Testing green coffee for ochratoxin A, part III: performance of ochratoxin A sampling plan

Green coffee shipments are often inspected for ochratoxin A (OTA) and classified into good or bad categories depending on whether the OTA estimates are above or below a defined regulatory limit. Because of the uncertainty associated with the sampling, sample preparation, and analytical steps of an OTA test procedure, some shipments of green coffee will be misclassified. The misclassification of lots leads to some good lots being rejected (sellers' risk) and some bad lots being accepted (buyers' risk) by an OTA sampling plan. Reducing the uncertainty of an OTA test procedure and using an accept/reject limit less than the regulatory limit can reduce the magnitude of one or both risks. The uncertainty of the OTA test procedure is most effectively reduced by increasing sample size (or increasing the number of samples analyzed), because the sampling step is the largest source of uncertainty in the OTA test procedure. The effects of increasing sample size and changing the sample accept/reject limit relative to the regulatory limit on the performance of OTA sampling plans for green coffee were investigated. For a given accept/reject limit of 5 microg/kg, increasing sample size increased the percentage of lots accepted at concentrations below the regulatory limit and increased the percentage of lots rejected at concentrations above the regulatory limit. As a result, increasing sample size reduced both the number of good lots rejected (sellers' risk) and the number of bad lots accepted (buyers' risk). For a given sample size (1 kg), decreasing the sample accept/reject limit from 5 to 2 microg/kg relative to a fixed regulatory limit of 5 microg/kg decreased the percentage of lots accepted and increased the percentage of lots rejected at all OTA concentrations. As a result, decreasing the accept/reject limit below the regulatory limit increased the number of good lots rejected (sellers' risk), but decreased the number of bad lots accepted (buyers' risk).     

Sampling grain shipments to detect genetically modified seed.

Using the binomial distribution, the effect of sample size on the variability among sample test results when sampling a lot with 1.0% genetically modified (GM) or biotech seed was evaluated. The coefficient of variation, cv, among 500-seed sample test results taken from a lot with truly 1.0% was computed to be 44.5%. Increasing sample size to 1000 seeds reduced the cv among sample test results to 31.5%. The effects of sample size and accept/reject limits on the buyer's risk (bad lots accepted) and the seller's risk (good lots rejected) was also evaluated assuming a tolerance of 1.0% GM seed. Increasing sample size decreases both the buyer's and seller's risks at the same time. Using an accept/reject limit below the regulatory tolerance decreases the buyer's risk, but increases the seller's risk. Using an accept/reject limit above the regulatory tolerance decreases the seller's risk but increases the buyer's risk.     

Testing shelled corn for aflatoxin, Part I: estimation of variance components

The variability associated with testing lots of shelled corn for aflatoxin was investigated. Eighteen lots of shelled corn were tested for aflatoxin contamination. The total variance associated with testing shelled corn was estimated and partitioned into sampling, sample preparation, and analytical variances. All variances increased as aflatoxin concentration increased. With the use of regression analysis, mathematical expressions were developed to model the relationship between aflatoxin concentration and the total, sampling, sample preparation, and analytical variances. The expressions for these relationships were used to estimate the variance for any sample size, subsample size, and number of analyses for a specific aflatoxin concentration. Test results on a lot with 20 parts per billion aflatoxin using a 1.13 kg sample, a Romer mill, 50 g subsamples, and liquid chromatographic analysis showed that the total, sampling, sample preparation, and analytical variances were 274.9 (CV = 82.9%), 214.0 (CV = 73.1 %), 56.3 (CV = 37.5%), and 4.6 (CV = 10.7%), respectively. The percentage of the total variance for sampling, sample preparation, and analytical was 77.8, 20.5, and 1.7, respectively.     

Sampling almonds for aflatoxin, part I: estimation of uncertainty associated with sampling, sample preparation, and analysis

Domestic and international regulatory limits have been established for aflatoxin in almonds and other tree nuts. It is difficult to obtain an accurate and precise estimate of the true aflatoxin concentration in a bulk lot because of the uncertainty associated with the sampling, sample preparation, and analytical steps of the aflatoxin test procedure. To evaluate the performance of aflatoxin sampling plans, the uncertainty associated with sampling lots of shelled almonds for aflatoxin was investigated. Twenty lots of shelled almonds were sampled for aflatoxin contamination. The total variance associated with measuring B1 and total aflatoxins in bulk almond lots was estimated and partitioned into sampling, sample preparation, and analytical variance components. All variances were found to increase with an increase in aflatoxin concentration (both B1 and total). By using regression analysis, mathematical expressions were developed to predict the relationship between each variance component (total, sampling, sample preparation, and analysis variances) and aflatoxin concentration. Variance estimates were the same for B1 and total aflatoxins. The mathematical relationships can be used to estimate each variance for a given sample size, subsample size, and number of analyses other than that measured in the study. When a lot with total aflatoxins at 15 ng/g was tested by using a 10 kg sample, a vertical cutter mixer type of mill, a 100 g subsample, and high-performance liquid chromatography analysis, the sampling, sample preparation, analytical, and total variances (coefficient of variation, CV) were 394.7 (CV, 132.4%), 14.7 (CV, 25.5%), 0.8 (CV, 6.1%), and 410.2 (CV, 135.0%), respectively. The percentages of the total variance associated with sampling, sample preparation, and analytical steps were 96.2, 3.6, and 0.2, respectively.     

Sampling hazelnuts for aflatoxin: effect of sample size and accept/reject limit on reducing the risk of misclassifying lots

About 100 countries have established regulatory limits for aflatoxin in food and feeds. Because these limits vary widely among regulating countries, the Codex Committee on Food Additives and Contaminants began work in 2004 to harmonize aflatoxin limits and sampling plans for aflatoxin in almonds, pistachios, hazelnuts, and Brazil nuts. Studies were developed to measure the uncertainty and distribution among replicated sample aflatoxin test results taken from aflatoxin-contaminated treenut lots. The uncertainty and distribution information is used to develop a model that can evaluate the performance (risk of misclassifying lots) of aflatoxin sampling plan designs for treenuts. Once the performance of aflatoxin sampling plans can be predicted, they can be designed to reduce the risks of misclassifying lots traded in either the domestic or export markets. A method was developed to evaluate the performance of sampling plans designed to detect aflatoxin in hazelnuts lots. Twenty hazelnut lots with varying levels of contamination were sampled according to an experimental protocol where 16 test samples were taken from each lot. The observed aflatoxin distribution among the 16 aflatoxin sample test results was compared to lognormal, compound gamma, and negative binomial distributions. The negative binomial distribution was selected to model aflatoxin distribution among sample test results because it gave acceptable fits to observed distributions among sample test results taken from a wide range of lot concentrations. Using the negative binomial distribution, computer models were developed to calculate operating characteristic curves for specific aflatoxin sampling plan designs. The effect of sample size and accept/reject limits on the chances of rejecting good lots (sellers' risk) and accepting bad lots (buyers' risk) was demonstrated for various sampling plan designs.     

Sampling almonds for aflatoxin, part II: estimating risks associated with various sampling plan designs

About 100 nations have established regulatory limits for aflatoxin in food and feeds. Because these limits vary widely from one country to another, the Codex Alimentarius Commission, working through the Codex Committee on Food Additives and Contaminants, has initiated work to harmonize aflatoxin limits and sampling plans for almonds, pistachios, hazelnuts, and Brazil nuts. Studies were developed to measure the uncertainty and distribution among test results for replicate samples taken from aflatoxin-contaminated almond shipments. The uncertainty and distribution information was used to develop a model to evaluate the performance of aflatoxin sampling plans so that harmonized sampling plans can be developed for almonds that reduce the misclassifying of lots in the export trade. Twenty lots of shelled almonds were sampled according to an experimental protocol in which sixteen 10 kg samples were taken from each lot. The observed aflatoxin distribution among the 16 sample test results was compared with 3 theoretical distributions. The negative binomial distribution was selected to model aflatoxin distribution among sample test results because it gave acceptable fits across all 20 observed sample distributions. By using the variance and distribution information, operating characteristics curves were developed to predict the effect of sample size and accept/reject limits on the probability of rejecting good lots and accepting bad lots.     

Sampling hazelnuts for aflatoxin: uncertainty associated with sampling, sample preparation, and analysis

The variability associated with the aflatoxin test procedure used to estimate aflatoxin levels in bulk shipments of hazelnuts was investigated. Sixteen 10 kg samples of shelled hazelnuts were taken from each of 20 lots that were suspected of aflatoxin contamination. The total variance associated with testing shelled hazelnuts was estimated and partitioned into sampling, sample preparation, and analytical variance components. Each variance component increased as aflatoxin concentration (either B1 or total) increased. With the use of regression analysis, mathematical expressions were developed to model the relationship between aflatoxin concentration and the total, sampling, sample preparation, and analytical variances. The expressions for these relationships were used to estimate the variance for any sample size, subsample size, and number of analyses for a specific aflatoxin concentration. The sampling, sample preparation, and analytical variances associated with estimating aflatoxin in a hazelnut lot at a total aflatoxin level of 10 ng/g and using a 10 kg sample, a 50 g subsample, dry comminution with a Robot Coupe mill, and a high-performance liquid chromatographic analytical method are 174.40, 0.74, and 0.27, respectively. The sampling, sample preparation, and analytical steps of the aflatoxin test procedure accounted for 99.4, 0.4, and 0.2% of the total variability, respectively.     

Testing shelled corn for aflatoxin, Part II: modeling the observed distribution of aflatoxin test results

The suitability of several theoretical distributions to predict the observed distribution of aflatoxin test results in shelled corn was investigated. Fifteen positively skewed theoretical distributions were each fitted to 18 empirical distributions of aflatoxin test results for shelled corn. The compound gamma distribution was selected to model aflatoxin test results for shelled corn. The method of moments technique was chosen to estimate the parameters of the compound gamma distribution. Mathematical expressions were developed to calculate the parameters of the compound gamma distribution for any lot aflatoxin concentration and test procedure. Observed acceptance probabilities were compared to operating characteristic curves predicted from the compound gamma distribution, and all 18 observed acceptance probabilities were found to lie within a 95% confidence band. The parameters of compound gamma were used to calculate the fraction of aflatoxin-contaminated kernels in contaminated lots. At 20 ppb, it was estimated that about 6 in 10,000 kernels are contaminated.     

External validation of EPIWIN biodegradation models

The BIOWIN biodegradation models were evaluated for their suitability for regulatory purposes. BIOWIN includes the linear and non-linear BIODEG and MITI models for estimating the probability of rapid aerobic biodegradation and an expert survey model for primary and ultimate biodegradation estimation. Experimental biodegradation data for 110 newly notified substances were compared with the estimations of the different models. The models were applied separately and in combinations to determine which model(s) showed the best performance. The results of this study were compared with the results of other validation studies and other biodegradation models. The BIOWIN models predict not-readily biodegradable substances with high accuracy in contrast to ready biodegradability. In view of the high environmental concern of persistent chemicals and in view of the large number of not-readily biodegradable chemicals compared to the readily ones, a model is preferred that gives a minimum of false positives without a corresponding high percentage false negatives. A combination of the BIOWIN models (BIOWIN2 or BIOWIN6) showed the highest predictive value for not-readily biodegradability. However, the highest score for overall predictivity with lowest percentage false predictions was achieved by applying BIOWIN3 (pass level 2.75) and BIOWIN6.     

Evaluation of sampling plans to detect Cry9C protein in corn flour and meal

StarLink is a genetically modified corn that produces an insecticidal protein, Cry9C. Studies were conducted to determine the variability and Cry9C distribution among sample test results when Cry9C protein was estimated in a bulk lot of corn flour and meal. Emphasis was placed on measuring sampling and analytical variances associated with each step of the test procedure used to measure Cry9C in corn flour and meal. Two commercially available enzyme-linked immunosorbent assay kits were used: one for the determination of Cry9C protein concentration and the other for % StarLink seed. The sampling and analytical variances associated with each step of the Cry9C test procedures were determined for flour and meal. Variances were found to be functions of Cry9C concentration, and regression equations were developed to describe the relationships. Because of the larger particle size, sampling variability associated with cornmeal was about double that for corn flour. For cornmeal, the sampling variance accounted for 92.6% of the total testing variability. The observed sampling and analytical distributions were compared with the Normal distribution. In almost all comparisons, the null hypothesis that the Cry9C protein values were sampled from a Normal distribution could not be rejected at 95% confidence limits. The Normal distribution and the variance estimates were used to evaluate the performance of several Cry9C protein sampling plans for corn flour and meal. Operating characteristic curves were developed and used to demonstrate the effect of increasing sample size on reducing false positives (seller's risk) and false negatives (buyer's risk).     

Qualitative aspects and validation of a screening method for pesticides in vegetables and fruits based on liquid chromatography coupled to full scan high resolution (Orbitrap) mass spectrometry

The analytical capabilities of liquid chromatography with single-stage high-resolution mass spectrometry have been investigated with emphasis on qualitative aspects related to selective detection during screening and to identification. The study involved 21 different vegetable and fruit commodities, a screening database of 556 pesticides for evaluation of false positives, and a test set of 130 pesticides spiked to the commodities at 0.01, 0.05, and 0.20 mg/kg for evaluation of false negatives. The final method involved a QuEChERS-based sample preparation (without dSPE clean up) and full scan acquisition using alternating scan events without/with fragmentation, at a resolving power of 50,000. Analyte detection was based on extraction of the exact mass (±5 ppm) of the major adduct ion at the database retention time ±30 s and the presence of a second diagnostic ion. Various options for the additional ion were investigated and compared (other adduct ions, M + 1 or M + 2 isotopes, fragments). The two-ion approach for selective detection of the pesticides in the full scan data was compared with two alternative approaches based on response thresholds. Using the two-ion approach, the number of false positives out of 11,676 pesticide/commodity combinations targeted was 36 (0.3 %). The percentage of false negatives, assessed for 2,730 pesticide/commodity combinations, was 13 %, 3 %, and 1 % at the 0.01-, 0.05-, and 0.20-mg/kg level, respectively (slightly higher with fully automated detection). Following the SANCO/12495/2011 protocol for validation of screening methods, the screening detection limit was determined for 130 pesticides and found to be 0.01, 0.05, and ≥0.20 mg/kg for 86, 30, and 14 pesticides, respectively. For the detected pesticides in the spiked samples, the ability for unambiguous identification according to EU criteria was evaluated. A proposal for adaption of the criteria was made.     

Uncertainty-based measurement quality control

According to a simple acceptance decision rule for measurement quality control, a measured value will be accepted if the expanded uncertainty of the measurements is not greater than a preset maximum permissible uncertainty. Otherwise, the measured value will be rejected. The expanded uncertainty may be calculated as the z-based uncertainty (the half-width of the z-interval) when the measurement population standard deviation σ is known or the sample size is large (30 or greater), or by a sample-based uncertainty estimator when σ is unknown and the sample size is small. The decision made based on the z-based uncertainty will be deterministic and may be assumed to be correct. However, the decision made based on a sample-based uncertainty estimator will be uncertain. This paper develops the mathematical formulations for computing the probability of acceptance for two sample-based uncertainty estimators: the t-based uncertainty (the half-width of the t-interval) and an unbiased uncertainty estimator. The risk of incorrect decision-making, in terms of the false acceptance probability and false rejection probability, is derived from the probability of acceptance. The theoretical analyses indicate that the t-based uncertainty may result in significantly high false rejection probability when the sample size is very small (especially for samples of size 2). For some applications, the unbiased uncertainty estimator may be superior to the t-based uncertainty for measurement quality control. Several examples from acoustic Doppler current profiler streamflow measurements are presented to demonstrate the performance of the t-based uncertainty and the unbiased uncertainty estimator.     

Determining the variability associated with testing shelled corn for aflatoxin using different analytical procedures in Louisiana in 1998

The number of elevator facilities with laboratories to test shelled corn for aflatoxin on site is increasing. The inherent difficulty in accurately determining the true aflatoxin concentration of a lot of corn may have serious implications. Deviations from the true value are of even greater significance at busy locations where a high throughput is desired. This study was instituted to measure (1) the differences in aflatoxin test results between elevator laboratories and the Louisiana Agricultural Chemistry (LAC) laboratory and (2) the variability in aflatoxin test results associated with sampling, sample preparation, and analysis of shelled corn at such locations. One hundred lots of shelled corn from 10 elevators in Louisiana were analyzed for aflatoxin using the Aflatest method (at elevators and at the LAC laboratory) and high-performance column liquid chromatography (HPLC; LAC laboratory only). Mean aflatoxin levels determined at elevator laboratories were significantly (P < 0.05) lower from those obtained in the LAC laboratory using the Aflatest method. Overall, Aflatest method results were lower than those obtained by HPLC. This difference may be attributed to analyst technical dexterity, difficulty in providing careful attention to detail in a high throughput environment, and/or substandard facilities found at elevators. The total variance was partitioned into the combined sampling plus subsampling variance and analytical variance. The sampling and sample preparation steps accounted for about 91.5% of the total variability. When using the HPLC analytical method, the analytical step contributed only 8.5% to the total variance.     

Enrichment of high-throughput screening data with increasing levels of noise using support vector machines, recursive partitioning, and laplacian-modified naive bayesian classifiers

High-throughput screening (HTS) plays a pivotal role in lead discovery for the pharmaceutical industry. In tandem, cheminformatics approaches are employed to increase the probability of the identification of novel biologically active compounds by mining the HTS data. HTS data is notoriously noisy, and therefore, the selection of the optimal data mining method is important for the success of such an analysis. Here, we describe a retrospective analysis of four HTS data sets using three mining approaches: Laplacian-modified naive Bayes, recursive partitioning, and support vector machine (SVM) classifiers with increasing stochastic noise in the form of false positives and false negatives. All three of the data mining methods at hand tolerated increasing levels of false positives even when the ratio of misclassified compounds to true active compounds was 5:1 in the training set. False negatives in the ratio of 1:1 were tolerated as well. SVM outperformed the other two methods in capturing active compounds and scaffolds in the top 1%. A Murcko scaffold analysis could explain the differences in enrichments among the four data sets. This study demonstrates that data mining methods can add a true value to the screen even when the data is contaminated with a high level of stochastic noise.     

Performance of sampling plans to determine aflatoxin in farmers' stock peanut lots by measuring aflatoxin in high-risk-grade components

Five 2 kg test samples were taken from each of 120 farmers' stock peanut lots contaminated with aflatoxin. Kernels from each 2 kg sample were divided into the following U.S. Department of Agriculture grade components: sound mature kernels plus sound splits (SMKSS), other kernels (OK), loose shelled kernels (LSK), and damaged kernels (DAM). The kernel mass (g), aflatoxin mass (ng), and aflatoxin concentration (ng of aflatoxin/g of peanuts) were measured for each of the 2400 component samples. The variabilities associated with measuring aflatoxin mass (ng) in OK + LSK + DAM, or A(OLD)ng, and in LSK + DAM, or A(LD)ng, and aflatoxin concentration (ng/g) in OK + LSK + DAM, or A(OLD)ng/g, and in LSK + DAM, or A(LD)ng/g, were determined. The variance associated with measuring aflatoxin in each of the 4 combinations of components increased with aflatoxin, and functional relationships were developed from regression analysis. The variability associated with estimating the lot concentration from each of the 4 combinations of components was also determined. The coefficients of variation (CV) associated with estimating the aflatoxin for a lot with aflatoxin at 100 ng/g were 90, 86, 94 and 96% for aflatoxin masses A(OLD)ng and A(LD)ng and aflatoxin concentrations A(OLD)ng/g and A(LD)ng/g, respectively. The performance of aflatoxin sampling plans using the combination of aflatoxin masses in OK + LD + DAM and LD + DAM components was evaluated with a 2 kg test sample and a 50 ng/g accept/reject limit.     

Predicting aflatoxin and fumonisin in shelled corn lots using poor-quality grade components

A study was conducted to determine if aflatoxin and fumonisin are concentrated in the poor-quality grade components of shelled corn. Four 1.0 kg test samples were each taken from 23 lots of shelled corn marketed in North Carolina. Inspectors from the Federal Grain Inspection Service divided each test sample into 3 grade components: (1) damaged kernels (DM), (2) broken corn and foreign material (BCFM), and )3) whole kernels (WH). The aflatoxin and fumonisin concentration was measured in each component and a mass balance equation was used to calculate the total concentration of each mycotoxin in each test sample. Averaged across all test samples, the aflatoxin concentrations in the DM, BCFM, and WH components were 1300.3, 455.2, and 37.3 ppb, respectively. Averaged across all test samples, the fumonisin concentrations in the DM, BCFM, and WH components were 148.3, 51.3, and 1.8 ppm, respectively. The DM and BCFM components combined accounted for only 5.0% of the test sample mass, but accounted for 59.8 and 77.5% of the total aflatoxin and fumonisin mass in the test sample, respectively. Both aflatoxin mass (ng) and aflatoxin concentration (ng/g) in the combined DM and BCFM components had high correlations with aflatoxin concentration in the lot. The highest correlation occurred when aflatoxin mass (ng) in the combined DM and BCFM components was related to aflatoxin concentration in the lot (0.964). Similar results were obtained for fumonisin. This study indicated that measuring either aflatoxin or fumonisin in the combined DM and BCFM grade components could be used as a screening method to predict either aflatoxin or fumonisin in a bulk lot of shelled corn.     

Determination of sample size for validation of allergen-screening methods

For various levels of confidence (i.e., 80 and 90%) and ratios (K = sigmap2/sigmaN2, where sigmap2 and sigmaN2 are the analyte variances for the positive and negative distributions, respectively), sample sizes sufficient to test the requirements that a given method detects > or = 90% of the positives (> or = 5 ppm of a given analyte) while misclassifying < or = 10% of the negatives (implying a specificity rate, true negatives that will be correctly classified, of 90%) were estimated by using a rationale that minimizes the cost of sampling.     

Measurement uncertainty as a tool for evaluating the ‘grey zone’ to reduce the false negatives in immunochemical screening of blood donors for infectious diseases

The risk of misclassifying infected individuals as healthy constitutes a crucial challenge when screening blood donors by means of immunoassays. This risk is especially challenging when the numerical results are close to the clinical decision level, i.e. in the ‘grey zone’. The concept of using measurement uncertainty for evaluating the ‘grey zone’ has previously not been systematically applied in this context. This article explains methods, models and empirical (top-down) approaches for the calculation of measurement uncertainty using results from a blood bank according to the internationally accepted GUM principles, focusing on uncertainty sources in the analytical phase. Of the different approaches available, the intralaboratory empirical approaches are emphasised since modelling (bottom-up) approaches are impracticable due to the lack of reliable model equations for immunoassays. Different methods are applied to estimate the measurement uncertainty for the Abbott Prism^® HCV immunoassay. The expanded uncertainty obtained at the clinical decision level from the intralaboratory empirical approach was 36 %. The estimated uncertainty was used to set acceptance and rejection zones following the procedure set in the Eurachem guideline, emphasising the need to minimise the occurrence of false negatives.     

Statistical approach for risk assessment of aflatoxin sampling plan used by manufacturers for raw shelled peanuts

Processed food manufacturers often use acceptance sampling plans to screen out lots with unacceptable levels of contamination from incoming raw material streams. Sampling plan designs are determined by specifying sample sizes, sample preparation methods, analytical test methods, and accept/reject criteria. Sampling plan performance can be indicated by plotting acceptance probability versus contamination level as an operating characteristic (OC) curve. In practice, actual plan performance depends on the level of contamination in the incoming lot stream. This level can vary considerably over time, among different crop varieties, and among locales. To better gauge plan performance, a method of coupling an OC curve and crop distributions is proposed. The method provides a precise probabilistic statement about risk and can be easily performed with commercial spreadsheet software.