Integrated Omic Analyses Identify Pathways and Transcriptomic Regulators Associated With Chemical Alterations of In Vitro Neural Network Formation

Abstract

Development of in vitro new approach methodologies has been driven by the need for developmental neurotoxicity (DNT) hazard data on thousands of chemicals. The network formation assay characterizes DNT hazard based on changes in network formation but provides no mechanistic information. This study investigated nervous system signaling pathways and upstream physiological regulators underlying chemically induced neural network dysfunction. Rat primary cortical neural networks grown on microelectrode arrays were exposed for 12 days in vitro to cytosine arabinoside, 5-fluorouracil, domoic acid, cypermethrin, deltamethrin, or haloperidol as these exposures altered network formation in previous studies. RNA-seq from cells and gas chromatography/mass spectrometry analysis of media extracts collected on days in vitro 12 provided gene expression and metabolomic identification, respectively. The integration of differentially expressed genes and metabolites for each neurotoxicant was analyzed using ingenuity pathway analysis. All 6 compounds altered gene expression that linked to developmental disorders and neurological diseases. Other enriched canonical pathways overlapped among compounds of the same class; eg, genes and metabolites altered by both cytosine arabinoside and 5-fluorouracil exposures are enriched in axonal guidance pathways. Integrated analysis of upstream regulators was heterogeneous across compounds, but identified several transcriptomic regulators including CREB1, SOX2, NOTCH1, and PRODH. These results demonstrate that changes in network formation are accompanied by transcriptomic and metabolomic changes and that different classes of compounds produce differing responses. This approach can enhance information obtained from new approach methodologies and contribute to the identification and development of adverse outcome pathways associated with DNT.

Over the last decade, there have been several important legal and/or policy decisions that have encouraged the development and use of in vitro approaches for identification of hazard associated with chemical exposure. These include legislative changes such as the REACH act in Europe (European Commission, 2006) and the Lautenberg Chemical Safety Act (United States, 2015) in the United States, and policy decisions such as the 2019 U.S. Environmental Protection Agency memo announcing the intention to develop a new approach methodologies (NAMs) workplan supporting the increased development and implementation of NAMs (Wheeler, 2019). These policy changes have in part been driven by the recognition that for some hazards, data on large numbers of chemicals are lacking, and the use of animal-based testing strategies is insufficient to address this data gap. Further, there is scientific recognition that, for a variety of reasons, animal-based toxicity studies often do not predict adverse outcomes in humans (NRC, 2007). As such, in the last 15 years, there have been significant efforts directed towards the development and implementation of in vitro approaches for hazard assessment.

Assessment of developmental neurotoxicity (DNT) hazard has relied heavily on animal-based protocols (EPA 870.6300 [U.S. EPA, 1998] and OECD TG 426 [OECD, 2007]) that assess numerous morphological and behavioral endpoints in offspring exposed, typically perinatally, to chemicals. However, a number of short-comings of in vivo DNT guideline studies have been identified (Tohyama, 2016; Tsuji and Crofton, 2012), including the high cost of conducting the study (approximately $1M/chemical), the large number of animals used (approximately 1000/study), and the amount of time it takes to conduct a DNT guideline study (up to 2 years). One review found that only 15 of 69 guideline DNT studies were used to set points of departure for risk assessments (Raffaele et al., 2010), whereas another review indicated that the learning and memory endpoints evaluated in the guideline DNT studies lack sensitivity (Vorhees and Makris, 2015). Further, these studies do not provide any mechanistic information regarding how a chemical may cause developmental neurotoxicity. Finally, because guideline DNT studies are “triggered” by other evidence of neurotoxicity under the Federal Insecticide, Fungicide and Rodenticide Act, and not at all required under the Toxic Substances Control Act, only a small fraction of chemicals used in the environment have been assessed for DNT (Crofton et al., 2012; Judson et al., 2009). Because of the large number of compounds for which no DNT hazard data are available, there is international agreement that novel approaches to DNT assessment, including in vitro screening approaches will be required in order to address this backlog, as current approaches are no longer sufficient (Fritsche et al., 2018).

The lack of data regarding the potential DNT hazard of thousands of compounds in commerce, coupled with the limitations DNT guideline studies discussed above has driven efforts to develop alternative assays for characterizing this hazard. Because development of the nervous system is complex, with multiple potential targets being expressed in different locations and at different times during development, a battery of in vitro assays has been proposed as a viable approach to assessing DNT hazard (Bal-Price et al., 2018). This battery consists of phenotypic assays that evaluate key neurodevelopmental processes such as neural progenitor cell proliferation and differentiation, neuronal and glial migration and differentiation, neurite outgrowth and synaptogenesis, and network formation (Bal-Price et al., 2018). This battery of assays consists of both animal (rodent) and human cell models. Although these assays are economical and efficient, they provide only limited information on potential mechanisms by which chemicals may be disrupting these critical neurodevelopmental processes. Information from these in vitro assays can be used in Integrated Approaches to Testing and Assessment to inform weight of evidence, fit-for-purpose, screening-level, and/or other regulatory decision scenarios (Paparella et al., 2020). The data from these in vitro assays can provide clues for the development of a more formalized framework by providing biological context to reduce the uncertainty of in vitro to in vivo extrapolation and ultimately lead to the development of possible mechanistic insights associated with DNT (Carlson et al., 2020). Such approaches may inform the development of adverse outcome pathways (AOPs, Watanabe et al., 2011) for developmental neurotoxicity, which are currently lacking (Bal-Price et al., 2015; Sachana et al., 2018).

To address this data gap, this study was conducted as a proof-of-concept, by examining transcriptomic and metabolomic responses following exposure to 6 reference chemicals using rat cortical cultures in a multiplexed format as part of the rodent neural network formation assay (NFA; Brown et al., 2016; Frank et al., 2017)). The NFA evaluates chemical effects on formation of functional networks of rodent neurons grown on microelectrode arrays (MEAs) from days in vitro (DIV) 0 to 12. To date, this assay has been used to evaluate the potential DNT of over 200 unique compounds (Brown et al., 2016; Frank et al., 2017; Shafer, 2019). It has high sensitivity and specificity (Shafer, 2019), and has been proposed as a component of the in vitro DNT assay battery (Bal-Price et al., 2018). The NFA collects multiple parameters describing chemical effects on the ontogeny of general activity, bursting activity, and coordinated network activity as well as viability (see Table 2 in Brown et al., 2016). These measures are multiplexed and, due to the noninvasive nature of MEA recordings, collected at 4 different time points during in vitro network development. This latter characteristic allows for the determination of toxicological “tipping points” (Frank et al., 2018; Shah et al., 2016), that demark when network development departs from normal trajectories. In this study, developing neural networks were exposed to concentrations of 6 chemicals and electrophysiological data were collected as previously described (Brown et al., 2016). At the end of exposure, media and cell lysates were immediately collected for metabolomic and transcriptomic analyses, respectively, in order to test the hypothesis that a combined -omic analysis would identify pathways, biomarkers, and/or critical genes that were altered by chemical exposure at concentrations that are known to alter neural network development. Such information could help to identify putative AOPs (Bal-Price et al., 2015; Sachana et al., 2018; Watanabe et al., 2011) by identifying relationships between key events (KEs) at the cellular and molecular level and linking them to the KE of altered network formation.

Table 2.

Differently Expressed Genes (DEGs) and Significant Metabolites

Compound	No. DEGs (14103; 14034 Mapped)	Total No. Metabolites
Cytosine arabinoside (CA)	1572	16
5-Fluorouricil (5FU)	304	17
Cypermethrin (CM)	8	6
Deltamethrin (DM)	10	5
Haloperidol (HP)	11	13
Domoic Acid (DA)	30	14

DEGs and metabolite counts after applying nervous tissue filter in IPA when selecting tissue and cell line settings for the expression analysis.

MATERIALS AND METHODS

All data files and scripts supporting this publication are available at DOI: 10.23719/1520859.

Rat primary cortical cultures.

Primary neocortical cultures were prepared from postnatal day 0 Long-Evans rat pups. All animal care and tissue harvesting procedures were approved in advance by the U.S. EPA, National Health and Environmental Effects Research Laboratory Institutional Animal Care and Use Committee. Detailed methods for the culture and plating density of cells used for this study have been previously described (Brown et al., 2016; Mundy and Freudenrich, 2000; Valdivia et al., 2014). Cortical cells were plated at a density of 150 000 cells per well on 48-well MEA plates (Axion M768-KAP-48) precoated with 0.05% polyethylenimine. Cells were placed in each well via a 25 μl media drop containing 20 ng/ml (Sigma-Aldrich, Cat No. L2020) laminin administered directly onto the MEA. Two hours after plating cells, 475 μl Neurobasal A media (Gibco, Cat No. 10888022) supplemented with B-27 (Gibco, Cat No. 17504044) was added to each well.

Compounds and dosing.

This study evaluated 6 compounds (Table 1) on 48-well MEA plates, all of which have evidence in the literature of DNT effects in humans or mammalian models (Mundy et al., 2015). The screening approaches used for this evaluation (Malo et al. 2006) tested each of the 6 selected “proof-of-concept” compounds a minimum of 2 times (Brown et al., 2016; Frank et al., 2017), each time using 3 different MEA plates from the same culture. Thus, each compound has been tested previously with 2 biological replicates consisting of 3 technical replicates each. Compounds (purity >95%) were stored as specified by the manufacturer (Table 1) and were prepared as previously described (Frank et al., 2017; Table 1). Briefly, a stock concentration of 30 mM was prepared and serially diluted by half-log increments to 1000× (between 0.03 and 10 mM) the desired final concentration, then diluted in-well on the 48-well MEA plate 2 h after cell attachment to make a 1× exposure solution (between 0.03 and 10 μM) with solvent maintained at 0.1% (vol/vol). Although all compounds were tested over the entire exposure range used in Frank et al. (2017), only selected concentrations were used for transcriptomic and metabolomic analysis due to limited resources for transcriptomic sample analysis. Solvent control wells containing dimethylsulfoxide or H₂O were included to monitor the normal development of cortical cells at final concentrations of 0.1% (vol/vol) and had no effect on network function (Brown et al., 2016). After dosing, the MEAs were returned to the incubator. After MEA recordings on DIV 5 and 9, the cells received a full media exchange with reexposure to the test compound.

Table 1.

Compounds Tested

Compound Name (Abbreviation)	Proposed Target	CAS No.	DTXSID	Tipping Pointa (μM [±95% CI])	Concentrations Tested (μM)	Min Viabilityb EC50 (μM)	Solvent Used	source
Cytosine arabinoside (CA)	DNA synthesis inhibitor	147–94–4	DTXSID3022877	0.046 [0.04–0.06]	1	0.84	DMSO	Sigma-Aldrich
5-Fluorouracil (5FU)	DNA synthesis inhibitor	51–21–8	DTXSID2020634	0.14 [0.10–0.17]	1	2.28	DMSO	Sigma-Aldrich
Domoic acid (DA)	Glutamate Receptor Agonist	14277–97–5	DTXSID20274180	0.21 [0.9–0.25]	0.3	NA	Water	Sigma-Aldrich
Cypermethrin (CM)	Voltage-gated Sodium Channel Modulator	52315–07–8	DTXSID1023998	0.28 [0.19–0.38]	3.10	NA	DMSO	Chem Service, Inc.
Deltamethrin (DM)	Voltage-gated Sodium Channel Modulator	52918–63-5	DTXSID8020381	0.05 [0.04–0.15]	3.10	NA	DMSO	Chem Service, Inc.
Haloperidol (HP)	Dopamine Receptor Agonist	52–86-8	DTXSID4034150	0.31 [0.15–0.49]	3	13.9	DMSO	Sigma-Aldrich

The table shows compound common name, CAS number, DTXSID identifier, the reported tipping point concentration, and the 95% confidence limits, concentrations tested in this study, minimum EC₅₀ values for viability (NA = less than 50% cytotoxicity was observed at the highest concentration tested, thus, no EC₅₀ value was determined), solvent used for dilutions, and compound source. The DTXSID identifier allows linking the compound to the Computational Toxicology Dashboard (https://comptox.epa.gov/dashboard; last accessed January 5, 2022), which provides further information on structure and other chemical properties.

Abbreviation: DMSO, dimethylsulfoxide.

^aFrank et al. (2017).

^bFrank et al. (2018).

Electrophysiology recordings.

A 15-min period of spontaneous network activity was recorded from each well on DIV 5, 7, 9, and 12 using a Maestro 768-channel amplifier, Middle-man data acquisition interface, and Axion Integrated Studio (AxIS) software v2.0 or later (Axion Biosystems, Atlanta, Georgia). Spontaneous activity was measured using a gain of 1200× and a sampling frequency of 12.5 kHz. The signal was filtered through a Butterworth band-pass filter (0.1–5000 Hz) prior to measuring online spike detection (threshold = 8× RMS noise) using the AxIS adaptive spike detector. All recordings were conducted at 37°C with 5% CO₂. Generally, root mean square (RMS) noise on each electrode was between 2 and 5 μV; electrodes with RMS noise exceeding 5 μV were grounded and excluded from recordings.

Normalizing area under the curve values from the MEA plates.

The normalized area under the curve (AUC) values for each compound, concentration, and endpoint were determined. The data used for activity were taken from the same cultures that were selected for sequencing and metabolomics. The AUC values were calculated as the trapezoidal AUC from DIV 2 to DIV 12 for each endpoint value from each well. The AUC values were normalized by dividing by the median AUC value of controls on the corresponding MEA plate (see Frank et al., 2017).

Metabolomic sample extraction, derivatization, and gas chromatography/mass spectrometry analyses.

Following the final DIV 12 MEA recording session, media samples were collected. A volume of 450 μl of media per well was transferred to prelabeled 1.5 ml centrifuge tubes and biological reactions quenched with 50 μl of acetone aliquoted into each tube on ice. The samples were then transferred to the −80°C freezer prior to gas chromatography/mass spectrometry (GC/MS) analyses. Media samples (either test or control) were thawed at room temperature for 1 h and metabolites were simultaneously extracted and derivatized for GC/MS analysis based on the methods of Tao et al. (2008). The organic layer was transferred to GC vials for analysis. GC/MS analyses were conducted on an Agilent 6890 GC interfaced to 5973 MS. Sample extracts were analyzed under full-scan (m/z 50–650), electron ionization-MS conditions.

Multivariate statistical analysis of metabolomic samples.

Following data collection, each individual chromatogram was exported as netcdf (.AIA) files. All files were imported into Metalign (Lommen, 2009) for spectral preprocessing and alignment following the developer’s guidelines for an Agilent GC/MS. Data were then filtered and truncated following the methods of Niu et al. (2014). Multivariate statistical analysis was performed by importing Excel spreadsheets of filtered chromatograms into Metaboanalyst (www.metaboanalyst.ca). Preliminary principal component analysis (PCA) was conducted for the entire dataset following Pareto scaled bins to test for outliers. The relative impact of a given chemical exposure was assessed by comparing score values for different treatment classes within a given PCA model. To make these comparisons, the score values for a given component (eg, PC1 scores in a PCA model) were tested for normality using the Kolmogorov-Smirnov test. If found normal, then a 1-way ANOVA test, with Tukey’s post hoc comparison, was conducted. This allowed the determination of significantly different means (95% confidence) of score values for different treatment classes. All statistical tests on score values were performed with Minitab 16 (Minitab, Inc., State College, Pennsylvania). Statistical significance was ascribed to score values with a p value <.05.

Univariate statistical analysis of metabolomics.

To determine the significance of metabolite changes resulting from the exposures, volcano plot analysis was conducted with Metabolanalyst coupling fold change (± 1.8) and t test (p value ≤.05) analyses. Spectral features identified using this approach were putatively identified using commercially available mass-spectral databases. (Supplementary Figs. 1–4).

Isolation of RNA and sequencing.

Total RNA was collected using the Qiagen RNeasy Plus Kit. Two to three wells of identically treated cultures were pooled, representing 1 sample. The experiment was repeated on 2 separate cultures. Samples were pooled by serially treating identically treated wells with a volume of 350 μl of lysis buffer with β-mercaptoethanol (Buffer RLT, Qiagen). This volume was added (as described in RNeasy Plus Kit instructions) to a well and mixed for complete lysis before being transferred to the next (identically treated) well(s) to be pooled together. The lysate sample of pooled wells was transferred to a clean 1.5 ml RNase-free plastic tube and the homogenized lysate sample was then taken through the gDNA Eliminator spin column, and RNA was then precipitated and eluted as described in the RNeasy Plus Kit instructions. The Qubit RNA BR (Broad-Range) Assay Kit was used to measure RNA concentration on the Qubit Fluorometer. Samples were stored in a −80°C freezer until transfer to the NHEERL Genomics Research Core for sequencing. 500 ng total RNA was loaded onto an Apollo 324 automated liquid handler to prepare polyA-selected RNA-seq libraries following manufacturer instructions for a PrepX PolyA mRNA isolation kit (Wafergen). mRNA libraries were denatured with NaOH and loaded onto an Illumina NextSeq 500 bench top sequencer at a final dilution of 1.8 pM with 5% PhiX. Reads were delivered in FastQ file format at an average coverage depth of 15.8 (range = 11.0–23.1) reads per sample.

Analyzing sequences.

The sequencer output data files (FASTQ files) were “trimmed” to remove sequence adapters and other low-quality bases from the ends of reads in Partek Flow (Build v4.0.16.0128, St Louis, Missouri). Read ends were trimmed by quality (Q-scores ≤ 7) and reads with length ≤ 25 or composed of any 1 base at > 95% frequency were removed by Partek Flow. The files were inspected for pre-alignment and postalignment quality control. The sequence was then aligned to a rat reference genome (Rn6) and transcripts were quantified to the transcriptome (RefSeq) by using Spliced Transcripts Alignment to a Reference (STAR) software (STAR2.4.1d, Illumina). Inspection of the postalignment QC metrics was conducted, the reads were mapped to the transcriptome to determine which genes the sequence represents, and annotation (gene level) was completed in the Partek Flow software. Transcript (gene) abundance levels from reads aligned to a rat genome were quantified using expectation-maximization algorithm (E/M) and normalized by dividing by counts per million (CPM) for the sample, adding 1 to all raw counts, and then log₂ transforming. After normalization, a batch effect due to different sequencing runs, which separated rep 1 from reps 2 and 3, was observed. To remove this effect, a general linear model was applied within the Partek Flow environment. Finally, differentially expressed genes (DEGs) were identified using the default Partek Gene-Specific Analysis algorithm on pairwise groups (control vs treated) followed by pathway mapping with statistical significance which was ascribed to score values with a false discovery rate (FDR) value < 0.05 and absolute fold change value of ≥ 2.0. No additional models were enabled to lognormally fit the data.

Integrated omic analysis.

To analyze both the metabolomic and the transcriptomic data, a combination of 2 supervised approaches was used within ingenuity pathway analysis (IPA; Qiagen; see Figure 1). First, an independent analysis of both genes and metabolites was conducted. DEG gene lists (1.8± fold; p value <.05) peak metabolite lists (± 1.8 fold; p value <.05) were also uploaded to IPA and mapped to pathways independently of each other for each compound. Second, following the independent core analysis for each compound, a combined list of genes and metabolites was created for each compound. Utilizing the canonical pathway and upstream regulator analysis functions in IPA, these lists were evaluated for canonical pathway enrichment and potential transcription factor regulators relative to neuronal development. It should be noted that for deltamethrin (DM) and cypermethrin (CM), different treatment concentrations were analyzed for transcriptomics and metabolomics due to a clerical error. As insufficient sample and resources were available to repeat the analysis, the results of the combined analysis used the available data. For independent and integrated analyses, IPA automatically detected entrez IDs for the transcriptomic data using the rodent database and the metabolomic data using the human metabolome database. In both approaches, we focused our analysis to tissues and cells specific to the nervous system, as well as human mapped pathways. Nervous system-related pathways and regulators included select IPA pathway categories like Neurotransmitters and Other Nervous System Signaling, Organismal Growth and Development, Intracellular and Second Messenger Signaling, Growth Factor Signaling, Cytokine Signaling, Cellular Growth, Proliferation and Development, Cell Cycle Regulation, and Disease Specific. These nervous system-related pathway categories were defined by low p values (<.05) and large magnitude absolute changes in z-score relative to the significance threshold. Fisher’s Exact test p value was used to find the probability that the combined gene, metabolite list would coincide with the IPA pathway list. Upstream analysis function of potential regulators of our lists was based on the number of differentially expressed downstream genes and metabolites in the list and quantified using a Z-score and p value. The Z-score is a correlation measure of how consistently the direction of expression changes in the gene metabolite list matches the direction of change from the annotated literature for targets in the regulatory or biological group.

Figure 1.

Experimental approach. Chemical concentrations for primary rat cortical cell treatment were selected based on network activity during microelectrode array-network formation assay. The media was used for metabolomic peak identification and RNA was extracted from the plated cells for transcriptomic analysis. Differentially expressed genes and metabolites were assessed independently and then secondarily in an integrated analysis for the 6 chemicals. Key molecular events were identified via ingenuity pathway analysis and were further investigated to determine relatedness in support of future putative adverse outcome pathway development.

RESULTS

The overall approach to this study is depicted in Figure 1. Following treatment of developing cortical neuronal networks with chemicals that previously were shown to alter network development (Brown et al., 2016; Frank et al., 2017), RNA was isolated from the cell pellet for transcriptomic analysis and media was saved for metabolomic analysis (Figure 1, step 1). At the highest level, we used IPA to probe neurodevelopmental disease (NDD) Categories associated with the responses to chemical treatment that resulted in altered network development. Separate transcriptomic and metabolomic, as well as integrated analyses were conducted (Figure 1, step 2). The results of analyses were then utilized to identify canonical pathways and upstream regulators associated with altered neuronal network formation following compound exposure (Figure 1, step 3). Last, we investigated the molecular relationships identified between the MEA and multi-omic analyses to provide information on the mechanisms underlying disruption of the neurodevelopmental process (Figure 1, step 4).

Compounds Disrupt Neuronal Network Formation at the Selected Concentrations

The compounds tested in this study were selected based on previous activity in the NFA (Brown et al., 2016), and the concentrations tested in this study are shown (Table 1) relative to their toxicological tipping point (Frank et al., 2018) and EC₅₀ value for cytotoxicity (Frank et al. 2017); these range from near the cytotoxicity EC₅₀ (cytosine arabinoside [CA] and 5 fluorouracil [5FU]) to intermediate between these concentrations (DM/CM/haloperidol [HP]) to near the tipping point concentration (domoic acid [DA]). The toxicological tipping point is the concentration at which network development no longer exhibits recovery from perturbations caused by chemical treatment (ie, homeostasis can no longer compensate for toxicological responses; Frank et al., 2018; Shah et al., 2016). Concentrations were chosen to fall between tipping point and cytotoxicity EC₅₀, as representative of network dysfunction before overt cell death. AUC values for 17 parameters of network formation (eg, mean firing rate, burst rate, interburst interval [IBI], network bursts, etc.) were altered by treatment with all 6 compounds (Figure 2). Overall, most, but not all parameters of network function were reduced by treatment, and distinct profiles of activity are evident across the different treatments. As these effects were generally consistent with previous results (Brown et al., 2016; Frank et al., 2017), no statistical analysis was conducted. Not surprisingly, CA, which was tested at a concentration near its EC₅₀ value for cytotoxicity produced the most robust changes (all decreases) in network function across the most network parameters. Interestingly, 5FU, which was also tested near its EC₅₀ value for cytotoxicity caused more modest changes in network parameters. Although changes induced by DM, CM, and HP were not as pronounced as for the DNA-synthesis inhibitors, there were clear patterns of response, and in the cases of the pyrethroids (DM and CM), showed concentration-dependence as well as a similarity in the profile of effects on network parameters. For example, CM and DM at the lower concentrations exhibited slight increases in IBI and interspike interval in network spikes (ISNetSP) and to a lesser extent, interspike interval (ISI), whereas HP depressed all parameters except for the percentage of spikes in network spikes (%SpNetSP). In contrast, DA, which was tested at a concentration close to its tipping point, showed small effects on the network function parameters, with the largest effect being an approximately 50% decrease in mutual information (MutInfo).

Figure 2.

Effects of compounds on network formation. Radar plots of the effects of the 6 compounds tested on neuronal network formation in rat primary cortical cultures. The DNA synthesis inhibitors cytosine arabinoside and 5-fluorouricil are plotted together as are the voltage-gated sodium channel modulators cypermethrin and deltamethrin. In each plot, the different parameters of network development are arranged around the perimeter of the plot, with the center of the plot being zero. The area under the curve for each parameter was determined as described in Frank et al. (2017), and is plotted as a percent of control from wells on the same MEA plate. As one proceeds around the plot in a clockwise direction, the parameters evaluate overall activity (MFR- #ABE), then bursting activity (ISI-IBI), Network spiking activity (#NetSp-Mean Sp in NetSp) and finally measures of connectivity (CorrCoeff and MutInfo). Abbreviations: MFR, Mean Firing Rate; BPM, Bursts per Minute; #AE, number of Active Electrodes; #ABE, number of Bursting Electrodes; ISI, interspike interval in a burst; %SIB, percent of spikes that are in a burst; Bdurn, burst duration; IBI, interburst interval; #NetSp, number of network spikes; NetSpPeak, Network Spike Peak; NetSpDurn, network spike duration; %SpNetSP, percent of spikes that are in a network spike; SDNetSpDurn, standard deviation of network spike duration; MeanSp in NetSp, mean number of spikes in network spike; CorrCoeff, correlation coefficient (r); MutInfo, normalized mutual information. See Brown et al. (2016) for additional details about network parameters. There are n = 7–10 replicates for each treatment/dose group.

These changes are consistent with effects previously demonstrated (Brown et al., 2016; Frank et al., 2017) for these compounds on metrics of neural network formation following exposure for 12 DIV. Following the recordings on DIV 12, media samples were collected for metabolomic analysis, and RNA was collected from cell lysates for transcriptomic analysis.

Identification of DEGs Corresponds to the Magnitude of Effects on Network Activity

Table 2 presents the number of identified DEGs associated with nervous tissue for each compound. The number of DEGs generally correlated with the magnitude of chemical effects on network function: 1572 and 304 genes were differentially expressed following treatment with DNA synthesis inhibitors CA and 5FU, respectively. These treatments also caused the most robust alterations in network formation (Figure 2). Although functional changes following DA treatment were subtle, 30 DEGs were identified following treatment with this glutamate receptor agonist. A total of 8 and 10 DEGs were identified following treatment with voltage-gated sodium channel (VGSC) modulators CM and DM, respectively, and 11 following treatment with the dopamine receptor antagonist HP.

Univariate Statistical Analysis Provides an Untargeted Approach to Identify Metabolites

Both PCA and partial least squares discriminant analysis (PLS-DA) were performed in MetaboAnalyst, to identify outliers and allow visualization of each compound by chemical class, respectively. The PLS-DA models showed concentration-dependent class separation. These metabolites were then identified by a Student’s t test (p ≤.05) comparison across each concentration compared with solvent matched controls tested as described earlier.

Transcriptomic, Metabolomic, and Combined Analyses Identify Molecular and Cellular Function and NDD Categories Associated With Functional Alterations in Network Development

Treatment with the 6 compounds used here resulted in different transcriptomic and metabolomic profiles (Figure 3). The IPA disease analysis tool of Ingenuity identified functional annotations for both DEGs and metabolites, and the integration of genes and metabolites, providing 3 major characterizations. The identified cellular and molecular events were classified by molecular and cellular function, and diseases and disorders, representing 2 hierarchical levels of biology from high levels (diseases and disorders) to molecular levels to aid in translation of results. These characterizations were captured for each compound relative to their metabolomic, transcriptomic, and integrated responses inclusive of both transcriptomic and metabolomic data entry. Because CA produced both the most robust changes in DEGs and metabolites, the presentation of results will focus on this exposure as an example and compare and contrast effects of the other compounds to CA. The analyses of all compounds are available in the Supplementary Materials.

Figure 3.

Clustering of each compound by chemical class. Principal component analysis (PCA) show concentration-dependent class separation of identified metabolites (left) and DEGs (right). Each data point represents the average response of solvent matched control cells subtracted from the individual response of exposed cells for each chemical. Note clear separation of treatment groups. See legend on each panel for compound identification.

Transcriptomic analysis in IPA identified 5 cellular and molecular functions that were associated with CA exposure during network development; cell death and survival, cell morphology, cell growth and proliferation, cellular development, and cellular movement (Figure 4). Like the transcriptomic analysis, metabolomic analysis also identified cellular growth and proliferation. However, the metabolomic analysis identified cell-to-cell signaling and interaction, amino acid metabolism, drug metabolism, and molecular transport. When transcriptomic and metabolomic information were combined, this analysis identified cellular development, cellular movement, cellular growth and proliferation, cell morphology, and cell death and survival as molecular and cellular pathways altered by CA.

Figure 4.

Results of transcriptomic, metabolomic, and combined analyses of cellular and molecular function categories. Statistically significant molecular and cellular function categories resulting from transcriptomic, metabolomic, and combined analysis following exposure to the indicated compounds that disrupt network formation. A common set of 5 disease and disorder categories were altered by the compounds. Values shown are the —log (p values), whereas the vertical dotted line represents the level of significance (p <.05).

The cellular and molecular functions and disease categories that were associated with treatment for the remaining 5 compounds had a high degree of overlap (Figure 4) with those identified for CA. In the combined analysis, cell-to-cell signaling and interaction was identified for 4 compounds (CM, DM, HP, and DA), while cell morphology (CA, CM, and DM), cell death and survival (CA, 5FU, and CM), and cell function and maintenance (5FU, DM, and HP) were each associated with 3 compounds.

When examining NDD categories associated with the transcriptomic and metabolomic responses to CA, transcriptomic responses identified 5 categories; neurological diseases, organismal injury and abnormality, psychological disorders, hereditary disorders, and skeletal and muscle disorders, whereas the metabolomic responses identified 3 of these 5 (neurological disease, psychological disorders, and organismal injury and abnormality) and the additional category of cardiovascular disease (Figure 5). In the combined analysis the disease categories of neurological disease, organismal injury and abnormalities, psychological disorders, hereditary disorders, and skeletal and muscle disorders were significant (Bonferroni’s test, p <.05; Figure 5). These categories were prevalent across all compounds in the combined analysis; with all 5 identified for 5FU, 3/5 for DM and DA, and 2/5 for CM and HP (Figure 5). Organismal injury and abnormalities were identified for all 6 compounds, whereas neurological disease was identified for 5/6 compounds (DA being the exception) and psychological disorder for 3/6 (CA, 5FU, and DA).

Figure 5.

Results of transcriptomic, metabolomic, and combined analyses of disease and disorder categories. Statistically significant disease and disorder categories resulting from transcriptomic, metabolomic, and combined analysis following exposure to the indicated compounds that disrupt network formation. A common set of 5 disease and disorder categories were altered by the compounds. Values shown are the —log (p values), whereas the vertical-dotted line represents the level of significance (p <.05).

Identification of Canonical Pathways Altered Following Exposure

Our analysis identified 5 canonical pathways associated with CA: CREB signaling in neurons, axonal guidance signaling, human embryonic stem cell pluripotency, hepatic fibrosis/hepatic stellate cell activation, and osteoarthritis pathway. It is at the level of canonical pathways that even more divergence is observed between the compounds tested (Table 3). The only pathway common between CA and 5FU was axonal guidance signaling (Figure 6). In contrast, this pathway was not among the top pathways identified for the remaining 4 compounds, but tRNA charging was common among these 4 chemicals and 5FU. DA, CM, DM, and HP were associated with several different biosynthesis and degradation pathways, although none of them were common across all 4 of these compounds (Table 3).

Table 3.

Canonical Pathways Associated With Altered Network Formation Using Integrated Pathway Analysis

Cytosine Arabinoside (CA)	5-Fluorouracil (5FU)	Domoic Acid (DA)	Cypermethrin (CM)	Deltamethrin (DM)	Haloperidol (HP)
• CREB signaling in neurons	• WNT/β-catenin neurons	• tRNA charging	• Alanine biosynthesis III	• (S)-reticuline biosynthesis II	• Phenylalanine degradation I
• Axonal guidance signaling	• Axonal guidance signaling	• Phenylalanine degradation IV (mammalian, via side chain)	• Glycine biosynthesis III	• Tyrosine degradation I	• Threonine degradation II
• Osteoarthritis pathway	• tRNA charging	• Glycine biosynthesis III	• Catecholamine biosynthesis	• 4-hydroxybenzoate biosynthesis	• Tyrosine biosynthesis IV
• Human embryonic stem cell pluripotency	• Hepatocyte Growth Factor (HGF) signaling	• Purine nucleotides de novo biosynthesis II	• tRNA charging	• 4-hydroxyphenyl-pyruvate biosynthesis	• Glycine biosynthesis III
• Hepatic fibrosis/hepatic stellate cell activation	• Nicotinamide adenine dinucleotide (NAD) signaling pathway	• Tyrosine biosynthesis IV	• Glycine Biosynthesis III	• tRNA charging	• tRNA charging
	• Hepatic fibrosis/hepatic stellate cell activation		• Thio-molybdenum cofactor biosynthesis

Figure 6.

Putative adverse outcome pathway linked to cytosine arabinoside exposure. Illustrated here is a putative adverse outcome pathway (AOP) for altered neurophysiology that is based on data (blue outline) from the microelectrode array network formation assay as well as other in vitro developmental neurotoxicity assays (see Harrill et al., 2018) after cytosine arabinoside exposure. Underneath the putative AOP are the cellular and molecular categories (orange), disease and disorder categories (light blue), and canonical pathways (light green) that were identified by -omic approaches following treatment with cytosine arabinoside. Overlap of the orange and light green bars is approximate. Key event relationships that are well established are indicated with solid arrows, while those that are postulated are indicated by dashed arrows.

Identification of Upstream Regulators Following Exposure

Analysis of upstream regulators for the test chemicals revealed both common regulators as well as regulators unique to each compound (Table 4). SOX2 was common to both CA and 5FU. Although LOC102724788/PRODH was common to CM and DA, otherwise, each compound had different profiles of upstream regulators. All of the identified upstream regulators have important roles in neurodevelopment, as indicated in Tables 5 and 6.

Table 4.

Top Upstream Regulators Associated with Chemical Exposures That Alter Network Formation

	Cytosine Arabinoside (CA)	5-Fluorouracil (5FU)	Domoic Acid (DA)	Cypermethrin (CM)	Deltamethrin (DM)	Haloperidol (HP)
Transcriptomic regulators	CREB1↑	SOX2↓	LOC102724788/PRODH	ARX	ATXN1	CREB1
	SOX2↓	TCF7L2↓	SLC1A2	LOC102724788/PRODH	GLS	CA9
	ZBTB17	PRMT1↓	ELAVL4	SLC1A2	HPRT1	CPT1B
	S100A8↑	PPP3CA	GLS	KLF9	ELAVL4
		NOTCH1		SLC23A2
Metabolomic regulators	β-estradiol		Kynurenic acid		K+	Afatinb GSKJ4

Predicted activation (↑) and inhibition (↓). Lack of a directional arrow indicates that activation or inhibition could not be predicted.

Table 5.

Transcriptomic Upstream Regulators and Their Roles in Neurodevelopment

Upstream Regulator	Gene/Protein Name	Neurodevelopmental Relevance	Reference
ARX	Aristaless-related homeobox	ARX is a paired-like transcription factor that serves as an essential regulator in a network of genes needed for normal forebrain patterning and growth	Lim et al. (2019)
ATXN1	Ataxin 1	Functions as a transcriptional repressor and modulates transcription that disrupts the dopaminergic pathway	Opal (2010)
CA9	Carbonic Anhydrase 9	Can influence the function of proton-sensitive membrane proteins involved in neuronal signaling	Ruusuvuori et al. (2013)
CREB1	CAMP Responsive Element Binding Protein 1	Nuclear transcription factor that regulates neuronal plasticity	Chang et al. (2009)
CPT1B	Carnitine palmitoyltransferase 1B	Enzyme required for the net transport of long-chain fatty acyl-CoAs and its dysregulation has been found to be a key player involved in neuronal function survival.	Zhang et al. (2015)
ELAVL4	ELAV Like RNA Binding Protein 4	Regulates the fate of neuronal mRNAs; implicated in neuronal plasticity, namely in recovery from axonal injury, learning and memory, and multiple neurological diseases	Bronicki and Jasmin (2013)
GLS	Glutaminase	Plays a role in excitatory synapses and participates in migration, proliferation, and differentiation of neural stem progenitor cells	Benavides-Rivas et al. (2020)
HPRT1	Hypoxanthine Phosphoribosyltransferase 1	Regulates molecular and cellular functions during neuronal differentiation as well as mechanisms that determine neuronal/glial cell fate decisions during neurogenesis	Kang et al. (2013)
KLF9	Krüppel-like factor 9	Inhibits neurite outgrowth/axon regeneration and promotes neuronal development/plasticity	Avci et al. (2012) and Bonett et al.(2009)
LOC102724788/PRODH	Proline dehydrogenase 2	Regulates proline catabolism, which is vital for normal Central Nervous System (CNS) functioning; and encodes proline dehydrogenase, which is involved in neuromediator synthesis in the CNS	Suntsova et al. (2013)
NOTCH1	Notch homolog 1	Inhibits neurite outgrowth in postmitotic primary neurons	Berezovska et al. (1999)
PPP3CA	Protein Phosphatase 3 Catalytic Subunit Alpha	Essential in the transduction of intracellular Ca2+-mediated signals	Myers et al. (2017)
PRMT1	Protein Arginine Methyltransferase 1	Significant regulator of glial cell function during postnatal brain development	Hashimoto et al. (2021)
SLC1A2	Solute Carrier Family 1 Member 2	Essential for glutamatergic neurotransmission	Penadés et al. (2020)
SOX2	SRY-Box Transcription Factor 2	Marker of neuroprogenitor status; expression inhibits neuronal differentiation	Graham et al. (2003)
SLC23A2	Solute Carrier Family 23 Member 2	Transporter that provides high ascorbate concentration in CNS; ascorbate is responsible for myelin formation, protection against glutamate toxicity and synaptic potentiation	May (2012)
S100A8	S100 Calcium Binding Protein A8	An anti-inflammatory protein that contributes to apoptosis and antiproliferation	Passey et al. (1999)
TCF7L2	Transcription Factor 7 Like 2	Master regulator that regulates different stages of postmitotic differentiation in the thalamus	Lipiec et al. (2020)
ZBTB17	Zinc Finger And BTB Domain Containing 17	Involved in cell cycle regulation	Lawir et al. (2017)

Table 6.

Metabolomic Upstream Regulators and Their Roles in Neurodevelopment

Metabolite/Chemical	Neurodevelopmental Relevance	Reference
B-estradiol	Plays a role in the regulation of synaptogenesis	Kretz et al. (2004)
Kynurenic acid	Impacts cortical development	Bagasrawala et al. (2016)
K+	Essential for neuronal signaling and sphingolipid synthesis; sphingolipids are key players in cell proliferation, differentiation, and transformation	Ferland (2012)
Afatinb	Potent and selective Epidermal Growth Factor Receptor (EGFR)-tyrosine kinase inhibitor; EGFR is differentially expressed in neurons and glia during development and is critical to cell migration, maturation, survival, and proliferation	Chen et al. (2019)
GSKJ4	Inhibits cellular proliferation through prompting cell death and by suppressing the cell cycle	Mandal et al. (2017)

DISCUSSION

This study demonstrates that a multi-omic approach can identify critical pathways involved in neurodevelopment in vivo at concentrations of compounds that induce changes in network formation in vitro. Exposure to different compound classes during in vitro network formation for 12 days resulted in different patterns of effect on network parameters. Further, different classes of compounds evoked different -omic responses, with compounds from the same class (eg, CA and 5FU) evoking similar responses while compounds from different classes (eg, DM and HP) evoked differential responses. This result suggests that these compounds are disrupting neural network formation through alteration of different pathways, but potentially with class-specific AOPs.

The need for novel methods to evaluate chemicals for DNT hazard is now well-established. In the last decade, a variety of in vitro assays based on the ability of chemicals to disrupt key neurodevelopmental events have been developed. Many of these neurodevelopmental events, such as proliferation, neurite outgrowth, synaptogenesis, etc., can be considered as KEs that could be involved in AOPs for compounds with DNT hazard potential. Among these assays, the neural NFA (Brown et al., 2016; Frank et al., 2017; Shafer, 2019) is unique in that it evaluates neuronal function, rather than structural or molecular endpoints. However, the NFA as currently performed does not provide any mechanistic information. Given that “alterations in network function” are listed as a KE in 5 AOPs found on the AOPWiki (https://aopwiki.org/events/386; last accessed January 5, 2022), understanding in more detail the pathways underlying alterations in network formation can serve to strengthen the existing AOPs involving this KE, and identify additional putative AOPs.

All 6 compounds tested in this study have been reported to cause DNT in mammals (Mundy et al., 2015). At the highest level of organization evaluated in this study (NDD), all 6 compounds were associated with a highly common set of categories for both cellular and molecular function as well as disease and development disorders (Figs. 4 and 5). The categories were consistent across the transcriptomic, metabolomic, and combined analyses, indicating common and robust responses to compound treatment. This information establishes an important linkage from the in vitro work conducted here to in vivo, human adverse neurodevelopmental outcomes by demonstrating that the -omic changes induced in vitro are disrupting some of the same processes that result in human NDD states. It is important to note that we are not postulating that exposure to these chemicals would cause a specific NDD such as (eg) schizophrenia, but rather that some aspects of disrupted neurodevelopment that occur following exposure to these chemicals during development have commonalities with the perturbations that result in NDD states. That chemical disruption of nervous system development and neurodevelopmental diseases has commonalities at the disease and developmental disorder level is not unexpected, nor is it surprising that common cellular and molecular categories were impacted by these chemicals. It is noteworthy that the overall functional response of the networks was also consistent for all compounds, being generally decreased in both the limited analysis in this study, as well as the more complete analyses in previous studies (Brown et al., 2016; Frank et al., 2017). Thus, there is general agreement between the functional and -omic responses at the highest levels.

When examining canonical pathways associated with the alterations in network function; however, some divergence begins to appear between the different classes of compounds. Axonal guidance signaling was associated only with the DNA synthesis inhibitors (CA and 5FU), whereas tRNA charging was common among 5 of the compounds. This could be a result of an integrated stress response (Kapur et al., 2017). Alteration of biosynthesis pathways was a common finding among the remaining 4 compounds (CM, DM, DA, and HP); however, all of the specific pathways identified were unique to each compound. Canonical pathway and upstream regulator identification for each compound may play a role in different KEs of neurodevelopment, which are all essential to formation of neuronal networks (Reiner et al., 2016). A similar theme was identified at the level of upstream regulators associated with compound exposure, with common regulators as well as unique regulators associated with each compound or among compounds with similar mechanisms. SOX2 was common to both of the DNA synthesis inhibitors, whereas LOC102724788/PRODH was common to CM and DA, otherwise the upstream regulators were unique for each compound.

Overall, these results help us build putative AOPs based on the multi-omic responses to compound exposures studied here. At the highest level of organization, they converge upon a common physiological response (decreased activity in network formation patterns) and disease categories (eg, neurological disorders, organismal injury, and abnormalities). As one begins to explore further the canonical pathways contributing to these disease categories, some common pathways are found (eg, axonal guidance), but more importantly different pathways begin to emerge that can be associated with different compound classes and their modes of action. This hypothesis is further supported by the analysis of upstream regulators, which show some commonalities across chemicals, but more broadly identified unique patterns of response for each compound. These results suggest that the actions on network function are the result of not a single AOP contributing to the decrease in activity, but rather a network of AOPs converging to impact physiological function as measured by the MEA. It is likely that decreases in network function following exposure to these compounds are the summation of several different events, such as decreases in neurite outgrowth, synaptogenesis, synaptic support from glia, and even differing levels of cytotoxicity/cell death. Thus, it could be expected that an interconnected AOP network will evolve with additional studies.

This information can then be combined with information from other sources to begin to develop putative AOPs for further exploration (Figure 1, step 4). An online analysis tool DAVID could be used as a complimentary source of interest in future projects to compare and potentially confirm which dysregulated signaling pathways are enriched relative to the IPA identified key upstream regulators and their predicted activation. Both CA and 5FU are known inhibitors of proliferating cells by inhibiting DNA synthesis, CA also causes oxidative stress in neurons (Geller et al., 2001). These may serve as initial KEs, where exposure of the cortical cultures used here results in decreased proliferation of the glial cells and increased oxidative stress in neurons. Further, according to Harrill et al. (2018), both CA and 5FU increased apoptosis and decreased proliferation in neuroprogenitor cells, and this also has been reported in primary cultures of cortical neurons (Geller et al., 2001). Neurite outgrowth has also been reported to be decreased by CA (Harrill et al., 2018) at concentrations similar to those used in this study. It has been demonstrated previously that compounds that decrease neurite outgrowth in cortical cells can also decrease network formation (Robinette et al., 2011). These effects on apoptosis, proliferation, and neurite outgrowth were nonselective, indicating that they were associated with loss of cell viability (Harrill et al., 2018). Both CA and 5FU also alter network formation, not only here but also in previous experiments (Brown et al., 2016; Frank et al., 2017). Thus, one can postulate an AOP as illustrated in Figure 6, wherein the actions of CA on neurons and glia in the cortical culture converge on decreased neurite outgrowth and ultimately result in decreased network formation. The different cellular and molecular and disease categories, as well as the canonical pathways can then be overlaid onto the putative AOP. Finally, one can begin to investigate this AOP in a more rigorous manner by manipulating the upstream regulators such as, SOX2, CREB1 (CA), NOTCH1 (5FU), and the other regulators in Table 4 to build confidence in the KE relationships (KERs) for cell proliferation, survival, neurite outgrowth, and network formation.

Such an approach can be taken for the other compound classes as well. At concentrations similar to those used here, HP was a selective inhibitor of neurite initiation in both hN2 and rat cortical neurons (EC₃₀ = approximately 3 μM), and nonselectively inhibited synaptogenesis at higher concentrations (Harrill et al., 2018). Data for DA and CM were not available from Harrill et al. (2018). In a recently released report, neither DA nor CM was active in a variety of in vitro DNT assays examining neurite outgrowth, proliferation, and glial function, suggesting its actions on network formation are relatively specific (Masjosthusmann et al., 2020). However, in a rat neural stem cell model, DA was reported to reduce the differentiation of stem cells into astrocytes, neurons, and oligodendrocytes at noncytotoxic concentrations (Gill and Kumara, 2019) Further, higher concentrations reduced axonal length (Gill and Kumara, 2019). DM was more selective than some of the other chemicals tested here in that it decreased only neurite maturation and synaptogenesis at concentrations (EC₃₀ = approximately 9 μM) that were about two-fold below cytotoxicity (Harrill et al., 2018). Putative AOPs for DM, HP, and DA are presented in Supplementary Figures 5–7, respectively. Note that for DA, the level of information regarding potential KEs identifies several important data gaps. All of these putative AOPs require additional investigation and further development. In particular, the levels of evidence for each proposed KE need to be further evaluated and KERs need to be established. However, knowledge of the upstream regulators identified in this study can be leveraged to conduct experiments that inform the KERs.

Several important caveats should be considered with respect to the current dataset. First, the samples for the -omic responses were collected only at the end of the exposure period over which network formation was evaluated with MEAs. Thus, the pathways and regulators identified here may not represent the initial response to perturbation of network development, but rather the downstream consequences of those initial responses. Although metabolomic time-course data could be collected easily by simply collecting media samples over time, collection of a transcriptomic response time-course would require the use of sister cultures such that cells could be lysed and RNA collected at each timepoint. Second, it should be recognized that the primary cortical culture used here is a mixed culture consisting of excitatory and inhibitory neurons as well as glia. Therefore, these individual cell types may respond differently, and obscure the ability to detect an influential response that is not occurring in the majority of cells or in a dominant cell population. Under the protocol used here, it would not be possible to discriminate differential responses of the cell types, and small responses in a single cell type might not be large enough to reach the limits of detection/significance, which could influence the identified pathways. Similarly, the metabolomics analysis was based on media samples, not intracellular metabolites. It is possible that such an approach favored detection of metabolites more associated with cellular death or altered membrane integrity over those that might reflect other biochemical processes. The advantage of this approach is that it would lend itself to repeated metabolomic analyses over time from the same well/population of cells. Third, the concentrations tested in this study were limited, and covered a different range of responses for the different compounds. DA was tested at the lowest concentration, which was near its toxicological “tipping point” concentration (Frank et al., 2018) where the homeostatic compensatory processes of the cell are overwhelmed. CM, DM, and HP were all tested at 10- to 100-fold higher than their respective tipping points, but were not cytotoxic (CM and DM) or were well below (HP) their cytotoxicity EC₅₀ values. In contrast, the DNT synthesis inhibitors were tested at concentrations that were close to or slightly exceeding their EC₅₀ values for cytotoxicity. Consequently, these differences may also contribute to the differential responses observed for each compound in the -omic responses. Since few DEGs went into the analysis for CM, DM, and HP, it was more difficult to identify commonly enriched pathways. Finally, the present experiments used primary cultures of rodent neurons, and therefore the results obtained here may not apply directly to humans. Additional studies are needed to resolve these questions, but the present results demonstrate that this approach can identify and support hypotheses regarding the mechanisms underlying alterations in neural network formation that can then be investigated further. Such an approach will provide valuable information for AOP development.

CONCLUSION

This study demonstrates the feasibility of this approach in obtaining -omic information from NFAs that incorporate rodent (or in the future, human) neuronal models. The transcriptomic and metabolomic responses help reinforce the functional MEA observations, suggesting that changes in these pathways may underlie the functional changes. This justifies further studies into the mechanisms of action in rat cortical cells and the relationship of these changes to -omic changes in neurodevelopmental disorders. Future studies could explore which DEGs and/ or metabolites are most predictive for network dysfunction.