Dear Authors,

As we received the support from Dr. Jamie Males, senior editor of PosOne, to communicate with you our concerns related to the publication entitled: “Metabolome disruption of pregnant rats and their offspring resulting from repeated exposure to a pesticide mixture representative of environmental contamination in Brittany”, please find hereafter the explanations and suggestions.

The article examines the potential for a cocktail of pesticides to alter the metabolomic status of pregnant rats and their offspring. We believe that this article contains a number of serious methodological and reporting deficiencies which critically undermine the utility of this publication as part of the ongoing debate around the safety of plant protection products (PPP).

In particular we would like to draw your attention to four major issues in this article, and the following suggestions for approaches to correct these issues:

1) That the delivered doses used in this study significantly exceeded the acceptable daily intake (ADI) values for some of the tested plant protection products and thus do not represent realistic exposure values as claimed by the authors. We propose that better justification for these dose levels be supplied by the authors, if this is not possible we suggest revising the sections discussing dose level justifications to highlight that the selected dose levels are unlikely to reflect realistic exposure values.

2) We believe that the biological sample collection and processing methodologies were sub-optimal and may have introduced unaccounted for flaws in this analysis. We suggest that the authors review their sampling and analysis methodology and revise the discussion of this article to better address the uncertainties arising from this section of the work.

3) Our view is that the statistical model used to conduct the multivariate data analysis was flawed due to a biased validation methodology, and that the outputs of this model are therefore highly questionable. We propose that the models be revalidated using a more rigorous methodology, and the results of these models be presented.

4) No attempt to verify the achieved dietary concentration of the tested doses was reported by the authors, this issue is of particular concern given the low doses used in this study. We propose that the authors present their diet analysis data to demonstrate the achieved concentration, stability, and homogeneity of the tested materials in the diet.

1-Relationship between doses used by the and realistic exposures to PPP

Whilst pesticide exposure is typically orders of magnitude lower than the ADI, the authors calculated a maximum theoretical dose via summation of the individual values (Σ = 447 μg/kg bw/day or 0.447 mg/kg bw/day). Administered diets contained the following active ingredients: acetochlor (246 μg/kg bw/day) + bromoxynil (12 μg/kg bw/day) + carbofuran (22.5 μg/kg bw/day) + chlormequat (35 μg/kg/day) + ethephon (22.5 μg/kg bw/day) + fenpropimorph (15.5 μg/kg bw/day) + glyphosate (12 μg/kg bw/day) + imidacloprid (12.5 μg/kg bw/day) (see Table 1). However, the relative proportions of the individual active ingredients within the mixture administered to pregnant animals was calculated based on the proportion (%) of environmental pesticide exposure in Brittany, using data from the registry of pesticide emissions in 2003. Consequently, delivered dietary doses were significantly in excess of the ADI for a number of PPPs, particularly acetochlor (68-fold), bromoxynil (1.2-fold), carbofuran (150-fold), and fenpropimorph (5-fold), and cannot be considered representative of general environmental contamination unless further justification is given.

*** MISSING TABLE - Table 1: EFSA Peer Review of Active Ingredients Administered by Bonvallot et al. (2018)***

2-Sample collection and processing methodologies

On gestational day (GD) 13, eight females per group were placed in metabolic cages and acclimatised for two days prior to a 24-hour urine collection on GD 15. The selection criteria was not specified, and it is unclear why two animals per group were excluded. At necropsy, maternal blood samples were taken from the facial artery, added to a glass vial containing heparin and placed on ice before centrifugation and freezing at -20°C. Endogenous metabolite profiling is susceptible to diurnal variation, consequently appropriate sample collection methods is vital to prevent misinterpretation and ensure that observed differences are not due to an experimental bias. The authors do not state whether or not necropsies and sample collection were randomised, appropriately distributing the time-points across treatment groups.
Plasma, urine and tissue samples were extracted or reconstituted in aqueous solution (D2O), which is not optimal for lipids due to low recovery, compromising the “lipid” nuclear magnetic resonance (NMR) data that Bonvallot et al. utilise in the discussion. Furthermore, NMR is less sensitive than MS, so less metabolites would be detected. Indeed, the authors inconsistently report 35 or 36 metabolites using NMR, whereas MS based metabolomics systems detect thousands. Whilst an advantage of NMR is the reproducibility of results batch to batch, this was not conducted and only 20 samples appear to have been run.

The authors conducted GC-FID analysis, in which fatty acyl chains are hydrolysed from their lipid form and quantified according to their chain length and saturation (number of double bonds), on liver samples. Whilst this GC method is good for comparing length and saturation, the full lipid is not evaluated and the location of fatty acyl binding is unknown. Consequently, there are significant uncertainties in the conclusions that can be made regarding metabolic pathways. Furthermore, whilst the authors used a C17 triglyceride as an internal standard, in accordance with the common misconception that only even numbered carbon chains exist in mammals, there are a small number of odd chain fatty acids incorporated into several different classes of lipids, so the internal standard amount (and therefore the hydrolysis step) is likely to be overestimated. We suggest that the authors review their sampling and analysis methodology and revise the discussion of this article to better address the uncertainties arising from this section of the work.

3-Statistical model validation

The raw NMR produced in this study data was evaluated using SIMCA processing. Whilst PLS (projection to latent structure) discriminant analysis is an appropriate method to find differentiating metabolites, the OPLS (orthogonal projection to latent structure) discriminant analysis provides significant advantages, as this separates the orthogonal variation not related to the two groups being discriminated. This aids data interpretation as other sources of variation, such as analytical drift, do not confound the results. The authors used SIMCA v12, which includes this model. The forerunner to this was using an OSC (orthogonal signal correction) filter, however SIMCA warns that this is prone to over interpretation and “fit”, spuriously identifying findings which have little significance. This filter has been used for several models in the paper.

The authors applied a PLS-DA method to identify potential metabolites corresponding to the bucked with variable importance in the projection (VIP) above 1. VIP is a measure of how influential a variable is, with higher values being more important to the model. A value of 1 is commonly used in metabolomics as anything above this is of “more than average importance”, but this is a rather relaxed value and using a value of 1 would lead to more metabolites being selected as differentiating (assuming the model validation was sound), when some may be quite weak. A VIP score of >2 would be more appropriate to give greater confidence in the identified metabolites.

As a supervised method PLS-DA requires proper validation to ensure the model is not showing a difference that is not true. This paper uses the model fit (R2Y), predictability (Q2Y) and permutations test to validate their models. For the permutations test, the samples are randomly assigned to each group and the model is rerun to see if any of the scrambled data sets give better results than in the original model. There are 2 points to check from this test, firstly that the y intercept of the Q2, which all pass and are reported in tables 3 & 6 of the article. However, it should also be confirmed that none of the scrambled models give better predictability than your original model. However, the authors have not conducted this confirmation. The CV-ANOVA is another validation test which should be used, which is a statistical significance test for the model with p<0.05 being accepted for statistically significant models. Again, this test does not appear to have been carried out. Finally, a cross validation plot should be viewed. During cross validation, part of the data is removed from the model and then the group it belongs to is predicted using the model. For a good model, the actual group and predicted group should be the same, however, many of the models tested showed up to 7 misclassifications indicating a poor fit for the data. The authors also inconsistently model and scale the data (tables 3 & 6 of the article). On reliable data, model type and scaling would provide very similar results, however, three different types of scaling and the application of the OSC filter were applied by the authors when previous models failed, suggesting an a priori bias.

Despite using the same software, the number of latent variables (components) reported varies between the paper and in-house created models. The software used by the authors auto-fits the appropriate number of components based on cross validation (removing part of the data and predicting it back in). Consequently, the software decides on the most appropriate number of components based on a compromise between the best fit and predictability. In the Bonvallot paper it appears that validation tests pass due to additional components have been forced by the authors. For example, dam plasma is reported in the paper with 3 latent variables (components), however only 1 is auto-fitted when using their data. The model with 1 component fails the permutations test, however the model with 3 passes. The CV-ANOVA the p value increases from 0.087 for the 1 component model (which qualifies as a fail) to 0.314 for the 3 component model. Failure to properly validate the model led the authors to incorrectly accept their model, in addition to introducing bias by forcing components that the software has excluded. In the absence of this validation, no meaningful conclusions should be drawn from these analyses. Using the 3 validation tests as described above, only 3 models should be interpreted: Dam brain and male foetus liver pass all tests. Dam liver passes using the scaling and model used by the paper, however it fails when modelling using more stringent criteria. We propose that that the models be revalidated using a more rigorous methodology, and the results of these models be presented.

4-Confirmation of administered dosage

The authors describe the administration of extremely low levels of all pesticides tested, with nominal administered doses ranging from 12 to 246 µg/kg/day. Given the low levels of exposure in this study it is essential to ensure that the administered doses were accurately characterised to confirm that diets were correctly prepared and as these data are not presented it is not clear that this validation step has been carried out. In addition, no assessment of the stability or the homogeneity of the test materials in the diet were presented by the authors of this publication. It should be noted that the stability and achieved concentration of the test material in the diet, and the homogeneity of the test material throughout the diet are standard requirements in internationally accepted test guidelines for mammalian toxicity assessment (e.g. OECD test guidelines 407, 408, 451, 452, and 453).


Thank you for your consideration, and we look forward to your response.


Yours Sincerely,


Dr. Stephanie Nadzialek,
On behalf of the European Crop Protection Agency.