (e.g., inclusion of small amounts of striatal tissue in cortical samples) or euthanasia protocols (i.e., animal sleep/wake state prior to euthanasia) could have large effects on protein detection (e.g., Shank3, which is highly expressed in striatum, or PSD95_SynGAP, which is activity-dependent). While we were careful to perform our dissections as consistently as possible, at a similar time of day and using metal brain molds to ensure consistent slicing, thinner vibratome slicing followed by a period of controlled slice recovery in ACSF, as for electrophysiology, may be a more optimal experimental strategy to ensure normalization of both activity and location.