An omics approach to study trace metals in sera of hemodialysis patients treated with erythropoiesis stimulating agents

Abstract

Hemodialysis (HD) represents a life-sustaining treatment for patients with end-stage renal disease. However, it is associated with several complications, including anemia. Erythropoiesis-stimulating agents (ESAs) are often administered to HD patients with renal anemia, but a relevant proportion of them fail to respond to the therapy. Since trace metals are involved in several biological processes and their blood levels can be altered by HD, we study the possible association between serum trace metal concentrations and ratios with the administration and response to ESA. For this study, data and sample information of 110 HD patients were downloaded from the UC San Diego Metabolomics Workbench public repository (PR000565). The blood serum levels (and ratios) of antimony, cadmium, copper, manganese, molybdenum, nickel, selenium, tin, and zinc were studied applying an omics statistical approach. The Random Forest model was able to discriminate between HD-dependent patients treated and not treated with ESAs, with an accuracy of 71.7% (95% CI 71.5–71.9%). Logistic regression analysis identifies alterations of Mn, Mo, Cd, Sn, and several of their ratios as characteristic of patients treated with ESAs. Moreover, patients with scarce response to ESAs were shown to be characterized by reduced Mn to Ni and Mn to Sb ratios. In conclusion, our results show that trace metals, in particular manganese, play a role in the mechanisms underlying the human response to ESAs, and if further confirmed, the re-equilibration of their physiological levels could contribute to a better management of HD patients, hopefully reducing their morbidity and mortality.

Graphical Abstract

Hemodialysis patients (HD) are at high risk of developing renal anemia. The study of serum trace metals in patients treated with erythropoiesis-stimulating agents could contribute to a better management of HD patients.

Introduction

Hemodialysis (HD) is a life-sustaining therapy for those patients who have severe renal disease with impaired or no kidney function. HD is a blood purification method able to remove metabolic waste products (‘uremic toxins’), restoring the normal intracellular and extracellular fluid environment.¹

However, chronic kidney disease (CKD), especially in its later stages, is associated with high mortality and morbidity, in particular, the complications observed with higher incidence are cardiovascular diseases, diabetes, anemia, and metabolic bone diseases.^2,3 Anemia occurs in >90% of the end-stage renal disease (ESRD) patients who are on HD treatment.^4–6 Erythropoiesis stimulating agents (ESAs) with adjuvant iron therapy and/or red blood cell transfusions represent the treatment of excellence for renal anemia, but a significant number of patients fail to respond to the therapy.^7–9 Moreover, several randomized trials have reported an increased risk of mortality and cardiovascular events in HD patients treated with ESAs with Hb targeted to normal ranges.^10,11

The principal mechanisms underlying renal anemia are inadequate erythropoietin (EPO) synthesis and EPO resistance. However, iron deficiency, chronic inflammation, high levels of uremic toxins, altered iron homeostasis, shortened red blood cell life span, vitamin deficiencies, and HD can also contribute to anemia onset.¹¹

Essential trace elements play a pivotal role in human health and disease, since they are involved in several processes: They function as cofactors in several enzyme reactions influencing directly the human metabolism; bind and diffuse oxygen, play a structural role for many metalloproteins; and as ions are present in body fluids affecting electrolyte balance, membrane potential, and charge.¹² Some studies suggest the importance of trace elements in CKD-associated erythropoiesis and anemia.^13–16 Indeed, both an altered dietary intake and the HD procedure itself put HD patients at risk for both accumulation and deficiency of trace elements.^17,18 The characterization of the relationship between trace elements and anemia could contribute to the elucidation of the biochemical mechanisms underlying the higher burden of morbidity and mortality observed in HD patients.¹⁹

In the present article, we decided to apply a chemometric omic approach to re-analyze the openly available data published by Brier et al.²⁰ with the aim of testing the hypothesis that serum trace metal concentrations and ratios could correlate with the response to ESAs treatment in a population of patients with ESRD. In the original publication, the authors find differences between prevalent and incident dialysis patients and identify associations between metals and hemoglobin (Hb), ESA dose and Hb-dose averaged response of patients.²⁰

Materials and Methods

Dataset description

In the present article, the metallomic dataset collected by Brier et al.²⁰ has been re-analyzed. Data and sample information were downloaded from the UC San Diego Metabolomics Workbench public repository²¹ (https://www.metabolomicsworkbench.org/) with the following Project ID number PR000565. For the reader's convenience, the most relevant pieces of information on the study population are reported below; however, for further details on sample collection procedures and trace metal analysis, we refer the readers to the original publication.²⁰

The study population includes 110 HD-dependent patients (69 males and 41 females); among them, 72 patients (40 males and 32 females) received EPO for anemia management (median months on dialysis 31.5, median Hb 10.7 g/dl). Whereas the remaining 38 patients (29 males, 9 females) were not treated with any EPO stimulating agents (median months on dialysis 1; median Hb 11.2 g/dl). EPO response index (ER) was normalized to achieved Hb >6 months as ESA dose/1000 × Hb and categorized into two groups according to the median value: ER < 2.6 (37 patients) and ER ≥ 2.6 (40 patients).

Data pre-processing

The levels of 14 metal ions [antimony (Sb), arsenic (As), cadmium (Cd), chromium (Cr), cobalt (Co), copper (Cu), lead (Pb), manganese (Mn), molybdenum (Mo), nickel (Ni), selenium (Se), tin (Sn), vanadium (V), and zinc (Zn)] were measured via high resolution inductively coupled plasma mass spectrometry (ICP-MS).²⁰ Only covariates with <25% missing data were considered, thus five elements (V, Cr, Co, Pb, and As) were removed from the present analysis. Missing data were imputed by means of a Random Forest (RF) approach as implemented in the R package ‘missForest’²² using 500 trees and default parameters. The imputation procedure was performed 100 times, and then the average matrix was calculated. All pairwise ratios of metals were calculated by means of the function ‘calc.rapports’, included in the R package ‘SARP.compo’. The matrix of ratios was unified with the concentration matrix, and the resulting 110 × 45 matrix was scaled and mean-centered using the basic R function ‘scale’. This final matrix was used as input for the following analyses.

RF analysis

The RF algorithm²³ was used for classification of the groups of interest. This algorithm has several strengths: (i) it can manage large numbers of predictor variables simultaneously; (ii) it shows good performances even in the presence of complex non-linear interactions ²⁴; (iii) it is almost immune from the overfitting due to the total number of variables in the data; (iv) it is relatively insensitive to noise; (v) it enables data visualization in a reduced discriminant space using the proximity matrix calculated during the process of forest growing; (vi) the percentage of trees in the forest that assign one sample to a specific class can be interpreted as a probability of class belonging; and (vii) it gives an unbiased estimate of the classification error using the out-of-bag (OOB) observations (avoiding the necessity of cross validation approaches).

RF is a classification algorithm that uses an ensemble of unpruned decision trees (forest), each of which is built on a bootstrap sample of the training data using a randomly selected subset of variables (metals or metal ratios in the present model).^25,26 In a typical bootstrap sample, ∼63% of the original observations occur at least once (Fig. 1). Observations in the original dataset that do not happen to occur in a bootstrap sample are defined as OOB observations. A classification tree is fitted to each bootstrap sample, but, at each node, only a small number of randomly selected variables are included in the model. The trees are fully grown, and each is used to predict the OOB observations. The predicted class of an observation is calculated by the majority vote of the OOB predictions for that observation, with ties split randomly. Accuracies and error rates are computed for each observation using OOB predictions and then averaged over all observations. Because the OOB observations were not used in the fitting of the trees, the OOB estimates are intrinsically cross-validated, and represent an unbiased estimation of the generalization error.²³ The percentage of trees in the forest that assign one sample to a specific class can be inferred as a probability of belonging to a given class.²⁷ In our case, the RF algorithm was used to build two distinct models: In one model, each tree is used to predict whether a sample comes from a patient treated with ESAs or from a patient not treated with ESAs; in the other model, each tree is used to predict whether a sample comes from a patient with a low or high EPO response index. The R package ‘RF’²³ was used to grow a forest of 1000 trees, using the default settings. To reduce the potential bias due to an unbalanced number of samples per group, the option ‘sampsize’ of the ‘RF’ function was used.

Figure 1.

Graphical representation of the Random Forest (RF) algorithm.

Accuracy, sensitivity, specificity, and the area under the receiver operating characteristic curve (AUROC) were calculated according to the standard definitions. The AUROC calculated on the RF score was assessed for significance against the null hypothesis of no prediction accuracy in the data by means of a 10² randomized class-permutations test (R-script in house developed).²⁸ In the permutation test the group labels of the samples are randomly permuted and a new classification model is calculated.²⁹ The performance of the model obtained with the permuted class vector is assessed, and the value of AUROC is expected to be lower than the one obtained for the original (unpermuted) data. By repeating this procedure N times (N = 10²), a null distribution of H0 for AUROC is obtained. H0 is then a distribution of diagnostic statistics of models that are expected to be not statistically significant.³⁰ Statistical significance of the RF model is then assessed by relating the value of the AUROC for the model calculated with the original data set to the H0 distribution of the AUROC values obtained for models calculated with the permuted data sets. P-value was calculated as

$$\begin{equation*} P - value = \frac{{1 + (AURO{C_{permuted}} \ge AURO{C_{original}})}}{N}. \end{equation*}$$

Univariate analysis

Each variable was divided into two groups using each median value as a threshold, and logistic regression models were performed to estimate the association between metals’ concentrations/ratios and ESA treatment.

$$\begin{equation*} {\rm{\ }}\log \frac{p}{{1 - p}} = {\beta _0}\ + \beta \left( m \right), \end{equation*}$$

where the first term is the logarithm of the odds ratio, β₀ is the intercept, and β(m) is the coefficient of a metal or of a metal ratio.

Additional multivariable models (where multivariable refers to a regression analysis for determining the unique contributions of various variables to a single outcome³¹) were calculated in order to adjust for two potential confounding variables: sex and ethnicity.

$$\begin{equation*}{\rm{\ }}\log \frac{p}{{1 - p}} = {\beta _0}\ + \beta \left( m \right) + \beta \left( {sex} \right) + \beta \left( {ethnicity} \right).\end{equation*}$$

Logistic regression models were computed using the function ‘glm’ (R package ‘stats’), and each model’s significance has been assessed through a Wald test.³²

Results

RF discrimination of ESA treatment

RF classification was used to investigate whether the metallomic profile could discriminate, in a predictive fashion, HD-dependent patients treated and not treated with ESAs. A discrimination accuracy of 71.7% (95% CI 71.5–71.9%), sensitivity of 73.5% (95% CI 73.2–73.7%), specificity of 68.5% (95% CI 68.0–68.9%), and AUROC 0.78 (with P-value 0.01) were obtained (Fig. 2), indicating that elements’ levels and their ratios are significantly involved in anemia. The variables that mainly contribute (Gini Index > 1) to the group discrimination in the RF model are Mn to Cd ratio, Mn to Mo ratio, Sn to Sb ratio, Mn, and Mn to Zn ratio.

Figure 2.

(A) Proximity plot of the Random Forest (RF) model discriminating hemodialysis (HD)-dependent patients treated (red dots) and not-treated (blue dots) with erythropoiesis-stimulating agents (ESAs) on the base of their metallomic profiles. (B) Variable importance plot of the RF model measured by Gini Index. The colored boxes on the right indicate the trends of each metal in the two groups (NT: not treated, T: treated with EPO).

The RF algorithm was also used to investigate possible differences associated with sex in the metallomic profiles of ESRD patients. From our analysis, no relevant clusterization emerged between male and female: 52.7% discrimination accuracy, 53.7% sensitivity, and 51.2% specificity.

Logistic regression analysis between metallomic features and ESA treatment

Concentrations of each metal were used to build logistic regression models for the evaluation of their net effect on treated and not-treated patients with ESAs. Multivariable models adjusted for sex and ethnicity were also calculated. Complete results are reported in Table 1. Mn decrement and Mo, Cd, and Sn increase are shown to be characteristic of patients treated with ESAs. Moreover, several of their ratios (Table 1) have significantly different distributions in the two groups. Mn, Mo, Cd, Zn/Cd ratio, Mo/Cd ratio, Mo Sn/ratio, Mo/Sb ratio, and Cd/Se ratio continued to be statistically significant also after correction for sex and ethnicity. Of note, the Zn/Sb ratio showed to be statistically relevant only in the multivariable model.

Table 1.

Association between metals and ESA treatment: univariate analysis and multivariable analysis adjusted for sex and ethnicity. For each variable, the reference is the group at the lower concentration

	Univariate analysis		Adjusted analysis
	Odds ratio	P-value	Odds ratio	P-value
Mn	0.214	0.001	0.25	0.007
Ni	2.267	0.047	1.841	0.218
Cu	0.851	0.689	1.124	0.811
Zn	1.175	0.689	0.882	0.794
Mo	3.213	0.006	3.566	0.018
Cd	3.857	0.002	2.361	0.087
Sn	2.692	0.018	1.153	0.782
Sb	0.851	0.689	1.017	0.972
Se	0.805	0.590	0.514	0.183
Mn/Ni	0.259	0.002	0.365	0.046
Mn/Cu	0.311	0.006	0.424	0.084
Mn/Zn	0.259	0.002	0.348	0.035
Mn/Mo	0.259	0.002	0.215	0.004
Mn/Cd	0.259	0.002	0.324	0.026
Mn/Sn	0.214	0.001	0.34	0.034
Mn/Sb	0.259	0.002	0.335	0.030
Mn/Se	0.259	0.002	0.351	0.040
Ni/Cu	1.917	0.111	1.61	0.324
Ni/Zn	2.267	0.047	2.072	0.143
Ni/Mo	0.441	0.047	0.249	0.010
Ni/Cd	1.625	0.230	1.769	0.237
Ni/Sn	0.615	0.230	0.9	0.832
Ni/Sb	2.267	0.047	1.569	0.357
Ni/Se	1.917	0.111	1.909	0.188
Cu/Zn	1.175	0.689	1.957	0.186
Cu/Mo	0.259	0.002	0.247	0.010
Cu/Cd	0.522	0.111	0.742	0.542
Cu/Sn	0.311	0.006	0.782	0.642
Cu/Sb	0.851	0.689	0.91	0.844
Cu/Se	1.175	0.689	1.624	0.330
Zn/Mo	0.311	0.006	0.226	0.006
Zn/Cd	0.724	0.423	0.737	0.525
Zn/Sn	0.371	0.018	0.688	0.452
Zn/Sb	1	1.000	0.594	0.294
Zn/Se	1.175	0.689	1.578	0.349
Mo/Cd	2.692	0.018	4.844	0.006
Mo/Sn	0.851	0.689	1.193	0.729
Mo/Sb	4.667	0.001	5.646	0.002
Mo/Se	3.213	0.006	4.898	0.004
Cd/Sn	0.441	0.047	0.562	0.237
Cd/Sb	3.857	0.002	3.625	0.011
Cd/Se	2.267	0.047	2.316	0.087
Sn/Sb	5.701	0.000	3.572	0.016
Sn/Se	3.213	0.006	1.902	0.200
Sb/Se	0.851	0.689	1.285	0.607

Exploratory analyses to evaluate associations between EPO resistance and the metallomic profile

RF classification was used to investigate whether the metallomic profile could discriminate between patients with low and high EPO response index. Patients who had not received ESA for 6 months were excluded from this analysis. The calculated model displays low discrimination power (53.3% accuracy, 57.5% sensitivity, and 48.7% specificity).

From logistic regression analysis, it emerged that a higher Mn to Ni ratio and Mn to Sb ratio are associated with higher EPO response index (ER) (odds ratio 0.325, P-value 0.02; odds ratio 0.325, P-value 0.02). Both these associations remain significant also after adjustment for sex and ethnicity. Complete results for all metals and metal ratios are reported in Table 2.

Table 2.

Association between metals and ER: univariate analysis and multivariable analysis adjusted for sex and ethnicity. For each variable, the reference is the group at the lower concentration

	Univariate analysis		Adjusted analysis
	Odds ratio	P-value	Odds ratio	P-value
Mn	0.406	0.054	0.402	0.060
Ni	0.947	0.906	0.863	0.760
Cu	0.623	0.304	0.725	0.500
Zn	2.2	0.090	1.887	0.186
Mo	1.167	0.736	1.052	0.918
Cd	1.29	0.578	1.495	0.400
Sn	1.29	0.578	1.394	0.484
Sb	1.776	0.213	1.969	0.161
Se	1.167	0.736	0.947	0.911
Mn/Ni	0.325	0.018	0.308	0.017
Mn/Cu	0.406	0.054	0.372	0.044
Mn/Zn	0.406	0.054	0.402	0.060
Mn/Mo	0.406	0.054	0.33	0.027
Mn/Cd	0.504	0.139	0.447	0.101
Mn/Sn	0.623	0.304	0.64	0.344
Mn/Sb	0.325	0.018	0.299	0.016
Mn/Se	0.406	0.054	0.402	0.060
Ni/Cu	0.769	0.566	0.704	0.472
Ni/Zn	0.947	0.906	0.993	0.989
Ni/Mo	0.769	0.566	0.853	0.735
Ni/Cd	0.504	0.139	0.493	0.138
Ni/Sn	0.769	0.566	0.743	0.531
Ni/Sb	0.623	0.304	0.654	0.373
Ni/Se	1.167	0.736	1.304	0.587
Cu/Zn	1.167	0.736	1.455	0.443
Cu/Mo	0.769	0.566	0.907	0.842
Cu/Cd	1.167	0.736	1.247	0.642
Cu/Sn	0.769	0.566	0.747	0.536
Cu/Sb	0.769	0.566	0.825	0.683
Cu/Se	0.769	0.566	1.01	0.984
Zn/Mo	0.769	0.566	0.814	0.675
Zn/Cd	0.947	0.906	0.869	0.768
Zn/Sn	0.769	0.566	0.773	0.583
Zn/Sb	0.769	0.566	0.617	0.322
Zn/Se	1.167	0.736	1.241	0.647
Mo/Cd	0.769	0.566	0.683	0.432
Mo/Sn	0.947	0.906	0.897	0.819
Mo/Sb	0.769	0.566	0.704	0.472
Mo/Se	1.167	0.736	1.174	0.744
Cd/Sn	0.769	0.566	0.747	0.535
Cd/Sb	0.769	0.566	0.796	0.628
Cd/Se	0.769	0.566	0.938	0.896
Sn/Sb	0.947	0.906	0.995	0.992
Sn/Se	0.947	0.906	0.995	0.992
Sb/Se	1.167	0.736	1.385	0.497

Discussion

Anemia is one of the more common complications in ESRD patients who are on HD treatment. ESAs are often administered to HD patients as a therapeutic approach, however there is a relevant proportion of patients that present hyporesponsiveness to EPO.^7–9 Since trace metals, as cofactors, condition the activity of numerous enzymes involved in all human metabolic and cellular processes, including erythropoiesis, and HD can alter their levels, investigating possible associations between trace elements and ESA treatment can provide powerful insights into the mechanisms of patients’ responses to ESAs. This hypothesis was investigated in the present study by analyzing trace metal concentrations and, for the first time, to the best of our knowledge, their ratios using a supervised machine learning algorithm such as RF and logistic regression analysis.

The RF analysis is shown to be able to discriminate between HD-dependent patients treated and not treated with ESAs, with an accuracy of 71.7% (95% CI 71.5–71.9%). The variables that mainly contribute to this discrimination are Mn to Cd ratio, Mn to Mo ratio, Sn to Sb ratio, Mn, and Mn to Zn ratio. These findings are further confirmed by logistic regression analysis that identifies alterations of Mn, Mo, Cd, Sn, and several of their ratios (Table 1) as characteristic of patients treated with ESAs with respect to not-treated patients.

No relevant discrimination emerged between patients with low and high EPO response index using the RF algorithm (53.3% accuracy); conversely, from regression analysis it emerged that higher Mn to Ni ratio and Mn to Sb ratio are associated with higher ER. It is interesting to note that not only Mn is higher in HD patients not treated with ESAs (Table 1), but higher levels of Mn (odds ratio: 0.406, P-value: 0.05) are also associated with lower risk of EPO resistance in ESA-treated patients. Manganese is an abundant trace element, which plays several important extracellular and intracellular functions³³: Mn serves a crucial role as a cofactor for many metabolic and antioxidant enzymes, including arginase, glutamine synthetase, pyruvate carboxylase, and Mn superoxide dismutase (Mn-SOD).^13,34 For its role in these metalloproteins, Mn is crucial for the maintenance of human homeostasis. Previous studies have demonstrated that renal dysfunction is prevalent in subjects with low blood Mn³⁵ and that blood Mn and Hb levels both correlate with kidney function.¹³ Moreover, a possible relationship between Mn homeostasis and hepcidin processing could be at the basis of anemia development in patients suffering from CKD.^13,36 The results here presented corroborate these hypotheses and suggest that Mn depletion could be involved in low responsiveness to ESA treatment.

From our analysis, it emerged that higher cadmium levels are characteristic of patients treated with ESAs, thus with pathological anemia. We detected also significant unbalances in its ratios with Mn, Mo, Sn, Sb, and Se. Cadmium blood concentrations have been found to be increased in HD patients with respect to controls,^37,38 and higher Cd levels have been associated with an increased risk of mortality in patients on maintenance HD.³⁹ Moreover, published data suggest that cadmium could be involved in anemia directly through hypoinduction of EPO synthesis in the kidneys and indirectly by the development of a state of functional iron deficiency and hemolysis.^20,40–42

In our dataset, molybdenum, nickel, and tin were shown to be increased in patients treated with ESAs. Moreover, several of their ratios were shown to be statistically altered in ESA-treated patients. Although elevated blood Mo levels have been already observed in HD patients³⁸ and a relationship between Mo and iron metabolism is known,⁴³ it is difficult to formulate a coherent hypothesis on the reason why Mo, Ni, and Sn increments are associated with anemia, and these observations need to be further confirmed by prospective longitudinal cohort studies.

Conclusions

The results here presented show that trace metals, in particular manganese, play a role in the mechanisms underlying renal anemia and response to ESA treatment in a population of ESRD patients. Not only blood metals’ levels are altered by these conditions, but also their interrelationships, as demonstrated by the alterations of several of their ratios. These data, if further confirmed, pave the way for future intervention studies in which to examine if the re-equilibration of metals’ physiological levels could contribute to an enhancement of patient response to ESA treatment and to a better management of HD patients, hopefully reducing their morbidity and mortality.

Conflicts of interest

The authors declare no conflicts of interest regarding the publication of this paper.

Data availability

Data are available at the NIH Common Fund's National Metabolomics Data Repository website, the Metabolomics Workbench, https://www.metabolomicsworkbench.org, where it has been assigned Project ID ST000565. The data can be accessed directly via its Project DOI: http://dx.doi.org/10.21228/M8DK6R.