Participant recruitment details are previously published.^{2–5, 11, 15–19} Briefly, all ALS patients meeting Gold Coast criteria seen in the University of Michigan Pranger ALS clinic are approached for enrollment and depositing of blood and urine samples into the University of Michigan ALS Biorepository, provided they are able to consent in English. Control participants are identified via population outreach using a University of Michigan research database, random address mailing, and Facebook advertisements. These population outreach efforts target individuals falling with the same sex and age distribution of the ALS participants. All participants provided informed consent for this University of Michigan IRB approved study (HUM28826) and demographic information, including age, race, education, military service history, medical history, and supplements/vitamins usage. Blood and/or urine samples collected at the time of recruitment were stored in the Michigan ALS Biorepository. Plasma samples were collected between February 26, 2012 - December 7, 2022 and urine samples were collected between September 23, 2015 - September 9, 2022. Additional information abstracted from medical records included ALS disease characteristics, including onset segment, age of symptom onset, age of diagnosis, ALS family history, El Escorial Criteria, and revised ALS functional rating scale (ALSFRS-R) score. The El Escorial Criteria and ALSFRS-R obtained upon initial presentation to the Pranger ALS clinic are utilized for analysis. C9orf72 expansion status was obtained from medical records and previously published data.²

Detailed methods are reported in online Supplemental File 1. Briefly, plasma and urine samples were analyzed by inductively coupled mass spectrometry (ICP-MS) at the Dartmouth Trace Element Analysis Core. The Dartmouth Trace Element Analysis Core did not provide normal reference ranges.

Two polygenic risk scores (PGS) were utilized. The first, an ALS-PGS, was constructed as previously reported² but updated with weights from a recent large-scale ALS genome-wide association study (GWAS, Supplemental Table 1).²⁰ PRSice-2 software was used for clumping and pruning SNPs using default settings, followed by weight application from the parent GWAS. No cutoff for SNPs (p-value threshold of 1.0) was used. Notably, the ALS-PGS was exclusively constructed for participants not included in the large ALS GWAS. A metal-PGS was constructed to represent metal metabolism (metal-PGS), utilizing relevant SNPs from literature related with higher body burden of metals, or poorer outcomes from xenobiotic exposures (Supplemental Table 2). As the selected SNPs originated from various studies with diverse outcomes, weights were not applied based on effect estimates; instead, the metal-PGS was constructed based on the count of risk alleles.

Online Supplemental File 1 shows genotyping analysis of sample batches. QC filtering strategies at the SNP level are detailed in Supplemental Figure 1, with sample-level filtering strategies illustrated in Supplemental Figure 2.

Measurements of metals in plasma and urine are subject to missingness due to detection limits. Five metals detected below the limit of detection (LOD) in over 40% of samples were excluded from ERS construction. Certain metals were manually excluded from ERS construction because the biofluid source was unsuitable for accurate and representative measurement.¹⁴ For plasma ERS, arsenic,¹⁴ cadmium,¹⁴ and silver were excluded.²¹ For urine ERS, manganese,¹⁴ silver,²¹ and strontium were excluded.²² Metal concentrations below LOD were imputed by LOD/2.

Certain adjustment covariates exhibited missing data. In the risk model, these covariates included military service (n=50 for plasma, and n=37 for urine). In the survival analyses, these covariates included family history of ALS (n=12 for plasma, and n=8 for urine), onset segment (n=2 for both plasma and urine), El Escorial criteria (n=17 for plasma, and n=20 for urine), ALSFRS-R (n=2 for plasma, and n=1 for urine), and time between symptom onset and diagnosis (n=5 for plasma, and n=4 for urine). Multiple imputation with chained equations was used for missing adjustment covariates and stratified by case-control status for each sample source. The control imputation model included metals (log-transformed), age at sample collection, sex, and military service. The case imputation model additionally included age at diagnosis, onset segment, ALS family history, El Escorial criteria, ALSFRS-R, smoking status, education, and cumulative hazard rate (Nelson-Aalen estimator) from diagnosis to last follow-up. Twenty multiple imputation datasets were generated using a predictive mean matching model.

For ALS risk models, single metal logistic regression models adjusted for age at sample collection, sex, and military service were constructed. Models were fit independently across 20 imputation datasets and results across models were pooled using Rubin’s rule. P-values were adjusted for multiple comparisons by Benjamini-Hochberg correction. To assess the potential confounding effects of education and smoking, sensitivity analyses were conducted by including education or smoking as additional covariates, respectively.

Environmental risk scores (ERS) were constructed to assess multiple metals concurrently. As metal levels differ between plasma and urine samples, separate ERS scores for plasma (ERS ^PR; plasma risk ERS) and urine (ERS^UR; urine risk ERS) were developed. An adaptive elastic net approach applied to the combined 20 imputed datasets using the R-package miselect²³ determined ERS weights. To optimize performance of the adaptive elastic net, hyperparameters were carefully tuned through a 5-fold cross-validation process to minimize deviance, based on the log likelihood for a logistic regression model.

To test for effect modification between genetic predisposition to ALS and elevated metal levels, logistic regression models evaluated the interaction between plasma/urine levels and ALS-PGS. To test for effect modification between altered metals metabolism genetics and elevated metal levels, logistic regression models evaluated the interaction between plasma/urine levels and metal-PGS. Adjustment covariates for both interaction models included age at sample collection, sex, military service, and the first five genetic principal components.² Of note, the risk model did not adjust for C9orf72 gene positivity status as there were no C9orf72 positive control cases.²

Cox proportional hazards models adjusted for age at diagnosis, sex, family history of ALS, bulbar vs non-bulbar onset, diagnostic El Escorial criteria, ALSFRS-R, and time between symptom onset and diagnosis were fit for each metal model post-diagnosis survival with subsequent Benjamini-Hochberg correction. Sensitivity analyses, in which smoking was included as an additional covariate, were conducted to assess the potential confounding effect of smoking on the results. To ensure that the definition of survival time did not impact the results, a sensitivity analysis was conducted by defining survival time from the date of sample collection to the event of interest.

Survival ERS was constructed via application of a Cox regression with ridge penalty to the combined twenty imputation datasets, following published methods.^{3, 4} Again, separate ERS were developed for plasma (ERS^PS; plasma survival ERS) and urine (ERS^US; urine survival ERS). Hyperparameters of the ridge penalty were tuned through a 5-fold cross-validation process with the aim of minimizing the deviance based on the log partial likelihood for a Cox model. Kaplan-Meier survival curves depicted the survival functions of quartiles based on ERS. Survival curves were adjusted for covariates via the inverse probability weights method. Propensity scores were computed using multinomial logistic regression models, with ERS quartiles as outcome variables and adjustment covariates (as stated above) as predictors.

Cox survival models assessed interactions between genetic factors (ALS-/metal-PGS) and plasma/urine metal levels. Models were adjusted for age at diagnosis, sex, family history of ALS, bulbar vs. non-bulbar onset, initial El-Escorial criteria, ALSFRS-R, time between symptom onset and diagnosis, the first five genetic principal components, and C9orf72 gene positivity.

Occupational¹¹ and non-occupational²⁴ data were generated as previously described. Briefly, participants provided a list of prior job titles and responsibilities that were sorted into standard occupational classification (SOC) codes (Supplemental Table 3) and answered questions regarding known metal exposures in their most recent occupation and non-occupational settings. SOC codes were separated into high- and low-metal exposure groups based on previously published results.¹¹ Distributions of ERS^PR and ERS^UR by high-/low-metal exposure groups were compared using Wilcoxon tests. Associations between occupational/non-occupational metal exposures and urine/plasma metals were assessed using the partial least squares method feature in the R-package “mixOmics”²⁵ and estimated associations were visualized in heatmap plots.

An analysis was conducted to assess the associations between the metal-PGS and log concentrations of metals in plasma and urine. The associations were examined using linear regression models adjusted for sex and two genetic ancestry principal components.

The full population cohort included 454 ALS and 295 control participants, with plasma samples from 387 ALS and 281 control participants (Table 1, Figure 1), and urine samples from 322 ALS and 194 control participants (Table 2, Figure 1). In both the plasma (Table 1) and urine (Table 2) cohorts, ALS individuals were older, more frequently male, and had lower educational attainment than control participants. ALS characteristics were consistent between plasma and urine cohorts with 93% and 91% of participants presenting with classical ALS, 31% and 27% meeting El Escorial Definite ALS criteria, and 26% and 27% having bulbar onset, respectively (Tables 1 and 2). Both plasma and urine samples were available for 255 ALS and 181 control participants (Supplemental Table 4).

Among the 42 familial ALS cases in our plasma sample cohort, 9 patients had a C9orf72 mutation, 7 patients had other mutations, 15 patients had no mutations identified, and 11 patients had no available test results (Supplemental Table 5A). In the 30 familial ALS cases in our urine sample cohort, 7 patients had a C9orf72 mutation, 5 patients had other mutations 12 patients had no mutations identified, and 6 patients had no available test results (Supplemental Table 5B).

Concentrations of the metals detected in plasma and urine samples are listed in Supplemental Tables 6 and 7, respectively. Metals found below the LOD in over 40% of plasma samples included aluminum, beryllium, mercury, molybdenum, and uranium, while in urine samples, these were beryllium, chromium, manganese, mercury, and silver. Among plasma metals, levels of cadmium, copper, lead, selenium, vanadium, and zinc were significantly (p_adjusted<0.05) higher in ALS versus control participants, while antimony, arsenic, barium, and strontium levels were lower (Supplemental Table 6). In urine, concentrations of aluminum, barium, cadmium, copper, iron, molybdenum, selenium, strontium, tin, uranium, vanadium, and zinc were significantly higher in ALS versus control participants (Supplemental Table 7). Correlations among metals were more pronounced in urine samples than plasma, and few metals showed significant interclass correlation between plasma and urine measures (Supplemental Figure 3).

The associations between metal concentrations and metal-PGS are summarized in Supplemental Table 10. Notably, an association was observed with beryllium in plasma, as well as inverse associations between aluminum and mercury in urine, with metal-PGS scores (p<0.05).

Recognizing the potential impact of supplement usage on metal levels, we compared metal levels characterized by participants’ supplement/vitamin usage. In plasma samples, significant differences in median metal concentrations between individuals using supplements/vitamins and those abstaining from them were only detected for selenium and vanadium (Supplemental Table 11). No significant differences were observed for urine measures (Supplemental Table 12). Overall, supplement/vitamin usage does not appear to substantially impact outcomes.

To address potential residual confounding related to education, a sensitivity analysis was conducted by including education as an additional covariate in the logistic regression models. The results were consistent with our original findings in both plasma samples (Supplemental Table 13) and urine samples (Supplemental Table 14). This suggests that the observed associations are robust and not significantly influenced by differences in education levels.

Additionally, to address the potential impact of smoking behavior on our findings, we included smoking status (Former Smoker, Smoker, Non-Smoker) as an additional covariate in both logistic regression and survival analyses. The findings continued to align closely with the original analyses for both plasma and urine samples, as detailed in Supplemental Table 15 (risk) and Supplemental Table 16 (survival) for plasma, as well as Supplemental Table 17 (risk) and Supplemental Table 18 (survival) for urine. Notably, the sensitivity analyses demonstrated significant associations between elevated levels of aluminum and antimony in urine with increased ALS risk, and higher selenium in urine with shorter ALS survival, which were not observed in the original analyses (p_adjusted< 0.05). Despite these changes, the ORs and HRs for these metals remained consistent between the original and sensitivity analyses, indicating that the overall strength and direction of the associations did not significantly change.

A sensitivity analysis using survival time defined as the time from the date of sample collection was conducted to address potential bias from the time gap between ALS diagnosis and sample collection. The results were consistent with our original findings in both plasma (Supplemental Table 19) and urine (Supplemental Table 20) samples, reinforcing the robustness of our initial analyses.