DNA Methylation of FAM50B/PTCHD3 Mediates the Relationships between Low Blood Lead Exposure and Neurobehavioral Development of 0–3 Aged Infants: A Prospective Birth Cohort Study in Southern China

Abstract

Lead, recognized by the World Health Organization as one of the 10 chemicals of major public health concern, ranks as the fourth leading environmental risk factor contributing to the global burden of disease. Nevertheless, the neurodevelopmental consequences of prenatal low-level lead exposure remain inadequately characterized, with limited understanding of its mechanistic underpinnings and a lack of robust biomarkers for susceptibility. In this birth cohort study, we quantified the concentrations of 27 metals along with DNA methylation levels at 12 gene regions in cord blood. Neurodevelopment was assessed longitudinally in infancy using the Ages and Stages Questionnaires (ASQ). Associations between metal exposures and neurodevelopmental outcomes were evaluated via three complementary statistical approaches: mixed-effect models, quantile g-computation, and Bayesian kernel machine regression. Among all metals examined, only lead (mean concentration of 15.3 μg/L) exhibited consistent and statistically significant negative associations with neurodevelopmental performance. A linear dose–response relationship was observed between lead levels and deficits in problem-solving, fine motor, and gross motor skills. Furthermore, we identified four CpG sites within FAM50B and PTCHD3 that mediate the effects of lead exposure and demonstrated strong predictive capacity for neurodevelopmental outcomes using a random forest model. Our results provide novel evidence that even low-level prenatal lead exposure adversely affects early neurodevelopment and implicate FAM50B/PTCHD3 methylation as both a promising biomarker of lead-related neurodevelopmental risk and a potential target for therapeutic intervention.

1. Introduction

Although global lead exposure has substantially decreased since the phase-out of leaded gasoline, it remains a pervasive public health challenge, particularly in low- and middle-income countries. , According to a 2020 Lancet report, an estimated 815 million children, approximately one-third of the global pediatric population, have blood lead levels (BLLs) exceeding 50 μg/L. The World Health Organization (WHO) has identified lead as one of the 10 chemicals of major public health concern. Based on the 2019 Global Burden of Disease, Injury, and Risk Factors Study (GBD), lead ranks as the fourth most significant environmental risk factor worldwide and is the single most impactful toxic substance on global health. − Lead exposure accounts for approximately 30% of cases of idiopathic intellectual disability globally, with its neurotoxic effects on child development being considered irreversible. , Extensive studies support a linear no-threshold (LNT) model for the dose–response relationship between BLLs and adverse health outcomes, indicating that even very low concentrations (1–10 μg/L) may cause harm, and risk increases proportionally with exposure levels. , Children are among the most vulnerable populations, with no established threshold below which lead exposure is considered safe. , Although direct evidence remains limited, the WHO concludes that no safe level of lead exposure has been identified. Nevertheless, critical uncertainties persist regarding whether low-level lead exposure significantly impairs neurodevelopment in children and which specific developmental domains are most susceptible. There is a pressing need for robust scientific evidence to substantiate the “as low as reasonably achievable” (ALARA) principle and inform policies aimed at minimizing lead-associated health risks. −

Accumulating epidemiological evidence highlights the neurodevelopmental risks associated with blood lead levels (BLLs) in the range of 18–50 μg/L. − A population-based study from North Carolina, USA, integrating blood lead monitoring with standardized educational outcomes, showed that BLLs as low as 20 μg/L were associated with deficits in mathematics and reading performance in children. Similarly, a study in South Korea reported a significant correlation between BLLs averaging 18.1 μg/L and attention deficits in school-aged children. Consistent with these findings, research involving Mexican children aged 6–13 demonstrated that BLLs around 34 μg/L were strongly linked to impaired attention. A cohort study from Shanghai Jintan Hospital in China further indicated that BLLs of 31.1 μg/L exert lasting adverse effects on intellectual development. The U.S. Centers for Disease Control and Prevention (CDC) has set a blood lead reference value of 22.5 μg/L for children. According to the latest National Health and Nutrition Examination Survey (NHANES), BLLs in infants aged 0–3 years in the U.S. range from 1.7 to 46.1 μg/L (https://wwwn.cdc.gov/nchs/nhanes/continuousnhanes/default.aspx). Nevertheless, comprehensive longitudinal studies evaluating neurodevelopment across the entire infant period (0–3 years) with BLLs consistently below the CDC reference level remain scarce. Moreover, coexposure to other heavy metals, such as cadmium, manganese, and mercury, may compound neurological risks, underscoring the need to investigate the combined effects of low-level lead within metal mixtures on early neurodevelopment and to identify which metals play the most critical roles. ,

Currently, there is a lack of quantitative susceptibility biomarkers for the objective assessment of neurodevelopment in infants aged 0–3 years. DNA methylation (DNAm), an epigenetic mechanism that is both reversible and stable under physiological conditions, offers considerable promise for early molecular diagnosis and intervention, particularly for neurodevelopmental impairments linked to heavy metal exposure. − DNAm-based biomarkers are already integral to clinical practice, especially in oncology, where they facilitate early screening, diagnostic confirmation, treatment monitoring, and prognostic evaluation. For instance, FDA-approved kits detecting methylation in SEPT9, NDRG4, and BMP3 are used for early colorectal cancer screening. − Assays targeting SHOX2, PTGER4, and RASSF1A methylation, approved by China’s National Medical Products Administration, aid in lung cancer diagnosis, − and MGMT methylation status predicts temozolomide response in glioma. , Similarly, methylation of PCDHGB7 and PAX1 serves as a diagnostic marker in cervical cancer, especially among HPV-positive women. − Despite these advances, few DNAm biomarkers are available for quantifying the neurotoxic effects of heavy metals in children. Current biomarker strategies predominantly target promoter methylation in genes associated with neurological disorders (e.g., FMR1, BDNF, MECP2), − as well as genetic variants (e.g., DRD2 polymorphisms, which modify the impact of lead on IQ), and expression changes in genes such as UNC5D whose levels correlate with neurodevelopmental outcomes. Although these markers do not yet enable quantitative risk assessment, they underscore the potential of DNAm as a powerful tool for evaluating environmental health risks and advancing the diagnosis and treatment of heavy metal-related neurodevelopmental abnormalities. This highlights a critical avenue for future research into predictive and quantitative epigenetic biomarkers of metal-induced neurotoxicity.

In our previous work, we established that methylation of the FAM50B/PTCHD3 fragments mediates the association between BLLs and IQ in children, demonstrating a linear dose–response relationship between DNAm status and cognitive performance. , However, those findings were derived from a human population with relatively high lead exposure (mean BLLs >40 μg/L), leaving it unclear whether they generalize to populations with lower exposure levels. Furthermore, as that study focused on children aged 9–11 years, the applicability of these epigenetic markers to earlier developmental stages remained uncertain. To overcome these limitations, the present study leverages a prospective birth cohort with repeated assessments conducted at 6 to 12 month intervals, establishing a clear temporal sequence between prenatal exposure and neurodevelopmental outcomes. This design allows us to evaluate the effects of low-level lead exposure on infant neurodevelopment and identify associated DNAm biomarkers, thereby providing a more comprehensive understanding of lead neurotoxicity across exposure gradients and developmental windows.

In this population-based birth cohort study, we measured metal concentrations and DNAm in cord blood samples collected at delivery. We examined the relationships between prenatal exposure to metals, DNAm profiles, and neurodevelopmental outcomes assessed using the Ages and Stages Questionnaires (ASQ). The study aimed to (1) identify which metals in a low-level exposure mixture most strongly influence neurodevelopment; (2) investigate whether DNAm at FAM50B and PTCHD3 mediates the association between metal exposure and neurodevelopmental outcomes; and (3) develop a predictive model of early neurodevelopment based on DNAm markers. An overview of the study design is presented in Figure . These findings broaden the utility of FAM50B/PTCHD3 methylation as a biomarker for assessing the environmental health risks of metal exposure on neurodevelopment and may facilitate early diagnostics or targeted interventions for neurodevelopmental impairments.

1.

Overview of study workflow. This study was based on the Maoming Birth Cohort Study (MBCS) and included 353 mother–child pairs. Twenty-seven cord blood metals were detected by ICP-MS, and twenty-one metals were detectable. The infants’ neurodevelopment was evaluated by the third edition of Ages and Stages Questionnaires (ASQ-3) when these infants were 3, 6, 12, 18, 24, and 36 months old. The number of participants for each follow-up was 332, 324, 307, 291, 270, and 249, respectively. The DNA methylation of 198 candidate CpG loci was detected in 353 cord blood DNA samples. Then, the associations of single metals or single CpG loci and ASQ were analyzed by mixed-effect models. The quantile g-computation and BKMR models were used to assess the joint effect of the metal mixture on ASQ. To explore the potential mediation effect of DNA methylation on the relationships between blood metals and ASQ, mediation analysis was performed. Finally, the prediction model of ASQ was constructed using 10 machine learning algorithms.

2. Materials and Methods

2.1. Study Population

Participants in this study were drawn from the Maoming Birth Cohort (MBC), an ongoing prospective cohort based at the Maoming Maternal and Child Health Care Hospital in Guangdong Province, China. The MBC was established to investigate the effects of early life factors, including environmental pollutant exposures, on birth outcomes, postnatal growth, and neurodevelopment, which was described in detail previously. The present study was conducted within the MBC between October 2015 and December 2018. Eligible participants were pregnant women recruited during their first trimester who had resided locally for at least three years and had no documented infectious diseases. A total of 11 258 women were enrolled in the baseline cohort. Among these, 1030 mother–infant pairs were selected for follow-up based on the availability of cord blood samples and willingness to participate in subsequent assessments. Structured questionnaires were administered at enrollment to collect information on demographic characteristics, socioeconomic status, lifestyle factors, dietary habits, housing conditions, and the neighborhood environment. Written informed consent was obtained from all of the participants. As of July 2020, the follow-up rate of the cohort was 81.0%, reflecting a strong participant retention. For the present analysis, we excluded participants who were lost to follow-up, did not provide cord blood specimens, were missing neurodevelopmental test data, or had poor-quality DNA samples. A total of 353 mother–infant pairs met the inclusion criteria and were retained for final analysis. Sample size estimation was performed using PASS software, and a minimum of 125 participants was determined to be sufficient. The demographic characteristics of the final study population were comparable to those of the overall cohort, supporting the representativeness of the analytic sample (Table ). The study was conducted according to the guidelines of the Declaration of Helsinki and approved by the Human Studies Committee of Sun Yat-sen University (Approval No. SYSU-2016-018). Informed consent was obtained from all parents or legal guardians of the infant participants.

1. Characteristics of the Study Population (n = 353) .

Characteristics	Study population (n = 353)	Total population (n = 1030)	P
Mother
Age (year)	28.7 ± 2.38	28.4 ± 5.50	0.133
Prepregnancy BMI (kg/m2)	20.98 ± 3.41	20.70 ± 2.7	0.126
Maternal education			0.697
≤High school	235 (66.57%)	682 (66.20%)
>High school	118 (33.43%)	348(33.80%)
Family income (CNY/year)			0.573
<30000	82 (23.23%)	247 (24.00%)
30000–100000	224 (63.46%)	657 (63.80%)
>100000	47 (13.31%)	126 (12.20%)
Smoking during pregnancy	144 (40.80%)	470 (45.60%)	0.082
Alcohol drinking	17 (4.82%)	31 (3.00%)	0.093
Child
Sex			0.480
Boy	195 (55.24%)	580 (56.30%)
Girl	158 (44.76%)	450 (43.70%)

^aNote: The table shows the population characteristics of this study and is compared with the total population of the birth cohort. The differences for age and prepregnancy BMI of the mother were tested using an MW-U test, and the remaining variables were tested by a Chi-square test. BMI: body mass index. CNY: Chinese yuan.

2.2. Blood Metal Detection

Twenty-seven elements in cord blood were measured by inductively coupled plasma–mass spectrometry (ICP-MS), including beryllium (Be), barium (Ba), aluminum (Al), silicon (Si), vanadium (V), chromium (Cr), manganese (Mn), cobalt (Co), nickel (Ni), copper (Cu), zinc (Zn), arsenic (As), selenium (Se), strontium (Sr), yttrium (Y), cadmium (Cd), tin (Sn), antimony (Sb), cerium (Ce), thallium (Tl), lead (Pb), uranium (U), iron (Fe), sodium (Na), magnesium (Mg), potassium (K), and calcium (Ca). The detailed method is presented in Supporting Information (SI) Text 1.

2.3. Neurodevelopment Test

Follow-up assessments were conducted at 3, 6, 12, 18, 24, and 36 months, with follow-up rates for each time point detailed in Table S1. Missing data were imputed in subsequent analyses using the appropriate statistical methods. Neurodevelopmental outcomes were assessed using the third edition of the Ages and Stages Questionnaires (ASQ-3), a validated screening tool for early childhood development. The ASQ-3 comprises domain-specific questionnaires, each containing six items that describe representative behaviors. Items are scored as “yes” (10 points), “sometimes” (5 points), or “not yet” (0 points), with subscale scores ranging from 0 to 60. The ASQ-3 has been validated in Chinese populations, demonstrating robust internal consistency, high test-retest reliability, and good construct validity. , The tools yield five developmental scores: gross motor (gmscore), fine motor (fmscore), communication (cscore), problem-solving (sqscore), and personal-social (psscore). Trained healthcare professionals administered the ASQ-3 in a dedicated quiet room, under emotionally stable conditions, with each assessment lasting approximately 30 min. Infants exhibiting fever, diarrhea, or signs of infectious disease at the time of assessment were excluded. Suspected developmental delay (SDD) was defined as any domain score falling below two standard deviations (mean – 2 SD) from the normative mean at the respective time point. A child was classified as having SDD if at least one of the five domain scores or the total ASQ score (ASQ-T) fell below the established threshold.

2.4. DNA Methylation Detection of Candidate Fragments

Genomic DNA was extracted from cord blood using a Blood DNA Mini Kit (Qiagen, Hilden, Germany) following the manufacturer’s protocol. DNA methylation (DNAm) levels in selected gene regions were quantified using the MethylTarget method, which is a multiplexed, targeted methylation enrichment sequencing technique. The human reference genome hg38 was used to define the positions of the CpG sites. Detailed descriptions of the methodology are provided in our previous publications. , For quality control, a human DNAm standard (Merck Millipore, Billerica, MA, USA) was used as a positive control, while 5-azacytidine (azaC; Sigma-Aldrich, St. Louis, MO, USA)-treated SH-SY5Y cells served as a negative control, with untreated SH-SY5Y cells included for reference. The average DNAm level across all CpG sites in the positive control was 97.33%, and the relative standard deviations (RSDs) between technical replicates ranged from 0.03% to 3.89%. For the negative control, DNAm levels were significantly reduced in azaC-treated cells compared with untreated cells (P < 0.05), confirming the reliability and sensitivity of the assay.

2.5. Statistical Analysis

Participant characteristics were summarized using means and standard deviations (SDs) for continuous variables and counts with percentages (n, %) for categorical variables. All analyses were performed using R and Python. Missing data in the Ages and Stages Questionnaires (ASQ) were assumed to be missing at random (MAR, independent of its own true values and dependent on other variables such as metals) and were imputed using the random forest algorithm via the “mice” package. To minimize confounding bias, a directed acyclic graph (DAG) was constructed using the “ggdag” package to inform the selection of covariates. Based on the DAG, the final adjustment set of covariates included child sex, maternal education and occupation, annual household income, and parental smoking and alcohol use. These variables were included as covariates in all regression models. Metals concentrations were standardized prior to analysis. To correct for multiple testing, the Benjamini–Hochberg procedure was applied to control the false discovery rate (FDR).

We evaluated the associations between individual metal exposures and ASQ outcomes, as well as those between DNA methylation (DNAm) at individual CpG sites and ASQ outcomes, using mixed-effect models (MEMs) implemented via the “lme4” package. For each association, we estimated the percent change in ASQ scores per unit increase in metal concentration or DNAm level along with corresponding 95% confidence intervals (CIs).

We assessed the joint effects of the metal mixture and the relative contribution of each metal to ASQ outcomes using quantile g-computation, implemented via the “qgcomp” package. This method combines the inferential simplicity of weighted quantile sum (WQS) regression with the flexibility of g-computation, providing a robust framework for estimating health effects of complex environmental mixtures. Given the substantial differences in concentration ranges among the 21 metals, all metal concentrations were standardized prior to model fitting.

To further assess the joint effects of the metal mixture, we applied Bayesian kernel machine regression (BKMR). This method combines Bayesian inference with statistical learning, leveraging a Gaussian kernel function to flexibly model exposure–response relationships. BKMR is particularly well-suited for capturing complex, nonlinear, and nonadditive interactions among multiple coexposures. Since there were thousands of rows of our data, it took dozens of hours to run the BKMR model on a single machine for a single ASQ indicator to get a complete result. Therefore, we used the computing resources of the National Supercomputing Center in Guangzhou to run the model, which could finish the analysis in half an hour. We estimated both the joint effect of the metal mixture and the marginal effect of individual metals by fixing the concentrations of all other metals at the 25th, 50th, and 75th percentiles. This approach enabled us to examine the influence of each metal while accounting for the mixture complexity.

To assess the potential mediating role of candidate locus methylation between prenatal metal exposure and infant neurodevelopment, we performed causal mediation analysis. Metal concentrations, infant neurodevelopmental scores (assessed via ASQ), and DNAm levels at each candidate locus were designated as the exposure, outcome, and mediator, respectively, adjusting for the covariates described above. The mediation effect was expressed as the percentage of the total effect (TE) attributable to the indirect effect (IE), calculated as (IE/TE) × 100. IE represents the effect of metals on neurodevelopment mediated through locus-specific DNAm, while TE is the sum of the IE and the direct effect of metals on neurodevelopment. Analyses were implemented in Python using the “mediation” module.

We also constructed predictive models to classify infant neurodevelopmental status using DNA methylation (DNAm) data, and details are provided in SI Text 2.

To address potential confounding, sensitivity analyses included additional adjustment for delivery method, birthweight, and preterm birth (gestational age <37 weeks). We evaluated effect modification by child sex through stratified analyses and examined cross-sectional associations between metal concentrations and continuous ASQ-3 scores. Participants with missing data were excluded to assess robustness of the metal–ASQ domain relationships via sensitivity models.

3. Results and Discussion

3.1. Characteristics of the Study Population

Table presents the demographic characteristics of the study population, which consisted of 353 participants, including 195 boys (55.24%) and 158 girls (44.76%). The mothers had an average age of 28.7 years, with a mean prepregnancy body mass index (BMI) of 20.98 ± 3.41 kg/m², indicating a generally normal weight status. Regarding maternal education, 179 mothers (57.01%) had an education level below high school, while 135 (42.99%) had completed high school or higher. Annual family income varied across the cohort, with 23.23% of families earning less than 30 000 CNY, 63.46% earning between 30 000 and 100 000 CNY, and 13.31% earning 100 000 CNY or more. Additionally, 40.80% of infants had relatives who smoked, and 4.82% had relatives who consumed alcohol. The demographic characteristics of the current study population were consistent with those of a previous cohort (n = 718), indicating that the selected sample is representative of the original research population and supports the generalizability of the findings of the study.

3.2. Cord Blood Metal Concentrations and ASQ Levels of the Participants

Be, Ba, Cr, Cd, Ce, and Y had detection rates below 30% and were therefore excluded from further analysis. The detection information on the remaining 21 metals is presented in Table S2. The average BLL was 15.30 μg/L, with only 1.13% of samples exceeding 50 μg/L, indicating relatively low exposure levels compared to both domestic and international standards. ,,, Similarly, As concentrations were within the normal range of 0.40–12.00 μg/L, with an average of 1.96 μg/L, which was close to the levels observed in a Nanjing and Beijing birth cohort. , Mn levels (51.21 μg/L) were higher than those in previous birth cohort studies, including a Canadian cohort from 2008–2011 (12.64 μg/L), a Shanghai cohort from 2013–2016 (18.20 μg/L), and a Beijing cohort from 2015–2016 (24.3 μg/L). − In contrast, concentrations of V, Ni, Sn, and Sb were notably lower than those in this Beijing cohort (1.75, 14.50, 33.20, and 6.38 μg/L, respectively). Al levels were significantly lower than a Lanzhou cohort (303 vs 1654 μg/L). Se levels were significantly below the normal range of 180–400 μg/L and those in 0–3 aged infants in the NHANES (153 μg/L), indicating potential selenium deficiency. Levels of Cu, Mg, and Fe were within normal ranges but on the lower end, suggesting marginal insufficiencies. K concentrations exceeded the normal range, averaging 2.18 mg/mL (normal: 1.17–2.11 mg/mL), while Na levels were below the reference range at 1.41 mg/mL (normal: 1.47–2.53 mg/mL). Ca (48.1 μg/mL) and Zn (1.56 mg/mL) concentrations were lower than the reference values (60.5–100.2 μg/mL and 4.43–9.81 mg/mL, respectively), indicating a possible Ca and Zn deficiency. These findings suggest that metal concentrations in this population were generally low, providing a relevant context for assessing the impact of low-level metal exposure on neurodevelopmental outcomes. The International Agency for Research on Cancer (IARC) has classified these metals as Group 1 carcinogens (determined as carcinogenic), including Al, Ni, As, U, and Sr-90; Group 2A carcinogens (probably carcinogenic), including inorganic lead compounds, Si carbide, Co, and trivalent Sb; and Group 2B carcinogens (possibly carcinogenic), including monomeric lead, V pentoxide, and monomeric Ni. These findings emphasize the complex exposure landscape within this cohort, characterized by varying levels of both essential and toxic metals, some of which may pose significant health risks, as classified by IARC.

Figure S1 presents the pairwise correlation coefficients (Spearman’s rank correlation) for metals. The majority of metals exhibited positive correlations. The strongest correlation was observed between Fe and K (r = 0.88). Other notable correlations included Mg with K (r = 0.74), Mg with Fe (r = 0.66), and Na with Mg (r = 0.62). These significant correlations underscore that these metals maybe derived from the same source. Table S3 shows that the proportions of abnormal ASQ results were 17.33%, 11.10%, 14.31%, 13.70%, and 18.65% across different developmental indicators, significantly higher than the 4%, 5%, 6%, 4%, and 4% reported in a Japanese birth cohort study (2011–2014) for the same indicators. The observed disparity may be linked to deficiencies in essential metals such as Zn and Ca, and higher nonessential metals such as lead among children in this region, suggesting that the neurodevelopmental health risks associated with low-level heavy metal exposure remain significant. These findings highlight the importance of considering multiple metal exposures when assessing neurodevelopmental risks in children, even at low exposure levels.

3.3. Associations between Blood Metals and ASQ

The MEM results revealed significant associations for four metalsPb, Sn, Si, and Uon infant neurodevelopment. Specifically, lead showed a consistent negative association across all five developmental indicators, while Sn was negatively associated with communication, gross motor, and fine motor skills. Si demonstrated a negative correlation with communication and problem-solving, and U was associated with lower scores in communication, gross motor skills, and personal-social behavior (Figure ).

2.

Associations between 21 metals and 5 ASQ outcomes based on mixed-effect models (n = 353). Models were adjusted for child sex, maternal education and occupation, annual household income, and parental smoking and alcohol use. ASQ results include five parts: cscore, fmscore, gmscore, psscore, and sqscore. cscore: communication score; gmscore: gross motor score; fmscore: fine motor score; sqscore: problem-solving score; and psscore: personal-social skill score. The y-axis represents 21 metals. The x-axis represents the percentage changes of communication, gross motor, fine motor, problem-solving, and personal-social skill scores of the infants under a standard deviation change of metal concentration. Red lines represent the significant results.

To evaluate the combined exposure effects of multiple metals on ASQ outcomes, we applied quantile g-computation. This approach included all 21 metals in the quantile g-computation models, with adjustments for the aforementioned confounders. We standardized the data to eliminate the influence of scale differences on the assigned weights. Figure S2 illustrates the weights of contribution of each metal to ASQ outcomes. The results revealed that Sn, Pb, and As had the largest weights for communication scores (−1.20, −0.61, and 0.29, respectively). For both gross and fine motor skills, Sn and Pb emerged as primary contributors (Sn: −0.48 for gross motor, −0.46 for fine motor; Pb: −0.31 for gross motor, −0.5 for fine motor). For problem-solving and personal-social behavior, Pb was the most influential factor (−0.91 for problem-solving and −0.39 for personal-social). These findings led us to focus subsequent analyses on the neurodevelopmental effects of blood Pb and Sn levels and on identifying DNAm biomarkers associated with these exposures.

3.4. Joint Effects of Blood Metals on ASQ

Based on the results of quantile g-computation, we selected As, Pb, Sn, and Sb (the contribution rates of the remaining 17 metals were <0.1) to analyze the interaction by the BKMR model. The results showed that higher concentrations of metal mixtures were associated with lower ASQ indicators, except for communication and fine motor (Figure A and Figure S3). When all other metals were fixed at a specific quantile (25th, 50th, or 75th percentile), Pb was negatively associated with three ASQ indicators (problem-solving > fine motor > gross motor sorting by effect size). As was negatively associated with four ASQ indicators (personal-social > problem-solving > fine motor > gross motor sorting by effect size). Sn was positively associated with fine motor. Sb was negatively associated with personal-social and problem-solving, whereas it was positively associated with gross motor (Figures B and S3). The dose–response curves suggested the negative associations between lead and ASQ indicators. The effects of lead on infants’ problem-solving, gross motor, fine motor, and personal-social abilities are linear or approximately linear, while the effect on infants’ communication ability is interval linear (linear at low and high concentrations, but fluctuating at certain concentrations). The dose–response curves between As and ASQ indicators showed both a U-shape first and then an inverted U-shape, indicating the similarity of the association modes for As on different neurodevelopmental indicators. Sb had an impact on infants’ problem-solving and personal-social abilities, but had almost no effect on communication, gross motor, and fine motor (the dose–response curve is almost horizontal). The dose–response curves of Sn and the five ASQ indicators were almost horizontal (Figures C and S3). These results suggest that different heavy metals have different modes of action on infant neurodevelopment and Pb plays the most important role.

3.

Relationship of exposure to the metal mixture (A) or a single metal (B) and the estimated differences in the problem-solving score (sqscore) as well as the dose–response (C) of metals on the problem-solving score (sqscore) (n = 353). Models were adjusted for child sex, maternal education and occupation, annual household income, and parental smoking and alcohol use. In (A), the x-axis shows the metal mixture quantile (ranging from 25th to 75th percentiles), and the y-axis shows the sqscore. Panel (B) shows the effect of a single metal on the sqscore, with all other metals fixed at a specific quantile (25th, 50th, or 75th percentile). Panel (C) shows the dose–response of a single metal on the sqscore; the x-axis shows the doses of each metal, and the y-axis shows the sqscore. The concentrations of the metals are the standardized concentrations in the figure.

3.5. Associations between Blood Metals and ASQ with DNA Methylation of Candidate Loci

Figure S4 presents the Spearman correlations between metal concentrations and DNAm at candidate loci. Among the 12 gene regions analyzed, 198 loci were found to correlate with various metals. Specifically, Pb, Se, and Sr were negatively correlated with the average DNAm (Ave-DNAm) of FAM50B1, while Pb was negatively correlated with the Ave-DNAm of PTCHD3, in contrast to Na, which showed a positive correlation. LCK exhibited negative correlations with As, Se, Sr, and Zn, while LOC101927932 was negatively correlated with Mn, U, and Zn and positively correlated with Sb. Furthermore, LOC283177 showed negative correlations with V and Zn, while PARD6G1 had a positive correlation with Cu. Zn was negatively correlated with the Ave-DNAm of GNAS. Among these significant associations, U had the highest number of correlated sites (123), while K had the fewest (56). Overall, the proportions of positive and negative correlations are roughly equal. Pb was significantly associated with 79 of the 198 sites, with 11 out of the 12 gene regions, excluding LCK, showing significant Pb-associated loci. Using a mixed-effects model (MEM), we examined associations between DNAm at 23 candidate loci and the five ASQ outcomes, controlling for previously mentioned covariates. With the exception of PTCHD3_5 (chr10: 27702924, P = 0.13), all remaining 22 sites were positively associated with at least one ASQ outcome (Figure S5). These findings align closely with our previous studies, which identified positive associations between FAM50B and PTCHD3 and children’s IQ. ,

In this study, elevated BLLs were associated with decreased levels of DNAm in FAM50B1, FAM50B2, CCDC144B2, PTCHD3, LOC101927932, LOC283177, and PARD6G1, while LCK showed an increase, aligning with the findings of our previous research. However, the DNAm patterns of GNAS, HTRA4, LRRC27, and KCNK3 deviated from those observed in earlier studies. Notably, 37 CpG sites were significantly correlated to lead exposure, displaying trends consistent with prior findings. Among these, FAM50B1 (80%) and PTCHD3 (100%) showed the strongest associations with lead, further supporting our earlier work. Detailed information about these sites and their lead correlation is presented in Table S4. To refine the analysis, we adjusted for relevant covariates and applied multiple linear regression (MLR) to examine the relationship between Pb exposure and the 37 CpG sites. This led to the identification of 23 significant sites: eight in FAM50B, 10 in PTCHD3, two in CCDC144B, and three in LOC283177 showed strong negative associations with Pb exposure (Table S5). This refinement strengthens our understanding of the influence of Pb on specific DNAm sites, underscoring key epigenetic markers for future research in Pb toxicity.

Both FAM50B and PTCHD3 were recognized or predicted to be imprinted genes and rarely studied. In our previous studies, we performed function enrichment analysis and found that these two genes played an important role in neuronal function and synaptic conduction. In SH-SY5Y cells, we explored the function of these two genes and found their DNA methylation affected lead-induced neurotoxicity via the PI3K-Akt pathway. Through database search, we also found that FAM50B is expressed higher in the brain (https://tissues.jensenlab.org/Search). These evidence suggested that FAM50B/PTCHD3 may be associated with brain and neurodevelopment of children.

3.6. DNA Methylation Mediates the Relationship of Lead and ASQ

Our preliminary research revealed that 10 segments exhibited hypomethylation under lead exposure, including FAM50B, PTCHD3, CCDC144B, and LOC283177. Notably, methylation levels of these segments were positively correlated with children’s IQ. Given this, the mediation effect opposes the total effect. Using Python’s “mediation” module, we screened four CpG sites that mediate the relationship between BLLs and infants’ neurodevelopment. The mediation pathways and contribution rates are depicted in Figure . Specifically, FAM50B_1_9 (chr6: 3849272) significantly mediated the relationships between fine motor, personal-social behaviors, and cord BLLs, each with a mediation contribution of 2.99% (0.79%–4.62%). PTCHD3_1 (chr10: 27702951) showed significant mediation effects for gross motor, personal-social behaviors, fine motor, and problem-solving, with contribution weights of 6.43% (0.99%–16.49%), 4.46% (0.08%–6.19%), 6.43% (1.30%–9.08%), and 6.43% (2.80%–10.31%), respectively. Additionally, PTCHD3_9 (chr10: 27702874) mediated associations between gross motor and problem-solving with contributions of 5.70% (0.26%–13.43%; 1.74%–12.88%) each, while PTCHD3_7 (chr10:27702904) exclusively mediated problem-solving with a 5.68% (2.16%–9.52%) contribution weight (Table S6).

4.

The mediation effects of 4 loci for cord blood lead and ASQ outcomes of fine motor (A), problem-solving (B), gross motor (C), and personal-social (D) (n = 353). Models were adjusted for child sex, maternal education and occupation, annual household income, and parental smoking and alcohol use. In (B), the proportion of mediation for PTCHD3 is the sum of PTCHD3_1 (6.43%), PTCHD3_7 (5.68%), and PTCHD3_9 (5.70%). In (C), the proportion of mediation for PTCHD3 is the sum of PTCHD3_1 (6.43%) and PTCHD3_9 (5.70%). FAM50B_1_9: chr6, 3849272; PTCHD3_1: chr10, 27702951; PTCHD3_7: chr10, 27702904; and PTCHD3_9: chr10, 27702874.

Compared to our previous study, the contribution weights in this analysis were relatively lower. Several factors likely account for this difference. First, the prior study focused on IQ as a health outcome, whereas this study focused on neurodevelopment. Moreover, children in this study were aged 0–3, contrasting with the 9–11 aged children in the previous study. Additionally, the BLLs in this study were substantially lower. Last but not least, there may be differences in the efficacy of epigenetic mechanisms under different exposure gradients, which is a direction for our future research. Nevertheless, our findings reinforce the negative impact of BLLs on the neurodevelopment of children, suggesting that specific CpG sites may play a pivotal mediating role in this process. Further investigation of these key sites could deepen our understanding of the mechanisms underlying low-level lead and its effects on early childhood neurodevelopment, thereby providing a scientific foundation for future interventions and prevention strategies.

3.7. Prediction Model of ASQ by FAM50B/PTCHD3 Methylation

We used all available data across follow-up periods to develop predictive models for infant neurodevelopment as assessed by ASQ scores. Accordingly, the models were not designed to predict ASQ outcomes at a specific time point but rather aimed to classify neurodevelopmental status over the entire 0–3 year age range. Ten classification algorithms were tested: random forest (RF), Bayesian network (BayesNet), C4.5 decision tree (DTr), support vector machine (SVM), k-nearest neighbors (kNN), neural network (NN), logistic regression (LGR), extremely randomized trees (Extra Trees), adaptive boosting (AdaBoost), and extreme gradient boosting (XGBoost). The data set was randomly split into a training set (n = 280, 80%) and a test set (n = 73, 20%). 10-fold cross-validation was conducted on the training set to evaluate internal performance, and external validation was performed on the test set. Detailed parameter settings for each model are provided in Supporting Information Text 3. Model performance was evaluated using five metrics: accuracy (ACC), area under the ROC curve (AUC), sensitivity, specificity, and F1 score. The results of the 10-fold cross-validation and external validation are presented in Figure S6 and Table S7. Among all models, the random forest algorithm achieved the highest performance, demonstrating strong internal robustness (ACC = 77%; AUC = 87%) and high external predictive accuracy (ACC = 72%; AUC = 75%). Models based on C4.5 decision tree, kNN, Extra Trees, and XGBoost also exhibited satisfactory performance, with ACC values ranging from 73% to 77% in cross-validation and 60% to 65% in external validation. Their AUCs ranged from 79% to 85% in cross-validation and 63% to 69% in external validation. In contrast, models based on logistic regression, SVM, NN, AdaBoost, and BayesNet showed poor performance, particularly in cross-validation, with specificity values below 40%. Based on these findings, the random forest model was selected as the optimal classification algorithm for ASQ prediction. In addition, we also performed covariate only models and calculated the additional gain of DNAm performance. The results showed that models with DNAm data were better than models without DNAm data (ACC and AUC values of RF model increased by 42–47% and 34–39%, respectively) (Table S8), indicating that the addition of DNAm has greatly improved the prediction performance. These results suggest that methylation at the FAM50B and PTCHD3 loci may serve as reliable biomarkers of early neurodevelopment in children exposed to low-level lead concentrations (∼15 μg/L).

In our previous study, we developed a linear model to predict children’s IQ by single CpG sites, which may introduce bias due to gene interactions. In contrast, the current model integrates four CpG sites to predict infants’ neurodevelopment in the context of low-level lead exposure. Several key differences distinguish the present model from the previous one. First, the health outcomes assessed differ: this study evaluates neurodevelopment via ASQ-3, while the previous study measured IQ using Raven’s test. ASQ-3 is more sensitive in reflecting early neurodevelopment, making it better suited for infants. Second, the level of lead exposure varies: the current model focuses on low lead levels (∼15 μg/L), while the previous model examined high exposure (>50 μg/L). Given that the average blood lead concentration in nonindustrial areas of China is around 20 μg/L, − this research on low-level lead exposure holds practical value. It can detect biomarkers sensitive to neurodevelopmental changes at low exposure levels, enabling early screening and risk assessment in vulnerable populations. This approach supports public health policy development to enhance lead exposure prevention and intervention. Third, we used a nonparametric machine learning model in this study, in contrast with the linear model used previously. Additionally, the age groups differ: this study focuses on children aged 0–3, while the previous one focused on children aged 9–11, broadening the biomarker’s applicability. Finally, the current model combines multiple loci, whereas the previous one averaged DNAm across fragments of FAM50B or PTCHD3. These differences may explain variations in the prediction outcomes. Nevertheless, FAM50B/PTCHD3 methylation remains a reliable indicator for predicting both IQ and neurodevelopment across varying levels of lead exposure. By identifying and utilizing these biomarkers, we can better assess the impact of lead exposure on the neurodevelopment of children, leading to more effective protective measures. That is, they maybe used to identify subpopulations at “safe” levels in routine lead screening but actually at high neurodevelopmental risk, thus achieving an earlier precise intervention.

3.8. Sensitivity Analysis

To evaluate the stability and reliability of our analysis, we conducted sensitivity analyses. After adjustment for additional variables such as delivery method, birthweight, and preterm birth, the overall results remained consistent with the main analysis, despite some observed changes. Excluding participants with missing ASQ data also yielded findings that aligned with those of the primary models. No evidence of sex-specific effects was observed in sex-stratified analyses (Table S9). Furthermore, cross-sectional associations between lead exposure and ASQ-3 scores were analyzed, and the main results remained consistent (Table S10). Through these sensitivity analyses, we confirmed the stability and reliability of the primary analysis results.

3.9. Strengths and Shortcomings

This study has several notable strengths. First, we utilized ASQ-3, a validated and standardized tool, to ensure reliable neuropsychological evaluations. Second, the face-to-face collection of numerous covariates enabled robust and comprehensive sensitivity analyses. Third, the dual validation of the identified biomarkers in both low and high blood lead concentration populations underscores their significance in assessing neurodevelopmental risks associated with lead exposure. This distinctive advantage demonstrates the robustness and generalizability of the biomarkers across varying exposure levels. Such findings not only enhance our understanding of the critical role of the biomarkers in neurodevelopmental assessments but also provide a solid foundation for future research and potential interventions aimed at mitigating the adverse effects of heavy metal exposure on children’s developmental outcomes. Fourth, mediation analysis allowed us to identify key CpG sites mediating the relationship between metal exposure and neuropsychological development, supporting the construction of a predictive model for children’s IQ based on DNAm. Lastly, we examined the joint effects of the metal mixture using quantile g-computation and the BKMR model using the National Supercomputing Center in Guangzhou, and we constructed a predictive model of ASQ using DNAm data.

However, several limitations of this study should be acknowledged. First, the half-life of blood lead is short (about 1–2 months); hence cord blood lead can reflect in utero exposure but not the children’s lifetime exposure. Second, despite controlling for multiple confounders, residual confounding from unmeasured factors, such as other environmental pollutants (e.g., particulate matter, PFAS, PAHs, bisphenol A), remains possible. − Third, despite careful covariate adjustment, our analysis did not necessarily represent a causal effect. Fourth, the sample size of this study was generally small, especially after subsequent machine learning models divided the training set and test set, which may reduce statistical power. Finally, this study was conducted in a single hospital and relied solely on ASQ assessments, which may restrict the generalizability of the findings.

4. Conclusions

In summary, in utero exposure to low levels of blood lead (∼15 μg/L) negatively affected the neurodevelopment of infants aged 0–3 years, with FAM50B/PTCHD3 methylation mediating the relationship between blood lead levels and neurodevelopment. Currently, no country has established a blood lead reference value below 15 μg/L, and there is no direct evidence supporting the safety of infants under this exposure level. To our knowledge, this is the first epidemiological study to explore the association between blood lead levels (BLLs) < 20 μg/L and neurodevelopment in infants aged 0–3 years while also identifying predictive DNA methylation biomarkers. Our findings have significant environmental implications for protecting children’s health and promoting global sustainable development. Additionally, blood lead showed an approximately linear relationship with ASQ indicators and exhibited both synergistic and antagonistic effects with arsenic (more than additive at low concentrations and less than additive at high concentrations). In this study, while 27 blood metals were measured, only blood lead was associated with all five neurodevelopmental outcomes, and this association was mediated by FAM50B/PTCHD3 methylation, which aligns with our previous findings. , Furthermore, the DNA methylation of four FAM50B/PTCHD3 loci demonstrated a strong predictive power for neurodevelopment. Our study underscores the need for ongoing attention to lead even at low exposure levels. The identified FAM50B/PTCHD3 biomarkers could aid in developing early diagnostic tools and targeted treatments for neurodevelopmental disorders as well as in quantitative environmental health risk assessments for heavy metals. These findings provide a scientific basis for public health policies and offer guidance to parents and society on preventing and mitigating lead exposure in infants.

Data will be made available on request.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/envhealth.5c00459.

• Results of the correlations between metals as well as between metals and DNAm of detected loci, the basic information on metals and ASQ, parameters and performance of machine learning models, and the basic information on the methylation sites associated with blood lead (PDF)

The authors declare no competing financial interest.