Prenatal Metal Exposures and Childhood Gut Microbial Signatures Are Associated with Depression Score in Late Childhood

Abstract

Background:

Childhood depression is a major public health issue worldwide. Previous studies have linked both prenatal metal exposures and the gut microbiome to depression in children. However, few, if any, have studied their interacting effect in specific subgroups of children.

Objectives:

Using an interpretable machine-learning method, this study investigates whether children with specific combinations of prenatal metals and childhood microbial signatures (cliques or groups of metals and microbes) were more likely to have higher depression scores at 9-11 years of age.

Methods:

We leveraged data from a well-characterized pediatric longitudinal birth cohort in Mexico City and its microbiome substudy (n=112). Eleven metal exposures were measured in maternal whole blood samples in the second and third trimesters of pregnancy. The gut microbial abundances were measured at 9-11-year-olds using shotgun metagenomic sequencing. Depression symptoms were assessed using the Child Depression Index (CDI) t-scores at 9-11 years of age. We used Microbial and Chemical Exposure Analysis (MiCxA), which combines interpretable machine-learning into a regression framework to identify and estimate joint associations of metal-microbial cliques in specific subgroups. Analyses were adjusted for relevant covariates.

Results:

We identified a subgroup of children (11.6% of the sample) characterized by a four-component metal-microbial clique that had a significantly high depression score (15.4% higher than the rest) in late childhood. This metal-microbial clique consisted of high Zinc in the second trimester, low Cobalt in the third trimester, a high abundance of Bacteroides fragilis, a high abundance of Faecalibacterium prausnitzii. All combinations of cliques (two-, three-, and four-components) were significantly associated with increased log-transformed t-scored CDI (β=0.14, 95%CI=[0.05,0.23], P<0.01 for the four-component clique).

Significance:

This study offers a new approach to chemical-microbial analysis and a novel demonstration that children with specific gut microbiome cliques and metal exposures during pregnancy may have a higher likelihood of elevated depression scores.

Graphical Abstract

INTRODUCTION

Depression is a major contributor to morbidity and mortality worldwide, affecting approximately 5% of adults globally (World Health Organization, 2023). In the United States alone, 2.7 million (4.4%) children suffer from depression (Bitsko, 2022), while nearly 25% of adolescents in Mexico experience one or more mental disorders, including depression and anxiety (Fleiz Bautista et al., 2012). While various determinants, such as temperament, biological and genetic factors, as well as social and physical environments, contribute to depression, the etiology is not entirely understood (Malhotra and Sahoo, 2018). Deficiencies in nutrient metals and exposure to toxic metals may also contribute to depression in children (Nguyen et al., 2022), and the gut microbiome (G.M.) may play a role along that pathway (Chen et al., 2021).

The composition and function of the human G.M. is a recognized contributor to human health and disease (Dinan and Cryan, 2017). It has been linked to a wide range of neurological health conditions, including depression (Aarts et al., 2017; Foster and McVey Neufeld, 2013; Hill-Burns et al., 2017; Vogt et al., 2017). While evidence on the association between overall measures of global gut microbial diversity and depression across the lifespan is mixed, there is evidence of an association between depression and increased abundance of pro-inflammatory bacteria and decreased abundance in short-chain fatty acid (SCFA) producing bacteria (Simpson et al., 2021). Evidence on the association between the G.M. and depression in children is limited; however, in a recent case-control study, Hao, et al. (2023) found that both bacterial and fungal taxonomic composition and inter-kingdom networks were altered in cases vs. controls (Hao et al., 2023). Further investigation is needed to understand the association and biological mechanisms between the G.M. and depression in children.

The composition and function of the G.M. can be influenced by exposures like drugs, diet, and environmental pollutants (Breton et al., 2013), including metal exposure, which can disrupt the G.M.’s composition, diversity, homogeneity, and structure (Eggers et al., 2019, 2023b; Kaur and Rawal, 2023). While the composition of the G.M. can alter due to these metal exposures, the G.M. also affects the availability of trace metals for host absorption by either competing with the host for absorption or modifying the metals to less bioavailable forms (Pajarillo et al., 2021). The interplay between gut bacteria and metals is influenced by factors like dosage, duration, timing of exposure, ingestion method, metal type, and bioavailability (Bist and Choudhary, 2022). Exposures to metals during the prenatal period may be particularly critical to both G.M. development and neurobehavior. In a study investigating the effects of prenatal exposure to Pb and stress on the G.M. and neurodevelopment in rats (Hua et al., 2023), findings reveal that both individual and combined exposures lead to deficits in hippocampal structures and learning/memory, with the combined Pb and stress exposure having the most pronounced effect. The study also shows changes in G.M., particularly in the presence of combined exposure, which appears to impact learning/memory negatively. Few epidemiologic studies have examined the influence of prenatal exposures on both G.M. and neurobehavior. Our group found that prenatal Pb was associated with G.M. composition, particularly reducing the abundance of Ruminococcus gnavus, Bifidobacterium longum, Alistipes indistinctus, Bacteriodes caccae, and Bifidobacterium bifidum (Eggers et al., 2023b). In other epidemiologic studies, prenatal Pb and other metal exposures have also been linked to adverse neurobehavior. Previous research investigating the combined impact of prenatal exposure to neurotoxic metals (Pb, Hg, and Cd) on infants’ neurodevelopment at six months found that concurrent Pb and Hg exposure during late pregnancy negatively affects neurodevelopment, particularly when Hg concentrations exceed the 50th percentile (Shah-Kulkarni et al., 2020). Taken together, these studies suggest a potential mechanistic pathway of the G.M. between prenatal metal exposures and neurobehavior; however, the importance of prenatal exposures on the combined developmental trajectory of the G.M. and neurobehavior is not yet understood. While these findings suggest complex interactions among metal exposures, G.M. composition, and neurobehavioral outcomes, we currently have a limited understanding of which metals and microbes interact to influence different aspects of neurobehavior and the timing of that interaction. Such information is critical to inform metal exposure mitigation strategies and potential interventions to protect neurobehavior and promote health. Epidemiologic studies have not yet uncovered these details because they have primarily focused on associations between single bacterial taxa or the global diversity measures of the entire microbiome. However, critical biochemical interactions occur between small groups of bacteria called microbial cliques. Metal exposure may influence Interactions between microbial clique members; thus, investigation is needed to understand metal-microbial cliques. Continuing upon our previous methods development work (Eggers et al., 2023b, 2023a; Midya et al., 2023a, 2023b; Midya and Gennings, 2023), we developed an analytical framework called Microbial and Chemical Exposure Analysis (MiCxA), which combines an interpretable machine-learning model with a downstream inference framework to identify metal–microbial cliques and estimates their associations with depression symptoms. This novel framework adds a new perspective to microbiome analysis, providing the flexibility of machine-learning algorithms while maintaining interpretability and allowing the simultaneous analysis of the G.M. with metal exposures using an Exposomic perspective. This study aims to identify prenatal metal and childhood gut microbial signatures associated with depression symptoms in late childhood using MiCxA.

METHODS

Study Design

All data came from the Mexico City-based Programming Research in Obesity, Growth, Environment and Social Stressors (PROGRESS) cohort. Between 2007 and 2011, a total of 948 pregnant women were enrolled in the study through the Mexican Social Security System. Study visits were completed in the second and third trimesters of pregnancy, at birth, every six months during infancy, and every other year after that by the mothers and their children. Each study visit consisted of physical exams with biospecimen collection, psychological and behavioral assessments, and surveys. Stool samples were collected from a subset of children between the ages of 9-11 (n=123) (Eggers et al., 2023b). Among the 123 participants with stool samples, 112 also had complete outcome data. The PROGRESS study protocol was reviewed and approved by the Institutional Review Board (IRB) at the National Institute of Public Health in Cuernavaca, Mexico, and the Icahn School of Medicine at Mount Sinai (ISMMS) in New York, New York, USA.

Metal Measurement

Venous whole blood samples taken during the second and third trimesters of pregnancy were used to measure metal concentrations. Metals were analyzed in a multi-chemical panel using a dynamic reaction cell inductively coupled plasma mass spectrometry (ICP-MS). Analysis was conducted in the Frank R. Lautenberg Environmental Health Sciences Laboratory at the Icahn School of Medicine at Mount Sinai in New York, NY, USA, using previously published methods (Levin-Schwartz et al., 2021; Tamayo y Ortiz et al., 2016). Metal concentrations used in this analysis were lead (Pb), arsenic (As), cadmium (Cd), manganese (Mn), cobalt (Co), zinc (Zn), chromium (Cr), cesium (Cs), copper (Cu), antimony (Sb), and selenium (Se). The laboratory elemental analysis quality assurance and quality control (QA/QC) procedures included the analyses of calibration standards in the range of 0.001 to 50 ng ml⁻¹, initial verification standards (ICVS), and continuous calibration verification standards (CCVS). QC samples consisted of mixed element standards (purchased from a different vendor as the calibration standards) at two different concentration levels, procedural blanks, 5% of samples prepared and analyzed in duplicate, in-house pooled blood samples at three levels (IHB) to monitor the accuracy, and lastly, the certified reference materials Seronorm Blood (L2, L3) - SERO AS, Billingstad, Norway, and National Institute of Standards and Technology (NIST) standard reference material (SRM) 955c (L1, L2, L3, L4) Toxic Metal in Caprine Blood, Gaithersburg, US. CCVS and in-house pooled samples were run after analysis of every ten samples. All Lab recovery rates for QC by this method were 90 to 110%, and inter-day and intra-day precision (given as a percentage of relative standard deviation) is <6% for samples with concentrations more than the limit of quantification. The method applied undergoes periodical proficiency testing accreditation by the QMEQAS (Quebec Multielement External Quality Assessment Scheme) and the New York State Biomonitoring Program for Trace Elements. The limits of detection (LOD) for each metal were: 0.391 ug/L for As, 0.113 ug/L for Cd, 0.117 ug/L for Co, 0.901 ug/L for Cr, 0.122 ug/L for Cs, 1.862 ug/L for Cu, 0.442 ug/L for Mn, 0.377 ug/L for Pb, 0.159 ug/L for Sb, 0.665 ug/L for Se, and 5.053 ug/L for Zn. We included all 11 metals, allowing us to take the data-driven microbial-metallomics approach and generate hypotheses to be tested.

Gut Microbiome Sample Collection and Processing

Stool sample collection procedures, library preparation, and data processing have been previously described (Eggers et al., 2023b). Briefly, whole stool samples were collected by participants at home and stored in their refrigerators for up to 24 hours. PROGRESS field staff picked up and delivered the samples to the ABC Hospital in Mexico City, where they were aliquoted using the FAST protocol (Romano et al., 2018), and stored at −70C. Frozen samples were shipped to the Microbiome Translational Center at ISMMS. They underwent DNA extraction and library prep, followed by shotgun metagenomic sequencing on an Illumina HiSeq, in two separate batches (n=50 and n=73). Raw sequencing reads were trimmed, and human DNA was removed using Trimmomatic (Bolger et al., 2014) and Bowtie 2 (Langmead and Salzberg, 2012). Microbial taxonomy was assigned using MetaPhlAn2 (Truong et al., 2015) and StrainPhlAn (Truong et al., 2017). This website https://www.ncbi.nlm.nih.gov/bioproject/PRJNA975184/ contains publicly available raw sequencing data. Note that we excluded all children who had any antibiotic use within the last month before fecal sample collection.

Depression Measurement

The Childhood Depression Inventory 2 Self-Report Short version (CDI) was used to score depression symptoms of 9- to 11-year-old pediatric participants by a trained researcher (Kovacs, 2015). The CDI is a questionnaire appropriate for participants aged 7-17 years and validated in Spanish (Cumba-Avilés et al., 2020). Items are scored on a scale normalized from 0 to 100, with higher scores indicating worse depression symptoms. The raw CDI scores were converted to t-scores with a range between 40 to 90.

Covariates

The Covariates considered in this analysis were child sex, maternal socioeconomic status (SES) during pregnancy, maternal age at birth, maternal body mass index (BMI) during pregnancy, child age at the time of stool collection, and metagenomic analysis batch. The mother’s height and weight were measured during the second trimester using professional digital scales and a stadiometer. From these measurements, the BMI was calculated and subsequently treated as a continuous covariate for regression analyses. To assess socioeconomic status during pregnancy, the 1994 Mexican Association of Intelligence Agencies Market and Opinion (AMAI) version 13*6 was employed. This classification system places families into six distinct levels based on responses to 13 questions concerning household characteristics. As most families in the study belonged to the low to middle SES bracket, these six categories were consolidated into three broader classifications: lower, middle, and higher (Sanders et al., 2022). Given the smaller sample size, we kept the set of covariates minimal (similar to the previous work (Midya et al., 2023b)); however, we utilized the idea of negative control outcome (Arnold and Ercumen, 2016) to address any selection bias and residual confounding (see Sensitivity analysis).

Statistical Analysis

All statistical analyses were conducted in R (version 4.3.0). The p-values were corrected for multiple comparison errors using a false discovery rate (FDR). Any two-tailed p-value < 0.05 is considered statistically significant.

Data Processing

The t-scaled CDI (t-CDI) score was log-transformed to meet distributional assumptions. The microbiome count data was transformed to relative abundance positive ratios that sum to 1 for each individual. The following analysis included only those microbial taxa with at least 5% relative abundance ratios across both batches to account for analytical batch effects. After screening, we chose not to rescale the relative abundance to reflect the contribution of the original distribution of the whole taxa. Further, all models were adjusted for a batch indicator variable. In this analysis, we only used the taxa-level relative abundance since it provides higher granularity. Any missing data in the covariates or metal exposures were imputed using the predictive mean matching implementation of the MICE package in R (Van Buuren and Groothuis-Oudshoorn, 2011). We measured the alpha and beta diversity measures using the Shannon diversity index (Shannon, 1948) and Bray–Curtis (Bray and Curtis, 1957) metric as implemented in the vegan R package (Oksanen et al., 2019).

Univariate Analyses

For all univariate analyses, the concentrations of metal exposures and the relative abundance values of each taxon were transformed into quartiles. Such transformation of relative abundance provides a balance in this generally skewed data. We constructed two forest plots to report the estimated associations and 95%CIs for each metal exposure and log(t-CDI) at each trimester. The association between each taxon and log(t-CDI) was presented through a volcano plot. Each model was adjusted with previously mentioned covariates.

Microbial and Chemical Exposure Analysis (MiCxA)

To identify metal-microbial cliques and estimate their associations and interactions with childhood depression symptoms, we developed MiCxA, combining an interpretable machine-learning tool with an inference framework. MiCxA follows the evolution from Microbiome Co-occurrence Analysis (MiCA), which aimed to identify only microbial cliques (Midya et al., 2023b). For interpretability and reproducibility, we provided details about this algorithm, similar to our previous paper. We conducted MiCxA in two stages: 1) identify metal-microbial cliques that predict log(t-CDI) scores and 2) examine the association (i.e., estimate the regression coefficients) between identified metal-microbial cliques and log(t-CDI) through an inference framework. We used the repeated holdout signed-iterated Random Forest (rh-SiRF) (Midya et al., 2023b, 2023a), treating relative abundances of the microbial taxa and concentrations of metal exposures during both the second and third trimesters as predictors and log(t-CDI) as the outcome. The rh-SiRF algorithm combines “Iterative Random Forests” with “Random Intersection Trees” to search for combinations of microbial taxa and prenatal metal exposures that predict the log(t-CDI) scores (Basu et al., 2018; Kumbier et al., 2018; Shah and Meinshausen, 2014). These predictive combinations of metal and microbial taxa were chosen following the branches grown in the decision trees. Instead of searching for all possible metal and microbial taxa combinations (3321 two-component, 88560 three-component), rh-SiRF teases out the most prevalent and predictive combinations on the decision path of its branches. A bagging step was also introduced to estimate the “stability” of the discovered combinations estimated through bootstrapped iterations. We added a repeated holdout step that randomly partitioned the data in training and testing sets to improve generalizability (Midya et al., 2023b). The rh-SiRF algorithm was repeated 1000 times, with a bootstrap step iterated 250 times on a training/test data partitioning of 60%/40%. All prevalent combinations with more than 25% stability were chosen for further downstream inference analysis. Lastly, it is important to note that the backbone of these algorithms doesn’t differ; the key difference between MiCxA and MiCA lies in how the chemical exposure and relative abundances are used in the algorithm. In this analysis, we simultaneously used both exposures and abundances as predictors in a single step; however, given a larger sample size, one can choose to use these predictors step-wise. Unlike the univariate analyses, the metal exposures and relative abundance values were not transformed into quantiles for MiCxA to provide the most flexibility while growing branches. Note that this part of the analysis provides combinations of the metals and microbes. The next stage will delve into converting those combinations to cliques.

From combinations to cliques

Among the list of most stable metal and microbial taxa combinations, we chose the top 1% combinations. These combinations are visualized through the Fruchterman-Reingold Layout of a forced-directed graph (Kobourov, 2012) implemented through the igraph package in R (Csárdi et al., 2023; Csardi and Nepusz, 2006). The chosen metal and microbial taxa combinations were transformed into cliques, which are essentially indicator functions that identify specific subgroups. We implemented a quantile-based threshold-finding algorithm that converts the combinations into indicators with respect to their threshold concentrations (in the case of metals) or relative abundances (in the case of microbial taxa). For example, consider a three-component combination of a metal-microbial clique consisting of metal A, metal B, and microbial taxa C, which was transformed into a three-component clique A+B+C− as an indicator function. Here, the clique A+B+C− implies higher concentrations of metals A and B (above certain thresholds) and a lower relative abundance of microbial taxa C in the sample. The binarized form of A+B+C− denotes an underlying sub-sample (or subgroup) satisfying the conditions of the clique. These interpretable cliques are then used in linear regression to obtain association estimates, adjusted for covariates and confounders. A schematic of this algorithm and R code with illustrations on a simulated dataset is provided on GitHub (https://github.com/vishalmidya/MiCA-Microbial-Co-occurrence-Analysis/blob/main/MiCA-vignette.md). Note that the p-values from the linear regressions were estimated without relying on any large-sample asymptotic arguments of normality by permuting the outcome 10⁴ times for each regression (Midya et al., 2023b).

Sensitivity analysis

We conducted multiple sensitivity analyses: (1) the rh-SiRF algorithm was repeated 1000 times by permuting the CDI outcome with the expected result that none of the two-component metal-microbial cliques would be detected with more than 0.1% stability; (2) the regression-based inference framework was repeated twice using separate thresholds, first increasing each threshold by ten percentile and second decreasing each threshold by ten percentile; we re-estimated the associations between metal-microbial cliques and the outcome (3) without imputing any missing covariate data, and (4) using a negative control outcome (Arnold and Ercumen, 2016) – the presence of any pet in the household at the time of CDI score measurement. We assumed that this negative control outcome would have similar potential sources of selection biases and residual confounding but would not be possibly associated with metal-microbial cliques; (5) without relying on large-sample asymptotic arguments, we conducted randomization-based inference by permuting the outcome 10⁴ times for each of the regressions and estimating the randomization-based p-value (Midya et al., 2023b); (6) to assess the extent of overfitting, we binarized the log(t-CDI) outcome and re-fitted the regression model with the same metal-microbial cliques obtained from the rh-SiRF with continuous log(t-CDI) outcome; lastly (7) we implemented a covariate balancing matched-sampling strategy that aimed to obtain similar covariate distributions within the two groups of binarized cliques – below and above the detected threshold (Greifer, 2023). Given the “similar” covariate distributions, we aimed to create “exchangeable” groups such that the clique was hypothetically and randomly assigned to each individual, and the covariates did not confound the clique assignment. We repeated the associations between each metal-microbial clique and CDI outcome after covariate balancing using subclass matching with the propensity score (Ho et al., 2011).

RESULTS

Study Population

Of the 112 study participants, there were slightly more males than females, the average age was between 9 and 10, and most were in the lower SES group (Table 1). The prenatal metal concentrations and the percentage detected above the limit of detection were presented in Supplemental Table 1. The metal exposures were more correlated within trimesters than between. Particularly, Cs, Cu, Sb, and Se were more highly and positively correlated within both trimesters (Pearson correlations > 0.45) (Supplemental Figure 1). When considering correlations between exposures to the same metals in different trimesters, Pb, Cd, and Mn were the most (positively) correlated across trimesters. The t-CDI scores ranged from 40-81, with a median of 51. Both the alpha and beta diversities had weak associations with log(t-CDI)[alpha-diversity: β = 0.03, p-value: 0.55, and beta-diversity: F=0.87, p-value:0.56].

Table 1.

Descriptive statistics of covariates, exposures, and the outcome from the study population stratified by alpha diversity (N = 112)

	Overall	Below median Alpha diversity	Above median Alpha diversity	p- value
Covariates
Child Sex				0.56
Male n(%)	68 (60.71)	36 (64.29)	32 (57.14)
Female n(%)	44 (39.29)	20 (35.71)	24 (42.86)
Maternal SES in pregnancy				0.99
Lower n(%)	61 (54.46)	30 (53.57)	31 (55.36)
Medium n(%)	41 (36.61)	21 (37.50)	20 (35.71)
Higher n(%)	10 (8.93)	5 (8.93)	5 (8.93)
Maternal age at birth (years) mean(Sd)	28.71 (5.82)	29.07 (5.68)	28.36 (6.00)	0.42
Maternal BMI in pregnancy (kg/m2) mean(Sd)	27.27 (4.47)	27.26 (4.30)	27.29 (4.67)	0.99
Child age at gut microbial sample collection (years) mean (Sd)	9.67 (0.88)	9.62 (0.75)	9.72 (1.00)	0.83
Metal Exposures (ug/L) median(IQR)
Second Trimester
Zn	5928.96 (1220.87)	5794.68 (1012.40)	6083.78 (1376.82)	0.39
Co	0.18 (0.12)	0.19 (0.11)	0.17 (0.12)	0.83
Third Trimester
Zn	6214.53 (1139.77)	6195.60 (1053.53)	6225.39 (1177.21)	0.39
Co	0.25 (0.20)	0.26 (0.21)	0.24 (0.18)	0.89
Outcome
T-scored Childhood Depression Inventory mean(Sd)	52.97 (8.14)	52.20 (8.01)	53.75 (8.26)	0.38

Zn: Zinc; Co: Cobalt; Alpha diversity was calculated using the Shannon diversity index.

Association between individual Metals and log(t-CDI)

In covariate-adjusted regression analysis, metal exposures in the second and third trimesters of pregnancy were significantly associated with log(t-CDI) scores at 9-11 years of age (Figure 1). In the second trimester, higher Zn exposure was associated with an increase in log(t-CDI) score (β[95% CI]=0.04[0.01, 0.06], p-value: 0.008), while higher Cr in the second trimester was associated with decreased log(t-CDI) in late childhood (β[95% CI]= −0.03[−0.05, 0.00], p-value: 0.05). Similarly, higher Co and As in the third trimester were associated with decreasing log(t-CDI) scores (β[95% CI]=−0.04[−0.07, −0.02], p-value: 0.002 and β[95% CI]=−0.03[−0.06, −0.01], p-value: 0.02, respectively), while Cr in the third trimester was associated with increasing log(t-CDI) (β[95% CI]=0.03[0.00, 0.05], p-value: 0.04). Only the FDR-adjusted p-value for association with Co in the third trimester was lower than 0.05.

Figure 1:

Association between individual prenatal metal exposures and log-transformed t-CDI scores among 112 PROGRESS subjects ages 9-11 years. A) shows associations (beta estimates and 95% CIs) with metal exposures in the second trimester and B) in the third trimester. The beta estimates imply a change in log(t-CDI) per quartile increase in each prenatal metal exposure. The dotted verticle line denotes the null value. All the analyses were adjusted for covariates. t-CDI, t scored Child Depression Index.

Volcano plot: relative abundance of microbial taxa and log(t-CDI)

Results from the volcano plot showed that several bacterial taxa were associated with log(t-CDI) (Figure 2). Relative abundance of Faecalibacterium prauznitzii was strongly associated with increased log(t-CDI) score (β=0.04, p-value=0.04), followed by Bifidobacterium longum (β=0.03, p-value=0.09), and Bacteroides fragilis (β=0.02, p-value=0.09). Relative abundance of Eubacterium eligens was negatively associated with log(t-CDI) score (β=−0.03, p-value=0.06). None of the FDR-adjusted p-values were lower than 0.05.

Figure 2:

Association between relative abundances of gut microbial taxa and log(t-CDI) among 112 PROGRESS subjects ages 9-11 years. The red and the black horizontal lines denoted log(p-values) at 0.05 and 0.1, whereas the red-dotted verticle line denoted the null value. The names of the bacterial taxa with p-value < 0.1 are presented in the volcano plot. The blue and the black dots represent negative and positive beta estimates, respectively. The beta estimate denotes change in log(t-CDI) per quartile increase in relative abundacnes per taxa. t-CDI, t scored Child Depression Index.

Repeated holdout SiRF and metal-microbe cliques

A total of 616 unique two-component cliques were detected, with only 2% having a stability (frequency of occurrence) of more than 1% (see Supplemental Table 2 for the list of all combinations with a stability of more than 1%). We chose the top 1% of cliques from this list as those constitute a closed-looped network. The interconnection of the chosen cliques is presented in Figure 3 through a network graph with Fruchterman-Reingold Layout (Csárdi et al., 2023). The closed-loop graph implied six 2-component, four 3-component, and one 4-component clique. These metal-microbe cliques were formed based on high Zn in the second trimester (concentration greater than 40^th percentile of the sample), low Co in the third trimester (concentration below sample median), high F. prausnitzii (relative abundance greater than 40^th percentile of the sample), and high B. fragilis (relative abundance greater than 60^th percentile of the sample). The most stable clique (with the highest frequency of occurrence) consisted of a high concentration of Zn in the second trimester and a low concentration of Co in the third trimester (35% of participants). A high relative abundance of F. pausnitzii and a high relative abundance of B. fragilis were identified as the most stable microbial clique, with almost 26% of participants. Finally, the four-component clique included high Zn in the second trimester, low Co in the third trimester, high F. prausnitzii, and high B. fragilis, with almost 12% of the sample.

Figure 3:

Network graph of metal-microbe cliques. The edges were weighted according to the frequency of occurrences (stability) of that particular clique based on the rh-SiRF results. High Zn: Concentration greater than 40^th percentile of the sample; Low Co: Concentration below sample median; High F. prausnitzii: relative abundance greater than 40^th percentile of the sample; High B. fragilis: relative abundance greater than 60^th percentile of the sample; Zn: Zinc; Co: Cobalt

Correlation between components of metal-microbial cliques

We estimated the Spearman correlations for the Co and Zn concentrations at the second and third trimesters and the relative abundances of B. fragilis and F. prausnitzii to understand the effect of correlation in forming the metal-microbial cliques (Figure 4). The Spearman correlations among the components remained minimal, with the highest absolute correlation of 0.2 between Zn and Co, implying a negligible effect of correlation in forming the cliques.

Figure 4:

Spearman correlation between components of metal-microbial cliques.

Associations between metal-microbe cliques and log t-CDI

The interpretable cliques were used in the regression framework to estimate the associations with log(t-CDI) scores, adjusted for previously mentioned covariates and confounders (Figure 5). We presented the individual associations (top row in Figure 5) to compare effect estimates with those of the higher-component cliques. Low Co in the third trimester had the largest association with increasing log(t-CDI) score (β[95% CI]=0.11[0.05, 0.16], p-value<0.001), with high F. pausnitzii a close second (β[95% CI]=0.08[0.02, 0.13], p-value<0.01). The two-component clique with low Co in the third trimester and high F. prausnitzii had the largest beta estimate among all six 2-component cliques (β[95% CI]=0.13[0.07, 0.18], p-value<0.001). Similarly, all four 3-component cliques were significantly and positively associated with log(t-CDI). The four-component clique, including a high Zn in the second trimester, a low Co in the third trimester, a high F. prausnitzii, and a high B. fragilis, had the largest effect estimate of all possible clique combinations (β[95% CI]=0.14[0.05, 0.23], p-value=0.003). To contextualize this finding, note that the mean(S.E.) of this estimated effect size was 7.7(2.5) in the non-log transformed scaled CDI. Considering the range of the t-scored CDI, it implies that the mean CDI score for almost 11.6% of participants with all four-component cliques was 15.4% higher than the rest of the children. Note that the cliques, being binary indicators, can characterize subgroups of the sample. All p-values from the metal-microbial clique associations remained statistically significant after correcting for FDR.

Figure 5:

Association between log(t-CDI) and individual metals, microbes, and metal-microbe cliques. Estimated associations (β[95% CI]) of each metal–microbe clique with log t-scored CDI. The y-axis denotes the cliques’ names and the proportion of the sample (in brackets); The individual associations (top) are presented to compare individual effect estimates with those of the cliques; High Zn: Concentration greater than 40^th percentile of the sample; Low Co: Concentration below sample median; High F. prausnitzii: relative abundance greater than 40^th percentile of the sample; High B. fragilis: relative abundance greater than 60^th percentile of the sample; Zn: Zinc; Co: Cobalt; 2T = 2nd trimester; 3T= 3rd trimester. Note that for consistency, individual associations for metals and microbes were framed as one-component cliques.

Sensitivity analysis

The results of the sensitivity analyses are presented as follows: (1) while permuting the outcome, the rh-SiRF algorithm result did not find any of the reported cliques more than 0.1% times (the metal-microbial clique of low Co in the third trimester and high F. prausnitzii occurred only once in 3819 times, while no other reported clique was identified in any of the iterations); (2) the directionalities of all the clique associations remained unaltered while each of the thresholds was increased and decreased by ten percentiles (Supplemental Figure 2 and 3); (3) the effect sizes of clique associations remained almost unaltered while repeating the analysis without imputing any missing covariate (Supplemental Figure 4); (4) all the cliques had nonsignificant associations with the negative control outcome, simultaneously representing a drastically different pattern of association, and, therefore strengthening the possibility of minimal selection bias and residual confounding (Supplemental Figure 5); (5) all the randomization-based p-values of the clique associations were lower than the large-sample asymptotic model-based p-values, except for the two 2-component cliques, low Co in the third trimester & high Zn in the second trimester and high Zn in the second trimester & high F. prausnitzii, where the randomization-based p-values were slightly higher, 0.06 and 0.01, from 0.03 and 0.002 (Supplemental Table 3); (6) the clique associations with binarized log(t-CDI) showed similar pattern and strength of associations as that of the continuous outcome (Supplemental Figure 6); and lastly, (7) the association estimates for cliques remained almost unaltered with that of results from Figure 4, implying clique assignment was not influenced by choice of covariates (Supplemental Figure 7).

DISCUSSION

To our knowledge, this is the first epidemiologic study to examine the combined associations of prenatal metal exposures and the effect of human gut microbial cliques on childhood depression in specific subgroups of children. We used a novel statistical approach (MiCxA) that combined interpretable machine learning tools with downstream regression-based inference. Using metal concentrations measured in second and third-trimester blood samples and gut microbiome taxa measured in 9-11 year age stool samples, our results suggest that a subgroup of children (11.6%) characterized by (1) high concentration of Zn in maternal blood samples during the second trimester, (2) low concentration of Co in the third trimester, combined with (3) a high abundance of both B. fragilis and (4) F. prausnitzii in the G.M. during childhood, had a mean CDI score 15.4% higher than the rest.

The word “clique” was first used in a graph-theoretical work to model groups of people who tend to know each other (Luce and Perry, 1949). Gradually, multiple problems in bioinformatics were solved using ideas related to cliques. Consequently, in microbiome research, multiple clique-based community detection algorithms were proposed and implemented throughout the years (Bhar et al., 2022; Kim et al., 2019; Mengucci et al., 2022). In most of these algorithms, the initial network was constructed using measures of correlations; therefore, the clique members were also correlated. However, clique members identified from MiCxA may not be necessarily correlated and were not correlated in this analysis because of their underlying tree-based algorithm, therefore conveying the possibility of nonlinear interaction among the clique members. While few, if any, comparable epidemiologic studies exist, to our knowledge, several studies using animal models have examined the mediating role of the G.M. between both nutrient and toxic metal exposures and neurobehavior, including depressive behaviors (Sauer and Grabrucker, 2019; Shen et al., 2022). Lead exposure has been shown to change the composition of the G.M., reduce SCFA production, and induce depression-like behavior in adult rats (Chen et al., 2022). Moreover, the introduction of probiotics mitigated the changes in SCFA production and depression-like behavior. Likewise, a study in adolescent mice found that exposure to Pb during brain development was also associated with changes to the G.M. as well as the onset of anxiety- and depression-like behavior and that probiotic intervention restored gut microbial function and reduced neurodivergent behavior (Z. Zhang et al., 2023). Considering nutrient metals, maternal Zn deficiency during pregnancy has been linked to preterm birth and pregnancy complications (Chaffee and King, 2012; Iqbal and Ali, 2021); however, prenatal Zn has been understudied in relation to child neurological outcomes (Cortés-Albornoz et al., 2021). Studies of Zn-deficiency and Zn supplementation during pregnancy have shown mixed results for depression and post-partum depression in mothers (Aoki et al., 2022; Fard et al., 2017; Hulsbosch et al., 2023; Rokoff et al., 2023), and while Zn intake in early childhood may be linked to neurodevelopmental outcomes (Ross et al., 2023), little is known about the effects of prenatal Zn on depression in children. Studies considering the G.M. as a mediator or modifier along this pathway are even more sparse. A study in a mouse model found that Zn-deficiency during pregnancy led to changes in the G.M., intestinal permeability, and neuroinflammation in mothers, which was ameliorated by Zn supplementation; however, health effects in pups were not examined (Sauer and Grabrucker, 2019). It may be difficult for experimental studies of Zn deficiency to examine children’s depression due to long onset time and severe Zn deficiency causing birth defects in offspring (Hurley et al., 1971). Our study is one of the first to indicate a negative association between prenatal Zn exposure and depression symptoms in childhood. However, it is important to note that our results do not indicate a negative association with Zn alone but only in combination with the other identified clique members.

Even less is known about prenatal Co exposure, the G.M., and depression in childhood. Co is a biologically necessary metal as a key component of Vitamin B12; however, high levels of Co exposure have been linked to adverse neurological, cardiovascular, and endocrine health (Leyssens et al., 2017). Co exposure has been associated with differences in gut microbial diversity and composition in human studies and animal models (Richardson et al., 2018; J. Zhang et al., 2023), while vitamin B12 level has been inversely correlated with several gut bacteria,(Qi et al., 2023) including Akkermansia muciniphila (Al-Musharaf et al., 2022), a species known for its antidepressant effects (Cheng et al., 2022; Ding et al., 2021; Sun et al., 2023). Taken together, these studies suggest a potential pathway from low Co to depressive symptoms through vitamin B12 and gut microbes such as A. muciniphila, although evidence of the association between Co, vitamin B12, and depression is limited. Lower levels of vitamin B12 in pregnancy have shown mixed associations with maternal perinatal depression (Chong et al., 2014; Dhiman et al., 2021; Peppard et al., 2019; Ramadan et al., 2022), however, there is evidence to suggest that Vitamin B12 supplementation may be an effective adjuvant strategy for improving depressive symptoms in pregnant women who are undergoing other forms of treatment (Borges-Vieira and Cardoso, 2023). Limited evidence also suggests that vitamin B12 deficiency in childhood may contribute to childhood depression symptoms (Esnafoglu and Ozturan, 2020), however, more research is needed. One study using a mouse model has shown that vitamin B12 deficiency in pregnant mice was associated with reduced social behavior in their offspring, but did not find a significant association with depression-like behavior (Xu et al., 2021). To our knowledge, our study is among the first to link low prenatal Co levels to depression symptoms in childhood, and considering the limited evidence of this association across the lifecourse, further research is warranted.

Few studies have examined the links between B. fragilis and depression. One study found B. fragilis to be more abundant in the G.M. of adults with Major Depressive Disorder (MDD) compared to controls. The same study administered B. fragilis to antibiotic-treated mice, inducing depression-like behavior, impairing hippocampal neurogenesis, and depleting hippocampal serotonin levels (Zhang et al., 2022). A key factor in depression etiology is the altered function of the GABAergic system (Prévot and Sibille, 2021), which has made the GABAergic system a promising target of therapeutics (Luscher et al., 2023). B. fragilis is a producer of gamma-aminobutyric acid (GABA), the primary inhibitory neurotransmitter in humans, which is both produced and consumed by members of the G.M. (Petroff, 2002; Strandwitz et al., 2019). F. prausnitzii is a consumer of GABA (Strandwitz et al., 2019). Human studies of F. prausnitzii in Bipolar Depression and MDD have shown mixed association results (Kovtun et al., 2022; Lu et al., 2019). F. prausnitzii administration in a rat model induced beneficial effects on depression-like behavior (Hao et al., 2019). While the direct impact of F. prausnitzii and B. fragilis on depression needs additional investigation, their opposite and complementary functions in GABA metabolism make their joint association with depression symptoms in our study likely to operate through the GABAergic system.

MiCxA provides several statistical advantages to previous methods, enabling the combination of multiple metal exposures and gut microbial abundances. The interpretable machine learning method efficiently circumvents the high dimensionality of exposures and microbes but still provides meaningful and interpretable cliques. Searching for a relevant metal-microbial combination without a prior hypothesis becomes difficult because of the extremely high number of potential combinations (3,321 two-component and 88,560 three-component). The use of machine learning techniques reduces this computational burden. The metal-microbial cliques are interpretable and can be used for later downstream analyses using simple regression or causal inference framework. Moreover, as shown in this analysis, the components of the metal-microbial cliques may not be correlated, implying the possibility of identifying nonlinear interactions. Chemical exposure mixture models like Bayesian Kernel Machine Regression and related tools (Bobb et al., 2018; Liu et al., 2022; McGee et al., 2023) can provide nonlinear pairwise interactions; however, those are qualitative, challenging to interpret, and do not provide thresholds. Similarly, methods for clustering and dimensionality reductions had been used previously on microbial data (Shi et al., 2022); however, interpreting the components or clusters is still challenging. Ensemble methods like MiCxA not only provide interpretability but also transportability for further analysis. Multiple previous methods have been used to discover combinations of interacting chemicals (Lampa et al., 2014; Stingone et al., 2017); however, these tools lacked the transportability that MiCxA provides.

Several limitations warrant consideration in the interpretation of this study’s findings. Firstly, the relatively small sample size of the study and the potential for batch effects might introduce variability and impact the generalizability of the results; however, we took multiple cautionary measures to ward off any bias due to batch effect. Secondly, while the study explores the intricate relationship between prenatal metal exposure, the G.M., and depression symptoms, the concurrent collection of microbiome and depression symptom data precludes a comprehensive mediation analysis. Furthermore, the simultaneous assessment of depression and microbiome data lacks a definitive timeline, raising concerns about the direction of the association and the possibility of reverse causality. The CDI score was based on a self-reported instrument, which may have resulted in response bias and inaccuracies due to social desirability or poor recall. Although diet is an important factor that alters the human microbiome (David et al., 2014), we did not adjust children’s diets in our regression models due to the limited availability of this data. Further, we acknowledge the possibility of residual confounding due to drug use, even though we limited our sample to children with no antibiotic use within one month before fecal sample collection. Lastly, the use of metals measured in maternal blood rather than directly in children introduces a level of indirectness in the exposure assessment, potentially affecting the precision and accuracy of the results. These limitations collectively highlight the importance of acknowledging the limitations inherent in the study design and emphasize the necessity of future research endeavors with larger sample sizes, more refined methodologies, and longitudinal approaches to advance our understanding of the intricate interplay among these factors.

In light of the study’s findings and limitations, two key considerations can strategically guide future research directions. Firstly, incorporating additional exposures and biomarkers, such as diet or other omics data, presents an opportunity to enhance the comprehensiveness of the analysis. Secondly, leveraging longitudinal study designs will allow us to examine the true mediation by G.M. By integrating these approaches, future research endeavors can build upon the study’s foundation, offering more profound insights into the dynamic interactions shaping the interwoven dynamics of prenatal exposures, the G.M., and mental health outcomes.

CONCLUSION

In this study, using a novel analytical method (MiCxA) for microbial exposomics, we found subgroups of children susceptible to increased depression scores, characterized by a unique combination of prenatal metals and components of the childhood gut microbial cliques. Future exposomic studies can broadly apply the MiCxA framework to consider multiple types of omics layers to characterize susceptible groups better, ultimately allowing a better understanding of their combined impact on health.

Highlights.

• Use of a novel machine-learning-driven method, Microbial and Chemical Exposure Analysis (MiCxA) to identify susceptible subgroups of children

• External and internal exposome components associated with childhood depression symptoms

• Children with specific gut microbes and prenatal metal exposures have increased depression symptoms

Data Availability Statement

Metagenomic data is publicly available at https://www.ncbi.nlm.nih.gov/bioproject/PRJNA975184/. All other data will be made available upon request to robert.wright@mssm.edu.