Machine learning reveals common transcriptomic signatures across rat brain and placenta following developmental organophosphate ester exposure

Abstract

Toxicogenomics is a critical area of inquiry for hazard identification and to identify both mechanisms of action and potential markers of exposure to toxic compounds. However, data generated by these experiments are highly dimensional and present challenges to standard statistical approaches, requiring strict correction for multiple comparisons. This stringency often fails to detect meaningful changes to low expression genes and/or eliminate genes with small but consistent changes particularly in tissues where slight changes in expression can have important functional differences, such as brain. Machine learning offers an alternative analytical approach for “omics” data that effectively sidesteps the challenges of analyzing highly dimensional data. Using 3 rat RNA transcriptome sets, we utilized an ensemble machine learning approach to predict developmental exposure to a mixture of organophosphate esters (OPEs) in brain (newborn cortex and day 10 hippocampus) and late gestation placenta of male and female rats, and identified genes that informed predictor performance. OPE exposure had sex specific effects on hippocampal transcriptome, and significantly impacted genes associated with mitochondrial transcriptional regulation and cation transport in females, including voltage-gated potassium and calcium channels and subunits. To establish if this holds for other tissues, RNAseq data from cortex and placenta, both previously published and analyzed via a more traditional pipeline, were reanalyzed with the ensemble machine learning methodology. Significant enrichment for pathways of oxidative phosphorylation and electron transport chain was found, suggesting a transcriptomic signature of OPE exposure impacting mitochondrial metabolism across tissue types and developmental epoch. Here we show how machine learning can complement more traditional analytical approaches to identify vulnerable “signature” pathways disrupted by chemical exposures and biomarkers of exposure.

Toxicogenomics has emerged as a powerful approach for screening and prioritization of potentially toxic chemicals, and to identify mechanisms of toxicity, molecular biomarkers of exposure, and adverse outcomes. RNA sequencing and other “omics” techniques produce massively parallel datasets with thousands of attributes. Such data poses significant challenges to traditional statistical analysis because multiple comparisons require strict correction to prevent false positives, known as the “curse of dimensionality.” (Chapman et al., 2014) Multiple analytical approaches have been developed to address this issue. The standard approach ranks significantly different attributes (genes) by their p-value, lowest to highest, then corrects for multiple comparisons to protect against false positives. One potential downside with this approach is that it regards each gene as a singular element within the transcriptome, and not as a node in a network of gene interactions. Typically, only the genes with the largest fold change are captured by this stringent analysis, whereas subtle changes in a transcriptional network (of which only one of the “significant” genes may be a member) may be lost to the correction process. This is particularly deleterious for some tissues. Notably, in brain, small but consistent differences between the sexes greatly impact function, and changes in gene expression of only 20% can have outsized impact on brain function and structure, changes which may be easily lost in traditional statistical analysis (Marrocco and McEwen, 2016; McCarthy et al., 2009). It also focuses on single genes rather than pathways or multiorgan systems which, particularly for epistatic genes, can make interpreting downstream or other higher-order outcomes challenging.

An alternative approach utilizes machine learning algorithms to extract a pattern of expression from the dataset elicited by the exposure and uses this pattern to predictively classify instances into exposure groups. The attributes (differentially expressed genes, DEG) that provide the most information to the classifier are considered the most reliable predictors of exposure condition. This approach avoids the problem of multiple comparisons, as it does not rely on inferential statistical tests or p-values. For transcriptomics, previous work in the machine learning field has recommended analysis with both methods to leverage the advantages of both (Chapman et al., 2014; Sullivan et al., 2015). Following this guidance and using both extant and newly generated transcriptome data from rat brain and placenta, we deployed multiple supervised ensemble machine learning methods to identify commonly impacted genes and biological functions following developmental organophosphate ester (OPE) exposure. We also sought to determine if there are any sex differences in the identified effects.

We first obtained and analyzed postnatal day 10 (P10) hippocampal punches from OPE-exposed and control rats of both sexes with traditional bioinformatics methods (q-value). We then compared hippocampal RNAseq results with previously analyzed and published RNAseq data from P1 cortex obtained from their siblings, and embryonic day 14 (E14) placentas from a different study (Rock et al., 2020; Witchey et al., 2022). We employed an ensemble multiple classifier system learning strategy composed of 4 supervised machine learning algorithms (support vector machines [SMV], multilayer perceptron [MLP] artificial neural networks, J48 decision trees, and RandomForest) to reanalyze these 2 datasets and identify common genes and biological functions impacted by developmental OPE exposure across all 3 tissues. These 4 algorithms were chosen for their diverse mathematical approaches to data classification, the supposition being that their convergence would reduce mean model variability and confirm optimal model performance though orthogonal mathematical approaches. Briefly the algorithms included (1) SMO SVM, which are nonprobabilistic, binary linear classifiers that can operate in infinite dimensions; (2) MLP artificial neural networks, which are nonlinear computations functionally designed based on decision processing of neural nets in biological systems; (3) decision trees (J48), which support hierarchical modeling analysis through decision nodes; and (4) RandomForest, which operates by constructing a multitude of decision trees that learn from each other at training time. Simply put, these are systems that can be trained to recognize certain data input patterns and then used to predict outcomes or classify data. The use of different approaches allows for orthogonal comparisons as an ensemble analysis because they each operate in a slightly different manner.

The placenta and brain are vulnerable to developmental OPE exposure

Investigating the potential neurotoxic effects of OPEs is an area of increasing urgency, given their global use as flame retardants (FRs) and plasticizers in commercial and household goods, their status as ubiquitous environmental contaminants, and nearly unavoidable human exposure. There is compounding evidence that OPEs, either individually or in combination, have deleterious effects in developing neural systems. For example, exposure to triphenyl phosphate (TPHP), the major component of the OPE mixture employed in this study (approximately 45% mass fraction) and the molecular backbone of the other isopropylated triaryl phosphate components in the mixture (Table 1), disrupts trophoblast development, and alters placental levels of estrogen, progesterone, chorionic gonadotropin, and testosterone in mice; hormones critical for fetal development and sexual differentiation (Hong et al., 2022a). Additionally, TPHP has been shown to cross the blood brain barrier, accumulate in brain, and increase apoptosis in mouse hippocampal and cortical neurons in vivo, as well as human neuroectoderm cells in vitro (Belcher et al., 2014; Liu et al., 2020). Structural and neurodevelopmental effects in offspring of dams exposed to OPEs during gestation include disruption to midbrain cytoarchitecture and developing axogenesis of dopamine (Newell et al., 2023) and serotonin (5-HT) neurotransmitter systems (Hong et al., 2022a; Rock et al., 2020). Postnatal TPHP exposure has been shown to decrease levels of synaptogenesis and axon guidance proteins in adult hippocampus (Zhong et al., 2021).

Table 1.

Chemical Components of OPE Mixture

Name (CAS No.)	Structure	Name (CAS No.)	Structure
TPHP (115-86-6)		B2IPPPP (69500-29-4)
2IPPDPP (28108-99-8; 93925-53-2; 64532-94-1)		B3IPPPP
3IPPDPP		B4IPPPP
4IPPDPP		T3IPPP (72668-27-0)

Comparatively less research has investigated the potential neurotoxic effects of the other isopropylated triaryl phosphate components in the OPE mixture used in this study. However, mono-isopropylated triaryl phosphates have been shown to induce both developmental cardiac and behavioral toxicity in zebrafish (Gerlach et al., 2014; Noyes et al., 2015). Additionally, we have previously shown this full mixture of OPEs accumulate in fetal placenta and disrupts the expression of genes that synthesize and transport 5-HT (Rock et al., 2020; Ruis et al., 2019) and that accumulation and severity of outcomes was higher in males. We have also demonstrated in rats that gestational exposure to this mixture of OPEs (via maternal ingestion) alters gene transcription regulating 5-HT synthesis, transport, and metabolism, and altered 5-HTergic (Rock et al., 2020) and dopaminergic axogenesis in fetal brain (Newell et al., 2023). Additionally, developmental OPE exposure disrupted social behaviors, and anxiety-like behaviors in adult rats (Witchey et al., 2020). We subsequently reported that gestational OPE exposure induces changes to transcription of genes responsible for ATP synthesis, electron transport chain activity, and axon guidance in P1 rat cortex in a sex-specific manner (Witchey et al., 2022).

Collectively, although published findings by us and others demonstrate that OPEs have broad, sex-specific, adverse effects on multiple organ systems during development, particularly the placenta and brain, consistent upstream effectors of the observed changes have been elusive. Here, using P10 siblings of the P1 animals in the Witchey et al. (2022) study, we sought to expand on our prior findings by first performing RNAseq in the P10 hippocampus. We chose P10 because it is a critical period of hippocampal circuit development (Supèr et al., 1998; Tamamaki, 1999). We then compared transcriptional changes in OPE exposed P10 hippocampus to previously published transcriptomic data generated by our lab from E14 placenta (Rock et al., 2020), and P1 cortex (Witchey et al., 2022) (siblings of the P10 animals), to determine if a biological “signature” of OPE exposure exists across brain and placental tissue, and if/how that “signature” is sex specific. Based on the known endocrine and developmental neurotoxicity of OPEs, we hypothesized that OPE exposure would induce common transcriptional changes in all 3 tissues, and that important differences in the capacity to detect these changes might be observed between traditional statistical approaches and the ensemble machine learning methodology. To test this hypothesis, we used 2 forms of transcriptomics analysis; a traditional approach based on p-values corrected for multiple comparisons (q-values), and a multiple classifier system machine learning approach to determine genes contributing information to the machine learning classifier. Both sets of results, including identified DEGs were then analyzed for gene ontology enrichment and compared.

Materials and methods

 

 

Animals and husbandry

The overall goal of this study was to compare transcriptional changes in P10 hippocampus with those in the P1 cortex and E14 placenta to identify common biological processes perturbed by OPE exposure across tissues, doses, and critical windows of development. To achieve this goal while minimizing animal usage and maximizing available resources, all tissues used for this experiment were generated in 2 previous studies, the details of which are published (Rock et al., 2020; Witchey et al., 2022). The P10 RNAseq dataset was specifically generated for the present study using hippocampal tissues obtained from brains already collected and stored from the siblings of the P1 animals (Figure 1).

Figure 1.

Experimental design of OPE exposure for placenta, cortex, and hippocampus. All exposures began at embryonic day 0 and continued until day of sacrifice. Milligram per kg bw/day dose was based on maternal mass and was approximately 8 mg/kg for E14 placenta, and 3.3 mg/kg bw/day for P1 cortex and P10 hippocampus.

Briefly, all protocols for animal care and experimentation were approved by North Carolina State University Institutional Animal Care and Use Committee (IACUC; Protocol Number: 17-070) and complied with the Animal Welfare Act and U.S. Department of Health and Human Services “Guide for the Care and Use of Laboratory Animals”. Naïve adult male and female Wistar rats were purchased from Charles River (Raleigh, North Carolina) and were housed and bred in temperature and humidity-controlled rooms on a 12-h light: 12-h dark schedule at the AAALAC-accredited Biological Resource Facility at NC State. As in all our prior studies, every effort was made to reduce potential confounding contamination from exogenous endocrine disrupting chemicals in housing materials and food, by using polysulfone caging, glass water bottles, soy-free food ad libitum (Teklad 2020), and wood chip bedding. For breeding, female rats were randomly assigned to control or OPE groups and weighed. Dosing was based on the average weight of that group. For example, for the P10 tissues, the average group weight of the dams was approximately 300 g, and this average weight was used to calculate OPE dosing. To generate the litters, female rats were paired with a male and checked for sperm plugs daily to establish day of conception and then housed individually in clean cages with clean bedding for the rest of the study.

OPE preparation and dosing

A mixture of OPEs commonly found in commercial FR mixtures was provided by Dr Heather Stapleton’s Laboratory at Duke University. The OPE mixture contained TPHP (Cas No. 115-86-6), 2-isopropylphenyl diphenylphosphate (2IPPDPP; Cas No. 28108-99-8; 93925-53-2; 64532-94-1), 3-isopropylphenyl diphenyl phosphate (3IPPDPP), 4-isopropylphenyl diphenyl phosphate (4IPPDPP), 2,4-diisopropylphenyldiphenyl phosphate (24DIPPDPP), bis(2-isopropylphenyl) phenylphosphate (B2IPPPP; Cas No. 69500-29-4), bis(3-isopropylphenyl)phenyl phosphate (B3IPPPP), bis(4-isopropylphenyl) phenylphosphate (B4IPPPP), and tris(3-isopropylphenyl) phosphate (T3IPPP; Cas No. 72668-27-0) in sesame oil vehicle to a concentration of 50 mg/ml. OPE solution was stored at 4°C in scintillation vials wrapped in foil until use.

The E14 placentas were obtained from dams exposed to 2000 µg OPEs per day (approximately 8 mg/kg bw/day) and the P1 and P10 brain samples obtained from dams dosed with 1000 µg total OPEs (approximately 3.3 mg/kg bw/day). The doses were based on past studies that demonstrated neurotoxic effects at exposures in this range (Baldwin et al., 2017; Gillera et al., 2021; Newell et al., 2023; Rock et al., 2018, 2020; Witchey et al., 2020). Inclusion of multiple doses was considered an advantage given the study goal of determining if there is an exposure “signature” that arises despite tissue, age at analysis, exposure duration, or dose. Dams were dosed orally daily with either OPE mixture or sesame oil. To achieve the correct dose, 20 µl of OPE mixture or sesame oil was pipetted onto 1/8 of a wafer cookie (Nilla, Nabisco) and allowed to fully absorb. Dam consumption of the entire wafer was observed to ensure full exposure. The use of a cookie for oral exposure reduces the significant stress induced by the common alternative, oral gavage (Bonnichsen et al., 2005; Brown et al., 2000; Cao et al., 2013; Walker et al., 2012). For all studies, exposure occurred at the same time of day, beginning once a sperm plug was detected and continuing daily until fetuses/pups were sacrificed, 4 h after final dosing.

P10 hippocampal tissue collection and processing

The day of birth was designated P0. On P10, 1 pup per sex was removed for necropsy from each litter across 12 litters (6 control and 6 exposed, 24 animals total). Litter was the statistical unit in this study. Sex was determined by anogenital distance and pups were anesthetized by carbon dioxide asphyxiation and euthanized by rapid decapitation. Brains were removed and frozen on dry ice, then stored at −80°C until they could be processed on a Leica CM1900 cryostat at −14°C. Each was cut to approximately bregma −2.92 mm using coordinates from The Rat Brain in Stereotaxic Coordinates (Paxinos and Watson, 2014). Three 1.25 mm punches (depth of 1 mm) covering CA1, CA3, and the dentate gyrus (DG) were made through dorsal hippocampus (Figure 2) and kept at −80°C until RNA extraction. Bilateral punches from the amygdala were also taken at this time for future analysis. Procedures were akin to what we have done for prior transcriptomics studies (Arambula et al., 2016, 2018; Witchey et al., 2022). RNA extraction from the hippocampal samples was performed using a Qiagen RNeasy Miniprep kit (Qiagen) following the manufacturer protocol. Extracted RNA was transferred to the NC State Genome Sequencing Laboratory for quality assessment and sequencing, following the same protocol as was used on P1 cortex in Witchey et al. (2022).

Figure 2.

Representative image of hippocampal punches from P10 rat brain extracted for RNA sequencing. Amygdala punches were taken for future studies and not included in the current work.

RNAseq data analysis

Data analysis was performed in consultation with Bioinformatics Core at the NC State Center for Human Health and the Environment. An average of approximately 44.7 million paired-end raw RNAseq data were generated for each replicate. The quality of sequenced data was assessed using the fastqc application, and 12 poor-quality bases were trimmed from the 5′-end. The remaining good-quality reads were aligned to the Rat reference genome (mRatBN 7.2 version 106) downloaded from the Ensembl database using STAR (Dobin et al., 2013) aligner. Per-gene counts of uniquely mapped reads for each replicate were calculated using the htseq-count script from the HTSeq (Anders et al., 2015) python package. The count matrix was imported to R statistical computing environment for further analysis. Initially, genes that had no count in most replicate samples were discarded. The remaining count data were normalized for sequencing depth and distortion, and dispersion was estimated using DESeq2 (Love et al., 2014) Bioconductor package in the R statistical computing environment (R Core Team, 1996). We fitted a leaner model using the exposure levels, and DEGs were identified after applying multiple testing corrections using the Benjamini-Hochberg procedure (Benjamini and Hochberg, 1995). The final significant genes were generated using p adjusted < .05. Results were deposited in GEO (accession number GSE230516).

P10 hippocampal machine learning analysis

Four machine learning algorithms were implemented in the analysis using The Weka Data Platform for AI (version 3.8.5), including sequential minimal optimization support vector machines (SMO SVM), an MLP artificial neural network, a J48 decision tree (J48), and a RandomForest decision tree. The algorithms were used to classify P10 hippocampal punches into 2 groups, controls or OPE exposed, based on mRNA transcripts per kilobase million (TPM) from the RNAseq data. Analysis was performed on male (6 control, 6 exposed) and female (6 control, 6 exposed) samples separately. For each model, 2 cross-validation methods were used to assess learning, a 10-fold stratified hold-out strategy and a percentage split (66% training, 34% testing) strategy, resulting in a total of 8 models. These 2 validation strategies were chosen to complement each other, as percentage split is routinely employed in artificial intelligence research but can be prone to inflated accuracy with small sample sizes. To offset this, we also used the 10-fold stratified hold-out strategy which is considered a conservative method of cross-validation and robust across different sample sizes (Chicco, 2017). By employing both methods and averaging their results, we sought to reduce potential variance and imprecision. All models were provided with the full dataset for each sex (12 replicates per sex) with no preprocessing step and were naïve to the testing data in each cross-validation. Learning was evaluated as percent testing data classified correctly. Accuracy was averaged across all 4 models for each cross-validation strategy to compare classifier performance. Area under receiver operating characteristic (AUROC) and the kappa statistic for each model was also calculated.

To avoid model “overfitting” and reduce data dimensionality, we sought to include only attributes which contributed information to the classifier. To accomplish this, the InfoGainAttributeEval function in Weka was used to calculate information gain for each attribute. The attributes with the highest information gain contribute most to classification, and thus information gain represents the relative importance of the attribute in building the machine learning model. High information gain attributes in the context of this experiment are DEGs which inform the classifier and allow it to predict OPE exposed tissue versus control. Thus we termed these attributes “high information genes.” We employed information gain data in recursive feature elimination by ranking the attributes by their information gain and sequentially removing them to optimize classification across the widest number of models.

Finally, as a negative control for machine learning, 10 randomized datasets based on the top information attributes of OPE and control animals in male and female groups were generated. In each dataset, the group labels (OPE and control) were randomly reassigned, thus there was presumably no pattern in the data for the algorithms to learn. The expected outcome of this negative control would be no learning. With 2 groups per sex in our experimental model, the probability of randomly guessing the correct classification of a particular instance is 50%, therefore 50% correct classification reflects no learning.

Gene ontology and pathway analysis

The web-based gene list enrichment analysis tool ShinyGO (Xijin Ge et al., 2020) (http://bioinformatics.sdstate.edu/go/) was used to perform gene ontology and biological pathway analysis on high information genes used in classification, and DEG genes q ≤ 0.05. We submitted Entrez gene symbols (and Ensemble IDs when lacking a gene symbol) and selected GO biological process, molecular function, cellular component, and KEGG pathway biological pathway analysis modules for further investigation in P10 hippocampal tissue.

Cross-tissue machine learning and enrichment analysis

To compare the effects of OPE exposure on the transcriptome of each of the 3 tissues of interest, the RNAseq data generated from E14 placentas and the P1 cortical samples was reanalyzed by 2 methods. First, lists of DEGs in males and females with q ≤ 0.05 were generated via “traditional” methods for male and female placental and male and female cortical tissue based on the data generated from those respective studies. DEGs shared among the OPE-exposed tissues were determined by comparing hippocampus to cortex, cortex to placenta, placenta to cortex and all 3 tissues to each other. Each list of common DEGs was then further investigated by gene enrichment analysis with ShinyGO.

In a second analysis, placental and cortical RNAseq results were also analyzed independently by the ensemble machine learning methods outlined above for hippocampus. Full details regarding husbandry and dosing for placental and cortical tissue can be found in the respective prior publications (Rock et al., 2020; Witchey et al., 2022). Briefly, for fetal placental tissues, dams were exposed to oral OPE from conception until E14, then sacrificed. Each fetus and placenta was collected, and the fetal side of the placenta from a pup of each sex per dam (5 control dams and 5 OPE; litter as the statistical unit), resulting in 20 total placentas for RNAseq analysis. For P1 cortical tissue, 12 dams (6 control, 6 OPE) were exposed to oral OPEs or vehicle from conception to P1. At P1, 1 pup of each sex per litter was collected, with only 5 OPE exposed P1 females used due to limited tissue availability, resulting in 23 P1 cortical samples for RNAseq.

Separate machine learning models to classify OPE exposed versus control tissue were then created for each sex/tissue combination (ie male placenta, female placenta, male cortex, female cortex) using the same multiple classifier system learning approach utilized for the hippocampal RNAseq data. The 4 machine learning algorithms (SMO SVM, MLP, J48, RandomForest) were trained on raw mRNA TPM placental and cortical RNAseq datasets. Model training was conducted independently for each tissue within each sex and tested on a subset of data from the respective sex/tissue dataset, resulting in 4 classifier models. Models trained on 1 sex/tissue combination were not tested on another, different sex/tissue combination (ie, the model for classifying OPE versus control male hippocampus was never tested on cortical or placental tissue of either sex). To assess learning respectively for each sex/tissue combination, 2 cross-validation strategies, k-fold stratified holdout and percentage split were used, and model performance was optimized as a function of percent instances classified correctly. High information genes were determined for each sex/tissue combination using InfoGainAttributeEval function in Weka.

To determine if a common “transcriptomic signature” of OPE exposure existed across all 3 tissues, the lists of high information genes derived from machine learning analysis of each sex/tissue combination were compared between tissues. This consisted of 4 comparisons in each sex, hippocampus to cortex, cortex to placenta, placenta to hippocampus, and across all 3 tissues. Once common high information genes present in each tissue were identified, these genes were used in gene ontology and pathway analysis using the web-based tool ShinyGO. Gene ontology modules GO biological processes and KEGG pathways were selected for P1 cortex, E14 placenta and cross-tissue analysis.

Results

OPE impact on hippocampal transcription

Machine learning analysis of hippocampus RNAseq data

Transcript per million data for all 14 523 genes were provided to the 4 machine learning algorithms for hippocampus transcriptome analysis. Percent instances classified correctly, kappa statistic and AUROC were calculated for both 10-fold stratified hold out and 66/34 percentage split cross-validation (Supplementary Table 1). Calculation of information gain revealed 133 high information genes informing classification in males, and 345 high information genes in females. Recursive feature elimination was performed based on information gain to reduce data dimensionality and optimize percent instances classified correctly (Figure 3), this maximum was reached when only the high information genes were provided in each sex (mean percent correct = 97.92% for stratified hold out in males and females, and mean percent correct = 100% in males and 93.75% in females for 66/34 percent split).

Figure 3.

Mean performance of 4 machine learning algorithms based on number of attributes in male (A) and female (B) OPE-exposed hippocampus. Average percent of correctly classified instances increases as attributes are eliminated for both validation methods (stratified hold out and percentage split). Through reduction of data dimensionality model performance approaches optimum with inclusion of only 133 top information attributes in males and 345 in females. Serial elimination of top 133 attributes in males, and 345 top information attributes in females resulted in an average drop classifier performance (Minus High Info).

Importantly, sequential elimination of high information genes in reverse order (ie No. 1 information gene to No. 133 or 345, respectively) led to a gradual decrease in model performance on average in both sexes (Figure 3). Incrementally decreasing the number of high information genes reduced the averaged percentage of instances classified correctly in both males in females. However, in the case of females, removing the top high information eventually increased percentage correct as the subgroup became a better fit due to a smaller number of highly predictive genes (Tmlhe, Maged2, Flii, Clasp1, Sap30l, Dgkg, Stk38, Uqcrq, Atp5pd, Lpl, Wnk3, Cbr3, Flcn, Phka1, Pacrgl, Pi4kb, Asxl1, Prim1, Vasn, Emilin3, Zfp358), though not above optimized performance levels achieved with recursive feature selection.

The list of high information genes in each sex was analyzed by gene enrichment analysis. In males, there was no significant enrichment in GO biological processes, cellular component, or KEGG pathways among the top 133 high information genes, similar to the results in male hippocampus with a q ≤ 0.05. However, in females, pathway analysis of the top 345 high information genes revealed significant enrichment in GO biological processes including cation and metal ion transport, and regulation of mitochondrial translation and gene expression. Significant enrichment of GO cellular components, specifically mitochondrion and catalytic complex was also identified (Figure 4), however, no significant enrichment was observed for KEGG pathways.

Figure 4.

Significantly enriched GO biological processes (A), molecular functions (B), and cellular components (C) pathways from top 345 information genes utilized by machine learning algorithms to classify female OPE exposed P10 hippocampus. Bar length represents fold enrichment of pathway, circle size represents number of genes within pathway, and color represents −log10(FDR) (blue = lower, orange = higher). (D) Representative selection of pathways and high information genes driving both enrichment and classifier performance.

As a negative control for machine learning, 10 sets of data based on the 133 and 345 high information genes in male and female hippocampus respectively, were randomized and model performance assessed. Average performance of each of the 4 algorithms and both cross-validation strategies was based on percent instances correctly classified, the kappa statistic and AUROC (Supplementary Table 2). In both males and females, no model achieved better than 50% correct classification, the maximum k = 0.05 ± 0.352, and the maximum AUROC = 0.483 ± 0.149.

Machine learning analysis of cortex RNAseq data

Transcription data generated in Witchey et al. (2022) derived from OPE exposed and control P1 cortex was analyzed using the same machine learning methodology employed for hippocampus. TPM values for all 32 884 attributes in OPE and control males and females were entered into the 4 machine learning algorithms and percent correct, kappa statistic, and AUROC were calculated for each algorithm using both stratified holdout and 66/34 percentage split cross-validation (Supplementary Table 3). Information gain was calculated for all attributes and 1303 high information genes in males and 660 high information genes in females were identified as driving model performance. Recursive feature elimination optimized percent instances classified correctly, kappa statistic, and AUROC (Supplementary Table 3). Percent instances classified correctly varied inversely with number of attributes (Figure 5) and maximum performance occurred at or near the number of high information genes in males (mean percent correct across all 4 algorithms = 97.92% stratified hold out, and 100% 66/34 percent split, Figure 5). Indeed, sequential removal of the 1303 high information attributes in males resulted in decreasing classifier accuracy. In females, average classifier performance increased then decreased with serial elimination of top 660 high information attributes with the stratified hold out validation method, whereas removing the same set of attributes decreased model performance in percentage split validation (Figure 5).

Figure 5.

Mean performance of 4 machine learning algorithms based on number of attributes in male (A) and female (B) OPE-exposed cortex. Average percent instances correct increase as attributes are eliminated for both validation methods (stratified hold out and percentage split). Performance approaches optimum with inclusion of only 1303 top information attributes in males and 660 in females (in stratified hold out). Serial elimination of top 1303 attributes in males resulted in decomposition of classifier performance in both validation methods. Removal of top information attributes in females lead to a slight increase then decrease in model performance for stratified hold out validation method, whereas the same lead to an abrupt decrease and then slight increase in percent split.

The set of high information genes driving classifier performance in both sexes were submitted for pathway analysis. In males, the top 1303 genes showed significant enrichment for GO biological processes related to mitochondrial function including oxidative phosphorylation, NADH dehydrogenase complex assembly, Aerobic electron transport chain, and oxidative phosphorylation, as well as significant KEGG pathways of oxidative phosphorylation and pathways of neurodegeneration (Figure 6). In females, the top 660 high information genes showed significant enrichment in biological processes of NADH dehydrogenase complex assembly, mitochondrial respiratory chain complex I assembly, and oxidative phosphorylation, and in KEGG pathways for oxidative phosphorylation, and pathways of neurodegeneration.

Figure 6.

Significantly enriched GO biological processes in males (A) and females (C) and KEGG pathways in males (B) and females (D). Pathways from top 1303 information genes utilized by machine learning algorithms to classify male and 660 female OPE-exposed P1 cortex. Bar length represents fold enrichment of pathway, circle size represents number of genes within pathway, and color represents −log10(FDR) (blue = lower, orange = higher), * pathways enriched in both males and females. Representative selection of pathways and high information genes driving both enrichment and classifier performance (E).

Ten sets of the high information attributes (1303 males, 660 females) were created, and the exposure groups (OPE or control) randomized and analyzed as a negative control for learning. Classifier performance (% correct), kappa statistic, and AUROC were calculated across all 4 algorithms and both validation strategies and averaged across the 10 sets (Supplementary Table 4). No model showed a percent classified correctly higher than 60.00 ± 24.15%, kappa = 0.31 ± 0.39, and AUROC of 0.93 ± 1.44.

Machine learning analysis of placenta RNAseq data

Data produced from RNAseq of E14 OPE and control exposed fetal placenta by Rock et al. (2020) were reassessed using the machine learning methodology employed for hippocampus and cortex. Transcript per million data for all 12 366 attributes in males and females were analyzed by 4 machine learning algorithms and measures of classification performance were calculated (Supplementary Table 5). Two hundred and sixty-three and 955 high information genes were identified in male and female fetal placenta, respectively. Model performance was optimized based on percent instances correct via recursive feature elimination (Figure 7). In males, optimal performance across all 4 algorithms with stratified hold out validation occurred with 500 high info genes and percent split with 200 (Figure 7). In females, mean classifier performance reached maximum for stratified hold out validation with 200 high info genes, and for percent split at 955 high info genes (Figure 7). In males, sequential removal of the top high information genes in males decreased model performance, whereas in females performance initially decreased as high info genes were removed and eventually increased as the list of genes decreased to a small set of highly predictive genes (Ehd3, Mlh1, Rpl36, Grin2d, Chfr, Aifm3, Tnfsf12, Ensa, and 2 Ensembl IDs without gene symbols, ENSRNOG00000004262 and ENSRNOG00000028330).

Figure 7.

Mean performance of 4 machine learning algorithms based on number of attributes in male (A) and female (B) OPE-exposed fetal placenta. Average percent instances correct increase as attributes are eliminated for both validation methods (stratified hold out and percentage split). Performance approaches optimum with inclusion of only 263 top information attributes in males and 955 in females. Serial elimination of top 263 attributes in males resulted in an overall drop in classifier performance (minus high info) in stratified holdout validation, whereas resulting in abrupt decrease then gradual increase in performance. Removal of top information attributes in females lead to modest increase and then decrease in model performance for stratified hold out validation method, whereas the same lead to a decrease and then slight increase in percent split.

For pathway analysis, the top 263 high information genes were selected in males based on the rough average between the 2 validation methods mean optimum performance (500 and 200 genes), whereas the top 955 high information genes in females were utilized. Pathway analysis of the top 263 info genes with ShinyGO revealed significant enrichment in GO biological processes of cytokinesis, proteasomal protein catabolic processes and cellular catabolic processes among others. For KEGG pathway analysis, only amyotrophic lateral sclerosis was significantly enriched (Figure 8). In females, GO biological processes of regulation of proteasomal ubiquitin-dependent protein catabolic processes, electron transport chain, and histone modification were significantly enriched (Figure 8). Additionally, KEGG pathways of thiamine metabolism, oxidative phosphorylation, and pathways of neurodegeneration were significantly enriched among others (Figure 8).

Figure 8.

Significantly enriched GO biological processes in males (A) and females (C) and KEGG pathways in males (B) and females (D) from E14 fetal placenta. Pathways from top 263 information genes utilized by machine learning algorithms to classify male and 955 female OPE-exposed E14 placenta. Bar length represents fold enrichment of pathway, circle size represents number of genes within pathway, and color represents −log10(FDR) (blue = lower, orange = higher), * pathways enriched in both males and females. Representative selection of pathways and high information genes driving both enrichment and classifier performance (E).

As a negative control for machine learning, 10 sets of the high information attributes (263 males, 955 females) were created, and the exposure groups (OPE or control) randomized and analyzed. Classifier performance (% correct), kappa statistic, and AUROC were calculated across all 4 algorithms and both validation strategies and averaged across the 10 sets (Supplementary Table 6). No model showed percent classified correctly higher than 61.00 ± 22.83%, kappa = 0.22 ± 0.46, and AUROC of 0.700 ± 0.43.

q-value analysis of hippocampal, cortex, and placenta RNAseq data

In male hippocampus, 1 DEG had a q ≤ 0.05, Foxp2. When correction for multiple comparisons was relaxed to p ≤ .01, 97 DEGs in male hippocampus were identified. Pathway analysis with ShinyGO showed significant enrichment for GO biological processes related to hippocampal function and cognition, including learning, learning or memory, and cognition, as well as response to amphetamine and glycerophospholipid catabolic processes (Table 2). There was no significant enrichment in molecular function, cellular compartment, or KEGG pathways. In female hippocampus, no DEG had a q ≤ 0.05. When the cutoff was relaxed to p ≤ .01, 141 DEGs were identified, however, pathway analysis revealed no significant enrichment. DEGs with q ≤ 0.05 in OPE exposed E14 placenta and P1 cortex were determined based on original RNAseq data generated in Rock et al. (2020) and Witchey et al. (2022), respectively (Table 3). The list of these DEGs were utilized in the present study only for cross-tissue gene analysis, and the specific analysis and discussion of the DEGs and significant GO pathways for each of those individual tissues can be found within those prior publications.

Table 2.

Significantly Enriched Biological Processes from 97 DEG between OPE and Control Male Hippocampus When p ≤ .01

Gene Ontology Term	Genes
Biological processes
Learning	Tpbg, Jph3, Drd1, Gpr88, Ppp1r1b, Foxp2
Learning or memory, cognition	Btbd9, Tpbg, Jph3, Drd1, Gpr88, Ppp1r1b, Foxp2
Response to amphetamine	Adora2a, Drd1, Ppp1r1b
Glycerolphospholipid catabolic proc.	Smpd4, Pla2g6, Abhd16a

Table 3.

Total Number of DEGs in Male and Female Tissues

Region	Genes q ≤ 0.05	High Information Genes
Males
Hippocampus	97*	133
Cortex	1037	1303
Placenta	3	263
Females
Hippocampus	141*	345
Cortex	808	660
Placenta	69	955

Total number of DEGs in male and female OPE exposed hippocampus, cortex, and placenta based on q-value correction for multiple comparisons and information gain.

^* p $\le$.01, as only 1 gene in males and zero genes in female hippocampus had a q ≤ 0.05.

Cross tissue gene analysis

DEGs with a q ≤ 0.05 were compared across the 3 tissues (for hippocampus, there were insufficient genes q ≤ 0.05, so the cutoff was changed to p ≤ .01) and common genes identified. Separately, high information genes informing classifier performance identified in the machine learning analysis (Table 3) were compared across the 3 tissues and submitted for pathway analysis.

Common genes across hippocampus and cortex

q value analysis

In males there were 23 DEGs between hippocampus and cortex, however, there was no significant enrichment in pathway analysis. In females there were 7 DEGs common across hippocampus and cortex with no significant enrichment GO pathway analysis.

Machine learning analysis

In males, there were 7 high information genes common across hippocampus and cortex, however, these 7 genes showed no significant enrichment. In females there were 15 high information genes common across hippocampus and cortex (Figure 9). These 15 genes showed significant enrichment in GO biological processes for oxidative phosphorylation, aerobic and cellular respiration, aerobic electron transport chain, mitochondrial ATP synthesis, and others. Additionally, KEGG pathway analysis was significantly enriched for oxidative phosphorylation and diseases of neurodegeneration. Of the 15 common genes between female cortex and hippocampus, 4 were responsible for significant enrichment. In cortex, all 4 of these genes were upregulated, and in hippocampus all 4 downregulated (Figure 9C).

Figure 9.

Number of common high information genes in male and female hippocampus and cortex (A). Significantly enriched gene ontology pathways (biological processes and KEGG) in females based on common high information genes utilized by machine learning classifiers in OPE-exposed hippocampus and cortex (B). High information genes common across female hippocampus and cortex are impacted by OPE exposure in opposite directions, green denotes increased expression, blue decreased expression (C). Common high information genes responsible for enrichment of select gene ontology pathways (D).

Common genes across cortex and placenta

q value analysis

In males there were 41 common DEGs with q ≤ 0.05 overlapping between cortex and placenta transcriptome, with significant enrichment for GO biological function of posttranscriptional reg. of gene expression (6 genes—Hnrnpd, Samd4b, Paip2, Fus, Zfp385a, Taf15), and no significant enrichment of KEGG pathway. In females, there was 8 common DEG, and 5 of these genes (Rpl12, Rpl14, Rpl18, LOC498555, Rps8) with significant enrichment of GO biological processes and KEGG pathways for ribosome.

Machine learning analysis

In males, there were 15 high information genes across cortex and placenta (Figure 10A). Pathway analysis of these 15 genes showed significant enrichment in GO biological processes for mitochondrial ATP synthesis coupled proton transport, ATP biosynthetic proc. among others. Although KEGG pathways were significant for oxidative phosphorylation (Figure 10B). In females, there were 66 high information genes common between female cortex and placenta. Pathway analysis revealed significant enrichment in GO biological processes for NADH dehydrogenase complex assembly, electron transport chain, oxidative phosphorylation, and others. KEGG pathway analysis of these 66 common genes revealed significant enrichment for oxidative phosphorylation and neurodegenerative diseases (Figure 10). In males, of the 15 common high information genes 3 were responsible for significant pathway enrichment, and the direction of their expression was in opposite directions in cortex and placenta (Figure 10C). In females of the 65 high information genes common across cortex and placenta, 12 were responsible for significant pathway enrichment, and were increased in all cases (Figure 10C).

Figure 10.

Number of common high information genes in male and female cortex and placenta (A). Representative sample of significantly enriched gene ontology pathways based on common DEGs in male and female OPE-exposed cortex and placenta derived from high information genes utilized by machine learning classifiers (B). Direction of expression of high information genes common across male and female cortex and placenta induced by OPE exposure, green denotes upregulation, blue down regulation (C). Common high information genes responsible for enrichment of select gene ontology pathways (D).

Common genes across hippocampus and placenta

q value analysis

In males, there were 2 DEG overlapping genes between male hippocampus and placenta and significant enrichment of GO biological function extraembryonic membrane development and KEGG pathway MAPK signaling pathway, however, this was only based on a single gene Map3k4. Although in females, there was zero common DEG q ≤ 0.05 between hippocampal and placental tissue.

Machine learning analysis

In males, there were 5 high information genes common across hippocampus and placenta, however, there was no significant pathway enrichment among these genes. In females, there were 17 high information genes common across hippocampus and placenta (Supplementary Figure 1), however, there was no significant GO biological or KEGG pathway enrichment among these 17 genes.

Common genes across hippocampus, cortex, and placenta

In both males and females, there were no overlapping genes among all 3 tissues following q-value analysis. Similarly, in males, there were no high information genes common across hippocampus, cortex, and placenta. However, in females, ubiquinol-cytochrome c reductase binding protein (Uqcrb) was the only high information gene identified across all 3 tissues.

Discussion

Results indicate: (1) machine learning classifiers can predict OPE exposure in all 3 tissues, (2) perinatal OPE exposure significantly alters gene expression in P10 hippocampus associated with voltage gated Ca²⁺ and K⁺ channels, and mitochondrial function in a sex-specific manner (greater impact in females), and (3) machine learning analysis identified common sets of mitochondrial associated genes altered by OPE exposure across hippocampus, cortex, and placental transcriptomes depending on sex. These results suggest early life OPE exposure has tissue-specific effects, but also impacts common biological pathways across tissues, sexes, dose, and developmental phase with mitochondrial disruption being a mode of action common to all. In parallel to machine learning, an analysis based on mean expression change corrected for multiple comparisons (q-value) was conducted, however, this second analysis reported no significant enrichment of GO pathways in male or female OPE exposed hippocampus. In analogous cross-tissue analysis, DEGs with corrected q-values showed significant enrichment only between cortex and placenta for ribosomal subunits. Collectively, these findings demonstrate that machine learning analysis is a powerful analytical tool to interrogate transcriptomics data, particularly in brain where expression changes may be small, to reveal biologically meaningful patterns in gene transcription that linear statistics methods may not detect, and provide orthogonal confirmation of findings derived from traditional bioinformatics analysis. The results suggest that OPEs have multiple, sex specific, overlapping modes of action, which was not unexpected given that the brain and placenta are highly sexually dimorphic, and we tested a “real-world” OPE mixture containing multiple components, each of which could have somewhat different effects.

 

Multiple classifier system optimizes number of high information genes

Optimum performance of each classifier varied based on different numbers of high information genes depending on sex and tissue (as seen in Figs. 3, 5, and 7). Additionally, optimum model performance was impacted by cross-validation strategy across algorithms. This variability was expected due to the diverse mathematical underpinnings of each model and the occasionally idiosyncratic interaction of the specific algorithms’ functionality and the data. Thus, no single algorithm is always the most reliable classifier for every dataset (or in some cases cannot be run on very large datasets due to technical limitations). To account for this expected variability, we utilized 4 diverse algorithms and 2 cross-validation methods, then averaged model performance across these methods to reduce prediction variance and allow a convergence of model performance on an optimum number of high information genes. As such, we enhanced precision and reduced uncertainty in the results.

OPE exposure impacts on hippocampal transcription are sex specific

All 4 of the machine learning algorithms utilized in this study were able to correctly predict perinatal OPE exposure in the P10 hippocampus with perfect accuracy, even though exposure had sex specific effects on the developing hippocampal transcriptome. When performance was optimized through recursive feature elimination, a small subset of approximately 10 genes that could be used as potential biomarkers were identified. In males these genes include Pla2g6, Ndufb10, Tmem132b, Man1b1, Clec3b, Tmem234, Ecd, Elavl1, and E2f3, and in females Plagl1, Pdpn, Tcf7l1, Lipt1, Hspb2, Nudt9, Sugp1, Slc16a7, Fam118b, and Tceanc2. Importantly, although these small subsets of genes were highly predictive of OPE exposure in male and female hippocampus respectively, they alone did not account for significant GO pathway enrichment in either sex (apart from Slc16a7). This suggests that although expression of these genes is reliably altered in OPE-exposed hippocampal tissue, they are not part of a common gene pathway. Similarly, when including all 133 high information genes in males derived from machine learning, or DEGs based on q-values, there was no significant enrichment for any GO pathways. In females, however, pathway analysis of the full 345 high information genes revealed significant enrichment across multiple pathways and functions, whereas q-value analysis yielded no significant enrichment. This likely indicates that the individual OPEs are acting via multiple mechanisms with intersecting pathways rather than a single, common one.

Sex differences in transcriptional response to OPEs likely arise due to the underlying processes of sexual differentiation occurring in the brain coincident with exposure (Marrocco and McEwen, 2016). A primary factor driving brain sexual differentiation is the perinatal androgen surge, which primarily masculinizes the male rodent brain through conversion to estradiol by aromatase, and activation of estrogen receptors. In the hippocampus, both estrogen and androgen receptors coordinate aspects of sexual differentiation and maturation. These steroid hormones elicit changes, including in hippocampus, through fundamental mechanisms of neural development including neurogenesis, cell proliferation, synaptogenesis, epigenetic reprogramming, and cell death (Forger, 2016; Kight et al., 2020; Stockman et al., 2022; Weinhard et al., 2018). It is on these 2 different hormonal backgrounds (androgen exposed vs no androgen exposure) in the perinatal hippocampus that OPEs are exerting their effects in each sex. In fact, there is evidence OPEs directly interact with sex and stress hormone receptors involved in neuroendocrine function in vitro, and in vivo in hypothalamus (Adams et al., 2020; Krumm et al., 2018). In contrast to previous studies, however, the present findings suggest males escape widespread OPE induced changes in transcription (no significant pathway enrichment), at least at the age examined. Whether early life androgens/estradiol exposure may blunt the effects of OPE exposure on gene expression in hippocampus specifically is a hypothesis requiring additional investigation. These sex-specific findings highlight the importance of sex as a biological variable in assessing response to a potentially toxic exposure, and provide additional evidence that the effects of endocrine disrupting compounds heavily depend on sex and developmental epoch at exposure and assessment.

OPEs alter ion channel and mitochondria transcription pathways in females

In female hippocampus, high information genes driving classifier prediction significantly enriched GO Biological processes of cation transport, metal ion transport, and mitochondrial translation and gene expression. Similarly, many of these same genes were responsible for enrichment of the GO molecular functions, voltage-gated ion channel activity and ion transmembrane transporter activity, and enrichment of GO cellular components, mitochondria, and catalytic complex. Examination of the specific genes enriching each pathway revealed components of 3 main actors, voltage-gated calcium and potassium channels, and electron transfer enzymes. OPE exposure decreased transcription of 3 protein subunits of electron transfer enzymes, Atp5pd, Atp6v1c1, and Cox7c, and increased expression of the voltage-gated calcium channel genes Cacng4 and Cacng6 as well as modulators of calcium release Slc35g1, Orai2, and Ano6. These findings are consistent with previous studies, as adult mice treated from P10-P70 with TPHP exhibited reduced expression of Atp5a and Cox5b and altered calcium-dependent synaptic proteins in hippocampus (Zhong et al., 2021). Voltage-gated calcium channel activity and energy metabolism are important features of synapse activity and altered expression of these genes may indicate changes in synaptic plasticity and hippocampal circuit activity.

Additionally in females, OPEs significantly decreased hippocampal expression of several voltage-gated potassium channel subunits and regulators, including Kcnq5, Kcnip3, Kcnj11, Kcne4, and Kcnc1. Potassium channel conductance is critical for typical action potential dynamics in neurons and has an important role in hippocampal circuit activity and related behaviors. For instance, the potassium channel subfamily protein KCNQ5 function is required for afterhyperpolarization current in mouse CA3 pyramidal neurons which limits neuronal firing frequency (Tzingounis et al., 2010), whereas KCNC1 conductance is required for high frequency spiking action potentials in interneurons of the hippocampus and cortex (Erisir et al., 1999; Lien and Jonas, 2003). Mutation of either of these protein subunits, which alter channel conductance, has been linked to intellectual disability, neurodevelopmental delay, and epilepsy in children (Park et al., 2019; Wei et al., 2022). Additionally, the Kcnip3 gene encodes the calcium sensing protein Calsenilin, which is highly expressed in voltage-gated potassium channels of hippocampal neurons and regulates channel conductance and long-term potentiation (Burgoyne, 2007; Lilliehook et al., 2002) a physiological process central to the formation of new memories. Voltage-gated potassium channel function also plays an important role in facilitating early neural network synchronization and circuit connectivity in the hippocampus (along with voltage-gated Ca²⁺ channels) during the first postnatal week (Safiulina et al., 2008). Significantly, these results recapitulate findings derived from traditional q-value analysis in P1 cortex, where regulation of cation channel activity was also a pathway downregulated in gestationally OPE exposed P1 male rat cortex, suggesting OPEs may interact with transcription of these genes in other brain regions across postnatal development (Witchey et al., 2022).

OPE exposure in female hippocampus also decreased expression of metabolic enzyme subunits localized to the mitochondrial catalytic complex such as Cox7c, Atp5pd, Uqcrb, Uqcrq, Pdhx, and Pdhb. These genes code for proteins involved in the activity of the electron transport chain and oxidative phosphorylation (Cox7c, Atp5pd, Uqcrb, and Uqcrq), and represent 3 of the 5 enzyme complexes involved in cellular respiration. Additionally, Pdhx and Pdhb encode subunits of pyruvate dehydrogenase which links glycolysis to the citric acid cycle, a process metabolically upstream of the electron transport chain. Our results suggest OPE exposure changes energy metabolism, an outcome linked with significant pathophysiological effects. Mutations in several of these genes, for example, have been linked to neurometabolic diseases such as Leigh syndrome (Pdhx, Pdhb), and mitochondrial complex III deficiency (Uqcrb, Uqcrq) (Fernández-Vizarra and Zeviani, 2015; Ganetzky et al., 2021). Mitochondria function is closely tied to pre- (Smith et al., 2016) and postsynaptic plasticity (Li et al., 2004), and pathology in mitochondrial anatomy and function has been demonstrated in many neuropsychiatric disorders (Kim et al., 2019). For instance, changes to oxidative phosphorylation are hypothesized to precede depression symptoms (Allen et al., 2018), and disrupted electron transport chain activity is associated with neurodevelopmental disorders including autism spectrum disorder and schizophrenia (Frye, 2020).

Importantly, our hippocampal results are concordant with our 2 prior, related publications, that used more traditional RNAseq analytical approaches. We had previously shown and published that developmental OPE exposure significantly increases gene expression in pathways of respiratory electron transport chain and ATP synthesis coupled electron transport in the cortex of male and female rats at P1 (Witchey et al., 2022), and that OPE exposure upregulates the mitochondrial dysfunction canonical pathway in E14 exposed placenta (Rock et al., 2020). Taken together, this suggests a wider sex- and region-specific dysregulation of mitochondrial function in response to developmental OPE exposure and bolsters the growing view that OPEs may play a role in the etiology of neurodevelopmental disorders.

OPE impacts on mitochondrial physiology in perinatal cortex and placenta

Notably, our machine learning results closely support and augment the previously published, related, studies. In both male and female P1 cortex, genes driving classifier prediction showed significant enrichment for mitochondrial physiology, including oxidative phosphorylation and electron transport chain, as well as pathways of neurodegeneration; outcomes in line with our previous findings (Witchey et al., 2022). Indeed, the specific genes enriching these pathways include subunits of 4 of the 5 catalytic complexes of the electron transport chain (eg Atp, Ndufa, Cox, Uqcr), which were also identified as DEGs in OPE exposed P1 cortex using conventional linear statistics (Witchey et al., 2022). Mitochondrial dysfunction is a hallmark of many neurodegenerative diseases (Johnson et al., 2021) and, accordingly, genes underlying mitochondrial physiology also drove enrichment in KEGG pathways for various neurodegenerative diseases, including Parkinson, Huntington, Prion, amyotrophic lateral sclerosis, and Alzheimer’s (Figs. 5C and 5D).

Genes involved in mitochondrial physiology also informed classifier performance in E14 placenta, particularly in females. Both GO biological processes and KEGG pathways for electron transport chain and oxidative phosphorylation, as well as pathways of neurodegeneration were significantly enriched. The high information genes responsible for enrichment in these cases were similar to those in cortex, including the same 4 catalytic enzymes (Atp Ndufa, Cox, and Uqcr, Figure 7E), and several of these were also present as high information genes in E14 males, however, in the context of pathways for intracellular transport and amyotrophic lateral sclerosis. Mitochondrial dysfunction was also identified as one of several pathways enriched in traditional linear statistical analysis of OPE exposed E14 tissue in females (Rock et al., 2020). Genes driving enrichment in this prior analysis included Atp5mc1, Sod2, Park7, and Ndufa13, all of which were also identified as high information genes in female OPE placenta by the present machine learning methodology. Our machine learning results demonstrate the efficacy of the approach to identify and contextualize important changes in transcription, and validate the conclusions of the previous studies.

OPE exposure impacts oxidative phosphorylation genes across tissues

Cross tissue comparison unequivocally identified oxidative phosphorylation and electron-transport chain function as biological processes impacted by OPE exposure common across tissue type and developmental timepoint. Although the specific set of high information genes overlapping across hippocampus/cortex, and cortex/placenta differ, a consistent, common signature of mitochondrial dysfunction emerged, particularly in females. In contrast, comparison of DEGs generated from q-values (the traditional bioinformatics approach) in these same tissues generated fewer common genes, and no significant overlap in enriched GO pathways based on these genes. This suggests that although traditional q-value analysis successfully detected at least some genes related to mitochondrial physiology in placenta and cortex independently (Rock et al., 2020; Witchey et al., 2022), this method lacked the sensitivity to detect these networks of genes across tissues, likely due to the stringency of the q-value cut-off and resulting exclusion of many genes in the pathway.

Comparison of common high information genes in female P10 hippocampus and P1 cortex revealed significant enrichment for oxidative phosphorylation, electron transport chain activity, and mitochondrial ATP synthesis. These common genes included Atp5pd, Uqcrb, Ndufs5, and Cox7c, coding for subunits in 4 of the 5 enzymes of the electron transport chain, several of which were identified in machine learning analysis of OPE exposed female hippocampus in isolation. This indicates that machine learning algorithms utilized the differential expression of these genes to predict OPE exposure in both P10 hippocampus and P1 cortex, suggesting significant overlap in OPE impact on mitochondrial processes between these tissues. Interestingly, OPEs impacted transcription of the common hippocampus/cortex genes in opposite directions. The 4 common genes were suppressed in hippocampus and upregulated in cortex, suggesting opposed metabolic stresses on these 2 functionally interlinked tissues. This is not entirely surprising given that each region has different developmental timing and at P10 the hippocampus is undergoing substantial differentiation and growth. Similarly, overlap of high information genes between OPE-exposed P1 cortex and E14 placenta occurred in both males and females, showing significant KEGG pathway enrichment for Oxidative phosphorylation in both sexes.

Interestingly, for hippocampus and placenta there was no significant enrichment of any GO or KEGG pathways, despite both tissues overlapping on these pathways with cortex and sharing many high information genes associated with mitochondrial function. This indicates that the specific set of overlapping genes in hippocampus/cortex and cortex/placenta is different. Indeed, there was only 1 high information overlapping gene between all 3 tissues in females, ubiquinol-cytochrome c reductase binding protein (Uqcrb), and zero in males. In addition to tissue-specific differences, this may be due to relatively distant ages (E14 vs P1 vs P10) at which the tissues were collected. That we identified some universality is significant, however, and may guide future investigations into mode of action and biomarkers of OPE exposure. It makes sense biologically that, although the same mitochondrial pathways are affected in each tissue, the specific transcriptional changes, including the individual genes affected in the transcriptome network, vary based on tissue, age, and sex. Ultimately, each dataset represents a snapshot in time and thus, whereas mitochondrial disruption may be a universal phenomenon that persists across time, the specific genes within those pathways impacted at the precise moment the tissues were collected, explicably vary.

Although the cell populations of E14 placenta, P1 cortex, and P10 hippocampus are heterogenous, each contain mitochondria, and each present common targets for molecular interference from OPEs. For example, fetal placenta, postnatal cortex, and hippocampus all highly express both ERα and β (Al-Bader, 2006; Pérez et al., 2003; Wang et al., 2019; Wilson et al., 2011) and the OPE TPHP, has been shown to have estrogen disrupting activity, including directly acting as an ERα and β agonist, activating the estrogen response element, and promoting estradiol synthesis from testosterone (Belcher et al., 2014; Ji et al., 2020, 2022; Kojima et al., 2016). Significantly, estrogen receptor activity is a powerful regulator of mitochondrial function, including in brain (Klinge, 2020; Rettberg et al., 2014), and both ERα and β activity are sufficient to alter the activity of many of the specific enzymes identified as being impacted by OPE exposure including pyruvate dehydrogenase, NADH dehydrogenase, and cytochrome oxidase of the electron transport chain (Irwin et al., 2012; Rettberg et al., 2014). Given the relationship between OPEs, estrogen activity, and mitochondria, paired with the consistent disruption of genes related to mitochondrial function across perinatal tissues, it seems possible that OPEs may be disrupting mitochondrial gene expression through an endocrine mechanism, a possibility meriting further exploration.

In addition to their endocrine disrupting properties, OPEs have also been shown to directly induce significant mitochondrial toxicity. For instance, exposure of mouse hepatocytes to OPEs in vitro (including TPHP) altered mitochondria morphology, reduced basal respiration rate, ATP production, and oxidative phosphorylation/glycolysis rate, all major indicators of impaired oxidative phosphorylation capacity (Le et al., 2021), whereas TPHP exposure similarly reduced basal respiration rate in zebrafish embryos (Lee et al., 2019). Although the mechanisms of this toxicity are unknown, disruption of mitochondrial-related gene transcription could underlie some of these observed physiological changes.

Normal mitochondrial function is critical for meeting the energetic demands of developmental processes such as neurogenesis (Khacho et al., 2019), synaptic transmission (Forgac, 2007; Pathak et al., 2015) and plasticity (Li et al., 2004; Smith et al., 2016). It therefore seems plausible that transcriptional changes to key enzymes in oxidative phosphorylation during critical periods might alter development of neural connectivity and associated behaviors. Developmental OPE exposure has been shown to have detrimental effects on a variety of behaviors, including socioemotional, motor, and anxiety behaviors in rodents (Oliveri et al., 2018; Wiersielis et al., 2020; Witchey et al., 2020). However, information regarding the effects of OPE exposure on behaviors specifically mediated by hippocampus is limited. Gestational exposure to TPHP impaired spatial navigation in minnows (Hong et al., 2018) and spatial memory, and contextual fear memory in rodents (Hong et al., 2022b; Zhong et al., 2021). However, many behaviors associated with hippocampal function including episodic memory and its components, pattern separation, and behaviors associated with affective disorders related to hippocampal activity such as depression have yet to be assessed. Similarly, the effects of OPEs on behaviors mediated by medial prefrontal cortex such as working memory, attention, and cognitive flexibility remain largely untested. Some measures of cognitive deficits in children, however, are suggestive of prefrontal involvement (Doherty et al., 2019; Vuong et al., 2020). For instance, decreased cognitive performance, attention, and impaired socioemotional behaviors in primary school children all correlated with increased, OPE exposure (Castorina et al., 2017; Choi et al., 2021; Hutter et al., 2013; Lipscomb et al., 2017), however, some studies have failed to show these effects (Percy et al., 2021). Given the important role of mitochondrial physiology in development of typical connectivity in these brain regions, and the impact OPE exposure appears to have on mitochondrial-related genes in these tissues, further investigation of potential OPE-induced deficits on these behaviors is of high priority, particularly in females.

Conclusions

Machine learning methods accurately and reliably predicted OPE exposure in P10 hippocampal, P1 cortical, and E14 placental transcriptome in male and female rats. Thus, it is a valuable approach for identifying unifying effects and biomarkers of exposure across sex, age, dose, and tissue type. Cross-tissue comparison of high information genes strongly implicated genes underlying oxidative phosphorylation and electron transport in mitochondria as common processes likely impacted by OPE exposure, particularly in females; an outcome highly concordant with the existing literature. Because normal mitochondrion physiology is critical to typical neural development and synaptic function, dysregulation of this process may lead to impaired circuit formation and behavior; outcomes already shown to be vulnerable to developmental OPE exposure. Significantly, mitochondria toxicity is implicated in several neurodevelopmental disorders including ASD and schizophrenia, thus our findings add weight to the concern that OPEs may be contributing to rapidly rising rates of these and related psychiatric disorders. Finally, these findings emphasize the complementary explanatory power of machine learning analysis when used in conjunction with traditional bioinformatics approaches. Because the machine learning methods utilized here do not rely on linear statistics, they provide orthogonal confirmation to traditional bioinformatics methods that have previously established disruption in genes related to oxidative phosphorylation and mitochondrial dysfunction in cortex and placenta and extends these results to hippocampus.

Supplementary data

Supplementary data are available at Toxicological Sciences online.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

National Institutes of Environmental Health Sciences - grant # ES028110 to H.B.P.