• Gilles-Eric Séralini^, ,
• Robin Mesnage,
• Nicolas Defarge,
• Steeve Gress

• University of Caen, Institute of Biology, EA2608, CRIIGEN and Network on Risks, MRSH-CNRS, Esplanade de la Paix, 14032 Caen Cedex, France

• Didier Hennequin

• University of Caen, Network on Risks, MRSH-CNRS, Bd Maréchal Juin, 14032 Caen Cedex, France

• Emilie Clair

• University of Caen, Institute of Biology, EA2608, CRIIGEN and Network on Risks, MRSH-CNRS, Esplanade de la Paix, 14032 Caen Cedex, France

• Manuela Malatesta

• University of Verona, Department of Neurological, Neuropsychological, Morphological and Motor Sciences, 37134 Verona, Italy

• Joël Spiroux de Vendômois

• University of Caen, Institute of Biology, EA2608, CRIIGEN and Network on Risks, MRSH-CNRS, Esplanade de la Paix, 14032 Caen Cedex, France

Available online 9 November 2012

• http://dx.doi.org/10.1016/j.fct.2012.11.007, How to Cite or Link Using DOI

• Permissions & Reprints

Refers to

• Lúcia de Souza, Leila Macedo Oda
• Letter to the editor
• Food and Chemical Toxicology, Available online 7 November 2012,
• PDF (141 K)      Universally Available

1. Introduction

Our recent research on the long term toxicity of a NK603 Roundup-tolerant genetically modified maize and of a Roundup (R) herbicide (Séralini et al., 2012) has provoked numerous positive and negative reactions throughout the world. This is the way the science moves forward, and here we answer to this intense debate. Our work is the most detailed study of an agricultural GMO life-long consumption, and especially on NK603 for which only a 90-day safety test existed on the same rat strain (Hammond et al., 2004). It is also the first long term detailed research on mammals exposed to a highly diluted pesticide in its total formulation with adjuvants. These adjuvants help to stabilize the active principles of pesticides, and promote a better penetration in organisms. R is the major herbicide of the world, tested from levels arising in tap water. Indeed in our study, its active principle glyphosate (G) was not studied alone, by contrast to the long term company regulatory experiment. As such, the debate is at the corner of science and regulatory issues on this topic. This fact can impact the economic development of such products, explaining severe comments within hours of the publication. This may explain why 75% of our first criticisms within a week, among publishing authors, come from plant biologists, some developing patents on GMOs, and from Monsanto Company owning these products.

We must firstly focus on science. Our work is a research study, it has not direct regulatory purposes and should not be considered as a final point in knowing the toxicological effects of NK603 and R. This is a first step in the iterative investigation of the long-term health effects on mammals of these commercial products that should be replicated independently (and also on developing animals). It has limits like any study, and here we carefully answer to all criticisms from agencies, consultants and scientists, that were sent to the Editor or to ourselves. These challenged our results and the validity of our protocol, some letters even requested the withdrawal of the publication from the journal. All remarks and answers are summarized in Table 1 and below some details are explained.

At this level, a full debate is biased if the toxicity tests on mammals of NK603 and R obtained by Monsanto Company remain non available for the scientific community and confidential. This is why, after several exchanges, we requested again from EFSA (European Food Safety Agency) on September 20th and October 18th 2012 the release on a public website of the raw data allowing commercialization of these products, in particular the longest study of NK603 on mammals (Hammond et al., 2004). We ask for a free and transparent exchange of scientific findings, mainly when these are related to public health and environmental risks. We recall that the tests on rats are usually considered as a model for mammalian health before clinical trials (for pharmaceutical drugs) or for a direct market release (for novel food and feed, pesticides or chemicals). Moreover, rat tests are also models for environmental risks assessment, since rats are models for other wildlife mammals. The public release of the raw data will indicate if significant differences observed in both studies are coherent, if the statistics are potent in both cases, before a renewal of the experiment by others, maybe through a protocol publically discussed.

2. Relevance of the scientific context

Some remarks emphasize a lack of context, claiming that the study was performed for non-scientific reasons. The establishment of this protocol was however the consequence of an intense debate about the biological relevance of numerous statistical differences toward controls admitted in 90-day feeding studies with agricultural GMOs (Spiroux de Vendomois et al., 2010). This is highly controversial, regulatory agencies having refuted findings which were validated by peer reviewed process in international journals (                  [EFSA, 2007] and [Séralini et al., 2007]). Indeed, they have their own criteria to judge of the biological relevance which does not fit with some recent knowledge (Doull et al., 2007), as for the case of sex specific non linear endocrine disruptions which were not admitted at a regulatory level but at a research level (Myers et al., 2009b). To overpass the divergence in biological interpretations of early signs of toxicity in biochemistry for GMO, one solution was to prolong 90-day tests to chronic periods. We have then chosen the NK603 GM maize because R tolerance is the trait of around 80% of agricultural GMOs (James, 2011) and because statistical differences in the 90-day feeding trial with this maize were admitted by both the petitioner and regulatory agencies (EFSA, 2009).

3. Originality and limits of the experimental design

Due to the economical and regulatory issues of this topic, it is not surprising that our research study was confounded with pre-commercial regulatory assessments. This is why the most common criticism is about the following of organization for Economic Co-operation and Development (OECD) guidelines. However, no guidelines really exist for GMO toxicity studies which are still not mandatory. Reviews confirmed that most of the studies did not follow specific guidelines or were contradictory (                  [Domingo, 2007] and [Domingo and Giné Bordonaba, 2011]). We compared our design (Table 1 of Séralini et al., 2012) to Hammond et al. (2004) inspired from OECD guidelines 408 for chemicals. We have thus replicated and improved their experiment measuring 3 doses instead of 2 on an 8 times longer period, with 11 blood and urine measures of around 50 parameters, 34 organs instead of 17, etc., in order to know if the statistical findings (in 90 days), were biologically relevant or not on a long term. We thus biochemically measured 10 rats per group as performed by Monsanto. Even up to two years, we had no reason to monitor biochemical effects on more than 10 animals per group as this is the number of animals measured in OECD 452 for chronic toxicity testing (OECD, 1997 was in application when the study started in 2008), even if 20 animals per group or more are possible.

The purpose of the addition of R treated groups was not to assess R long term carcinogenesis which needs to follow OECD 453 guideline with at least 50 rats per group. It was to test R in similar conditions than the GM maize to understand if the R feed residues can explain the possible pathologies evidenced. Two main hypotheses were indeed tested: effects from the GM maize itself, treated or not by R, and/or from herbicide residues alone in drinking water using 3 doses by treatments. We noticed that the initial test published by Hammond et al. used 2 doses by treatments although 3 were requested for the OECD guideline 408 they had followed. Also, one of the criteria for biological relevance for Monsanto or critics is the linearity of the effects according to the dose. This cannot be claimed with only 2 doses (NK603 initial assessment, Hammond et al., 2004). It is thus surprising that the relevance of Monsanto’s and agencies conclusion of safety was not challenged due to protocol insufficiencies. A recent review is often cited as a proof of the safety of GMOs consumption on a long term (Snell et al., 2012). However, on the 24 studies they evaluated, only 2 are long term on rodents, since 2 years on pigs or cows are not life-long experiments. The 2 rodent studies are from Sakamoto et al. (2008) in a Japanese journal with not all rats fed transgenic soya analyzed, and Malatesta et al. (2008a) in mice on the same plant with hepatic effects. Moreover, on the 24 studies, 16 did not mention the use of the closest isogenic line as control, many did not describe the methods in detail, and they had other deficiencies too (Snell et al., 2012). However, all these studies were accepted as proofs of safety regardless of these remarks. Conclusions of safety seem to need less requirements than conclusions of toxicity. Scientifically however, toxicity is easier to conclude than safety.

Our controls were also questioned and this needs some clarification. Some claimed that controls are lacking for 4 groups (GMO + R and GMO alone at 11% and 22%). We compared all treated groups to the control group containing 33% of the closest available isogenic maize, as all diets were equilibrated to 33% maize (for instance the 11% GM maize diet was supplemented with non-GM control to reach 33%). More accurately the closest available isogenic line was the DKC2675 variety compared to the DKC2978 GM maize (NK603). Regulatory agencies also questioned the culture conditions. One R treatment was applied 4 months before harvest. Fungicides were applied similarly. We couldn’t use the same R formulations on the field (Canada) and in the drinking water of the rats (France) because authorized formulations are different between countries. The diet was nutritionally equilibrated from substantially equivalent maize and was then PCR checked for GMO content. A major concern was the presence of mycotoxins. Fumonisin B1 and B2, zearalenone, deoxynivalenol (DON), nivalenol, 3-acetyl-DON, 15-acetyl-DON, fusarenone X, T2 toxin, HT2 toxin and diacetoxyscirpenol were all under recommended limits in food/feed. We didn’t detail each of these points when no particular changes affecting the understanding of the results were noticed.

As a research protocol, the Good Laboratory Practice (GLP, OECD, 1997; 2004/10/EC regulation) environment was respected, meaning that housing conditions, manufacturing process, diet composition and storage, stability of solutions and diet contaminants have been assessed by approved laboratories. Anatomopathology was performed in a blind manner (without knowing the treatments) by professional anatompathologists approved for regulatory purposes. An electronic chip was in each rat for identification, but the technicians did not know the nature of the diets and of drinks prepared independently, or the control group. The cages were moved regularly and similarly for all animals. The blood (1 mL) and urine samples were coded and the biochemical parameters measurements were also blind, like the decisions of euthanasia, according to precise regulatory ethical rules (hemorrhages, impossibility to drink and eat, large tumors over 25% body weight because they provoke mortality). All the animals were monitored during the experiment by professional veterinarians. The statistical study was also accomplished on coded groups. However we have made research amendments adding additional analyses (tissular and biochemical parameters) adapted to the findings preventing the request of regulatory GLP agreement. This is classical that a regulatory agency does not take into account research studies because they are not GLP (Myers et al., 2009a). By nature, a research protocol is rarely compatible with GLP agreements. GLP agreement is a good tool to normalize regulatory assessment but research studies needs more degree of freedom, in test protocols, models, etc.

4. Rat strain

We would like to explain the choice of the rat strain. This is another redundant remark. We recall that OECD norms (408, 452 and 453) are not prescriptive for the strain of rats. Sprague Dawley rats (SD) are subject to spontaneous neoplasms and this was supposed to invalidate the model for carcinogenesis. This strain develops tumors, hence it is preferentially used by some agencies such as for the National Toxicology Program (King-Herbert et al., 2010). This would be a non-sense to study pathologies in a strain insensitive to these. Long term OECD guidelines 452 even state that rats and mice have been preferred experimental models because of their susceptibility to tumor induction. The same reasoning is used for chronic progressive nephropathies (CPN) developed by SD rats. The fact that the strain developed spontaneous CPN with age (Hard and Khan, 2004) does not invalidate the model as we looked at the difference in the chronology, age, number and severity of CPN in comparison to controls.

To assess the biological relevance of results, many authors refer to historical data of control rats, within the lab or the manufacturer, which greatly enhances the control variability and the risk of false negative findings. It is now established that this concept should be used with caution. There are several reasons for that. Control diets for rats are generally not monitored, neither for pesticides (Hayes, 2004), nor for chemicals leaching from cages or other environmental sources (Howdeshell et al., 2003). This artificially enhances the background effects. The supplier even recognized that their historical data came from rats potentially fed GMOs since this was not controlled (Harlan communication), except in our experiment. Thus it was not appropriate to use historical data; it is also why we didn’t use reference groups eating different non-substantially equivalent diets, as they increase the standard deviation of the control groups, hiding differential effects due to treatments.

Many non relevant remarks have also been emitted. Among others, some criticized the use of ad libitum feed to explain the increase of tumor incidence. Guidelines on the design and conduct of chronic toxicity studies state that rodents should be fed and watered ad libitum (OECD, Guidance Document No. 116). The hormonal imbalances were criticized to be due to diurnal or cyclic variations. However the samplings were performed at the same time in the morning.

5. Focus on statistics

Statistics do not tell the truth, but may help the understanding. The biological interpretations and the crossing of methodologies are the key. We have applied the most modern statistical methods (OPLS-DA, see below) multivariate data analysis of around 50 parameters measured 11 times for 200 rats. This allowed in a blinded manner to obtain results significantly discriminant at 99% confidence levels. These discriminant biochemical markers were for instance sexual hormones (at 95% for females at month 15) when the differences in hormone-dependent tumor incidences with the control group began. The disabled pituitaries were the second most reached organ as certified independently by the pathologists in a blinded manner in treated female groups. This is known to elicit mammary tumors in rats and pituitary is a target of endocrine disrupting chemicals (Wozniak et al., 2005). The pathologists explained most of the mortality in females by the tumors, which leaded to euthanasia independently of the grade of cancer. This is why we did not detail the grades of tumors in our research, and the cancerous nature of major ones are described in our study (Fig. 3 legend and results section (Par. 3.2)). These observations together with microscopy reinforced our conclusions.

All this was for us more pertinent than the study of non-powerful Kaplan–Meiers’ curves on survival (because of the groups of 10 animals dying progressively) that cannot allow to conclude on a mortality linked or not to the treatment. Taking into account these limits, we decided to be factual and thus to describe the chronology and incidence of tumors and deaths. The statisticians could test in comparison the power of the statistical analyses of the tests of Hammond et al. (2004) concluding on safety and used for market release. For us, it is extremely low to conclude to safety.

In this case, PLS-regression (Projections to Latent Structures by means of partial least squares) is of particular interest because, unlike the conventional multivariate data analysis methods, it can analyze data with variables more numerous than observations which can be strongly correlated (Wold et al., 2001). In the case of Orthogonal Partial Least Squares Discriminant Analysis (OPLS-DA) there is separation between the inter-group variation (represented on the predictive component) and the intra-group variation (variability of the samples, represented on the first Y-orthogonal components). OPLS-DA is thus not a method to compare mean differences; however for biochemical understanding, we have presented those in Table 3 of our study, with highly discriminant parameters in bold. OPLS-DA renders possible to identify which variables are responsible for the separation of the groups. For instance, we also indicate in Fig. 5B that estradiol and testosterone are significantly discriminant at 95% confidence levels in some groups (not at 99% like other parameters presented).

Moreover, the SIMCA-P (V12) software (UMETRICS AB Umea, Sweden) for the multivariate analysis of biochemical data uses a method of validation of models which is a cross-validation k-fold. The Q²(Y) parameter which measures the predictive ability of the models is calculated according to this cross-validation method. Only valid models with a satisfactory predictive quality Q² index were retained for the selection of the discriminant variables (in bold, Table 3). Furthermore all models retained are significant (CV-ANOVA test with p-value <5%). One of us previously used this method and published in international peer-reviewed journals (                   [Malzert-Freon et al., 2010a] and [Malzert-Freon et al., 2010b]).

6. Pertinence of the results

First major criticisms were made about the results and their presentation. A scientific publication is limited in figures/tables and only shows the data necessary to understand and discuss the conclusions. This is why behavioral studies, ophthalmology, microbiology in feces and in infectious nodules, G in tissues, body and organ weights, feed and water consumptions, transgene in tissues, time effects will be the subject of future publications. Their addition does not improve the understanding of this one. As a matter of fact, the peer review process has controlled the logic of this body of data, and additional data were included in the revised publication.

The second major claim for the results is that we attached too much importance to mortality and tumor findings relatively to their scientific pertinence. We are aware of the limit of these findings as discussed previously for statistics. The body of evidences for our conclusion comes from the converging methodologies and data (See Section 5). The variability of the mortality can indeed, if interpreted alone, be expected by chance, but in fact the statistics are not powerful enough to conclude that or the contrary. This is why we have described raw data. For instance, males presented up to 4 times (2 times in mean) more large palpable tumors than controls, like in females. As this may represent a potential risk for the human population, this cannot be simply disregarded so rapidly with non-potent statistics. This is also why we emphasize statistically discriminant biochemical effects at the 15th month, were most of animals were alive (in treated groups 90% males, 94% females, and 100% controls). The two nephroblastomas in GMO fed groups linked to the deaths were criticized for bringing confusion in results, because these tumors are often of embryonic and/or genetic origins. However, these tumors are also known to be promoted by pesticide exposure (Fear et al., 1998).

The summary of the major histopathological findings in Table 2 was subject to the same criticisms. In fact, we indicated the severities of the CPN and only marked or severe CPN were shown. Indeed, elderly rats are subject to CPN and taking into account all CPN could hide interesting differences. The power of statistics may be discussed as for Figs. 1 and 2, however the body of evidences has to take into account all of these data as explained previously.

For electronic microscopy findings, it is important to recall the context. In several published researches, ultrastructural abnormalities in livers of mice fed with GM soybean were found (Malatesta et al., 2002). The pattern was reproduced by adding the herbicide R directly on rat hepatocytes (Malatesta et al., 2008b). We thus wanted to test if the same disruptions can be seen in the liver in our experiment. This was the case, and it is also explained looking at the in vitro effects of R in the literature including in our articles. Glycogen dispersion or appearance in lakes found in electronic microscopy was attributed to feeding state by some external remarks. However, differences in feed consumptions have not been noticed, and the author has experience in this domain. Not only glycogen appearance in lakes was noticed, but also a reduced rate of transcription of mRNA and rRNA which is not known to be due to feeding state, but rather to a toxicity. Patterns of electronic microscopy were coherent with a detoxifying activity increase in liver, and this is corroborated by difference in cytochrome activities.

A major gap in some toxicological assessments is the negligence of endocrine disrupting effects. As noticed previously, the central dogma in toxicology is that effects vary linearly to the dose. This is true for poison intoxication. However, endocrine responses curves in U, inverted U or J shapes appear to be common for biological answers to environmental pollutants (Vandenberg et al., 2012). The endocrine disturbance is supported by bibliography in human cells (Gasnier et al., 2009) and rat testicular cells, for R residues (Clair et al., 2012), and in this work it is demonstrated by statistical sexual hormonal imbalances and disabled pituitaries. Moreover, doses varied from 50 ng/L to 2.5 g/L, thus a factor 50 million, we cannot expect linear effects with such a wide range of doses tested. The kidneys and the livers are also sensitive to endocrine disruptors. As the two major detoxifying organs, containing cytochromes P450 or other enzymes involved in xenobiotic or sex steroid metabolism, they often react with sex hormone steroids and related compounds (Pascussi et al., 2008).

Last but not least, we have identified phenolic acids as biomarkers of metabolic disturbances in the GM diet. We have also measured is flavones in the diet even if the maize does not produce these compounds. Rats indeed didn’t eat only maize but also other plants in an equilibrated diet. Even OECD 452 guidelines on chronic toxicity asked for testing the content in phytoestrogens. Importantly, phenolic acids decrease is a good indication for change in the metabolism of the GMO leading to a lack of protection against the pathologies observed in the animals fed this GMO, but we don’t exclude other toxic effects of the GMO alone which have not been explained in the experiment.

7. Discussion: findings in regard with the contradictory hypotheses

Critics claimed that no argument exist for R to be a sex endocrine disruptor, based on a review by Williams et al. (2000), where most of the studies cover G effects alone and not R ones. We recall again that G is never used as such, but in formulations with other ingredients allowing the toxicity, both on target and non-target species. This is extensively described for G-based herbicides, but also for other pesticides (Eddleston et al., 2012). This is why in our opinion, all discussions of our study referring to testing G alone is not relevant. And worst, we notice non-scientific assertions justifying R innocuity by the structural homology of G with non-toxic aminoacids, and adjuvant innocuity by a comparison of their activity to soap. This is not scientifically relevant to predict with certainty toxicological effects or safety. The fact that G alone is neither a carcinogen nor an endocrine disruptor in regulatory tests is not a proof of the whole formulations safety, especially when some formulations contained toxic compounds (Cox, 2004). Unexpected new active principles for human cell toxicity in G-based herbicides have challenged the relevance of testing G as the active principle in R (Mesnage et al., 2012). R has been already described as an endocrine disruptor in vivo (                   [Dallegrave et al., 2007],                   [Oliveira et al., 2007],                   [Romano et al., 2012] and [Romano et al., 2010]) with the mechanistic underneath in vitro.

Several studies have shown endocrine disrupting effects of R, such as a significant decrease in hCG-progesterone production and decreased levels of Steroidogenic acute regulatory (StAR) mRNA production in MA-10 mouse Leydig cells (Walsh et al., 2000), a significantly decrease in aromatase activity and mRNA levels in JEG3 cells and placental and equine testicular microsomes with R (                   [Richard et al., 2005] and [Benachour et al., 2007]), inhibitions of transcriptional activities of androgen and of both estrogen receptors alpha and beta in cells (Gasnier et al., 2009), and decrease in testosterone production in rat Leydig cells (Clair et al., 2012). All these studies reinforce our findings.

Some critics emphasize that no effect was reported on farm animals fed an unknown part of GMOs, neither in the human population of the USA. First, we should keep in mind that farms animal are not designed to live as long as their physiology could allow, and usually are not able to develop long-term chronic diseases, unlike the rats in our life-long experiment. If this happens on lactating cows, biological analyses performed on cattle are far less complete than the one that are done in regulatory tests and in our study. We also recall that no labeling of GMOs exists in the USA, so the amount of GMOs in feed and in food are unknown, and no “control group” exists. Thus, without a clear traceability or labeling, no epidemiological survey can be performed.

8. Ethical issues

A lot of critics argue against our non-release of raw data. This is very unusual when other papers have to be published out of those. Our study is not a regulatory one. However, due to the social impact and for full scientific understanding of the NK603 and R risks, we will release our raw data if the regulatory agencies that have taken industry data into account release the data pertinent for environmental and health risk assessments, in particular their longest toxicological tests on mammals, as we have indicated to EFSA. As a first step, we have communicated the raw data of figures 1 and 2 to the French food safety agency (ANSES), and answered their questions on experimental design and results, including analysis of food composition and mycotoxins content, etc.

Most of the criticisms on this topic have dealt with animal welfare, some thinking that we overpassed the threshold in size of tumors above which animals should be euthanized, with the purpose of taking shocking photographs. We recall that in a GLP environment, the animal welfare is of major concern and that we fully respected the threshold in tumor sizes before euthanasia. Pictures of every animal and organ were taken. We presented those related to the most observed pathologies, including microscopic ones, in an illustrative purpose in Fig. 3, with rats representative of each group.

Some critics concern the role of funders, and possible conflicts of interests. Of course, the funders had no role in the design and conduct of the experiment, neither on its interpretation. The data remained confidential for the funders also. We recall that in the assessment of GMOs, chemicals and medicines, the companies test their products by themselves, often in their own laboratories. Hereby, conflicts of interest exist in these cases because of the rapid market release research. Our study does not aim at the request of commercializing a new product. We wanted to estimate the health risk of these products. It is the most detailed test independent from the biotech and pesticide companies. We encourage others to replicate such chronic experiment, with more statistical power. Now, the burden of proof has to be obtained experimentally by studies independent from industry. This was recommended by regulatory agencies that have assessed our work in France, even if it their objective is more to regulate products than to review research. GM NK603 and R cannot be regarded as safe as long as their safety is not proven by further investigations.

Conflict of Interest

The authors declare that there are no conflicts of interest.