STATISTICAL THEORY AND ANALYSIS OF GMO ENFORCEMENT (STAGE) DEFRA PROJECT CB0209 FINAL PROJECT REPORT

STATISTICAL THEORY AND ANALYSIS OF GMO ENFORCEMENT (STAGE) DEFRA PROJECT CB0209 FINAL PROJECT REPORT JANUARY 2005 CENTRAL SCIENCE LABORATORY, SAND HUTTON, YORK, YO41 1LZ

CONTENTS Contents...1 EXECUTIVE SUMMARY...4 1. INTRODUCTION...9 2. REVIEW OF LITERATURE AND CURRENT PRACTICE...11 2.1 Outline of current sampling methods...11 2.1.1 Certified seed lot size...12 2.2 Sampling size and frequency...12 2.2.1 International Seed Testing Association (ISTA) Rules [2]...12 2.2.2 ISO 13690: 1999 Cereals, pulses and milled products sampling of static batches [1]...15 2.2.3 ISO 6644 Flowing cereals and milled cereal products - automatic sampling by mechanical means [8]...16 2.2.4 EN ISO 542: 1995 (BS 4146: 1991) Oilseeds sampling [7]...17 2.2.5 Commission Recommendation on technical guidance for sampling and detection of genetically modified organisms and material produced from as or in products in the context of Regulation (EC) No. 1830/2003 [3]...18 2.2.6 American Association of Cereal Chemists (AACC) method 64-70A and method 64-71 [22]...19 2.2.7 USDA Grain Inspection, Packers and Stockyards Administration (GIPSA) 19 2.2.8 European Enforcement Project (EEP)...21 2.2.9 Kernel Sampling Technique Evaluation (KeSTE) [10] / Kernel Lot Distribution Assessment ( KeLDA)...22 2.3 GM testing methods...22 2.3.1 Commercial GM testing methods and procedures...22 2.3.2 Quality control of GM testing methods...22 1

2.4 EU recommended methodology and detection limits...24 2.4.1 Background EU legislation...24 2.4.2 Theoretical detection limits of polymerase chain reaction (PCR)...25 2.4.3 EU recommended methods for PCR tests...27 2.5 Review of seed audits...28 3. DATA COLLATION AND ANALYSIS...30 3.1 Database description...30 3.2 Statistical analysis of database...33 3.2.1 Data description...34 3.3 Results of statistical analysis...35 3.3.1 Is there a difference in performance between before and after extraction methods were changed?...35 3.3.2 Do some primer/ positive control sets perform better than others?...37 3.3.3 Is there any change in performance over time?...39 3.3.4 What is the estimated false negative detection rate for unknown samples? 42 4. SIMULATION OF SAMPLING AND ANALYSIS OF GM SEEDS...44 4.1 Model Design...44 4.1.1 Stratified-binomial model for heterogeneity...47 4.1.2 Beta-binomial model for heterogeneity...47 4.1.3 Sampling from a working sample...49 4.1.4 Features of the stratified-binomial and beta-binomial models...49 4.1.5 Comparing beta-binomial and stratified binomial models to a physical model for diffusion of GM seeds into a lot...51 4.1.6 Simulating the detection of GM seed in an analytical sample...52 4.1.7 Analytical replication in the model...53 4.2 Outputs from the model...54 2

4.3 Input values for model parameters...55 4.3.1 Commodity parameters...56 4.3.2 Lot parameters...57 4.3.3 Sampling parameters...57 4.3.4 Analytical parameters...57 4.3.5 Seed lot / as grown scenario...58 4.4 Results from the model simulations...59 4.4.1 Seed lot / as grown scenario...59 4.4.2 Grain as commodity scenario...62 4.4.3 Effect of heterogeneity on the detection of 0.1% GM seeds in a lot...64 4.4.4 Model sensitivity analysis...66 5. CONCLUSIONS...68 5.1 Future work...68 References...70 3

EXECUTIVE SUMMARY 1. This project was commissioned to research the theory and practice of sampling and detection for genetically modified organisms (GMOs) and to develop an enhanced understanding of our statistical confidence in results obtained. The project has draw together practical experience, data and statistical theory to address the scenarios commonly encountered in the work of the GM Inspectorate (GMI). This has enabled us to state clearly and with supporting evidence the confidence we have in the whole testing process and to propose how to deal rationally with situations such as re-tests and false positives. 2. A novel statistical simulation model has been developed that combines sampling and testing for GMOs in one framework, incorporating the uncertainties in each. This model offers policy makers and regulators an evidence-based tool that can be used to inform decision taking. The model has been used to critically review the performance of current practice in sampling and testing to detect reliably the presence of genetic modification (GM) in oilseed rape at permitted thresholds under a range of commonly encountered situations. 3. Model outputs and application. The model produces operating characteristic curves that show the probability of detection of a GMO as a function of the true mean proportion in the lot. Other summary statistics relating to the repeatability of various stages in the process are also produced. The model directly incorporates all parameters with significant impact on uncertainty and limits of detection for GM sampling and testing; it can therefore be used to comprehensively define the minimum parameter bounds within which good practice (i.e. fitness-for-purpose reliability) is achievable. Thus, the model can inform guidance that should be issued to underpin the enforcement of European Union (EU) and national legislation. The model can also be used to optimise the allocation of resources to greatest effect to meet cost or logistic constraints. The model could be used in the future to assist the GM Inspectorate in its audits of the procedures of seed importers and producers by providing objective standards against which to judge the supplied information and, through model sensitivity analysis, an understanding of what data are critical to know in order to be confident in the fitness for purpose of the submitted procedure. False positive 4

and negative rates are sometimes provided in supporting data these can be incorporated in the model to assess their impact on the claimed results. The model could also be used to inform decisions and provide an objective basis upon which to propose further sampling and/or testing in any investigation where a potential breach of consent or legislation is suspected. 4. Model construction. The model structure explicitly includes all stages of the sampling and testing process. Its mathematical and statistical calculations are derived from the best available and generally accepted theory. The input parameters of the model relate directly to information commonly available in the course of collecting samples and conducting laboratory tests. The model is implemented in Microsoft Excel VBA and runs readily and quickly on a moderately powerful personal computer. 5. Modelling heterogeneity. Heterogeneous distribution of GM seeds in bulks is an issue that has been given much consideration. Heterogeneity is important because, as the degree of heterogeneity increases, the choice of appropriate sampling plan to ensure samples are representative of the material being analysed becomes increasingly important. Heterogeneity can be modelled by considering the bulk/lot as composed of a mixture of sub lots each characterised by a mean proportion of GM seed, and specifying how the lot is partitioned into these sub lots. Two methods of taking account of heterogeneity in the sampled bulk were considered in the development of the model used in this study: first, a model that partitions the bulk into two components with respectively zero or constant proportion of GM seeds (called stratified-binomial ) and, second, a more general model where each sub lot conforms to the binomial distribution and the mixture of these (i.e. whole lot) is modelled by the beta distribution this is the beta-binomial model. Heterogeneity is input to these models as the proportion of the bulk / lot expected to contain 100% or 95%, respectively, of any GM seeds present. This is an intuitive parameter that could be estimated by expert opinion where measurements are not available. Results from simulations based on these two methods for heterogeneity showed that the stratified-binomial approach generally gives results with sharp discontinuities in the probability distribution 5

functions for number of seeds in each increment. The results from the beta-binomial model give plausible continuous probability distributions for GM throughout the whole bulk / lot. The stratified-binomial model assumes no mixing between the GM and non-gm sub lots ; whereas, the beta-binomial model gives results that are consistent with a simple physical model for diffusion and seems more realistic given that there is inevitably some mixing as seeds are handled and transported. 6. Model inputs. Model input data were derived from a review of international protocols that define sampling and testing processes and relevant literature together with an analysis of data held at the Central Science Laboratory (CSL) collected in the course of quality control for GM testing. The model has been applied to a range of situations that relate to the work of the GM Inspectorate which include sampling and testing scenarios such as certified seed lots, seeds as traded commodity, and as-grown seeds prior to certification. 7. Sensitivity and specificity. False negatives (presence of GM seeds in the bulk / lot not detected) can arise even where the testing process is fully under control and are typically expected to be close to the limits of detection. Scenario-specific false negative rates (i.e. sensitivity) can be estimated from the model or can be an input parameter where sufficiently detailed information on the testing process is not available. False positives (i.e. positive results for samples that should be negative) can only occur where the testing process is out of control and are caused by extraneous factors such as cross-contamination of analytical aliquots. Hence, estimation of false positive rates (i.e. specificity) is not included in the model. Laboratory protocols reject all test results from a batch where negative controls are found to be positive and this provides effective protection against spurious results. In the dataset used in this study, the rate of failure of negative controls (false positive rate) was 0.7% and was approximately two orders of magnitude less than the false negative rate. False positive rates can be included in the model where desired, for example when modelling decision rules for replication to reduce the risk of reporting false positive results. 6

8. Results from simulations. The application of the model has confirmed the interdependence of sampling and testing in determining the reliability of the combined process. Results from the range of practical scenarios for oilseed rape modelled show that current best practice is generally reliable in assuring that existing thresholds down to 0.1% can be met by the analysis of samples (of 3000 seeds) from a small number of increments taken from a lot, even in the presence of heterogeneity, provided that suitable analytical replication is employed (e.g. duplicate analytical samples with duplicate DNA extractions). However, if no analytical replication is employed then the presence of 0.1% GM will not be reliably detected by the analysis of samples (of 3000 seeds) based on a large number of increments, even if taken from a homogenous lot; any heterogeneity in the lot will make this less reliable. The model has also been used to show that the level of replication (analytical sample, DNA extraction, polymerase chain reaction (PCR)) is at least as important as the number of replicate analyses. The effects of bulk / lot heterogeneity can be ameliorated to a large degree by employing replicate analytical samples, but carrying out replicate PCR determinations of a single analytical extract generates little improvement. 9. Potential extensions. The model structure and flexible output provision will allow it to be readily adapted to new situations (e.g. quantification by Most Probable Number based on splitting samples into pools for testing). In addition, a further quantitative module could be developed to estimate the risks (for producers and consumers) attached to labelling or rejection decisions based on the entire sampling and analytical process whether the analysis is semi-quantitative (e.g. number of positive seed pools) or fully quantitative (e.g. real-time PCR). 10. Conclusions. It is essential to combine both sampling and testing uncertainty in any model of the reliability of GM seed testing. Heterogeneity is a crucial factor in such a model and should be incorporated in a physically-plausible manner. Results from simulations under a range of commonly-encountered scenarios for GM inspection and enforcement of oilseed rape seed show that good practice can ensure the reliability of the sampling and testing process in meeting the requirements of legislation on thresholds and confidence levels. This study has 7

objectively demonstrated the information required to ascertain the reliability of claimed results. 8

1. INTRODUCTION The regulation of adventitious presence of GMOs in seeds, crops and foodstuffs requires sampling and testing processes that can provide a predictable minimum level of confidence in their results. This is essential if test results are to be used to inform decisions on labelling and/or rejection of commodities due to their GM content, and to maintain public confidence in the food supply chain. The issue under investigation in this study is the confidence that can be placed on the combined performance of sampling and testing processes to detect (not quantify) GM at low level. It is particularly relevant, therefore, to 'unauthorised'gmo adventitious presence for which acceptable thresholds are low ( 0.1%). The overall aim of this project was to examine the whole process of GM testing, and its component stages, in order to estimate the total confidence that can be placed on final test results. By considering various different elements from sampling through to laboratory testing, the study has also identified the processes and parameters most critical to the overall test confidence levels and the sensitivity of the process to each of its component elements. As an example, we have investigated the sampling and testing of oilseed rape (OSR). There is no single internationally accepted protocol for sampling and GM testing as an integrated process. Accepted methods are commodity specific and can vary regionally. International standards do exist for sampling methods, which also can be specific to commodity type e.g. International Standards Organisation (ISO) Standards [1] for grain and International Seed Testing Association (ISTA) Rules [2] for seed; some are specific to GM testing e.g. EU legislation guidance [3]. In many situations, none of these sampling methods need necessarily have been employed. It is important to be able to estimate the effect that any sampling scheme could have on the confidence in a test result. Testing methods for GM presence in OSR are currently restricted to PCR-based methods. Lists of recommended and/or validated tests are available [4] but they do not stipulate the entire procedure to be used (including sampling) or precisely describe the parameters specifying the laboratory methods. As a result of these uncertainties, it is difficult for regulators, producers and consumers to have a known confidence in test results or known level of risk attached to decisions based on those results. 9

Hence, a tool for modelling the entire measurement process of sampling and PCR-based GM detection by computer simulation has been produced. The model has been used to examine the detection of GM seeds in oilseed rape certified seed lots and oilseed rape grain lots using parameters that are the most realistic and up-to-date at this time. The framework of the model, however, allows the modelling of many different scenarios and inputs to provide several key outputs that can be used to assess the robustness of any test result, given sufficient information on how that result was obtained. The outputs can be used to estimate several commonly requested parameters and information such as the expected false-negative rate and effectiveness of further analytical replication. 10

2. REVIEW OF LITERATURE AND CURRENT PRACTICE 2.1 Outline of current sampling methods The purpose of sampling is either to obtain a sample corresponding in characteristics and composition to the lot from which it was taken, or to detect unwanted hidden contamination. This review summarises and discusses the various sampling protocols for seed and grain that have been published (or are pending) in the EU and elsewhere. Particular emphasis is given to oilseed rape, but other commodities are also covered. At present there are very few published protocols designed specifically for sampling for the detection of adventitious GMOs. Exceptions are the United States Food and Drug Administration s (USFDA) Cry9C protocol, which provides the agricultural sector with guidelines for sampling maize for Cry9C protein residues from Starlink GM maize; also a draft protocol produced on behalf of the European Enforcement Project (Reed, 2002) [5]. Kay and Paoletti (2001) [6] suggest employing ISO 13690 (cereals) [1] or ISO 542 (rapeseed) [7] when sampling in order to detect GM presence in static bulks. For flowing grain, they suggest using ISO 6644 [8]. In their crop guidance documents the CSL GM Inspectorate recommend using International Seed Testing Association (ISTA) rules when sampling for adventitious GM presence, although it must be borne in mind that these rules have been developed in order to fulfil quality criteria that fall short of the 0% level for unauthorized GM presence. For example, in the case of oilseed rape the varietal purity standard is 99.7% for certified seed (and only 99% for fodder rape). The New Zealand Ministry of Agriculture and Forestry also stipulate that samples are collected using ISTA (or Association of Official Seed Analysts) methodology. However, very few of the current sampling standards have been devised for the purpose of GM detection (including ISTA), and little, if any, validation has been conducted on them to determine whether they are fit for this purpose. Guidelines from various EU and US sources for the detection of GMOs are summarised in a draft document by Paoletti et al., (2003) [9]. The authors conclude that most sampling protocols are based on the false assumption of homogeneity or random distribution, and propose a validation protocol to identify appropriate sampling strategies by simulation studies based on data from EU member states (Kernal Lot Distribution Assessment, KeLDA) [10]. 11

2.1.1 Certified seed lot size The fundamental basis of seed certification is to enable certain parameters, e.g. purity and germination, to be stated for a particular lot with a high degree of confidence. Statements on seed quality are based on tests carried out on samples from each lot. Confidence in these statements requires that the samples are representative, which in turn is based on the lot being homogeneous and the samples being drawn randomly. The issue of samples being representative is particularly important when considering GMOs as the tolerable levels are very low, adventitious GM presence is unlikely to be inherently uniformly distributed and it may be necessary to distinguish between two lots derived from the same as grown crop. There are two issues of concern in the literature for this project. First, none of the studies cited specifically deals with oilseed rape. Second, as the studies tend to be precipitated by pressure from the industry to increase maximum lot sizes, there are no hard data on long-established lot sizes although it is expected under current maxima that 1% of lots will be expected to be heterogeneous [11]. 2.2 Sampling size and frequency 2.2.1 International Seed Testing Association (ISTA) Rules [2] The first issue of the ISTA rules in 1931 included maximum seed lot sizes and ISTA continue to set maximum seed lot sizes as a precautionary measure to limit heterogeneity in seed lots. ISTA tests [2] for heterogeneity are based on a variance ratio method [12], which compares measured variance with expected variance from a binomial distribution assuming equal means and sets acceptable limits on heterogeneity. A modification of this method [13] gives the statistic D which, although related to heterogeneity, is independent of the number of samples and can be used to combine data from disparate studies. This has led to a number of studies ([13], [14], [15]) and comprehensive reviews ([16], [17]) demonstrating that as lot size increases both the number of heterogeneous lots and the average heterogeneity (H value) increase in a linear fashion. ISTA stipulates sampling frequencies for seed lots up to 40t (±5%) in mass (this is the maximum lot size and applies to maize). For oilseed rape, the maximum lot size is 12

10t ±5%. There is no specified minimum mass for a seed lot. Legislation in England [18] currently sets the maximum lot size for oilseed rape at 10 tonnes; these regulations stem from EC Directive 2002/57 [19], which, in turn, was informed by the ISTA rules [20]. Actual sampling frequencies range from 3 primary samples per container (Table 1) for small containers, to 1 sample per 700kg for large containers or seed in bulk (Table 2). Table 1: Sampling requirements (following ISO standards) for seed in sacks or similar sized containers, containing at least 15kg but not more than 100kg of seed. Number of containers in the lot Minimum number of primary samples to be taken 1 4 3 primary samples from each container 5 8 2 primary samples from each container 9 15 1 primary samples from each container 16 30 15 primary samples from the seed lot 31 59 20 primary samples from the seed lot 60 or more 30 primary samples from the seed lot Table 2: Sampling requirements (following ISO standards) for large containers or seed in bulk. Lot mass Minimum number of primary samples to be taken Up to 500kg At least 5 primary samples 501 3,000kg One primary samples for each 300kg but not less than 5 3,001 20,000kg One primary samples for each 500kg but not less than 10 20,001kg and above One primary samples for each 700kg but not less than 40 For sampling containers holding more than 100kg of seed, and for seed in bulk, primary samples must be taken from different horizontal and vertical positions systematically selected or at random. The ISTA rules are periodically updated to take account of increases in current knowledge. The latest edition is Edition 2004 (valid from 01/01/04) [2]. Containers may be sampled systematically or at random and primary samples drawn from top, middle or bottom. The position from which the seed is taken is to be varied from container to container. 13

Where seed lots are held in containers holding less than 15kg of seed, a 100kg mass of seed is taken as the basic unit and the small containers are combined to form sampling units not exceeding this mass (e.g. 10 packages of 10kg, 8 packages of 12kg). For sampling purposes, each unit is regarded as one notional container and the sampling procedures prescribed in Table 1 are used. For example, a seed lot consists of 3.84t divided into 320 packages of 12kg each: 8 12kg = 96kg, 320/8 = 40 containers, therefore the number of primary samples to be taken from the lot is 20. The sampling instrument must be capable of sampling all parts of the seed lot. Samples are generally taken using spears, and these range from small sack spears to large multi-chambered instruments. Samples may also be drawn from a seed stream during processing, using an approved automatic sampling device. In this case portions of seed are taken at regular intervals throughout the processing of the lot using the same sampling intensity as for seed in bulk (Table 2). At the time of sampling the lot must have been subjected to appropriate mixing, blending, and processing techniques so that it is as uniform as practicable. If there is evidence of heterogeneity then the seed lot should not be sampled. In cases of doubt, the ISTA rules state that heterogeneity can be determined using an equation that takes account of the expected and observed variance of the particular attribute being tested. However, there are likely to be a number of quality characteristics where it is not possible to visually differentiate between a homogeneous and a heterogeneous lot and therefore there will be no cue to conduct heterogeneity testing GMO presence falls into this category. Primary samples should be of approximately equal size and, whilst no indication is given of the desired mass, this factor will be influenced in practice by the actual seed mass and the various test requirements. Once all the primary samples have been taken and appear to be uniform, they are combined to form the composite sample. The sample submitted for testing is obtained by mixing and reducing the composite sample (if necessary) to an appropriate size using a riffle or centrifugal divider. For OSR, the minimum mass of a submitted sample is 200g. In the laboratory the composite sample is once again thoroughly mixed and further reduced to obtain the working sample for each specific test. 14

ISTA sampling methodologies have been adopted by the CSL GM Inspectorate when sampling to detect adventitious GM presence and are recommended in the guidance issued by the Inspectorate to UK seed producers and importers. 2.2.2 ISO 13690: 1999 Cereals, pulses and milled products sampling of static batches [1] ISO 13690 concerns the sampling of static bulks of cereals, pulses and milled products. This protocol covers manual or mechanical sampling up to a depth of 3m and mechanical sampling up to 12m. For bulks exceeding 12m depths, ISO 6644 must be used [8]. Consignments are considered in lots of 500t maximum and there are recommended sampling rates for lot sizes from 500 to 10,000 tons. Samples must consist of at least 3000 grains each and samples must be tested individually in order to fall within expected detection limits and to reveal potential heterogeneity in the lot. ISO 13690 does not apply to seed grain or when sampling to test for hidden infestation. a) Sampling from bags: increments are taken from different parts of the bag (e.g. top, middle, bottom) using a sack/bag spear. The sampling rate is shown in Table 3, however there is no indication of the maximum or minimum size of the bag. Table 3: Sampling requirements for seed in bags according to ISTA rules. Number of bags in consignment Number of bags to be sampled Up to 10 Each bag 10 to 100 10, taken at random More than 100 Square root (approx.) of total number, taken according to a suitable sampling scheme a a For example, divide into (n-1) groups containing n (or n-1) bags; the remaining bags constitute a group. For a consignment comprising 200 bags, 200 = 14.142, therefore n=14. Make up 14 groups of 14 bags (total 196 bags); draw up a list of 1 to 14, and cross out one number (e.g. 7); sample the 7 th bag from each group; sample 1 bag at random from the remaining group. A total of 15 bags have therefore been selected. b) Sampling from rail or road wagons, lorries, barges or ships: each laden wagon is sampled and a suggested pattern is provided for increments depending on the lot size (Figure 1) 15

Figure 1: Sampling from rail or road wagons, lorries, barges or ships [7]. Up to 15t: 5 sampling points 15 to 30t: 8 sampling points 30 500t: minimum of 11 sampling points c) Sampling from silos, bins or warehouses: grain is sampled using a grid system (e.g. one similar to b)). The minimum number of increments is determined as follows: square root of tonnage divided by 2 and rounded to the next whole number (e.g. 500t = 12 samples; 10,000t = 50 samples). The composite sample is formed by combining the increments and mixing them thoroughly. The laboratory sample is obtained by coning and quartering or by using a riffle or centrifugal divider. The size of laboratory sample(s) is determined by the type and requirements of the tests, but a minimum of 1kg is generally recommended. 2.2.3 ISO 6644 Flowing cereals and milled cereal products - automatic sampling by mechanical means [8] This method is applicable for all depths of bulk cereals/milled cereal products. It describes an approach whereby a mechanical sampling device is used to take an increment or series of increments from a lot, either continuously or intermittently, and repeatedly. The sampling device must be capable of taking increments from the entire cross section, or as much of it as possible. Consignments are considered in lots of 500t maximum, and a maximum increment size of 1kg is recommended for 16

intermittent sampling. A maximum composite sample size of 100kg is recommended giving a sampling rate of 1 sample per 5t. 2.2.4 EN ISO 542: 1995 (BS 4146: 1991) Oilseeds sampling [7] ISO 542 specifies general conditions relating to the sampling of oilseeds for the assessment of quality. Consignments are considered in lots not exceeding 500t and the material may be in bulk or in bags. Increments are taken either from the flowing product (preferred method) or, in the case of lorries and wagons by sampling at least 5 different positions according to the size of the consignment. Recommended sampling apparatus includes spears, scoops and steam samplers. It is necessary to take a sufficient number of increments to provide a representative bulk sample. Products in bags increments must be taken from 2% of the bags forming the lot, with a minimum of 5 bags sampled. Products in bulk increments taken from flowing products must be taken across the whole section of the flow at time intervals depending on the rate of flow. In the case of wagons or lorries increments shall be taken at 3 levels at least, at the points specified in Figure 1 (ISO 13690 (see 2.2.2, above), although ISO 542 specifies only 30 to 50t for this last sampling plan. Bulk samples are mixed and divided to obtain the required number of laboratory samples (e.g. using quartering iron, conical or slot divider). Recommended sample sizes are shown in Table 4. Table 4: Recommended mass of samples of oilseeds [7] Product Increment mass (kg) Bulk sample mass (kg) Laboratory sample mass (kg) Medium- and 0.5 100 2.5 to 5 large-sized seeds Small seeds 0.2 50 1 to 2 17

It is apparent that for small seeds such as oilseed rape 250 increments of 0.2kg should be taken per lot to achieve a bulk sample of 50kg. This sample is then reduced to obtain a laboratory sample of 1 to 2 kg. 2.2.5 Commission Recommendation on technical guidance for sampling and detection of genetically modified organisms and material produced from as or in products in the context of Regulation (EC) No. 1830/2003 [3] Regulation 1830/2003 [21] describing Member States obligations in respect of the traceability and labelling of GMOs requires that technical guidance on the sampling and detection of GMOs is established. This draft guidance document was accepted in October 2004. It should be noted that this guidance document has been developed to facilitate compliance with labelling requirements and is not concerned with unauthorised GM events although the sampling issues for both are analogous. This recommendation states that sampling of seeds must be carried out in accordance with internationally accepted methods and that these will normally be the latest ISTA methods. The general principles and methods of sampling should be in accordance with the ISTA rules and the associated ISTA Handbook on Seed Sampling. The recommendation also states that systematic sampling be used as at this time there is no substantial information on the distribution of GM seed in conventional seed lots. In this context, the associated risks and corresponding quality levels are defined in relation to the thresholds for genetically modified seeds and relate to the percentage of GM seeds in the total seed. The associated risks are defined as follows: the risk to the producer (α risk) shall be no greater than 5% that their seed lot will be rejected if it has a true GM content (%) below the Acceptable Quality Level (AQL) (this gives a 95% confidence level to the producer that a lot below the AQL is accepted); the risk to the consumer (β risk) shall be no greater than 5% that they receive a seed lot that has a true GM content (%) above the Lower Quality Level (LQL) (this gives a 95% confidence level to the consumer that a lot above the LQL is rejected). For thresholds greater than 0%, the Acceptable Quality Level (AQL) is half the threshold and the Lower Quality Level (LQL) is twice the threshold, according to normal analytical practice, but not exceeding 0.9%. For a threshold of 0%, the Lower Quality Level 18

(LQL) is 0.1%. The minimum number of seeds to be examined is set at 3000 and is determined on the basis of the LQL of the 0% threshold for unauthorised events being 0.1% and a β risk = 5%. Sampling of bulk commodities (grains) should be carried out in accordance with the relevant ISO standards (e.g. ISO 6644 and ISO 13690 for cereals and pulses) where available. Where such standards do not exist, guidance may be sought from the European Network of GMO Laboratories (ENGL) but all sampling strategies should ensure that (a) any sample-sized unit within a bulk should have the same probability of being selected as any other, and (b) grain size is taken into account. It cannot be assumed that grain lots are homogeneous with respect to GM content, thus the sampling plan must aim to provide a number of samples the test results of which approach as closely as is practicable the true mean and variance of the entire lot. Samples must consist of at least 3000 grains each and samples must be tested individually in order to meet expected detection limits and to reveal potential heterogeneity in the lot. The recommendation states that a sequential testing plan may be adopted and therefore further sampling and testing may be necessary (e.g. if the test result is sufficiently close to the threshold to preclude a labelling decision with acceptable risk). 2.2.6 American Association of Cereal Chemists (AACC) method 64-70A and method 64-71 [22] Method 64-70A concerns the manual sampling of grains and suggests sampling plans but does not recommend sampling rates other than for moving streams where it is suggested to take 1 sample (size unspecified) for every 500 bushels (approximately 14.6t wheat). Method 64-71 recommends diverter sampler activation rates of once every 5.5t. 2.2.7 USDA Grain Inspection, Packers and Stockyards Administration (GIPSA) GIPSA s Grain Inspection handbook Book 1 Sampling [23] discusses the mechanics of sampling grain and provides various sampling plans for barges, trucks, and hopper cars, although the principles would presumably apply equally to static bins of similar dimensions. Grain in sacks is also covered, albeit briefly, with a recommended sampling intensity of 72 samples for lots over 10,000 sacks. No indication of 19

sampling intensity is given for lots with less than this number of sacks, and the size of the sacks and samples is not stipulated. Whether the sampling regime is designed for regulatory or market requirements is not specified, and although the book emphasizes the need for representative sampling it seems to rely on number and position of sampling points per container rather than number of samples per mass of material. 2.2.7.1 GIPSA s Practical application of Sampling for the Detection of Biotech Grains [24] This document [24] briefly discusses sampling theory with reference to the detection of GM maize, although the principles it refers to can be applied to other commodities. Although the sampling plan assumes that a single kernel can be detected in a sample regardless of the size of the sample, it does acknowledge that in practice analytical methods have limits and it suggests that large samples should be subdivided and each part tested separately. The following formula and Table 5 give sample sizes for various lot concentrations for different probabilities: N = log(1-(g/100))/log(1-(p/100)) Where: n is the sample size (number of kernels) G is the probability of rejecting a lot concentration, and P is percent concentration in the lot Table 5: Requisite sample sizes and associated probability levels for maize according to different lot concentrations [24] Lot concentration (%) Number of kernels required 99 % rejection 95 % rejection 90 % rejection 0.05 % 9209 5990 4605 0.10 % 4603 2995 2302 0.50 % 919 598 460 2.2.7.2 GIPSA s Sampling and testing recommendations for the detection of Cry 9C protein in hybrid seed corn [25] This document [25] prescribes the use of the sampling protocols set out in the GIPSA Grain Inspection handbook to obtain a representative sample when testing for Starlink corn. The method involves systematically dividing the sample until 2400 kernels remain. These are then tested using lateral flow test kits (e.g. the 20

Cry9C QuickStix test). GIPSA has validated several of these test kits and claim reliable detection limits down to one StarLink kernel in 800 (0.125%) (see www.usda.gov/gipsa/biotech/starlink/191.pdf). The original document also gives the probabilities of accepting a lot for varying concentration levels based on a detection limit of 0.2% (Table 6). Table 6: Probability of accepting maize at various GM concentrations (% of kernels containing Cry9C protein) [25] Lot concentration (%) Probability of acceptance (%) 0.0 100 0.05 30.1 0.09 11.5 0.1 9.1 0.19 1.0 0.3 0.1 2.2.8 European Enforcement Project (EEP) 2.2.8.1 Protocol for seed sampling for testing for GM presence This standard operating procedure has been produced on behalf of the EEP and states that sampling of seed should be in accordance with Article 7 of Council Directive 66/400/EEC for beet seed, or the equivalent article in the applicable directives for other crops, and that the general principles and methods of the ISTA rules (1999) [26] (latest edition [27]) and associated ISTA Handbook on Statistics in Seed Testing (revised version 2002) [28] should be followed. 2.2.8.2 Protocol for bulk grain sampling for testing for GM presence [5] The EEP bulk sampling Standard Operating Procedure (SOP) is based on ISO 950, with samples taken from 5-11 points for lorries of 15-50t (sample size 1kg for rapeseed or 2kg for soya or maize). Samples from ships holds, docksides, stores or silos should be from lots up to 500t, and the suggested rate for flowing grain is 1kg/100t for cereals (wheat and barley) and double this for maize and soya. The 21

ISO 950 approach for static bulks is 10 samples (about 5kg) per 100t from different depth and positions. ISO 950 has since been superseded by ISO 13690. 2.2.9 Kernel Sampling Technique Evaluation (KeSTE) [10] / Kernel Lot Distribution Assessment ( KeLDA) The aim of the KeLDA project is to estimate the distribution of GM material in soybean kernel lots (i.e. whole soya) imported within EU Member States, evaluate currently adopted sampling strategies for the detection of GM material in lots of bulk raw materials and provide recommendations for implementing sampling strategies. The project has been initiated on the basis that in the past there has been a general oversimplification of the distribution of GM material (or other contaminants) in bulk lots. According to Lischer (2001) [29, 30] the degree of heterogeneity of a lot can be described as a scalar function, the homogeneity of which is zero. The current status of the project is that the sampling has been completed and the majority of analyses have been carried out. A trial sampling model has been developed and is available to participants on CD (Excel format). 2.3 GM testing methods 2.3.1 Commercial GM testing methods and procedures There is little information available on the testing procedures of commercial organisations involved in the analysis of GM plants. These organisations are in competition with each other and closely guard their methods and procedures. The methods and procedures used by the Central Science Laboratory GM testing service (http://www.csl.gov.uk/prodserv/ana/foodauthentication/foodauthentication.cfm#gm) provided much of the information and data required to mathematically model the testing process in this study. 2.3.2 Quality control of GM testing methods The literature contains many methods for the detection and more recently the quantification of GM plants. Methods applicable to the detection and quantification of GM OSR are referred to e.g. by Block et al (2003) [31], Demeke et al (2002) [32], Zeitler et al (2002) [33] and James et al (2003) [34]. These recommended methods 22

are all based upon polymerase chain reaction (PCR) analysis of GM DNA. PCR is a sensitive and robust technology that is particularly applicable to the detection of GM plants due to its versatility and rapid application. 2.3.2.1 Specificity The methods in the literature must, however, be used with caution. Recently PCR-based methods for the detection and quantification of OSR have been published [33], [35]. These methods, as reported, include OSR-specific PCR systems, however, when they were tested at CSL, it was found that they amplified numerous Brassica species, and were not therefore OSR specific (unpublished observations). This has particular implications for the detection and quantification of GM OSR where there is the possibility of mixed Brassica seed lots. Many Brassica family seeds look identical and cannot be distinguished visually from each other. Testing methods would not identify the proportion of OSR seeds within a mixed Brassica sample and could conceivably underestimate the quantity of GM OSR present. 2.3.2.2 Calibration standards PCR-based analyses must be controlled using standards against which results can be measured. To ensure quality and reproducibility, standards must be maintained across all analyses. Currently there are certified reference standards only for RoundUp Ready soya and BT11, Bt176 and MON810 maize produced and certified by the Institute for Reference Materials and Measurements (IRRM) (European Commission Joint Research Centre, Geel, Belgium). There are currently no certified reference standards for OSR, which means that it is not possible to compare methods used by different laboratories and monitor and maintain quality of the analysis. Additionally most laboratories will be unable to calculate a true limit of detection and will, therefore, be unable to determine the sensitivity and robustness of their methods. A certified reference material for OSR will be difficult to make due to the stability problems inherent in the oily meal produced when OSR seeds are milled. An alternative may, however, be produced in the form of DNA, either genomic DNA 23

isolated from OSR or, more likely, plasmid DNA. Plasmids are small circular pieces of DNA found in bacteria and can be produced on a large scale, stored for long periods of time without loss of quality and can be spiked into different DNA backgrounds. They can therefore be used as gold standards during GM analyses. A range of plasmids could be produced, covering the whole range of GM elements included in GM plants and the repertoire of plasmids could easily be extended to include new genetic elements. CSL is currently developing plasmid calibration standards under the remit of the Defra-funded project Plasmid Standards for Real Time PCR and UKAS Accreditation of GM Enforcement Testing. 2.4 EU recommended methodology and detection limits The following sections summarise and discuss publications that are part of, or have arisen from, European Commission regulations on GMO release and monitoring. In particular, how issues arising from these publications impact on the statistics of GMO detection methodologies for OSR. To date, only methods that use PCR to amplify and detect GM DNA are available for OSR. This review does not aim to examine the quantification of GM DNA as performed by real-time PCR or sequential seed bulk tests although the properties of these methods in GM detection are discussed. 2.4.1 Background EU legislation Two articles of EU legislation are of particular importance to statistical issues in the detection of OSR: Directive 2001/18/EC [36] details the procedures required in Member States for the authorisation of release and marketing of GMOs and their appropriate labelling. In conjunction with this directive, the Scientific Committee on Plants (SCP) published an opinion [37] which addresses several questions in the remit of the 2001/18/EC Committee. Most pertinent of these is the opinion on 'zero tolerance'in which the SCP states that for routine analysis and reasons of the limit of analytical sensitivity, the detectable threshold for GMOs should be 0.1%. Following this opinion, and sampling constraints as shown below, 0.1% has been adopted widely as the limit of detection (LOD) for analytical methods. 24

Regulation 1830/2003 [21] describes Member States obligations for traceability and labelling of GMOs. The regulation requires the development of EU guidance on sampling and detection (see section 2.2.5) [3]. The guidance gives sampling and analytical protocols for food products and grain commodities with particular reference to methods recommended by the European Commission s Joint Research Centre (JRC). Included are specific recommendations for sample sizes and procedures but it does not give minimum requirements for LOD or analytical method performance other than the requirement that the JRC should approve such methods. Minimum performance requirements for sampling and detection are therefore not set by this legislation. Specific EU legislation on GMOs in seeds does not yet exist. However, following the SCP opinion on GMO thresholds in seeds [37], a committee (Working Group on Seed Legislation, Sampling and Detection) has written draft recommendations for sampling and detection of GMOs in seeds. Currently these recommendations follow ISTA rules for sampling [2] and statistics [28]. ISTA rules do not currently specifically address adventitious presence of GM seed, however the rules and statistics applying to 'other seeds content'have been used as these satisfy the requirements for GM seed detection. Under the draft seed recommendations a minimum sample size of 3000 seeds is required. Following the binomial distribution, this ensures at least 95% probability that the sample will contain GM if the 'real'level in the lot is at least 0.1%. However, it should be noted that this does not mean that there is a 95% probability that the level in the 3000 seed sample is 0.1%, only that there is at least 1 GM seed in the sample. To ensure 95% confidence of detection in the sample, the analytical method should therefore display a 100% detection rate for the presence of 0.033% GM DNA (1 seed in 3000) or better. Different testing plans are under discussion for inclusion in the recommendations for sampling and detection of GMOs in seeds. As is shown in the following section, the choice of testing plan can have profound effects on the statistics of GM detection. 2.4.2 Theoretical detection limits of polymerase chain reaction (PCR) In theory PCR has the ability to detect one copy of a target GM DNA molecule and amplify it exponentially to give a strong positive signal. However, the occurrence of a single target in a PCR test is governed by binomial (or Poisson) probability [28]. In 25

order for a test PCR solution to contain at least one target copy of the GM DNA, the mean number of copies must be three for 95% confidence and five for 99% confidence. A common known variable in the PCR test is the total amount of DNA that is added to the reaction. Given the constant haploid genome size for OSR (one genome = 1.225 pg [38], the total number of haploid genomes in the PCR can be estimated. Table 7 shows the theoretical detection limit for a GM test PCR containing different amounts of test DNA. Table 7: Limits of Detection (LODs) for differing amounts of test DNA (with 95% confidence). The LODs are shown for a homozygous and heterozygous GM event. Theoretical and empirical LODs [39] are shown Total DNA amount in PCR 10 ng 25 ng 50 ng 100 ng Theoretical LOD, homozygous GM Theoretical LOD, heterozygous event Empirical LOD, homozygous GM Empirical LOD, heterozygous event 0.037 % 0.015 % 0.0075 % 0.0037 % 0.074 % 0.03 % 0.015 % 0.074 % 0.22 % 0.088 % 0.044 % 0.022 % 0.44 % 0.176 % 0.088 % 0.044 % Hughes and Totten [39] used a parametric regression of PCR test data to estimate PCR sensitivity. In their study, 95 % sensitivity (i.e. 95 % probability of a positive result when >1 copy of target is present) required a mean target copy number of approximately 18. This empirical result greatly increases the theoretical LOD for OSR as shown in Table 7. However, these LODs, while providing a useful guide, are unlikely to be applicable generally to all OSR GM tests. It is advisable that each LOD should be determined individually using thorough methods similar to those of Hughes and Totten [39]. This is because DNA extracts from different plant tissues will contain variable amounts of compounds that inhibit PCR. Different plant tissues will also contain variable amounts of organelle DNA (extranuclear DNA that does not contain GM material but can significantly contribute to the total amount of DNA extracted). 26

Using the LODs in Table 7 as a guide, it can be seen that PCR tests may often fail to detect the minimum content in a 3000 seed bulk (0.033 % if homozygous or 0.0165 % if heterozygous). The 3000 seed bulk size is recommended in EC food/feed and seeds testing guidance [3] for test plans that use real-time PCR techniques. The use of smaller bulk sizes and increased numbers of PCR tests, such as the recommended sequential seed tests would therefore be advisable to increase the likelihood of detection of GM at low levels. 2.4.3 EU recommended methods for PCR tests The EU guidance for food and feed GM testing recommends using the JRC methods database [4]. For OSR, 13 methods are listed [31], [32], [33], [34] (Table 8). To date none of these methods have been validated to international criteria or evaluated in ring-trials. The methods detect a variety of GM elements, either individually or in multiplex, but none are event-specific, i.e. identify a unique junction of plant and GM DNA. Table 8: JRC listed methods for GM detection in OSR. For definitions of GM elements, see BATS report [40]. Method, Reference Matrix type GM element LOD Mass of DNA used in PCR (ng) 1. Block et al [31] seed P35S-BAR 2.8-21 copies - 2. Block et al [31] seed P35S-BAR - 50 3. Demeke et al [32] seed CTP2 0.1 % 50 4. Demeke et al [32] seed EPSPS 0.1 % 50 5. Demeke et al [32] seed P35S 0.1 % 50 6. Demeke et al [32] seed Pnos 0.1 % 50 7. Demeke et al [32] seed BXN 0.1 % 50 8. Demeke et al [32] seed multiplex 0.5 % 50 9. Zeitler et al [33] seed EPSPS 0.05 % - 10. Zeitler et al [33] seed PAT 0.05 % - 11. Zeitler et al [33] seed P35S 0.05 % - 12. James et al [34] leaf multiplex - 75 13. James et al [34] leaf NPT - 75 Variable LODs are reported for the methods in Table 8. Block et al. [31] give only a theoretical LOD in terms of estimated target copy numbers in the PCR. Other methods report the lowest GM level tested which gave positive results as the LOD. It can be seen that multiplex methods i.e. those where more than one GM element is 27

targeted in a single PCR, have a higher LOD. Where seed was the test matrix, the methods used mixtures of known proportions of GM and non-gm seed. Currently no certified reference material (CRM) is available for any OSR. Researchers have therefore relied on GM seed provided by the producer. The high oil content of OSR seed has made attempts to make stable CRMs difficult (IRMM, personal communication). Unlike other GM seeds (maize and soya), OSR cannot be easily dried or frozen once it has been ground into mixtures of known GM content. It is likely therefore that any future CRM for OSR will take the form of 100% GM seed or DNA based standards that can be 'spiked'into controls. See also paragraph 2.3.2.2 above. In order to distinguish real negative PCR results from PCR failure due to presence of inhibitors or experimental error, the above methods include control reactions that amplify native OSR genes (endogenous controls). These control tests can be performed separately or, for increased confidence, in the same reaction as the test PCR. For these endogenous controls to be accurate, the targeted OSR gene or anonymous DNA sequence should be present in the same or similar copy number as the target GM. This ensures that the endogenous control and the GM target have an equal probability of amplification. Block et al. used S-glucosyltransferase as their endogenous control, which they reported to be present in at least two copies. Demeke et al. used the mitochondrial gene Nicotinamide Adenine Dinucleotide (NAD), which is likely to be present in very high copy numbers (>1000). Zeitler et al. used the phosphenolpyruvate carboxylase gene, which has an unknown copy number in OSR, but it is likely to have several copies due to its ubiquitous cytoplasmic functions in plants. 2.5 Review of seed audits A review of data from the GM Inspectorate s seed audits of oilseed rape revealed that most laboratories that undertake testing for GMOs on behalf of seed companies give very little general information on their testing protocols and almost nothing on the statistical aspects of the tests. The majority of laboratories quote a limit of detection (usually 0.1%) although some will quote a limit for quantification, e.g. 0.1%, with a lower limit for detection, e.g. 28

0.01%. Some laboratories state that sampling and testing are in line with the guidance as issued by the GMI, although sampling is usually the responsibility of the seed company, who will usually quote ISTA rules and/or use a sub-sample of seed taken for seed certification purposes. Some laboratories also quote the size of the "initial" sample and the "test" sample. These terms, however, are undefined and the initial sample varied in size from around 50g to 2kg, whilst the test sample was commonly given as 2 1g. It is usually unclear whether duplicates were made before or after grinding of the seed. Two laboratories have stated that their PCR reaction is guaranteed to include >20,000 genomes and has a detection limit of 20 GM genomes. One laboratory has clearly stated that, for each test, 5 replicates of 1000 seeds were ground separately and has quoted detection/confidence limits of 0.06% GM with 95% confidence and 0.09% GM with 99% confidence. Only one company has provided a comprehensive protocol, stated compliance with a recognised standard [41], and is regularly involved in ring tests. A further company has published a press release detailing excellent results in a proficiency test run by the United States Department of Agriculture for detection of GM maize and soya but presents no results for oilseed rape. Finally, none of the certificates presented clearly state how the % GM is calculated (i.e. relative to genome or mass). As the majority of tests are real-time PCR, it must be presumed that units are percent GM genomes/total homologous genomes but this remains unclear. 29

3. DATA COLLATION AND ANALYSIS A database of recent test results for oil seed rape has been created by collating the anonymous data available from PCR tests undertaken by the Central Science Laboratory GM testing service. This database allows the import of PCR test electrophoresis worksheets and, in order to gain information on the frequency of control failures and replicate disagreements, also includes results of repeated tests where initial tests did not meet quality assurance standards. The data are stored in a secure Microsoft SQL Server 2000 database at Central Science Laboratory, York and commercial confidentiality is strictly enforced. A statistical analysis of the results of the positive controls and test samples has been performed, to address the following four questions: 1. Is there a difference in performance between before and after extraction methods were changed? 2. Do some primer / positive control sets perform better than others? 3. Is there any change in performance over time? 4. What is the estimated rate of false negative results for unknown samples? 3.1 Database description To access the database, a graphical user interface application was developed in Java (Figure 2, below). Single or multiple batches (here batch corresponds to an electrophoresis worksheet) can be imported into the database from worksheets that have previously been saved in Microsoft Excel format. Data are imported into the relational database and stored in a number of linked tables, which can be viewed in the application through selecting from either the appropriate toolbar buttons or file menus: i. Batch table retrieves information relating to an electrophoresis worksheet e.g. PCR primer name, how many duplicate samples were run, number of positive and negative controls, date batch was run, etc.. 30

ii. Selecting Negative Control or Positive Control tables retrieves information relating to controls e.g. PCR blank, extraction blanks, positive reference name, results of controls, etc.. iii. Selecting the Samples table retrieves information relating to a particular sample e.g. the I.D. of the sample, results and type of sample test material e.g. seed, grown-on and leaf, etc.. Figure 2: Screenshot of the graphical user interface developed to access/view the data held within the database. 31

Summary statistics of false positives control results, false negative control results, and replicate disagreements are generated and updated automatically as data are entered into the database; these can be viewed in a separate table (Figure 3). Figure 3: Screenshot of statistical results generated automatically from database. All the table views can be sorted in order by clicking the header of each column within the table view and saved by selecting the appropriate toolbar buttons or file menus. A copy of the original electrophoresis worksheet can be generated from data held within the database and displayed within the application (Figure 2) by selecting the batch in the table view and clicking the retrieve button. The software comes with full in-built searchable context-sensitive help, which is accessed from the help menu (Figure 4). 32

Figure 4: Screenshot of help systems built into the database software. 3.2 Statistical analysis of database The use of PCR to detect and quantify the presence of GM in seeds is well established. Each reaction must be accompanied by a set of appropriate controls by which the analyses can be verified. However, the verification of results is hampered by the lack of suitable positive controls because OSR-based certified reference materials are yet to be produced. At CSL, positive control material has historically been derived from a variety of sources including GM OSR provided by commercial companies, OSR samples which have previously been the subject of analyses and have been found to be positive for a particular trait, other non-osr samples which have been found to be positive for a particular trait, and certified reference material for other GM plants. By combining all of the above, there has been comprehensive coverage of the range of GM OSR promoters, terminators and traits for the detection 33

of all known GM OSR lines to date. However, the strategy of using samples of unknown provenance can be problematic because the level of GM in these samples is unknown and can, in theory, be very low and possibly below the limit of detection. The presence of GM material can sometimes be detected where it is present at quantities below the limit of detection, but subsequent analyses of such control materials may produce a high proportion of negative responses. Therefore, when repeating analyses using these samples as positive controls, there may be an increased chance of obtaining a quality assurance failure for the whole test run even when the results generated by the unknown samples may be correct. This part of the study was designed to look at the range of positive control samples used in order to determine their efficacy. The method for DNA extraction from oilseed rapeseeds was changed in November 2002 in light of new statistical analysis. Prior to that date approximately 1000 seeds were used for DNA extraction. However, this would not allow a theoretical detection limit (with 95% confidence) of 0.1%. After November 2002 the number of seeds extracted was increased to 3000, therefore giving a theoretical limit of detection of 0.1%. At the same time the method of addition of the DNA was altered; previously 1µl of DNA extract was added per PCR analysis. However, pipetting so small a volume could be prone to inaccuracy so the DNA extracts were diluted and the volume pipetted increased to give the same amount of DNA extract to be added to the PCR. The effect of these changes on the performance of positive controls has been assessed together with the performance of the controls over time and the individual performance of primer and positive control sets. 3.2.1 Data description The information provided on positive controls in the database includes an identifier for the positive control, the date and result of the analysis, and the batch number. Through the batch number, the positive control can be associated with the samples tested alongside the control and the primer set used. There is no direct information in the database about the extraction date but this can be inferred from other information. Several of the positive control extracts have been initially produced from test samples. For these the time of first analysis has been used as an estimate of the extraction date. This date could be determined for roughly half of the samples. For the remaining 34

samples the laboratory notebook was consulted. Some samples were extracted at several different times. These samples were treated as a separate group in addition to the classification into original and revised extraction methods, because for these samples no long-term decline is expected. This group is referred to as multiple extractions. There are 696 entries for positive controls in the database, which cover the time from 22/07/2002 to 20/09/2004, and refer to 47 control samples. Four samples of these have been excluded because it was not possible to identify from the records now available which extraction method was used. The data used in the statistical analyses are summarised in Table 9. Table 9: Summary of the number of control samples and the number of related analyses in the database for each extraction method used in the statistical analyses. Extraction method Number of samples in the database Number of samples used on more than 1 day (used for comparison of extraction method) Number of samples used for longer than 4 months (used for detecting a time trend) Multiple 8 samples (178 analyses) 4 samples (172 analyses) 4 samples (172 analyses) Original Revised Unknown 31 samples (462 analyses) 4 samples (36 analyses) 4 samples (5 analyses) 22 samples (446 analyses) 13 samples (315 analyses) 3 samples 1 sample (34 analyses) (9 analyses) - - 3.3 Results of statistical analysis 3.3.1 Is there a difference in performance between before and after extraction methods were changed? For this analysis all single extraction samples, which were used on more than one day, have been used for the analysis. In the boxplot below (Figure 5) the results for multiple extraction samples are also included for reference. 35

Figure 5: Comparison of the performance of positive control samples between different DNA extraction methods, with the performance of each sample measured as the percentage of analyses with positive results. Percentage of analyses with positive results 0 20 40 60 80 100 Original method Revised method Multiple extractions The data have been tested using logistic regression. No statistically significant difference between the two extraction methods was detected (F 1,23 = 1.85, p>0.05). The model-derived prediction for the correct identification rate was 89.9 % (S.E. 3.1 %) for the original extraction method and 70.6 % (S.E. 17.1 %) for the revised extraction method. It should be noted that there are far fewer samples for the revised extraction method (3 samples based on 34 runs) than for the original extraction method (22 samples). This is reflected in the much larger Standard Error (S.E.) for the performance estimate for the revised method. This analysis will be repeated when further data from the revised method become available. 36

3.3.2 Do some primer/ positive control sets perform better than others? A second logistic regression has been fitted for the primer and positive control set. Many of the positive controls contain a limited number of GM elements and are therefore closely related to the primer sets with which they are used. Table 10 below illustrates the number of analyses in which the positive control was correctly identified. Table 10: Number of analyses in which the positive control was correctly identified, according to the primer set used. Total number of results Number of positive results Percentage correct classification 37 Standard error (derived from logistic regression estimates) Bar 54 51 94.44 5.165 Barnase 12 12 100.00 0.002 Barstar 57 51 90.00 6.417 CmoVB 49 49 100.00 0.001 CRT 66 51 77.27 8.548 E93 44 44 100.00 0.001 GOX 47 45 95.74 4.879 NptII 76 62 81.58 7.369 OSR 56 52 92.86 5.703 p35s 61 52 85.25 7.524 Pnos 71 55 77.46 8.217 Tnos 56 54 96.42 4.109 No statistically significant differences were found between primers (F 11,81 =1.62, p>0.05), between methods (F 1,81 =1.97, p>0.05) and also no significant interaction of method with primer (F 3,78 = 2.02, p>0.05). This analysis should be interpreted with caution since the sample size in some of the groups is very low and the large sample approximation of the test statistic may not therefore be valid. Again, this analysis should be repeated when more data from the new method are available. A third analysis has been performed on the complete dataset that includes samples with multiple extraction information. To avoid a large number of empty or nearly empty classes, information about extraction methods has been ignored. Differences between primers are not statistically significant (F 1,23 = 1.79, p>0.05).

Despite the apparent lack of statistically significant differences in this data, the results are still consistent with the current opinion in the GM Testing Service that some positive control materials perform better than others (Figure 6). The positive control material used for E93, CMoVB and GOX is the same. Their combined percentage of correct identification is 98.6%, whereas the positive controls used for CRT or pnos are only 77.3 and 77.5% respectively. Note that the positive control material used for E93, CMoVB and GOX came from a sample sent for analysis by CSL UKAS-accredited 1 GM testing service and is not oilseed rape. Figure 6: Comparison of the performance of positive control samples between different primers, with the performance of each sample-primer combination measured as the percentage of analyses with positive results. Percentage of analyses with positive results 0 20 40 60 80 100 Bar Barnase Barstar CMoVB CRT E93 GOX NptII OSR p35s pnos tnos Primer 1 UKAS - the United Kingdom Accreditation Service. 38

These results highlight the need for effective positive controls, which will allow analyses to be conducted with the near certainty (certainly above 90%) that the positive controls will perform satisfactorily. The positive controls most urgently needed are for CRT and pnos. These could be provided under the remit of the Defra project Plasmid Standards for Real Time PCR and UKAS Accreditation of GM Enforcement Testing. Better use of alternative GM plant certified reference materials, specifically for NPTII and p35s, should lead to a lower proportion of quality assurance failures. 3.3.3 Is there any change in performance over time? The analysis of control samples gives a binary result: detected / not detected. Hence, to compare performance over time, data have to be grouped over appropriate time intervals. This time interval has to be large enough to include sufficient numbers of runs within each interval to estimate the success probability, whilst at the same time the intervals need to be small enough to have a sufficient number of time intervals to estimate change over time. Hence, the results were grouped into 4-month intervals, starting at the first time of use since the extraction date is not available for all samples. Table 11 gives an overview of the results and Figure 7 shows the percentage of runs with positive results over time. 2 2 Apart from sample B201 all values of 0% and 50% are failures from at most two runs. 39

Table 11: Results for control samples grouped into 4-month periods (number of positive analyses / number of negative analyses in each 4-month period). Group / Sample Original extraction method 4-month periods after first use 1 2 3 4 5 6 7 +ve/-ve +ve/-ve +ve/-ve +ve/-ve +ve/-ve +ve/-ve +ve/-ve A291 4/0 1/0 B9 87/5 25/5 4/0 B10 2/0 3/0 B62 12/0 12/2 B101 1/0 9/0 B121 1/0 1/1 0/2 B122 33/2 4/0 B129 1/1 1/0 5/0 6/2 B171 7/1 6/4 B177 2/0 1/0 B201 2/9 0/4 1/1 B279 8/0 10/0 1/0 5/0 B281 1/0 24/1 Revised extraction method B362 3/3 2/1 Multiple extractions 0.1% GM soya 5/0 5/0 5/0 0.5% GM maize 4/1 1/0 0.5% GM soya 1/0 2/1 10/2 16/0 B9 Wheat 21/4 6/2 4/0 12/1 26/0 8/1 30/4 40

Figure 7: The performance of each positive control sample over time, grouped into 4-month intervals since first use as control. 1 2 3 4 5 6 7 Percentage of analyses with positive results 100 80 60 40 20 0 100 80 60 Multiple extractions AC AB AD CD D D D C B D C Original extraction method GO QR HK EJ GM QE LJ MQ F Q O FL R N H F M N M K P D S 0.1% GM soya 0.5% GM maize 0.5% GM soya B9 Wheat A291 B9 B10 B62 B101 B121 B122 B129 B171 B177 B201 B279 B281 B362 Revised extraction method S A B C D E F G H J K L M N O P Q R S 100 80 60 40 40 P 20 20 0 P K 0 1 2 3 4 5 6 7 1 2 3 4 5 6 7 Time period after first use (in 4-month intervals) No indication of any decline in performance for the positive controls in either of the groups was observed. Unfortunately, most samples were used for only a short time, which reduces the (statistical) sample size compared to the previous (statistical) analyses. In particular, for the revised extraction method only few data are available for the time trend analysis (1 sample). No formal statistical analysis has been performed due to the limited sample size and the clear lack of a visible trend; also, there were not enough data to split this analysis by primer. 41

With the future availability of the database it will become possible to monitor this aspect of performance on a regular basis. When more entries for the new extraction method become available in the database, the analysis will be repeated. 3.3.4 What is the estimated false negative detection rate for unknown samples? An estimate of the false negative detection rate for unknown samples was calculated from the within-run accordance 3 displayed by the measurement results. The false positive detection rate was much lower than the false negative detection rate (Figure 3), hence the false negative rate could be estimated as the proportion of true-positive samples that could be expected to give a negative/negative upon duplicate analyses by assuming that all positive results came from positive samples. Hence, the following estimates for the probability that a positive sample gives a (single) positive result (p) and the probability that a positive sample gives one positive and one negative result (pq) were gained from the results shown in Table 12. Table 12: Within-run agreement for samples Results of duplicate analyses Count Negative / Negative 1088 Positive / Negative 120 Positive / Positive 357 p 2 357 = 477 + X 120 2 pq = 477 + X where X is the unknown number of positive samples that gave a negative response 357 q 60 p = 0, p + q = 1 q = 60 417 2 q 0. 02 3 Accordance is the (percentage) probability that two identical test materials analysed by the same laboratory under repeatability conditions will both give the same result (i.e. both found positive or both found negative) 42

Hence, by assuming that the variation between duplicate results that disagree is representative of the variation associated with all measurements, an estimated false negative response rate of 2% was calculated. While this figure was reassuringly low, it could not be used to model such quantities as limit of detection, or the probability that a particular sample would generate a false negative response because no information about the quantity of GM material in samples used to generate the estimate is available. It is necessary to consider the quantitative aspects of the detection method (such as mass of DNA extracted from seeds, number of gene copies required to ensure detection, etc.) in order to provide the information necessary to reliably estimate the probability of detecting a given quantity of GM material. 43

4. SIMULATION OF SAMPLING AND ANALYSIS OF GM SEEDS Modelling of the sampling and testing process in its entirety as described here has not previously been attempted. The sampling process alone has been previously modelled by the project KeSTE [9]. In this model, the effect of lot heterogeneity on the sample variance was investigated, but the probability of obtaining a correct test result was not explored. Remund et al.[42] investigated the operating characteristics of testing systems for seed purity, but addressed only the quantification method using seed pools and not the sampling error. Their program, SeedCalc (see http://www.seedquest.com/best/spreadsheet) does permit the use of an empirically estimated global (i.e. not varying with proportion of positive seeds in lot) false negative and global false positive rate. 4.1 Model Design The computer model developed in this study was written in Microsoft Excel VBA. The model was designed to be as accessible as possible and was, therefore, written in a language and output format available to most PC users; it does not require excessive processor time or a computer more powerful than a desktop PC. The first part of the model simulates incremental sampling of seeds from a heterogeneous lot and the production of laboratory and analytical samples. The second part simulates the analysis (by PCR) of the sample produced by the first part of the model. A schematic representation of the whole model is shown in Figure 8. For a well-mixed, i.e. homogenous population, the binomial distribution describes the sampling process well and standard sampling theory can be used to determine the sample size required to detect at least one positive unit for a given proportion of positive units in the target population with a given level of confidence. In the case of GM sampling, a homogenous sample means that the GM is well spread throughout the sample 4 [43]. If it is expected that the proportion of GM could vary with location 4 A lot is homogenous relative to a given characteristic is uniformly distributed according to a given probability law throughout the lot. A lot is heterogeneous relative to a given characteristic if the characteristic is NOT uniformly distributed throughout the lot. 44

in the lot, then the assumptions for binomial sampling are violated and the lot must be treated as heterogeneous. 45

1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0 0 0 500 500 1000 1000 1500 1500 2000 2000 2500 2500 3000 3000 3500 3500 Figure 8. Schema of model. Model of lot Mass of each increment; Mass of 1000 seeds; Mean proportion GM; Heterogeneity Monte Carlo Number of increments Working Sample Number of GM seeds; Total number of seeds Mass of laboratory sample Laboratory Sample Number of GM seeds; Total number of seeds Replicate analytical sample Mass of analytical sample Analytical Sample Replicate extraction Number of GM seeds; Total number of seeds Mass of genome; Mass DNA in PCR; Extraction variation Analytical Extract for each PCR replicate from each extract for each extraction replicate from each analytical sample for each replicate analytical sample Replicate PCR Expected number of GM copies in Extract PCR Solution Number of GM copies in PCR PCR GM detected? Number of GM copies required for 95% detection rate Model Outputs Proportion detected; Proportion of analytical samples containing 1 GM seed; PCR accordance; Extraction Accordance; Analytical sample accordance 46

In order to describe how the heterogeneity of the presence of GM seeds affects the distribution of GM seeds in increments, and hence the probability that the presence of GM seeds is detected, it is necessary to incorporate a model for the form of heterogeneity. While there may be a large number of possible models, for this study two forms of model have been considered that are based on commonly used models for heterogeneity in similar contexts. These are: a stratified-binomial model, and a beta-binomial model. 4.1.1 Stratified-binomial model for heterogeneity This is a very simple model that divides a seed lot into two distinct parts and is defined by the mean proportion in the lot and the proportion of the lot that contains all the GM seeds (a portion s, the hotspot, of the lot contains all the GM seeds in the whole lot, and the remaining 1-s of the lot does not contain any GM seeds). If the lot contains a proportion p of GM seeds, then the probability that a seed taken from at random from the hotspot is GM is p s. The probability that a seed is GM when taken at random from outside the hotspot is zero. Increments that contain seeds taken from both within the hotspot and from the remainder of the lot (i.e. across the boundary of the hotspot) can contain a proportion of GM seeds constrained to lie between zero and p s. 4.1.2 Beta-binomial model for heterogeneity A statistical distribution often used to model sampling from heterogeneous populations is the beta-binomial distribution. The beta-binomial distribution can be regarded as a mixture of binomial distributions each with mean proportions varying according to the beta distribution. The statistical parameters of the beta-binomial distribution are the sample size, the overall proportion in the population with a positive response (here GM seed content), and the variance of this proportion within the population. However, a parameterisation that is analogous to that of the stratified-binomial model is used in this study to communicate the degree of heterogeneity in a more intuitive way. The beta-binomial model is parameterised here by the mean proportion in the lot and the proportion of the lot that contains 95% of the GM seeds. Let a proportion p of GM seeds be present in a lot, and the GM seed be heterogeneously distributed such that a proportion y of the total GM seed is in a 47

proportion s of the lot (y>s for a heterogeneous lot). For example, in a very heterogeneous lot 95% of the GM seeds might be present in only 10% of the lot; for a less heterogeneous lot 95% of the GM seeds might be present in 70% of the lot; in a perfectly homogenous lot 95% of the GM seeds would be present in 95% of the lot. Then, the beta distribution describing the lot heterogeneity can be parameterised with the shape parameters ν and ω (ν > 0, ω > 0) simultaneously satisfying Equation 1 and Equation 2. ( x, ω ) dx = s 1 Betadist ν, Equation 1 q ( x, ν, ω) dx = yp 1 x Betadist Equation 2 q Where and Betadist ( x, ν, ω ) ν 1 x = Beta ω ( 1 x) ( ν, ω ) 1 ( ν ) Γ( ω ) ( ν + ω ) 1 ω 1 ν Γ Beta ( ν, ω ) = u 1 ( 1 u) =. 0 Γ For a fixed value of p, the parameterisation is achieved by finding the appropriate value of the variance associated with p (var(p)) that satisfies Equations 1 and 2 via Equation 3. ( 1 p) ( p) 2 p v = p, ω = var v p v Equation 3 [44] For an increment of n seeds, selected from a random location in a lot, the probability of the increment containing k GM seeds is then given by the beta-binomial distribution with the shape parameters ν and ω determined by solving Equations 1 and 2. p ( k) n Γ = k n n! where = k k!( n k)! ( v + ω ) Γ( v + k) Γ( ω + n k) Γ( v) Γ( ω ) Γ( v + ω + n) Equation 4 [45] 48

Hence, a cumulative probability distribution can be quickly calculated for the number of GM seeds in an increment that has been randomly selected from a lot by calculating logarithms of the terms in Equation 4 for values of k from 0 to n. The production of a bulk working sample made up of m increments, each containing n seeds, taken from random locations in a lot can then be simulated by the taking the sum of m values randomly selected from the beta-binomial cumulative distribution to give the number of GM seeds in the bulk sample (containing a total of m n seeds). If each incremental sample is distributed according to the beta-binomial, the percentage of GM seeds in the working sample can then be described as a sum of beta-binomially distributed random variables. Here simulation methods are used for inference. 4.1.3 Sampling from a working sample Sampling from the working sample can be described by the hypergeometric distribution, since the working sample is well mixed and can be expected to be homogenous; the binomial distribution is not used because sub-samples remove a significant proportion of the working sample. Thereafter any further sub-sampling can also be simulated using the hypergeometric distribution (with the realistic assumption that samples are properly mixed prior to sub-sampling as is universal practice). 4.1.4 Features of the stratified-binomial and beta-binomial models Figure 9 shows inverse cumulative distribution functions for the probability p that a grain selected at random from an increment is GM. The distribution functions describe a heterogeneous lot containing 0.1% GM seeds; for the stratified-binomial distribution s=0.6 (i.e. GM seeds are contained within 60% of the lot), for the betabinomial distribution s=0.6 and y=0.95 (i.e. 95% of the GM seeds present are contained within 60% of the lot. 49

Figure 9: Inverse cumulative distribution functions for stratified-binomial and betabinomial models of incremental (of mass i) sampling from a lot (of mass l) containing 0.1% ( p )of GM seeds in a hotpsot that consists of 60% (s) of the volume of a lot. The stratified-binomial model is characterised by sharp steps. A proportion s p (s p <s) of increments are taken from a volume of the lot where p = p s, a proportion 1-s 0 (1- s 0 <1-s) are taken from a volume of the lot where p = 0, and a proportion s 0 -s p of increments include seeds from inside and outside the hotspot. For a 3-dimensional lot: s 0 s p = k s 2 i l 1 / 3 where l is the mass of the lot i is the mass of an increment s is the proportion of the lot containing the GM seeds k is a constant (for spherical increments taken from a spherical lot k 6) For example, for a certified seed lot of 10 tonnes containing GM seed within 60% of its volume from which increments of 0.1 kg are taken, 91% of model increments will be taken from portions of the lot where p = p s or p = 0 and the remaining 9% of 50

model increments will include seed from portions of the lot where p = p s and seeds from portions of the lot where p = 0 (Figure 10). Figure 10: Section through a stratified-binomial model of a spherical lot (10 tonnes) showing the portions of the lot where increments (0.1 kg) are taken entirely from volumes where p=0, increments are taken entirely from volumes where p = p s, and where increments include contributions from both portions of lot. The beta-binomial model gives a smooth cumulative distribution function for p that extends beyond p s. Hence, the proportion of GM seeds near the middle of a beta-binomial model hotspot is higher than near the edge of the hotspot. Also there is no sharp edge to the hotspot. It is hard to imagine physical circumstances in which there would be no mixing at all in the course of handling and transport of seed so we consider it is more realistic to use a model that has no sharp discontinuities. 4.1.5 Comparing beta-binomial and stratified binomial models to a physical model for diffusion of GM seeds into a lot One plausible physical model for the distribution of GM seeds in a lot that has not been deliberately put through a process of mixing is that of diffusion. This physical model assumes that a discrete body of GM seeds is present at harvest, or introduced to a lot, and that this body undergoes diffusion into the lot as the result of many small 51

disturbances. The density of GM seeds at a point in a three-dimensional lot can then be described by the spherical Gaussian distribution. For the spherical Gaussian distribution with 95% of GM seeds contained within 60% of a (spherical) lot with mean proportion of 0.1% GM seeds, 1% of increments would be taken from a portion of the lot where p>0.0068. For the beta-binomial model 1% of increments are taken from a portion of the lot where p>0.0057. Whereas, for the stratified-binomial model, the whole hotspot has p=0.0017 and higher values are not possible unlike the other models postulated. Hence, both the stratified and beta-binomial models could be used to give a model hotspot of the right size, but only the beta-binomial gives a realistic representation of the variation that may be expected within hotpots and across their boundaries and hence was the model used in this study. 4.1.6 Simulating the detection of GM seed in an analytical sample Where the analytical sample (usually consisting of at least 3000 seeds) is finely ground and thoroughly homogenised, it can then be considered (from the point of view of the very large number of particles present) as equivalent to a solution of GM DNA in unmodified DNA. The processes of extraction, clean-up, and dilution leading to the extract, from which the final aliquot presented to the PCR reaction for detection is taken, can then be modelled as the transfer of a portion of the solution from the homogenised analytical sample to the extract. The outcome of these combined processes is expressed here as the expected mass of DNA in the final extract and the variation associated with mass of DNA (and hence the expected number of GM copies and the variation associated with expected number of GM copies). The usual model for the variation associated with such processes is the Gaussian distribution [46]. The final sampling process in the analytical method, that of taking a small aliquot from the extract for PCR involving the sampling of a relatively small number (10 1000) GM copies from the extract, gives a simulated actual number of GM copies presented to the PCR. The variation associated with this kind of process is usually described by the Poisson distribution. Finally, the probability of detecting that number of GM copies, as presented to the PCR, can be calculated from the detection limit of the PCR, expressed as the number of copies required to give a 95% probability of detection. 52

For an analytical method that begins with the homogenisation of an analytical sample of n s seeds, containing n g modified seeds, with extraction and dilution steps leading, on average, to a final PCR solution containing a mass m PCR of DNA, then the production of the extract can be simulated by taking a random value from the Gaussian distribution: ( Mean RSD) E = RandomGaussian, Equation 5 where ng mpcr Mean = Equation 6 n m s gene and RSD given by the Relative Standard Deviation observed for the mass of DNA in the PCR extract for the extraction method employed. The number of copies of modified genes in the PCR aliquot is then simulated by taking a random value from the Poisson distribution with expected value E ( E) N = RandomPoisson. Equation 7 Finally, detection is simulated by taking a random number from the uniform distribution [0,1] and comparing it to the probability of detection P: GM detected if RandomUniform[ ] P Where N N ( 1 0. ) LOD 0,1 Equation 8 P = 1 95 Equation 9 and N LOD is the number of gene copies required for produce a positive result for 95% of PCR reactions. 4.1.7 Analytical replication in the model Analytical replication can improve the detection ability of a method if it is used appropriately. Three hierarchical levels of replication are examined in the model: i) Replicate PCR. A single analytical sample is homogenised and a single extract is produced from the homogenised sample. Replicate aliquots are taken for PCR. This is simulated in the model by taking replicate random samples from 53

the Poisson distribution (Equation 7) and passing each value to the detection test (Equation 8). (This is the lowest level in the hierarchy of replication.) ii) iii) Replicate Extraction. A single analytical sample is homogenised. The homogenised sample is split and replicate extractions are performed. This is simulated by taking replicate random samples from the Gaussian distribution (Equation 5) and passing the values to the Poisson sampling and detection Equations (Equations 7 and 8). Replicate analytical sample. Successive analytical samples are taken from the laboratory sample (without replacement) and analysed separately. This is simulated by sampling from a laboratory sample (without replacement) to give replicate values for the parameters n g and n s in Equation 6, which are used in the simulation of DNA extraction (Equation 5). Hence, separate sets of parameters are passed to the Poisson sampling and detection equations (Equations 7 and 8). (This is the highest level in the hierarchy of replication considered here.) 4.2 Outputs from the model Estimates for the values of five quantities are produced by the model: probability of detection, probability of analytical sample containing at least one seed, replicate PCR accordance (probability that duplicate analyses give the same qualitative result) [47] extraction accordance, and analytical sample replicate accordance. The value of these output parameters can be used to decide whether the performance of the measurement system is fit for purpose, and to provide information on the best way to improve the measurement system. For example, if the probability of detection is too low, and the probability that an analytical sample contains at least one seed is very similar to the probability of detection, then the simplest route to improve the method is through increasing analytical sample size (if possible), or the number of increments (if affordable). If the probability of detection is much lower than the proportion of analytical samples containing at least one seed, then performance could be improved by increasing the 54

number of replicate analyses using the lowest, and hence least expensive, level of replication (PCR<extraction<analytical) that displayed a low accordance. 4.3 Input values for model parameters Each simulation of the sampling and analytical process requires that values be supplied for 16 parameters. These fall into four categories: commodity, lot, sampling and analysis. In each case, the values or ranges used here were chosen from the most up-to-date and relevant sources that could be found within the scope of this study. The model, however, allows for different parameter values to be readily explored should users require. Tables 13 and 14 show a summary of fixed and variable parameters respectively, their type, variability, and values or ranges used. Table 13: Summary of simulation model parameters that are constant across scenarios. Type Parameter Value / Range Commodity Lot Analytical seed mass haploid genome mass 4 5 g /1000 seeds 1.225 pg proportion GM 0.1 to 0.9 % heterogeneity 60 % mass of analytical sample 12 g PCR DNA mass 50 ng PCR DNA variability RSD 20% replicate analytical samples 1 6 usually 1 replicate DNA extractions 1 4 usually 1 replicate 1 4 usually 1 55

Table 14: Summary of simulation model sampling parameters that vary with scenarios. Scenario Parameter Value / Range Seed lot Grain As grown number of increments 5 30 mass of increments 100 750 g mass of lab sample 50 100 g mass analytical sample 12 g number of increments 10 100 mass of increments 500 g mass of lab sample 1000 2000 g mass analytical sample 12 g number of increments 5-30 mass of increments 100 750 g mass of lab sample 50 100 g mass analytical sample 12 g 4.3.1 Commodity parameters The commodity parameters used in this study to represent OSR were seed mass and mass of haploid genome. Seed mass must be considered because sampling standards (e.g. ISO methods) determine sample increment size in terms of mass, not number of seeds. The number of seeds sampled will therefore vary depending on the commodity. The haploid genome size determines the absolute number of DNA copies that are sampled for a particular sample mass and/or amount of DNA in the final PCR test, i.e. the larger the genome size, the fewer copies of target DNA are sampled per gram of sample or per nanogram of DNA in the PCR. The minimum unit of genome size is the haploid genome, consisting of one copy of each of the normal complement of chromosomes for the species studied. The haploid genome of a single copy GM event will, therefore, contain one copy of the target DNA molecule. However, it is also important to note that one such target copy is a double helix and actually provides two single stranded DNA targets for initiation of the PCR reaction. The most recent and comprehensive survey of crop plant genome sizes is Arumuganathan & Earle (1991) [38], which cites the range 2.34-2.56 pg per diploid cell for oilseed rape. The haploid mean, 1.225 pg, was used in the model. 56

4.3.2 Lot parameters The lot parameters describe the amount and distribution of GM seeds in the sampled population. The size of the population is not relevant where incremental samples are very small compared with the size of a lot. Two lot parameters are used. The first is the true proportion of GM of the lot in terms of % GM DNA (as defined in [48], which here encompasses the range 0.1 to 0.9% GM seeds: 0.1% being the practical detection limit quoted by many testing methods and the threshold for unauthorised GM presence; 0.3% being the currently proposed EU interim threshold for GM in seeds; and 0.9% being the EU threshold for GM in food / feed (including grain). The second lot parameter is heterogeneity, which is expressed in this study as the proportion of the lot that contains 95% of the total GM. The values used were based on estimates from FSE gene flow data [49], applicable to as grown and, as a worst case, to seed lots, and the KeLDA project [10], applicable to grain as a commodity. KeLDA and FSE data exhibited approximately 60% heterogeneity (i.e. 60% of the total mass contained 95% of the total GM). (Note although the KeLDA data are for soya, they are the only measurements of heterogeneity with respect to GM in any commodity that have ever been made to date.) 4.3.3 Sampling parameters The sampling parameters were determined by the range in number of sample increments, their mass and treatment (e.g. mixing, sub-sampling) as published in sampling standards (ISO [1], ISTA [2], EU recommendations [3]). The sampling process was treated as the process of obtaining the analytical sample from the lot. 4.3.4 Analytical parameters The analytical parameters pertain to the extraction of DNA and PCR testing process after the analytical sample has been obtained. Two mass parameters are used: the mass of the analytical sample (i.e. mass of seeds homogenised for DNA extraction), and the mass of DNA added to each PCR. The uncertainty associated with the PCR DNA mass is considered as a separate parameter. Determinations of DNA concentration can vary considerably between protocols. Some published methods do not determine the DNA concentration (and hence the mass of DNA added to each PCR) but consider it to be fairly constant. Other methods may measure concentration 57

and require specified DNA amounts but the variability of these amounts is not considered. Associated with the PCR DNA mass is the sensitivity of the PCR test, given as number of target DNA copies required for a 95% detection rate. Theoretically, the binomial or Poisson distribution governs PCR sensitivity [50], however, empirical determinations have shown this may not be the case in practice [39]. Different PCR tests are likely to have different sensitivities due to sequence differences, competition, and inhibitor effects. The ideal approach would seem to be the empirical determination of sensitivity for each PCR test system used. In this study, we have used the sensitivity value of 18 target copies as previously determined [39]. A parameter proportion of false negative responses is available to be used as an alternative input to the model, but is not used in this study. Here, the false negative response rate is estimated and output from the model through directly modelling the effect of the parameters mass of DNA presented to PCR, number of gene copies required for 95% detection rate, and variability associated with extraction of DNA. Where values for these parameters are not available an empirical false negative rate could be used. The effect of the parameter false positive response rate has not been examined directly in this study, which focuses on detection ability. Gaining accurate estimates for false positive response rates is difficult as only indirect measurements or inferences can be made, hence it is necessary to build a replicate / decision system that that is tolerant of a suitably high false positive rate in individual determinations (i.e. higher than is likely to be observed) while maintaining sufficient detection ability. This model can be used to build such a system. Arguably the most important aspect of the system of analytical parameters in this model is the procedure applied to replication. Three levels of replication were applied to the model as described in Figure 8 and detailed above. Because different combinations of these replication types can be applied, they can be regarded as a variable parameter of the model. 4.3.5 Seed lot / as grown scenario A value for the heterogeneity of the distribution of GM seeds in seed lots and as-grown lots was derived from the results of the Genetically Modified Herbicide 58

Tolerant (GMHT) Farm-Scale Evaluations (FSE) geneflow project [49] (95% of GM seeds in 60% of the lot). It represented the estimated heterogeneity in a simulated seed lot harvested from a field of conventional OSR (3.3 Ha square field 10 tonne OSR) grown immediately next to an equally sized field of GM OSR (i.e. without any significant separation) and with no subsequent mixing. This can be considered an extreme worst case for seed lot and as grown crop heterogeneity that could only arise if existing guidance/rules on minimum separation distances (at least 100m in the EU [51] had been ignored. 4.4 Results from the model simulations 4.4.1 Seed lot / as grown scenario Results of the simulated sampling and analysis (without analytical replication) are shown in Table 15 and Figure 11. The proportion of positive results (produced without analytical replication) for each true mean GM proportion will lie between the 5 increment line (minimum used for testing seed lots) and the 30 increment line (maximum used for testing seed lots). The infinite increment line shows the theoretical maximum possible improvement that could be achieved by increasing the number of increments and also the performance that would be achieved for the detection of GM in a homogenous (well mixed) seed lot by the analytical system employed in the model. The 2.5 and 97.5 percentiles of the simulation results are shown in thinner lines for each increment line. 59

Table 15: Performance of the simulated detection system at critical %GM seed thresholds for OSR lots with levels of heterogeneity (measured as the percentage of the lot that contains 95% of the GM seeds) = 60% (as observed in GMHT Farm-Scale Evaluations Geneflow and KeLDA projects). Scenario Number of sample increments Replicate analyses (1) Mean proportion GM (%) Probability of 1 GM seed in 1 analytical sample (2) Probability of detection (2) Seed / as grown 5 1,1,1 0.1 0.87 0.89 0.84 0.86 0.3 0.98 0.99 0.98 0.98 0.9 1.00 1.00 1.00 1.00 30 1,1,1 0.1 0.93 0.95 0.90 0.91 0.3 1.00 1.00 1.00 1.00 0.9 1.00 1.00 1.00 1.00 5 2,2,1 0.1 0.87 0.89 0.96 0.97 0.3 0.98 0.99 1.00 1.00 0.9 1.00 1.00 1.00 1.00 5 1,1,4 0.1 0.87 0.89 0.87 0.89 0.3 0.98 0.99 0.98 0.99 0.9 1.00 1.00 1.00 1.00 Grain 10 1,1,1 0.1 0.91 0.92 0.88 0.89 0.3 0.99 1.00 0.99 0.99 0.9 1.00 1.00 1.00 1.00 Notes to Table 15: 100 1,1,1 0.1 0.94 0.95 0.91 0.93 0.3 1.00 1.00 1.00 1.00 0.9 1.00 1.00 1.00 1.00 10 2,2,1 0.1 0.90 0.92 0.98 0.99 0.3 1.00 1.00 1.00 1.00 0.9 1.00 1.00 1.00 1.00 (1) x,y,z : z aliquots are taken from each of y replicate extracts and delivered to the PCR test. y replicate extracts are produced from each of x replicate analytical samples, giving x y extracts and x y z PCR test aliquots. (2) 95% confidence interval for estimate of probability. 60

Figure 11: Probability of detection of GM seeds in a heterogeneous (95% of GM seeds in 60% of the lot) seed lot / as grown lot analysed without replication for various numbers of sample increments. Thin lines are 2.5 and 97.5 percentiles of simulation results.!"# #$ #$ #$ For seed lots containing 0.1% GM, some analytical replication is always necessary to achieve a 95% probability of detection because the model predicts that the analysis of a sample will not give a positive result 95% of the time even when taken from a perfectly mixed lot (Figure 11) the best achieved is approximately 93%. For lots containing a higher proportion of GM, a 95% predicted probability of detection for the analysis of samples based on 30 increments occurred at approximately 0.13% GM and, for samples based on 5 increments, at approximately 0.18% GM. With sufficient analytical replication (in this case duplicate extracts from duplicate analytical samples giving four replicate results), the analysis of a sample based on as few as five increments gave a predicted probability of detection over 95% (Figure 12, Table 15) for heterogeneous (95% of GM material contained within 60% of the lot) seed lots containing 0.1% GM. However, it is important to choose the correct form (i.e. hierarchical level) of replication. Four replicate results produced by taking four aliquots from a single extract and repeating just the PCR analysis stage gave only an 88 % probability of detection for a sample taken from a lot containing 0.1% GM. 61

Figure 12: Probability of detecting GM seeds in a heterogeneous (95% of GM seeds in 60% of the lot) seed lot / as grown lot with replicate analysis (duplicate extracts from duplicate analytical samples) of a laboratory sample based on five increments. Thin lines are 2.5 and 97.5 percentiles of simulation results.!"# 4.4.2 Grain as commodity scenario A value for the heterogeneity of the distribution GM seeds in grain lots (95% of GM seeds in 60% of the lot) was derived from measurement results (unpublished) generated during the KeLDA [10] project. This value represented the heterogeneity observed between the results of quantitative measurements of GM in 100 increments of soya beans taken from a number of large shipments. It represented a best available estimate of the heterogeneity that may be observed in large grain lots. Results are shown in Figure 13 and Table 15. The proportion of positive results (produced without analytical replication) for each true mean GM proportion will lie between the 10 increment line (minimum used for testing grain lots) and the 100 increment line (maximum number used for testing grain lots). The infinite increment line shows the theoretical maximum possible improvement that could be achieved by 62

increasing the number of increments and also the performance that should be achieved for the detection of GM in a homogenous (well mixed) grain lot. Figure 13: Probability of detecting GM seeds in a heterogeneous (95% of GM seeds in 60% of the lot) grain lot analysed without replication. Thin lines are 2.5 and 97.5 percentiles of simulation results.!"# #$ #$ #$ For grain lots containing 0.1% GM, some analytical replication is necessary to achieve a 95% probability of detection because the model predicts that the analysis of a sample will not give a positive result 95% of the time even where taken from a perfectly mixed lot (Figure 13) the best achieved is approximately 92%. For lots containing higher proportion of GM a 95% predicted probability of detection was achieved for the analysis of samples based on 100 increments at 0.12% GM, and for the analysis of samples based on 10 increments at 0.15% GM. With sufficient analytical replication (duplicate extracts from duplicate analytical samples) the analysis of a sample based on 10 increments gave a probability of detection over 95% (Figure 14, Table 15) for heterogeneous grain lots containing 0.1% GM. 63

Figure 14: Probability of detecting GM seeds in a heterogeneous (95% of GM seeds in 60% of the lot) grain lot with replicate analysis (duplicate extracts from duplicate analytical samples) of a laboratory sample based on ten increments. Thin lines are 2.5 and 97.5 percentiles of simulation results.!"# 4.4.3 Effect of heterogeneity on the detection of 0.1% GM seeds in a lot Figure 15 shows the predicted probability of detection for the analysis (without replication) of samples based on 5, 10, 30 and 100 increments taken from a lot containing 0.1% GM seed where heterogeneity varies between 95% of GM in 5% of the lot (high heterogeneity) to 95% of GM in 95% of the lot (no heterogeneity). The predicted probability of detecting 0.1% GM seeds in samples based on 100 increments was affected to only a small degree by heterogeneity. Whereas, the predicted probability of detecting 0.1% GM seeds in a sample based on 5 increments rapidly diminished as lots become more heterogeneous. Increasing the number of increments had little effect on the predicted probability of detection where heterogeneity was low (95% of GM seeds in 80% of lot). 64

Figure 15: The effect of heterogeneity (measured as the proportion of lot that contains 95% of the GM seeds) on the ability to detect 0.1% of GM seeds in a lot. Thin lines are 2.5 and 97.5 percentiles of simulation results.!"# #$ #$ #$ #$!#%! Figure 16 shows the effect of lot heterogeneity on the predicted probability of detecting 0.1% GM in lots where replicate analyses are carried out (duplicate analytical samples each extracted twice). Compared with the results shown in Figure 15, increasing analytical replication did not greatly reduce the detection false negative rate for scenarios where heterogeneity was high and the number of increments was low. However, analytical replication did lead to a significant improvement in predicted detection rate for samples based on large numbers of increments, or taken from lots with low heterogeneity. The effect of increasing the number of increments or increasing analytical replication can be summarised from Figures 15 and 16 thus: increasing analytical replication increases the maximum detection rate towards which scenarios tend as heterogeneity is reduced; increasing the number of increments decreases the extent to which the detection rate is reduced as heterogeneity increases. 65

Figure 16: The effect of heterogeneity (measured as the proportion of lot that contains 95% of the GM seeds) on the ability to detect 0.1% of GM seeds in a lot where replicate analysis is employed (duplicate extracts from duplicate analytical samples). Thin lines are 2.5 and 97.5 percentiles of simulation results. #$!"# #$ #$ #$!#%! 4.4.4 Model sensitivity analysis Figure 17 shows the effect of varying input parameters by ±5% (±1 for LOD as number of copies) on the probability of detecting (without analytical replication) the presence of 0.1% GM seeds in a heterogeneous seed lot (95% of GM in 60% of the lot) by taking a sample made up of 5 increments. The simulation results were most sensitive to small changes in the size of seeds (mass per 1000 seeds), lot heterogeneity and analytical sample size. Hence, the uncertainty associated with these parameters needs to be reduced where possible or at least well characterised. Figure 17 also shows the effect of varying input parameters by ±5% (±1 for LOD as number of copies) on the probability of detecting (with analytical replication) the presence of 0.1% GM seeds in a heterogeneous grain lot (95% of GM in 60% of the lot) by taking a sample made up of 100 increments. The simulation results were not sensitive to small changes in the size of any of the parameters because the probability of detection was close to 100% for this scenario. Variation of, or uncertainty about, parameters will not greatly affect the proportion of positive results for this scenario. 66

Figure 17: Sensitivity of certified seed / as grown simulation results to ±5% variation in input parameters. Error bars represent 95% Confidence Intervals on simulated results. 100 increments. Duplicate extracts produced from duplicate analytical samples!"# 5 increments. Single extracts produced from single analytical samples & $$% $$$"$ '% $$# $$ $$ # $$(&)*+,-(./#0 0" 67

5. CONCLUSIONS The most important conclusions that can be drawn from this study are that the reliability of a result for the detection of the presence of GM seeds is a function of both analysis and sampling, and that the fitness for purpose of a sampling plan is dependent on the performance of the detection system and vice versa. The modelling tool presented in this study draws together the statistics of sampling and the statistics of detection by PCR to enable users of results to form objective well-informed conclusions about the reliability of results, and to enable the design of fit-for-purpose measurement (sampling + analysis) systems. The modelling tool has been used to show that it is possible to meet the goal of reliably detecting the presence of 0.1% GM seeds in heterogeneous seed lots and grain lots by the analysis of samples (of 3000 seeds) based on a small number of increments if appropriate analytical replication is employed (e.g. duplicate analytical samples with duplicate DNA extractions). However, if no analytical replication is employed then the presence of 0.1% GM will not be reliably detected even by the analysis of samples (of 3000 seeds) based on a large number of increments taken from a homogenous lot. The model has also been used to show that the level of replication (analytical sample; extraction; PCR) is at least as important as the number of replicate analyses. The effects of lot heterogeneity can be ameliorated to a large degree by employing replicate analytical samples, but carrying out replicate PCR determinations of a single analytical extract generates little improvement. 5.1 Future work The model described here does not explicitly address the problems presented by detection of different types of mixtures of GMOs in seed or grain lots. However, the effect of independent events can be estimated as the combined effect of individual events (probability of detecting A or B = 1-probability of detecting neither A nor B). This can readily be done using the model with some adaptation of the output. The statistics of detecting the presence of an unauthorised 'stack'of events, which has arisen by crossing between two or more different GM events, has not been explored in 68

this study. In particular, the probability of detecting the presence of (unauthorised) stacked events in the presence of other individual events could be problematic because of the ambiguity between detected events being common to individual seeds or to the whole sample. However, the model presented in this study could be used to estimate the probability of unambiguously detecting stacked events, with some further work on interpretation of results from samples divided into pools for testing purposes, by comparing the rates of detection for different concentrations of GM material in small (seed pool) samples (e.g. comparing the probability of detecting only A with the probability of detecting only B, or A and B). This version of the model, which simulates qualitative results, employs internal estimates of the quantity of GM material collected in the initial sampling increments taken from the bulk lot and (eventually) carried through to the final PCR detection step. Hence, the model could readily be adapted to examine the effect of bulk heterogeneity on quantitative GMO measurement by the addition of modules describing the statistics of real-time PCR and the statistics of seed pool testing / determination (i.e. estimation of Most Probable Number ). In addition, a further quantitative module could be developed to estimate the risks (for both producers and consumers) attached to labelling or rejection decisions based on the entire sampling and analytical process whether the analysis is semi-quantitative (e.g. number of positive seed pools) or fully quantitative (e.g. real-time PCR). 69

REFERENCES 1. International Standards Organisation (1999). "ISO 13690. Cereals, pulses and milled products - sampling of static batches." See http://www.iso.org/iso/en/cataloguelistpage.cataloguelist. 2. ISTA (2004). "International Rules for Seed Testing." International Seed Testing Association, Zurich, Switzerland. See http://www.seedtest.org/en/product---1--1082--203--1.html. 3. Commission of the European Communities, (2004). "Commission Recommendation of 4 October 2004 on technical guidance for sampling and detection of genetically modified organisms and material produced from genetically modified organisms as or in products in the context of Regulation (EC) No 1830/2003 2004/787/EC." Official Journal of the European Union, L 348, 18-26. See http://europa.eu.int/eurlex/lex/lexuriserv/site/en/oj/2004/l_348/l_34820041124en00180026.pdf. 4. IHCP-JRC Biotechnology and GMOs Unit. "GMO METHODS DATABASE." See http://biotech.jrc.it/methodsdatabase.htm. 5. P. Reed (2002). "Protocol for bulk grain sampling and testing for GM presence. Draft SOP produced on behalf of the European Enforcement Project." Central Science Laboratory, York. See http://www.csl.gov.uk/prodserv/cons/plant/certification.cfm. 6. S. Kay & C. Paoletti (2001). "Sampling strategies for GMO detection and/or quantification." European Commission Directorate General, JRC. See http://biotech.jrc.it/doc/euroreport_sampling_strategies.pdf. 7. International Standards Organisation (1995). "EN ISO 542. Oilseeds - Sampling." See http://www.iso.org/iso/en/cataloguelistpage.cataloguelist. 8. British Standards Organisation (2002). "BS ISO 6644. Flowing cereals and milled cereal products - Automatic sampling by mechanical means." See http://www.bsi-global.com/. 9. C. Paoletti, M. Donatelli, S. Kay & G. Van den Eede (2003) Simulating kernel lot sampling: the effect of heterogeneity on the detection of GMO contaminations. Seed Science and Technology, 31, 629-638. 10. IHCP-JRC Biotechnology and GMOs Unit. "Kernel Lot Distribution Assessment (KeLDA)." See http://biotech.jrc.it/sampling_kelda.htm. 11. ISTA (2001). "Position paper concerning the increase or repeal of ISTA maximum lot size." International Seed Testing Association. See http://www.seedtest.org/upload/cms/user/34a4intmi200134positionpaperons eedlotsize.pdf. 12. S. R. Miles (1962) Heterogeneity in seed lots. Proceedings of the International Seed Testing Association, 27, 407-413. 13. J. G. Tattersfield (1977) Further estimates of heterogeneity in seed lots. Seed Science and Technology, 5, 443-450. 14. J. Jorgensen & K. Kristensen (1990) Heterogeneity of grass seed lots. Seed Science and Technology, 18, 515-523. 15. M. Kruse & A. M. Steiner (1995) "Variation between seed lots as an estimate for the risk of heterogeneity with increasing ISTA maximum lot size." in Seed Symposium Abstracts, ISTA 24th Congress. Copenhagen, Denmark. 16. R. M. Coster (1993) Seed Lot-Size Limitation as Reflected in Heterogeneity Testing - a Review. Seed Science and Technology, 21, 513-520. 70

17. M. Kruse (1999) "Uniformity of seed lots as a quality trait - the statistical aspects." in Book of Abstracts, World Seed Conference. 139-143. Cambridge, UK. 18. UK Government, (2002). "The oil and fibre plant seed (England) regulations." HMSO. See http://www.legislation.hmso.gov.uk/si/si2002/20023174.htm. 19. Commission of the European Communities, (2002). "Council Directive 2002/57/EC of 13 June 2002 on the marketing of seed of oil and fibre plants 2002/57/EC." Official Journal of the European Union, L 193, 74-97. See http://europa.eu.int/eurlex/pri/en/oj/dat/2002/l_193/l_19320020720en00740097.pdf. 20. ISTA (1993) "International Seed Testing Association Rules." in Proceedings of the International Seed Testing Association, Vol. 21 (supplement). International Seed Testing Association, Zurich, Switzerland. 21. European Parliament and the Council of the European Union, (2003). "Regulation (EC) No 1830/2003 of the European Parliament and of the Council of 22 September 2003 concerning the traceability and labelling of genetically modified organisms and the traceability of food and feed products produced from genetically modified organisms and amending Directive 2001/18/EC." Official Journal of the European Union, L 268, 24-28. See http://europa.eu.int/eurlex/pri/en/oj/dat/2003/l_268/l_26820031018en00240028.pdf. 22. American Association of Cereal Chemists (AACC). "American Association of Cereal Chemists (AACC) method 64-70A and method 64-71. Mechanical Sampling of Dry, Free-Flowing Granular material." See http://www.aaccnet.org/approvedmethods/toc.htm. 23. Packers and Stockyards Administration) GIPSA (Grain Inspection, USDA (1995). Grain Inspection handbook - Book 1 Sampling. United States Department of Agriculture, Grain Inspection, Packers and Stockyards Administration, Federal Grain Inspection Service. See http://www.usda.gov/gipsa/reference-library/handbooks/graininsp/grbook1/bk1.pdf. 24. GIPSA, United States Department of Agriculture (2003). "Practical sampling for the detection of biotech grains." See http://www.usda.gov/gipsa/biotech/sample1.htm. 25. Packers and Stockyards Administration) GIPSA (Grain Inspection, United States Department of Agriculture. "Sampling and testing recommendations for the Detection of Cry 9C Protein in Hybrid Seed Corn." See http://www.usda.gov/gipsa/biotech/starlink/cry9cdetection.htm. 26. ISTA (1999). "International Rules for Seed Testing." ISBN 3-906-549-27-5. International Seed Testing Association, Zurich, Switzerland. See http://www.seedtest.org/en/product---1--1082--203--1.html. 27. ISTA (2005). "International Rules for Seed Testing." International Seed Testing Association, Zurich, Switzerland. See http://www.seedtest.org/en/productdetail---1--1082--203--1.html. 28. J. Bányai & J. Barabás (2002). Handbook on Statistics in Seed Testing, Edn. Revised Version. ISTA. See http://www.seedtest.org/upload/prj/product/stahandbk2002.pdf. 29. P. Lischer (2001) Sampling procedures to determine the proportion of genetically modified organisms in raw materials. Part I: Correct sampling, 71

good sampling practice. Mitteilungen aus Lebensmitteluntersuchung und Hygiene, 92, 290-304. 30. P. Lischer (2001) Sampling procedures to determine the proportion of genetically modified organisms in raw materials. Part II: Samplingfrom batches of grain. Mitteilungen aus Lebensmitteluntersuchung und Hygiene, 92, 305-311. 31. A. Block & G. Schwarz (2003) Validation of different genomic and cloned DNA calibration standards for construct-specific quantification of LibertyLink in rapeseed by real-time PCR. European Food Research and Technology, 216, 421-427. 32. T. Demeke, R. W. Giroux, S. Reitmeier & S. L. Simon (2002) Development of a Polymerase Chain Reaction Assay for detection of three Canola transgenes. Journal of the American Oil Chemists'Society, 79, 1015-1019. 33. R. Zeitler, M. Pietsch & H. U. Waiblinger (2002) Validation of real-time PCR methods for the quantification of transgenic contaminations in rape seed. European Food Research and Technology, 214, 346-351. 34. D. James, A. M. Schmidt, E. Wall, M. Green & S. Masri (2003) Reliable detection and identification of genetically modified maize, soybean, and canola by multiplex PCR analysis. Journal of Agricultural and Food Chemistry, 51, 5829-5834. 35. M. Hernandez, A. Rio, T. Esteve, S. Prat & M. Pla (2001) A rapeseed-specific gene, Acetyl-CoA carboxylase, can be used as a reference for qualitative and real-time quantitative PCR detection of transgenes from mixed food samples. Journal of Agricultural and Food Chemistry, 49, 3622-3627. 36. European Parliament and the Council of the European Union, (2001). "Directive 2001/18/EC of the European Parliament and of the Council of 12 March 2001 on the deliberate release into the environment of genetically modified organisms and repealing Council Directive 90/220/EEC." Official Journal of the European Union, L 106, 1-39. See http://europa.eu.int/eurlex/pri/en/oj/dat/2001/l_106/l_10620010417en00010038.pdf. 37. Scientific Committee on Plants (2001). "Opinion of the Scientific Committee on Plants concerning the adventitious presence of GM seeds in conventional seeds. (Opinion adopted by the Committee on 7 March 2001)." European Commision, Health & Consumer Protection Directorate-General. See http://europa.eu.int/comm/food/fs/sc/scp/out93_gmo_en.pdf. 38. K. Arumuganathan & E. D. Earle (1991) Nuclear DNA content of some important plant species. Plant Molecular Biology Reporter, 9, 208-218. 39. J. P. Hughes & P. Totten (2003) Estimating the accuracy of polymerase chain reaction-based tests using endpoint dilution. Biometrics, 59, 505-511. 40. S. Bruderer & E. Leitner (2004). "Genetically Modified (GM) Crops: molecular and regulatory details." Centre for Biosafety and Sustainability (BATS). See http://www.gmo-watch.org/gvo-report140703.pdf. 41. Association Francais de Normalisation (AFNOR). "Foodstuffs - detection and quantification of genetically modified vegetal organisms and derived products. Part 1 : guidelines and requirements for methods based on the detection of nucleic acids." See http://www.boutique.afnor.fr. 42. K. M. Remund, D. A. Dixon, D. L. Wright & L. R. Holden (2001) Statistical considerations in seed purity testing for transgenic traits. Seed Science Research, 11, 101-119. 72

43. European Committee for Standardisation (2004). "Foodstuffs - Methods of analysis for the detection of genetically modified organisms and derived products - Sampling (pren ISO 22174)." Comité Européen de Normalisation, Technical Committee 275: Food Analysis - Horizontal Methods. Working Group 11: Genetically Modified Foodstuffs. See http://www.cenorm.be/cenorm/businessdomains/technicalcommitteeswo rkshops/centechnicalcommittees/wp.asp?param=6256&title=cen%2ftc +275. 44. M. Evans, N. Hastings & B. Peacock (1993). Statistical Distributions, Edn. 2nd. John Wiley & Sons Inc, New York. See http://eu.wiley.com/wileycda/wileytitle/productcd-0471371246,desccdauthorinfo.html. 45. P. M. Altham (2002) An aspect of discrete data analysis: fitting a betabinomial distribution to the hospitals'data. Developments in biologicals, 107, 77-83. 46. Eurachem/CITAC (2000). "Quantifying uncertainty in analytical measurement." Eurachem. See http://www.measurementuncertainty.org/mu/guide/index.html. 47. Langton S. D., Chevennement R., Nagelkerke N. & Lombard B. (2002) Analysing collaborative trials for qualitative microbiological methods: accordance and concordance. International Journal of Food Microbiology, 79, 175-181. 48. CEN/TC275/WG11N0039 (2002) "Foodstuffs - Methods for the detection of genetically modified organisms and derived products - sampling." in 9th meeting of WG 11. Berlin. 49. R. Weekes, Deppe, C., Boffey, C., Morgan, D., Morgan, S., Bilton, M., Daniels, R. and Henry, C. (2004) Crop-to-crop gene flow at farm-scale evaluation sites of oilseed rape (Brassica napus) in the UK. Transgenic Research, submitted, under review. 50. S. Kay & G. Van den Eede (2001) The limits of GMO detection. Nature Biotechnology, 19, 405. 51. Anne-Katrin Bock, Karine Lheureux, Monique Libeau-Dulos, Hans Nilsagård & Emilio Rodriguez-Cerezo (2002). "Scenarios for co-existence of genetically modified, conventional and organic crops in European agriculture." Institute for Prospective Technological Studies - Joint Research Centre. See http://www.jrc.cec.eu.int/download/gmcrops_coexistence.pdf. 73