Identification of the common neurobiological process disturbed in genetic and non-genetic models for autism spectrum disorders

Abstract

Autism spectrum disorders (ASD) are neurodevelopmental disorders. Genetic factors, along with non-genetic triggers, have been shown to play a causative role. Despite the various causes, a triad of common symptoms defines individuals with ASD; pervasive social impairments, impaired social communication, and repeated sensory-motor behaviors. Therefore, it can be hypothesized that different genetic and environmental factors converge on a single hypothetical neurobiological process that determines these behaviors. However, the cellular and subcellular signature of this process is, so far, not well understood. Here, we performed a comparative study using “omics” approaches to identify altered proteins and, thereby, biological processes affected in ASD. In this study, we mined publicly available repositories for genetic mouse model data sets, identifying six that were suitable, and compared them with in-house derived proteomics data from prenatal zinc (Zn)-deficient mice, a non-genetic mouse model with ASD-like behavior. Findings derived from these comparisons were further validated using in vitro neuronal cell culture models for ASD. We could show that a protein network, centered on VAMP2, STX1A, RAB3A, CPLX2, and AKAP5, is a key convergence point mediating synaptic vesicle release and recycling, a process affected across all analyzed models. Moreover, we demonstrated that Zn availability has predictable functional effects on synaptic vesicle release in line with the alteration of proteins in this network. In addition, drugs that target kinases, reported to regulate key proteins in this network, similarly impacted the proteins’ levels and distribution. We conclude that altered synaptic stability and plasticity through abnormal synaptic vesicle dynamics and function may be the common neurobiological denominator of the shared behavioral abnormalities in ASD and, therefore, a prime drug target for developing therapeutic strategies.

Supplementary Information

The online version contains supplementary material available at 10.1007/s00018-022-04617-3.

Introduction

Autism spectrum disorders (ASD) are neurodevelopmental disorders defined by social and communication impairments and restricted and repetitive behaviors [1]. Genetic, epigenetic, and environmental factors have been associated with the etiology of ASD [2–4]. The causative role of several of these factors has been confirmed and investigated through the generation of mouse models [5, 6]. Behavioral phenotyping of genetic and non-genetic mouse models confirmed the presence of impairments and abnormalities characteristic of ASD in humans. However, on the cellular and molecular level, a variety of different effects have been described. Nevertheless, the presence of the common and specific behaviors in mice, and humans, which provide the basis for the ASD diagnosis [1], suggests a convergence of the different causative factors on a common biological process; the neurobiological base of ASD.

In line with this, we hypothesize that shared alterations on gene and protein expression level can be detected amongst different mouse models of ASD. “Omics” approaches are highly suited for discovering common motives, as they investigate biological systems in a hypothesis-free approach. While the initial characterization of several mouse models was based on hypotheses such as the presence of synaptic deficits, neuroanatomical abnormalities, or neuroinflammation [7], the data likely contain more information that was not pursued during these focused studies. The unbiased analysis of transcriptomics and proteomics data of these mice will allow the identification of affected cellular processes and signaling pathways in a bottom-up approach, potentially revealing shared dysregulated processes not previously considered. Thus, we have selected several publicly available genetic and non-genetic mouse models that commonly show ASD-like behavior. To ensure the broadest coverage within genetic mutations associated with autism, we selected models from both single nucleotide variations (SNV) and copy number variations (CNV). The genetic models of engrailed-2 (En-2) mutant mice represent SNV [8, 9], and CNV models are represented with 16p11.2 mutant mice [10, 11], SH3 and multiple ankyrin repeat domains protein 2 (Shank2) and Shank3 mutant mice [12, 13], and fragile X mental retardation 1 (Fmr1) knockout (KO) mice [14]. The mechanism through which these models result in ASD like behavior are discussed in further detail below.

EN-2 has been identified as an ASD candidate gene in linkage studies [15]. The gene encodes a homeodomain-containing protein that plays a role in pattern formation during neurodevelopment. Altered epigenetic modification of the EN-2 promoter was found in the cerebellum in individuals with ASD, correlated with levels of EN-2 expression [16]. Therefore, the model represents a non-syndromic form of ASD that may be based on (epi) genetic changes.

The 16p11.2 CNV of the human genome encompassing 30 genes has been linked to ASD [17]. Mice with genetic alteration of the 16p11.2 synonymous region on mouse chromosome 7 have been reported to display ASD-like behavior and, therefore, represent ASD associated with a CNV.

The family of Shank genes has been linked to syndromic and non-syndromic forms of ASD [18]. SHANK proteins encode for postsynaptic scaffold proteins at glutamatergic synapses [19] but have been described in many tissues outside the CNS as well, with various functions. SHANK3 is a major ASD candidate gene, with mutations found in 1–2% of individuals with ASD [18]. Deletions of or within SHANK3 likely are the most significant contributor to Phelan McDermid syndrome, characterized by deletions in 22q13.3 that almost all include the locus of the SHANK3 gene. Several Shank2 and Shank3 KO mice have been generated in recent years, and their ASD-like behavior has been confirmed [20].

FMR1 encodes the fragile X mental retardation protein (FMRP). ASD and fragile X syndrome (FXS) are frequently co-diagnosed [21]. In line with this, Fmr1 KO mice display several ASD-like features [22] along with known ASD co-morbidities such as hyperactivity and seizures. The model, therefore, is representative of a syndromic form of ASD.

In addition, mouse models for ASD have been generated by manipulating non-genetic factors. For example, in utero exposure to valproic acid (VPA) [23], maternal immune activation (MIA) [24], and prenatal zinc deficiency [25–27] have all been reported to result in ASD-like behaviors. However, hypozincemia has been shown in rats exposed to VPA [28], and balancing zinc levels can prevent behavioral deficits in MIA and VPA mice [28, 29]. Therefore, altered zinc levels may underpin the phenotypes of these non-genetic models of ASD [29].

Here, we have performed brain region-specific proteomics analysis of prenatal zinc deficient (PZD) mice and compared the data to published results from En-2 [30, 31], 16p11.2 CNV [10], Shank2 [31], Shank3 [32], and Fmr1 (KO) mice [33]. Additionally, we included data from a study that analyzed cortical neuron cultures with genetic knockdown of the ASD candidate genes: Mecp2, Mef2a, Mef2d, Fmr1, Nlgn1, Nlgn3, Pten, and Shank3 [34]. We have identified synaptic vesicle release and recycling as a common biological process altered across all models centered around set of key proteins; VAMP2, STX1A, RAB3A, CPLX2, and AKAP5.

Methods and materials

Unless otherwise indicated, all chemicals were obtained from Sigma-Aldrich (Merck Life Science, Ireland). Primary antibodies were purchased from the following companies: Abcam (anti-Akap5, #ab204082), BioLegend (anti-Snap25, #899909; anti-Synaptophysin, #837104; anti-PSD-95, #MMS-5138), Merck Life Science (anti-β-Actin, #A2228), and Synaptic Systems (anti-MAP2, #160004, anti-Homer, #160004; anti-Shank3, #162304). Alexa Fluor-conjugated secondary antibodies were obtained from Invitrogen/Life Technologies Europe. Secondary HRP conjugated antibodies were purchased from Dako. PKA antagonist KT5720 (#1288) was purchased from Tocris, and PKA activator 8-Bromo-cAMP sodium salt (#HY-12306) was obtained from MedChemExpress (MCE).

Animals

PZD mice were generated as described previously [25–27] using C57BL/6 mice purchased from Janvier Labs. After birth, offspring from PZD and CTRL mice received milk from dams on a standard laboratory diet and were fed standard laboratory food after weaning. The control and PZD cohort contained the same number of female and male mice (5 male and 5 female per group). Mice were sacrificed at 11 weeks old. Animal experiments were performed in compliance with guidelines from the Federal Government of Germany for the welfare of experimental animals and approved by the Regierungspräsidium Tübingen and the local ethics committee (Ulm-University) ID:Number:1239.

Protein biochemistry

Protein fractionation

Mouse brain tissues were weighed and homogenized in buffer-A (320 mM sucrose, 5 mM HEPES, pH 7.4, containing protease inhibitor mixture) using a pulse sonicator (sonic-dismembrator-120, Fisherbrand). Cell debris and nuclei were removed by centrifugation at 3200 rpm for 15 min at 4 °C. The protein concentration of the soluble fraction was determined using Bradford assay, and S1 was used for western blotting.

Western blotting (WB)

Proteins were separated by SDS-PAGE and blotted onto nitrocellulose membranes. Primary antibodies for Actin (2 µg/ml) and AKAP5 (2 µg/ml) were used for protein detection. Immunoreactivity was visualized using HRP-conjugated secondary antibodies and the SuperSignal detection system (Pierce, Upland, USA).

WB quantification

Evaluation of bands from WBs was performed using ImageJ. Three independent experimental blots were imaged using an Alliance Q9 Advanced by Uvitec. Individual band intensities were quantified by integrated density and normalized to the loading control with subsequent ratio averaging.

Proteomics

Brain samples of 10 Ctrl and 10 PZD mice were dissected into different brain regions. For proteomics analysis, cortex, hippocampus, and striatum samples of PZD and Ctrl mice were lysed in Buffer A, homogenized with a sonic dismembrator (Fisherbrand), and sent on for further analysis. Label-free Swath analysis of protein lysates was performed at the BSRC Mass Spectrometry and Proteomics Facility (St. Andrews, Scotland).

Omics meta-analyses

Proteomics and transcriptomics data sets investigating genetic ASD variants were collected from National Center for Biotechnology Information (NCBI) and Proteomics Identification Database (PRIDE) using search terms such as; autism spectrum disorder and ASD. As a result, six ASD studies using genetic models were selected for further analysis (listed in Table 1).

Table 1.

Selected studies of genetic ASD models for comparison with proteomics data from PZD mice

Data set accession	Model system	References
GDS6016	Engrailed-2 (En-2) mutant mouse	[30]
GDS4430	16p11.2 copy number variation mouse model	[59]
GSE79824	Shank2 knockout (KO) mice	[31]
GSE71034	Fragile X mental retardation 1 (Fmr1) knockout (KO) mice	[32]
GDS4759	E16 primary cortical neuron cultures transduced with shRNA constructs of austism spectrum disorder (ASD)-implicated genes (ACGs): Mecp2, Mef2a, Mef2d, Fmr1, Nlgn1, Nlgn3, Pten, and Shank3	[34]
PXD005192	Hippocampus and striatum of Shank3Δ11 knockout (KO) mice	[32]

Functional analysis of the genes identified to be up- and downregulated in genetic and PZD mouse models was first performed using the Bingo plugin (3.0.4) [39] in Cytoscape open-source platform (v3.7.1). Overrepresentation tests were employed with Fisher’s exact tests to calculate p value and the Benjamini–Hochberg procedure to calculate false discovery rates (FDR). FDR cutoffs of < 0.05 were utilized [40]. To reduce the amount of redundant and overlapping terms, REVIGO was used (http://revigo.irb.hr/), set to a similarity cutoff of 0.5. Following this filtering method, the overrepresented GO terms were then visualized as a network using Cytoscape EnrichmentMap (3.3.1) plugin [41]. This method of visualization allows the FDR (color of node), gene set size (size of node), and the similarity of GO terms (thickness of edges) to be displayed.

Individual genetic and non-genetic ASD data sets were merged into a single “meta matrix” on common gene name using Perseus open-source software (v1.6.14.0) [35]. To minimize inherent intraexperimental variability of each study, resulting meta-matrix was normalized using NormalizedDE (v3.12) using R (v4.0.2). Following the suggested criteria in [38], Medl-G normalization was selected for the proceeding analysis steps. During Medl-G normalization, each variable’s intensity in a given sample is divided by the median of intensities of all variables in the sample and then multiplied by the mean of “median of sum of intensities of all variables in all samples.” The normalized data are then transformed to Log2.

Following normalization, we used volcano plot analysis with Perseus to identify significantly modulated genes across all ASD models. Genes of interest were selected by threshold values set for each volcano to -Log10(p) = 0.05 on the y-axis and to the interquartile (IQR) range of housekeeping gene expression described by Eisenberg and Levanon on the x-axis [36] in the data set, assuming housekeeping gene expression act as internal controls and are a more accurate thresholding approach for meta-omics analysis. Resulting upregulated and downregulated genes/proteins were used to identify the core ASD genes/proteins modulated across all genetic studies by overlay analysis (Venny 2.1), using a master list of ASD-linked genes, with an inclusion criterion of being present ≥ 3 studies.

The Gene list identified during the overlay analysis was used as a training set for the machine-learning algorithm. The prediction model implementation was carried out in WEKA open-source platform [37] using its built-in Auto-WEKA function [38]. Auto-WEKA considers the problem by simultaneously selecting a learning algorithm and setting its hyperparameters, going beyond previous methods that address these issues in isolation. Auto-WEKA does this using a fully automated approach, leveraging recent innovations in Bayesian optimization. In this instance, Auto-WEKA selected Multilayer Perceptron with the following hyperparameters; learn rate = 0.8, momentum = 0.4, and validation threshold = 20. Subsequently, tenfold cross-validation was used to evaluate the accuracy of predictions. This divides the entire data set into 10 equal parts. Prediction is repeated 10 times, each timekeeping one of the 10 parts as test data and the other 9 parts as training data to build the prediction model. Finally, test results on all the 10 parts are accumulated together to calculate the average prediction accuracy.

To account for the potential skew of the training set and its effect on the prediction model performance percentage accuracy, F-measures, and area under the receiver operator curves (ROC) were used to assess model performance (Fig. 4B, C). Following machine learning classification, resulting gene lists were passed through a confidence thresholding using WEKA built-in cost–benefit analysis, resulting in a smaller gene set with a high confidence of classification membership. This graphical analysis of the benefits of using the model summed up and the costs associated with it subtracted. 95% prediction accuracy was chosen as a cutoff for ASD up- and downregulated gene sets used in subsequent analysis.

Fig. 4.

Machine learning approach to identify new pathways and genes potentially regulated in ASD. A Overview of the steps taken from suitable data set identification and their formatting to the machine learning analysis. B Receiver operator curves (ROC) for the downregulated and upregulated gene identification. C Metrics such as f-measure and ROC area under the curve for upregulated and downregulated genes, where the range is 0–1. In both groups, these metrics are close to the value of one, showing that the prediction accuracy is high. D Downregulated and upregulated pathways and their associated biological processes obtained post machine learning. Regulation of synaptic transmission and vesicle-mediated transport is affected by downregulated proteins, immune system response, cell junction formation, and neurogenesis that are identified through upregulated proteins, FDR ≤ 0.05

Neuronal cell culture

Primary rat hippocampal neurons (Gibco Thermofisher, A36513) were cultured per manufacturers’ recommendations. Therefore, neurons were plated onto poly-L-lysine (PLL) (0.1 mg/ml) coated Screenstar 96 well microplates (Greiner Bio-One GmBH) at a density of 3 × 10⁴ cells per well in neurobasal media supplemented with 2% B27 supplement and 1% penicillin/streptomycin. After 24 h post-seeding, 50% of media was replaced with fresh neuronal growth media. For zinc-deficient conditions, neurons were changed into growth media additionally supplemented with 50 µM CaEDTA. CaEDTA was shown to affect extracellular as well as intracellular zinc levels [42], and does not significantly affect the levels of other metals [43]. 50 µM CaEDTA has been reported to affect synapses previously [44, 45]. It is sufficient to chelate the approx. 2.6 µM free zinc contained in neurobasal medium [46] and considers that zinc is released into the synaptic cleft with synaptic activity, reaching local concentrations that may routinely exceed 10–20 μM [47]. Cells were cultured at 37 °C and 5% CO₂ for 11 days for live imaging and 14 days for immunocytochemistry.

For vesicle staining in live cells, hippocampal neurons were cultured until DIV11. To induce zinc deficiency, neurons were cultured in the presence of 50 µM CaEDTA and for high zinc conditions, cells were treated with 125 µM ZnSO₄ for 30 min at 37 °C. For live-cell staining, conditioned media of all conditions was removed, stored until further use and cells were treated with 5 µM Synaptogreen C4 (Sigma Aldrich) for 30 min at 37 °C in fresh neurobasal growth media supplemented with 2% B27 supplement and 1% penicillin/streptomycin (with an additional 50 µM CaEDTA for zinc-deficient conditions or 125 µM ZnSO₄ in high zinc condition). Cells were counterstained with DAPI for 5 min at 37 °C. Afterward, vesicle dye was removed, cells were washed with 1 × PBS, and conditioned media was reintroduced. For time point 1 (T1), pre-dye-loaded cells were immediately imaged on ImageXpress Micro Confocal in live-cell mode (chamber 37 °C, 5% CO₂). For the second time point T2, cells were reimaged in the conditioned media after 30 min in the incubator on the ImageXpress Micro Confocal in live-cell mode (chamber 37 °C, 5% CO₂).

Immunocytochemistry and fluorescent microscopy

Primary hippocampal neurons on DIV14 were treated with antagonist KT5720 at 2 µM for 2 h at 37 °C or PKA agonist 8-Bromo-cAMP at 500 µM for 2 h at 37 °C in zinc-deficient conditions. After the incubation period of antagonist and agonist of PKA in wells treated with CaEDTA, all conditions were fixed with 4% paraformaldehyde solution (PFA)/4% sucrose/PBS for 15 min at RT before subsequent immunofluorescent staining., Afterward, neurons were washed with 1xPBS with 0.2% Triton-X for 10 min at RT, and blocked with 10% FBS/1 × PBS for 1 h at RT. Afterward, cells were incubated with primary antibody diluted in blocking solution (BS) for 2 h at RT. After 3 × 5 min washes with 1xPBS, the secondary antibody, also diluted in BS, was applied for 1 h at RT. After that, cells were washed with PBS (5 min), counterstained with DAPI, and washed with aqua bidest. Screenstar imaging plates were imaged on the ImageXpress Micro Confocal High-Content Imaging System (Molecular Devices) with a 40X S Plan Fluor ELWD objective.

Image processing and quantification

Image analysis was performed using an in-house protocol for Cell Profiler analysis software [48]. Here, we used a variation of a protocol established previously [49, 50]. Briefly, image sets corresponding to various fluorescent channels in each field of view were subjected to various normalization and analysis pipelines. We identified neuronal structures labeled with DAPI and either 1. Homer and Synaptophysin, 2. AKAP5 and SNAP25, 3. STX1A, SHANK3, and MAP2, or 4. AKAP5, PSD95, and MAP2 as required. Initially, image sets were normalized between experiments against their bulk fluorescent intensity properties. Next, segmentation of the nuclei according to DAPI labeling, with a median size of 62.5 pixels of each cell in the field of view, was identified, corresponding to an arbitrary fluorescence intensity above background and shape (circular). Cell bodies were then identified through their size, fluorescent intensity in channels associated with Synaptophysin, SNAP25 or MAP2 as appropriate, and their encompassment of a nuclei object and designated cell body’. Next, neurites were identified from regions not inclusive of cell bodies that displayed a spindle-like morphology and fluorescence intensity above the background corresponding to the labeling of Synaptophysin, SNAP25 or MAP2 in associated labeling. These objects were enhanced using the neurite enhance module within CellProfiler by the “Tubeness,” resulting in segmented objects designated “Neurites” corresponding to the neurite network in a field of view that did not include the cell body. Restricting further segmentation to the identified neurite network as region of interest (ROI), punctae with a typical diameter range of 2–15 pixels in channels corresponding to Homer, Synaptophysin, AKAP5, SNAP25, STX1A, SHANK3, or PSD95 was undertaken. Synapses were subsequently identified in areas where two “punctae” from two different labeling combinations, Homer & Synaptophysin, STX1A & SHANK3 or AKAP5 & PSD95 co-localized in their corresponding channels. Similarly, presynaptic assemblies were identified through the colocalization of punctae of SNAP25 & AKAP5. The area of the neurite network was derived from the pixel area of the segmented “Neurite” object occupied per image. Data per field of view were aggregating per condition with biological repeats treated separately. Identified punctae, synapses, and presynaptic assemblies were counted and expressed as a frequency per cell or per neurite area as indicated. In addition, the mean fluorescence intensities within synapses or presynaptic assemblies for each labeling under interrogation were determined for each object. The values associated with Homer & Synaptophysin marked synapses and SNAP25 & AKAP5 presynaptic assembles were aggregated across experiments and plotted on an x–y scatter density plot (generated in Perseus software) [35] to visualize the relative composition of the synapse for the proteins investigated or analyzed per experiment by geometric means ± SEM of each protein under investigation and expressed as bar graph histograms.

Data associated with AKAP5 & PSD95 or STX1A & SHANK3 marked synapses were aggregated across experiments with individual data points representing the median object value (count or fluorescent intensity) per field of view (FOV) prior to aggregation per condition across experiments. These data points are shown overlaid with median and interquartile range whisker plots.

The average number of neuron cells per condition per experiment was 2850. The median punctae per condition per labeling per experiment measured were in excess of 80,000 in full Zn and 50,000 in low Zn. The median synapses and presynaptic assemblies per condition per labeling per experiment measured were in excess of 19,000 in full Zn and 17,000 in low Zn. In conditions pertaining to low Zn treated with the PKA agonist, the median number of punctae was in excess of 48,000 per experiment measured while the median number of synapses was in excess of 25,000 per experiment. In the conditions pertaining to low Zn treated with the PKA antagonist, the median number of punctae was in excess of 7500 per experiment measured while the median number of synapses was in excess of 3300 per experiment measured.

Image analysis of the live imaging cultured cells was performed with a similar modified analysis protocol using Cell Profiler software. In short, experimental conditions were normalized as above against their bulk fluorescent intensity properties. Neuronal structures were identified using DAPI and Synaptogreen. Objects identified in the same channel were differentiated based on varying levels of fluorescent intensity, which was rescaled to allow for the correct identification of cell bodies, neurites, and presynaptic vesicles, respectively. After identification, these segmented objects were “masked” onto the original images for fluorescent intensity quantification used in analysis. The DAPI channel, as above, was used to identify nuclei cell bodies that were identified through their size, shape, and fluorescent intensity in the channel associated with Synaptogreen. Neurites were then identified from regions not inclusive of cell bodies or nuclei that displayed a spindle-like morphology and fluorescence intensity above the background as labeled again by Synaptogreen. These objects were enhanced using the neurite enhance module within CellProfiler by the “Tubeness,” resulting in segmented objects designated “Neurites” corresponding to the neurite network in a field of view that did not include the cell body. This created a neurite network defining our region of interest in which to identify presynaptic vesicles marked by Synaptogreen which were identified as speckles having a mean intensity fluorescence above the background intensity. The median fluorescence intensities for identified presynaptic vesicles were reported and the number of presynaptic versicles was interrogated at time points 1 and 2. Data were aggregated per condition inclusive of a control, low Zn, and Zn supplemented, respectively. In control conditions, the median number of presynaptic vesicles per experiment measured was in excess of 93,000. In conditions pertaining to low Zn and high Zn, the median number of presynaptic vesicles measured per experiment was in excess of 18,000 and 120,000, respectively. Quantifications considered each conditions’ spontaneous signal reduction in fluorescent intensity above the background to establish the relative fold change difference between time points 1 and 2 as well as the overall number of presynaptic vesicles identified between the time points.

Statistics

Statistical analysis was performed with GraphPad Prism 6. Data are shown as mean ± SEM, geometric mean ± SEM, or median ± IQR as indicated. Outlier data were detected and removed using ROUT outlier detection with Q = 5%. For comparisons of two independent groups, Student’s t tests were used, whereas ordinary 1 way ANOVA and Tukey multiple comparison correction were used where more than two independent groups were present. Statistically significant differences are indicated in the figures by *p ≤ 0.05, **p ≤ 0.01 and ***p ≤ 0.001.

Results and discussion

Prenatal zinc-deficient mice show brain-region-specific alterations in the expression of proteins, with a significantly altered subset across all brain regions

ASD displays a complex genetic architecture with a multifactorial etiology involving 1200 + genetic associations [51]. This complexity has confounded clarity around the underlying causes of the disorder [52], especially given the variety of genetic CNVs or SNVs, in addition to insults derived from environmental sources. With that said, convergence on measurable behavioral traits has prompted researchers to describe this disorder by identifying common molecular pathways that gather contributing genetic associations into categories based on their impact on common biological processes [53–55]. These approaches, viewing individual contributions in the context of the networks perturbed, have been remarkably successful in illuminating cornerstones of the disorder. Indeed, analysis in this regard has identified common root pathways that have been implicated as pivotal in several brain disorders that were previously described as unrelated [56–58]. Notably, little work has been undertaken examining the environmental factors that give rise to ASD phenotypes, where there is a paucity of hypothesis-free data-driven assessment. Accordingly, we performed proteomics analysis of three brain regions (hippocampus, cortex, striatum) of PZD mice, a non-genetic mouse model with ASD-like behavior, and their controls (n = 10 per group) (Fig. 1A). All selected brain regions harbor zincergic synapses (glutamatergic terminals with zinc-containing neurotransmitter vesicles), allowing us to identify the common altered processes across brain regions with this type of zinc-signaling and to compare those with the biological processes that are affected in other models for ASD. We undertook an integrated approach to downstream analysis for evaluation to identify and prioritize key candidate ASD-associated genes disrupted by prenatal zinc deficiency. Firstly, we statistically evaluated each brain region in isolation to detect those proteins that have been significantly altered (Fig. 1B). We generated differentially expressed (DE) gene sets per region from the resultant proteomics data. While several proteins are changed in a brain-region specific pattern (Suppl. File 1), a subset of these belong to the same protein network that is changed across all brain regions (Fig. 1B, C, and Suppl. File 1). This protein cluster was identified as part of the vesicular transport and fusion machinery where AKAP5, BIN1, RAB3A, CPLX2, VAMP2, STX1A, STX1B, SYT1, YKT6, and MYO5A were DE in the PZD mouse brain regions in comparison with controls. Next, we subjected the DE gene sets of the PZD mouse model to over-representation pathway enrichment analysis [41] (Fig. 1B and Suppl. File 1). Consistently, among the most significantly altered biological processes we detected, membrane dynamics were heavily represented including: “Vesicle trafficking” (striatum and hippocampus), “Exocytosis” (striatum), “Synaptic vesicle transport” (striatum), “Vesicle-mediated transport” (striatum and cortex), “Membrane fusion” (cortex), “Vesicle fusion” (cortex), “Neurotransmitter transport” (cortex), and “Transmission across synapses” (hippocampus).

Fig. 1.

Identification of commonly affected pathways by three different analysis approaches. A Overview of experimental procedures. Three brain regions were prepared from PZD mice and controls. Protein lysates were analyzed using label-free mass spectrometry. B Normalized expression values were used to identify significantly regulation proteins in each region, threshold values of p = 0.05, and the interquartile range (IQR) of housekeeping gene expression used on the x and y axes, respectively. Proteins associated with synaptic vesicle recycling machinery including AKAP5, BIN1, VAMP2, STX1A, RAB3A, YKT6, SYT1, CPLX2, and STX1B were found to be modulated. C Hierarchical clustering analysis based on gene expression similarities identified the same pathways to be regulated with repeating genes highlighted. D Gene set enrichment analysis (GSEA) identified SNARE interactions in vesicular transport as the significantly modulated pathways common between cortex and hippocampus, with similar core genes being YKT6, STX1A VAMP2, STX1B, and two new genes identified SEC22B and STX12. A–D Red represents significantly upregulated, blue downregulated proteins. Shades of blue and red reflect the magnitude of regulation (based on FDR; light blue to dark blue = increasing significance; light red to dark red = increasing significance)

We further analyzed our data through unsupervised hierarchical clustering, again with downstream pathway enrichment analysis and gene-set enrichment analysis (Fig. 1C, D). Both statistical methods differ in their approaches, the former relying on clustering on shared expression, whereas the latter incorporates a defined gene set for comparison generated through a priori knowledge of biological processes [59]. Importantly, both approaches complement DE gene expression analysis as they include the entire expression data set rather than only those determined as significantly regulated. Again, this analysis yielded region-specific differences, with clusters of genes centering on pathways involved in synaptic vesical dynamics (Fig. 1C, D).

Lastly, we undertook pathway network analysis of our DE gene sets, grouping related clusters of shared biological processes. Congruently, gathering these networks together in shared modules revealed a striking concentration of processes around synaptic vesical dynamics, resulting in highly interconnected networks highlighting these pathways’ dysregulation (Fig. 2A, B). The identified alteration of vesicle trafficking and fusion in the hippocampus of PZD mice was also confirmed in western blot experiments with protein lysate using the upregulation observed for AKAP5 in mass-spec analyses as a representative protein (Supplementary Fig. S1).

Fig. 2.

Overarching pathways upregulated and downregulated in the PZD mouse model. A Key upregulated proteins mediating synaptic vesicle transport and synaptic signal transmission (i), shown in the associated protein network (ii, right panel), including key proteins identified in this study (Fig. 1C) such as YKT6, BIN1, VAMP2, STX1A, and AKAP5 (FDR ≤ 0.05). B Within the same pathway (vesicular transport) and in addition, neurogenesis/axonogenesis (i), interacting proteins (ii, right panel), such as STX1B, STX7, CLPX2, and RAB3A, are downregulated as well, hinting at a general dysregulation of the underlying biological process (FDR ≤ 0.05). A, B Shades of blue and red in A(i) and B(i), respectively, reflect the magnitude of regulation (based on FDR; light blue to dark blue = increasing significance; light red to dark red = increasing significance). The size of the circles reflects the amount of proteins defining the process that have been significantly altered (Gene set size). The colors assigned to proteins in panels (ii) show their association with biological pathways displayed in A(i) and B(i)

In addition, “Regulation of axon branching,” “Regulation of neuron apoptosis,” “Presynaptic organization,” and Postsynaptic organization” were affected in PZD mice (Fig. 1B). These findings align with previous studies reporting that zinc is an anti-apoptotic agent with an essential role in neurogenesis [48, 49]. Further, pre- and postsynaptic alterations through zinc deficiency have been mechanistically linked to impaired SHANK protein scaffold formation [50–52]. Thus, a picture emerges where prenatal zinc deficiency triggers a synaptic pathology during brain development. This occurs through processes that ultimately alter neurogenesis and synaptic (signal) transmission via abnormal synaptic vesicle and vesicular transport processes (Fig. 2A, B (i)), defined by the alteration of key proteins in PZD mice (Fig. 2A, B (ii)).

Functional assays confirm the identified synaptic phenotype and provide a novel drug target

Our data have revealed a dysregulation at the neuronal synapse within the PZD mouse, thereby implicating zinc levels as a critical regulator of this compartment, particularly focused on synaptic fusion machinery. To evaluate whether this effect is mediated directly in neurons, we exposed a rat neuronal cell culture model to low zinc conditions and evaluated by high content imaging (HCI) (Fig. 3A) the impact on neuronal synapse formation (using Synaptophysin, Homer, PSD95, and Shank3) and known synapse associated fusion proteins involved in synaptic vesicle fusion (STX1A, AKAP5) [60–62]. Consistent with previously published findings identifying zinc as a key regulator of the synapse, our HCI analysis revealed that there are fewer synapses formed as identified by HOMER/Synaptophysin (Supplementary Fig. S2A) and STX1A/SHANK3 co-labeling under low zinc (Fig. 3Ci). Interestingly, this effect was not observed for synapses marked by AKAP5/PSD95 (Fig. 3Bi), potentially associated with the late-stage of hierarchical disassembly of PSD95 in mature synapses and their reported insensitivity to Zn signaling [63]. A previous study has shown late recruitment of PSD95/SHANK1 to synapses rendering them less zinc sensitive, while early in synapse formation, synaptic SHANK2 and SHANK3 scaffolds predominate, which are regulated by Zn-binding [63]. Consistently, when we investigated whether the relative protein concentration in these compartments was altered in response to zinc availability, there was no change in the overall recruitment of HOMER, Synaptophysin, and PSD95 (Fig. 3Bii and Supplementary Fig. S2C). However, alterations were detected in the recruitment of AKAP5, STX1A, and SHANK3, in accordance with findings previously reported [26, 63] or presented here, further signifying their sensitivity to Zn signaling (Fig. 3Biii, Cii, and Ciii). Under low zinc availability, there were increased levels of AKAP5 recruited to synapse marked by PSD95 in accordance with overall levels of AKAP5 being higher under low zinc (Fig. 3Biii; Fig. 1B; Supplementary Fig. S1). Similarly, STX1A and SHANK3 levels in synapses were also increased under low Zn (Fig. 3Cii, 3Cii; Fig. 1B; Supplementary Fig. S1), though the overall number of detected synapses was decreased (Fig. 3Ci). Thus, these findings are in line with previously published data hypothesizing that zinc deficiency increases the selection pressure on synapses [63]. As a result, the pool of mature, potentiated, and mostly zinc-insensitive synapses is less affected, while immature synapses that are generally lower in the content of certain PSD proteins, like SHANK3, are more susceptible of low zinc levels. The elimination of these immature synapses results in a decrease in the total number of synapses. However, since mature synapses remain, it is likely that the detected STX1A and SHANK3 positive synapses label this pool of mature and stable synapses with, on average, higher levels of these proteins, thereby leading to an overall increase in protein detected.

Fig. 3.

Detection and relative quantification of neuronal synapses. Rat hippocampal neurons were either grown in full growth media (CTL) for 14 DIV or changed into growth media additionally supplemented with 50 µM CaEDTA for 14 DIV (CaEDTA). Additionally, as indicated, Zn deficient cells were treated with antagonist KT5720 at 2 µM for 2 h at 37 °C (CaEDTA_KT) or PKA agonist 8-Bromo-cAMP at 500 µM for 2 h at 37 °C (CaEDTA_cAMP) prior to measurement by fluorescent microscopy and downstream image analysis. A Overview of the image analysis pipeline created using Cell Profiler 3.1.4 software. All treatment conditions were analyzed using the same pipeline with exact same parameters to allow relative quantification of immunoreactive signals. B (i) Microscopy image quantification of the median number of synapses per neurite area defined by co-localizing AKAP5&PSD95 segmented punctae enclosed in MAP2 defined neurites. (ii) Microscopy image quantification of the PSD95 median fluorescence intensity within defined synapses under normal and CaEDTA + conditions. (iii) Microscopy image quantification of the AKAP5 fluorescence intensity of synapses detected under normal and CaEDTA + conditions. C (i) Microscopy image quantification of the median number of synapses per neurite area defined by co-localizing STX1A&SHANK3 segmented punctae enclosed in MAP2 defined neurites. (ii) Microscopy image quantification of the SHANK3 median fluorescence intensity within defined synapses under normal and CaEDTA + conditions. (iii) Microscopy image quantification of the STX1A fluorescence intensity of synapses detected under normal and CaEDTA + conditions. Data from three experimental repeats (n = 3) were analyzed for outliers through ROUT analysis Q = 5%, pruned and resultant data aggregated. Results are displayed by plotting the median value per field of view within each condition with the overall data median and interquartile range of these values represented by whisker plots. A minimum of 200,000 synapses and 140 FOVs were analyzed per condition. Data were analyzed by Student’s t tests comparisons of two independent groups, or ordinary 1 way ANOVA + Tukey multiple comparison correction for more than two independent groups: *P < 0.05; **P < 0.01; ***P < 0.001; n.s. not significant

Additionally, AKAP5 localization to presynaptic assemblies as marked by SNAP25, was reduced under low Zn. This indicates a dysregulation in the recruitment to nascent synapse formations (Supplementary Fig. Ai, S2Bii and S2C), reinforced by the lack of reduction in overall AKAP5 punctae and an increase in SNAP25 punctae. Together, these data identify AKAP5 trafficking and recruitment to synapses or presynaptic assemblies as being severely altered under low Zn. Intriguingly, AKAP5 is a key regulator of synaptic dynamics of the presynaptic compartment through PKA recruitment and subsequent regulation of the synaptic fusion machinery activity through STX1A phosphorylation [64, 65], linking it to the phenotypes observed for STX1A and SHANK3 synapse formation. To evaluate whether the effects observed on this fusogenic protein network could be regulated through PKA, we treated cells under conditions of low Zn availability with a PKA agonist (cAMP) and a PKA antagonist. Interestingly, activation of PKA was sufficient to rescue the effects of reduced synapse numbers present when Zn is absent and further increased both STX1A and SHANK3 localization to the synapse (Fig. 3Ci, Cii, and Ciii), whereas inhibition of PKA was not sufficient to offset the impact of low Zn availability.

Together, our data implicate zinc availability as a driver of the dysregulation of the presynaptic compartment, likely through a signaling axis involving PKA activity. These data further identify these pathways underpinning the aberrant regulation seen when environmental factors drive autistic-like phenotypes.

A subset of genes/proteins is significantly altered across all genetic models for ASD

Given that our data provided striking consistency in implicating the presynaptic compartment dynamics as a potential biological process dysregulated in the environmental origins of ASD, we reasoned that the processes are also the underlying causes of ASD in genetic models. Accordingly, to understand if these alterations are also present across various genetic origins of ASD, we investigated published data sets of ASD generated through genetic manipulations. Here, aligning with the postulate that there is a degree of convergence of affected molecular processes and pathways in ASD model systems with shared phenotypes, we identified both proteomics and transcriptomics data representing diverse genetic causes in reported models [30–34, 65]. We compared transcriptomics and proteomics data from six published studies by processing the data with a machine learning evaluation pipeline to generate a unified DE gene set (Fig. 4A). Machine learning algorithms for metadata analysis frequently yield greater coverage and accuracy than strict overlay analysis in identifying DE gene sets, due to the relative tolerance of this approach to the absence of data in any individual experimental data set. The performance of these methods is largely dependent on the quality of the classifier training set [66, 67].

Conversely, DE gene sets identified through overlay analysis represent a ground truth of genes or proteins common across all experimental data sets. Thus, here we implement a hybrid approach, where we exploited the stringency of overlay analysis to identify a core subset of DE genes common in ASD. These genes subsequently served as the training data for our machine-learning algorithm (Fig. 4A). This analysis generated a classification algorithm that showed exemplary performance as determined through AUC, ROC, and F-measure evaluation (Fig. 4B, C). We next subjected these metadata-derived DE gene sets to pathway network analysis, grouping related clusters of shared biological processes together in modules as we did previously. Significantly, these analyses yielded network modules dysregulated consistent with those observed in the brain of PZD mice, identifying networks of genes involved in synaptic dynamics and neurogenesis in addition to networks of biological processes previously reported as perturbed in ASD (Fig. 4D) [68–71].

Biological processes that are altered in genetic models of ASD overlap with those found in the environmental model

To further evaluate the associations between environmental and genetic causes, we compared the identified processes dysregulated in the PZD mouse with those identified through genetic models in a meta-analysis (Fig. 5A; Supplementary File 1). Again, these overlay analyses yielded common pathway processes centering on synaptic dynamics and synaptic vesicle fusion. Lastly, we compared the genetic and environmental DE gene sets with data from a list of genes generated through the evaluation of numerous ASD studies in addition to literature curated gene candidates as being most highly associated with ASD [72]. To allow for integration, protein IDs were converted back to the corresponding gene ID as appropriate. The DE gene lists associated with our meta-analysis that utilized both proteomics and transcriptomics data generated in-house or from public repositories. Overlaying the master list of ASD-associated genes with our own meta-analysis of both up- and downregulated genes from both genetic and non-genetic models of ASD, was able to identify common gene signatures. Intriguingly, commonly dysregulated genes generated highly interconnected interaction networks. These consistently focused on proteins known to be involved in biological processes regulating synaptogenesis and synaptic vesicular dynamics such as STX1A, VAMP2, BIN1, AKAP5, and CPLX2 (Fig. 5B, C).

Fig. 5.

Comparison of common pathways identified in genetic and non-genetic models and common proteins. A Identification of common upregulated and downregulated pathways, with the explicit list presented in Supplementary file 1, (FDR ≤ 0.05; Supplementary File 1). B, C Common upregulated (B) and downregulated (C) proteins between genetic mouse models and the prenatal zinc deficient (non-genetic) model, overlaid with a master list of genes generated through evaluation of numerous ASD studies (genome-wide association (GWAS) in addition to literature curated gene candidates as being most highly associated with ASD [72]. The identified proteins are represented in a network, and the associated biological pathway is shown. Reoccurring significantly modulated proteins such as STX1A, VAMP2, BIN1, AKAP5, and CPLX2 are highlighted. D, E Rat hippocampal neurons were either grown in full growth media (CTL) for 11 DIV or changed into growth media additionally supplemented with 50 µM CaEDTA for 11 DIV (CaEDTA) or cultured in normal media for 11 DIV and supplemented with 125 µM ZnSO₄ for 30 min prior to measurement by fluorescent microscopy and downstream image analysis D Representative images of a field of view and zoomed insets (ROI indicated with yellow box) that were quantified in E. Panels display segmented neurites, cell bodies, vesicles, and merged overlays as indicated at time point 1 and time point 2 under each of the experimental conditions tested E Microscopy image quantification of (i) the median fluorescent intensity per vesicle localized to neurites at time point 1 (T1) divided by median fluorescent intensity per vesicle localized to neurites at time point 2 (T2) and (ii) the median number of vesicles per neurite area at time point 1 (T1) divided by median number of vesicles per neurite area at time point 2 (T2). Data were aggregated across experimental repeats (> 2) and represents in excess of 36,000 vesicles quantified per condition. Data were analyzed for significance by ordinary 1 way ANOVA + Tukey multiple comparison correction: *P < 0.05; **P < 0.01; ***P < 0.001; n.s. not significant

Synaptic vesicular dynamics are regulated through Zn availability

Given that both genetic and environmental models for ASD converge on dysregulation of synaptic vesicular dynamics, we reasoned that Zn might have a functional impact on synaptic vesicle release. We again recapitulated our in vivo mouse model that develops under low zinc status for the whole duration of pregnancy by subjecting neurons to low zinc availability from the beginning of the experiment (DIV0) until the time when first neurons are mature (around DIV14) [73]. In contrast, given that brain zinc levels are tightly controlled and studies have shown that even high zinc intake for weeks does not substantially increase zinc levels in the brain [74], we performed a transient increase in zinc availability. Previous experiments have shown that the addition of zinc for 30 min results in zinc reaching synapses, inducing changes that are similarly seen after transient activity-dependent release of endogenous synaptic zinc [26]. Indeed, under low or high Zn availability, there was increased or reduced synaptic vesicular activity, respectively, consistent with the effects detected in proteins levels of the related fusogenic network (Fig. 5D, Ei, ii). However, given that we detect the average signal from multiple vesicles in the presynaptic compartment, it remains an open question whether a specific type or pool (reserve, ready-releasable, both) is affected.

Conclusion

In summary, here, we performed the first comparative analysis of genetic models for ASD in comparison with an ASD model generated through environmental manipulation. We identified specific proteins and biological processes affected by prenatal zinc deficiency that remain permanently altered in these mice despite their zinc status being quickly normalized after birth. Several of these alterations occur in a brain region-specific manner. However, interestingly, many proteins altered were found to be members of a biological network affected by prenatal zinc deficiency across all analyzed regions. Surprisingly, after comparing several genetic models for ASD that represent syndromic and non-syndromic forms of autism, we found protein members of the same network regulating synaptic vesicle dynamics that is also altered in each of the genetic mouse models. The same set is found altered in cell culture systems modeling ASD by manipulating the same genes targeted in the mouse models as well as additional ASD candidate genes. In particular, AKAP5-mediated PKA-dependent presynaptic vesicle exocytosis may be an important convergence point that has been so far underreported in ASD.

Here, we have identified the commonly altered process in the brain of all analyzed models characterized as ASD models through the common behavioral impairments they share with humans with ASD. We conclude that the identified process and protein network is key to the ASD pathology across a large share of ASD models and likely, individuals with ASD and, thus, represents a prime target for developing future pharmaceutical interventions.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

SM, AKS, AMG, and KMcG contributed to the study conception and design. Material preparation, data collection, and analysis were performed by SM, AKS, MF, AMG, and KMcG. AMG and KMcG wrote the first draft of the manuscript, and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.

Funding

The Autism Research Institute funds AMG. SM and SEH are funded by the Irish Research Council (IRC) postgraduate grants; GOIPG/2019/3693 and ESPG/2020/88, respectively.

Data availability

The data sets and analysis pipelines generated for this study are available from the corresponding author upon reasonable request.

Declarations

Conflict of interest

SM, AKS, SEH, MF, AMG, and KMcG declare that there is no conflict of interest regarding the publication of this paper.

Ethical approval

Animal experiments were performed in compliance with guidelines from the Federal Government of Germany for the welfare of experimental animals and approved by the Regierungspräsidium Tübingen and the local ethics committee (Ulm-University) ID:Number:1239 during AMG’s previous appointment at Ulm University.

Consent to participate

Not applicable.

Consent to publish

Not applicable.