JP5147084B2 - Method for detecting defects in DNA microarray data - Google Patents

Description

The present invention relates to a method for detecting hybridization problems in DNA microarray data.

Hybridization is the basis of microarray analysis and is widely used, but it is not without technical problems. For example, hybridization may form a donut-like geometric pattern around the center of the chip image. Such a pattern appears similar to surface scratches due to dust contamination and often results in a reduced signal in certain areas of the chip. Although an analysis program for recognizing such a problem has been proposed, this method is catastrophic because it cancels the entire array chip data when a large defect exists. The dChip package implements an automatic algorithm that recognizes and removes outliers in model-based data standardization. This algorithm finds patterns in the response between perfect match (PM) probes and mismatch (MM) probes for each gene and identifies cells and probe sets that do not match the resulting pattern as outliers. However, this approach is based on a series of mathematical models derived from a simple point of view of the combination of biological fundamentals and data. In addition, this model, which inevitably includes parameters dealing with noise, has no objective index indicating how well it represents the experimental system, so it is difficult to strictly check the validity of this model and calculation method. is there.

One reason for the recognition of hybridization defects ad hoc is that even if such a problem is seen in a large part of the chip area, a signal that reflects the transcript level, i.e. scaled. It is believed that it has no effect on the probe value. In GeneChip (registered trademark), transcripts are measured by about 10 pairs of adjacent PM cells and MM cells, and these pairs are distributed over the entire chip. Thus, failure occurs simultaneously in both related pairs of PM and MM probes, and no more than one probe pair fails for a gene. Signals are found by several computational algorithms based on different principles, but many focus on outliers due to probe failure. For example, Affymetrix MAS5 finds the signal as a weighted average between probe pairs, and RMA finds the signal by median polish of PM values.

In order to prevent loss of accuracy in signal data, it is desirable to recognize such problems and remove them from the data before analysis. A weighted average or median is only robust if outliers occur in both directions (ie positive and negative) with the same frequency. In practice this is rare because most of the problems generate outliers that reflect the cause. For example, a bright spot appears if the problem is caused by a fluorescent material, and a dark spot appears if the chip surface is damaged. These defect types affect the results by reducing the robustness of the calculation. If the target is cell data rather than genes, such defects directly affect the analysis, as is the case when analyzing mRNA variants processing. The microarray preparation problem presents a barrier to advanced analysis of GeneChip data.

As described above, microarray data often contains problems caused by non-uniform hybridization and dust contamination. In order to prevent degradation of analysis accuracy and resulting false positives, it is necessary to remove such problems prior to analysis.

The present invention seeks to find a problem as a local tendency of cell data in comparing each array to the ideal criteria for hybridization. Prior to data normalization, the cell at the location recognized as a problem is canceled. This cancellation does not affect the original distribution of array data. This is because cancellation is independent of signal intensity. As a result, the remaining data can be used for analysis.

In one aspect of the present invention, a method for detecting defects in DNA microarray data comprises:
Providing target DNA microarray data comprising a set of cell values obtained from a DNA microarray;
Providing reference data comprising a set of reference values, each reference value corresponding to each cell value of the DNA microarray data;
Obtaining a difference value between each cell value of the DNA microarray data and each reference value of the reference data;
Replacing each cell value of the DNA microarray data with each difference value to obtain a pseudo image;
Calculating a value representative of a small area corresponding to a predetermined number of cells in the pseudo-image based on a difference value of the predetermined number of cells, wherein Obtaining a set of values representing a small region by repeating the calculation while moving one cell at a time on the image;
Detecting one or more subregions containing outlier representative values based on a comparison between an expected normal distribution of the representative value set and a distribution of the representative value set, wherein the detected 1 One or more subregions contain defective cell values;
Consists of.

In one preferred embodiment, the target DNA microarray data and the reference data are standardized. Specifically, the cell value and the reference value are logarithmic values.

In one preferred embodiment, the reference value is a value representative of each cell obtained from a plurality of standardized DNA microarray data. In one aspect, the value representing each cell is a measure of central tendency, including the mean, median, and mode. In a preferred example, the representative value is a trimmed average, median, or weighted average.

A plurality of standardized DNA microarray data is acquired by the same type of DNA microarray as the target DNA microarray data. The plurality of DNA microarray data is, for example, 6 to 10 sets of DNA microarray data. In one aspect, a DNA microarray data set for the reference data is obtained based on the same tissue. In one aspect, a DNA microarray data set for the reference data is obtained based on a plurality of different tissues. In the latter, preferably a lot of DNA microarray data is prepared for different diverse tissues.

In one aspect, the size of the window, i.e. the small area, is 3x3 cells to 10x10 cells. In one preferred embodiment, the size of the window, i.e. subregion, is 5x5 cells. In one aspect, the value representative of the window is a measure of central tendency. Specifically, the representative value is a median, trimmed average, or weighted average.

In one preferred embodiment, a set of values representing a small region is standardized to obtain a set of indicators, with an indicator that exceeds a pre-determined critical limit based on an expected normal distribution of the set of indicators. One or more small areas are detected.

In one aspect, the index and the rejection limit are z scores.

In one aspect, the cell values of cells belonging to one or more detected windows are rejected.

The present invention relates to a computer program for causing a computer to execute a method for detecting a defect in the DNA microarray data.

The present invention relates to a computer program for causing a computer to execute a method for detecting a defect in the DNA microarray data and removing the defect.

The present invention relates to a computer readable medium storing the above program.

FIG. 1 is a histogram of the median standard deviation of the moving window. The mode was 0.31, which was larger than the expected value of 0.25. FIG. 2 shows the coincidence of two sets of ideal criteria for leaves analyzed at two different laboratories. The distribution of differences between hybridization and standards is shown. A straight line y = x indicates a normal distribution. The data is denser at the center of the plot, only the 2.3%, 0.1%, 0.003% data has a z-score greater than 2, 3, 4. The distribution of differences between hybridization and standards is shown. A straight line y = x indicates a normal distribution. The data is denser at the center of the plot, only the 2.3%, 0.1%, 0.003% data has a z-score greater than 2, 3, 4. The distribution of differences between hybridization and standards is shown. A straight line y = x indicates a normal distribution. The data is denser at the center of the plot, only the 2.3%, 0.1%, 0.003% data has a z-score greater than 2, 3, 4. Indicates the distribution of index values. Indicates the distribution of index values. Indicates the distribution of index values. Reproducibility in repeated experiments is shown. Combinations of experiments are shown in each column. Left: Original data, Center: Remaining data, Right: Canceled data. PM data (n = 10000) randomly selected from the indicated array pair is shown. The expected cancellation value is 2 windows. FIG. 6 shows multiple canceled data for two and 20 expected value windows (50 and 500 cells, respectively). FIG. 7 shows the standard deviation of the cell data difference in the reproducibility measurement. FIG. 8 shows the reproducibility of the data processed by the dChip package. The result using PM-only model is shown. The corresponding original data is shown in FIG. 5 (left). Left: remaining data, right: canceled data. PM data (n = 10000) randomly selected from the indicated array pair is shown. FIG. 9 shows the position of the canceled window on the chip. Four typical experimental results with the indicated expected values are shown. Upper left: hybridization with a relatively small number of cancellations, upper right: heterogeneous hybridization, lower left: regular shape with linear boundaries, lower right: clusters of symmetrical positions. FIG. 10 is a diagram illustrating generation of reference data. FIG. 11 is a diagram illustrating acquisition of a difference value between DNA micro data and reference data. FIG. 12 is a diagram showing scanning of the window on the pseudo image. FIG. 13 is a view showing the window of the present invention provided in the pseudo image. FIG. 14 is a flowchart illustrating the present invention.

A General Description of the Present Invention A method for detecting and eliminating problems caused by non-uniform hybridization and dust contamination will be described with reference to FIGS.

A hardware configuration for executing the method of the present invention is a computer device (not shown), and the computer device is an input device, an output device, a display device, a hard disk, a storage device, a computer-readable medium, or other storage means. Possible storage devices and a processor. Numerous data of the present invention including measurement data and calculation data are stored in a storage device. A number of calculations are performed by the processor. Optionally, a number of data including measurement data and calculation data may be displayed on the display device in a number of formats.

Target DNA microarray data is prepared (S1 in FIG. 14). Target DNA microarray data is a set of cell values. The target DNA microarray data is initially obtained as a set of signal intensities of DNA microarray probe cells. In one preferred embodiment, each cell value is a standardized logarithmic value (z score) obtained by taking the logarithm and z-normalizing the logarithmic value. Median-based standardization is used. The target DNA microarray data is stored in a storage device.

Standard data is prepared. The reference data is a set of reference values. Each reference value corresponds to each probe cell of the DNA microarray. The reference value is hypothetical data or reference data, and a set of reference values is typically obtained by calculation results. Ideally, the reference data is a set of expected virtual values that are the most average or most probable values. Typically, the reference value set is prepared as a z-score so as to correspond to the z-normalized cell value of the target DNA microarray data. The reference data is stored in the storage device.

As shown in FIG. 10, in one embodiment, the reference data is obtained from multiple standardized data sets (eg, 6-10 sets) obtained from the same type of target DNA microarray. Each reference value is obtained by calculating a value representative of each cell value of the plurality of standardized data sets. Typically, the representative value is a measure of central tendency including an average, median, and mode value. In a preferred embodiment, the representative values are trim average, median, weighted average. In one preferred embodiment, the reference data is obtained from a plurality of standardized DNA microarray data sets for the same tissue used for the target DNA microarray data. The organization used for reference data is not limited to the same organization. Alternatively, the reference data can be obtained from representative values (eg, trim average, median, weighted average) of a number of standardized DNA microarray data sets for a variety of different tissues. Median-based standardization can be used. If the target DNA microarray data is GeneChip data, it is desirable that the reference data is also created from a plurality of GeneChip data. If there is complete DNA microarray data with no defects or errors, one DNA microarray data can be used as reference data.

The reference data is not limited to that obtained based on actual measurement. Each reference value of the reference data may be the same value. Each reference value of the reference data may be zero. In this case, the difference value and the cell value of the target DNA microarray data are the same. Alternatively, the reference value may be a set of pseudo-random numbers with a small variance.

The standardization method of the target DNA microarray data and the reference data is not limited to the median-based method. Other approaches that allow standardized data (typically z-score) to be compared between the two can be used. For example, as a standardization, a three-parameter method (Konishi, T., Three-parameter lognormal distribution ubiquitously
found in cDNA microarray data and its application to parametric data treatment.
BMC Bioinformatics, 5: 5, 2004, incorporated herein by reference). At this stage, the background value by the three parameter method may be ignored. Other standardization techniques known to those skilled in the art can also be used.

As shown in FIG. 11, the difference value between each cell value of the target DNA microarray data and each reference value of the reference data is acquired (S2 in FIG. 14). At this stage, large difference values cannot be canceled. This is because a large difference value may be a biologically meaningful value. The difference value may include a ratio of each cell value to the reference value. The difference value is calculated by the processor and stored in the storage device. Each difference value corresponds to each cell of the DNA microarray.

A pseudo image is obtained by replacing each cell value of the DNA microarray data with a difference value (S3 in FIG. 14). That is, if the DNA microarray consists of M × N probe cells (also known as spots), the pseudo-image also consists of M × N cells, and each cell has a corresponding difference value. I have. As shown in FIG. 13, the pseudo image is an image in which each cell has a difference value Δz ₁ , Δz ₂ , Δz ₃ , Δz ₄ , Δz ₅ ,. The pseudo image may be displayed on the display device, but it is optional.

A window of a predetermined size corresponding to a predetermined number of cells of a small area in the pseudo image is prepared. As shown in FIG. 12, the window moves one cell at a time on the pseudo image, and sequentially calculates the representative value of each window based on the difference value of the predetermined number of cells to obtain a set of representative values of the window. (FIG. 14 S4). Although FIG. 12 shows an outline of a moving window, the moving (scanning) direction of the window may be any direction including a horizontal direction and a vertical direction on the pseudo image. As shown in FIG. 13, the window W moves horizontally on the pseudo image for each cell (W _t , W _{t + 1} , W _{t + 2} ...). The representative value of the window is calculated by the processor and stored in the storage device.

The window algorithm or window manipulation of the present invention is itself similar to the neighborhood or local operation in image processing. That is, paying attention to each cell C in the pseudo image, a representative value of a neighborhood (small region) including the cell C in the pseudo image is calculated. However, in the window operation according to the present invention, it is not necessary to update the value of the cell C with the representative value. In the present invention, the purpose of the window operation is to obtain a value representative of a small area in the pseudo image. Further, in the present invention, the cell C is not necessarily located at the center of the window, that is, the small area. The position of the cell C in each small area may be any position determined in advance in the small area.

The representative value of a window, i.e. a small area consisting of a predetermined number of cells, is a representative value (measurement value) including the mean, median and mode.
of central tendency). In a preferred embodiment, the representative value is a median, trimmed average, or weighted average. In one aspect, the size of the small area (window) is 3 × 3 to 10 × 10 cells. In one preferred embodiment, the small area (window) size is 5 × 5 cells. In FIG. 13, the window W corresponds to a small area of 5 × 5 cells. A representative value (for example, median) of 25 cells of the pseudo image is acquired, and the acquired value represents a focused window (small region). The representative value of the small area (for example, 5 × 5 cells) is calculated while moving the window W for each cell on the pseudo image. The representative value acquired for each window (small area) is stored in the storage device. The calculation of the representative value of the window does not require actual display of the pseudo image on the display device.

Here, even if the probe cells of the DNA microarray are adjacent probe cells of the DNA microarray, they are randomly arranged so as not to have any biological significance, so the window (from a predetermined number of cells) The value representing the small area should be normally distributed according to the central limit theorem.

A set of values representative of the window is standardized to obtain a z-score of the representative value and compared with the rejection limit value (S5 in FIG. 14). The z score of the representative value serves as an index for comparison with a preset rejection limit value, that is, a cutoff value (these are prepared as a z score). When standardizing the representative value, it is necessary to obtain the width of the distribution of the representative value set. The width may be obtained indirectly from the width of the distribution of the difference value set. The width of the difference value distribution is acquired, and the acquired width is corrected by the compensation coefficient. The compensation factor may vary depending on the representative value. For example, when the representative value is an average, the compensation coefficient is 1√n. The compensation coefficient may be obtained using a simulation such as a Monte Carlo method. The width of the distribution of the representative value set may be obtained directly from the distribution. For example, IQR (Interquartile Range) or MAD (Mean Absolute Deviation) may be used as the distribution width. The width may be obtained from the slope of a linear regression that approximates the Q-Q plot of the representative value set. The width may be determined in advance based on various actual measurements. Specifically, a set of widths acquired from various measurements may be prepared, and the mode value of this set may be used as a predetermined width. Standard deviation may be used for standardization of the representative value set.

One or more windows containing possible defective cell values are detected by comparing each index value with a rejection limit value predetermined based on a predicted normal distribution of the index. A window whose index value exceeds a predetermined rejection limit value is regarded as a small region including a defective cell value (S6 in FIG. 14). The rejection limit value is determined in advance as a z-score by the operator according to a normal distribution. For example, when it is desired to cancel two windows from an ideal normal distribution, 4.61 can be determined in advance as the z score. However, if a rejection limit value of 4.61 is set, more than two windows are usually detected.

All the cell values of the detected window or windows are canceled or rejected (S7 in FIG. 14). For example, if a window corresponds to 25 cells (5 × 5), 25 cells are discarded per window. If the two detected windows are separated, 50 cell values are discarded. If the two detected windows overlap by 5 × 4, the 5 × 6 cell area becomes the discard area. The data remaining after the rejection and / or the rejected data may be displayed on the display device in various formats such as a chart or an array image of the probe cell. Or you may correct | amend instead of canceling the cell value of the detected 1 or several window.

In an exemplary embodiment of the invention, a predetermined size window is used. However, in other embodiments, different sized windows may be used. For example, one window corresponding to 3 × 3 cells and another window corresponding to 7 × 7 cells are used to integrate the results. That is, the window is inspected based on a set of two representative values. The detected windows are compared between the results of different sized windows and the overlapping cells are canceled. For example, if the detected windows completely overlap, 3 × 3 cell data is rejected.

In the following chapter, we describe an algorithm for finding and removing problems. Problems are identified from biological effects by means of data distribution. The algorithm is based on several demonstrable assumptions, and the validity of these assumptions was tested in the results section using non-limiting exemplary GeneChip data. The effectiveness of the algorithm and the effect of data cancellation were tested using GeneChip data obtained from a series of experiments. The algorithm greatly improved the reproducibility of the measurement and showed that only a small amount of defect-free data was removed.

B Specific Methods and Experiments B-1 Algorithm The proposed method for identifying the microarray problem is hereinafter referred to as a parametric scanning algorithm and will be described below.

The present invention provides a parametric scanning algorithm that detects defects based on data distribution characteristics. The entire cell data is scanned using a window algorithm, and windows with index values that exceed a critical value (also known as a threshold) are recognized as defects and are removed from the array data. The index is determined from the difference between the target and the ideal criterion obtained as the trimmed average in multiple experiments and represents the statistical center of the multiple differences in each zone. The threshold is derived at the screening level specified by the operator, but has only a limited effect on data cancellation.

A reference ideal array is selected and an index representative of the size of the unique area at each chip is determined. An area having an index larger than the threshold value compared to the reference is recognized as a problem area.

The standard is obtained as a set of trim averages in multiple hybridizations. The experimental results are normalized by taking the logarithm of each median value (including PM and MM cells). A trimmed average of the data for each cell in the array is calculated and the resulting set of averages is taken as an ideal criterion for hybridization. If the average is calculated using a sufficiently large number of array data, the value can be considered stable and suitable as a reference. No special distribution is expected in the ideal standard.

The difference between simply standardized array data and the reference is obtained for each cell. These differences will represent both biological responses and experimental noise. It is predicted that the difference distribution is generally normal. This is because the logarithm of appropriately measured and standardized biological changes follows a normal distribution. The difference is z-normalized using a robust estimator of the distribution parameters and the distribution is checked on a quantile-quantile (Q-Q) plot. Difference normalization is for analysis of difference features and is an optional step in the present invention.

The indication is obtained using a median of z-standardized values of differences in a plurality of neighboring cells corresponding to a plurality of cells in a window on the array. The matrix of differences is rearranged to reflect the physical alignment of the chips, and data is collected by moving windows that simulate scanning across the pseudo image of the chips to determine the median. The window median is robust to biological responses. This is because adjacent cells on the chip have no biological relationship. In contrast, experimental problems of hiding and adding signals in the window can affect the window median. The window median will follow a normal distribution in a strict sense, according to the effects characterized by the central limit theorem. This model does not expect a specific distribution for the problem, but the affected window will produce outliers in the normal distribution of the matrix median.

The indicator is obtained by standardizing the matrix median. Standardization is difficult. This is because the width of the matrix median distribution is not robust to the problem. In fact, the width can increase as the number of problems increases. If the distribution is simply z-normalized, the number of recognized problems will be reduced. However, this effect can be easily avoided by finding the width from the width of the distribution of the differences between the cells. As a rule, the current study of an average of 25 cell windows predicted a width of 0.25. Here, the width of the cell difference distribution is robust to the problem. Because many problems generate outliers, they do not affect the distribution of the central quantiles. In practice, the distribution of cells is not a perfectly normal distribution but has a long tail due to systematic additive noise in the data. However, the appropriate width can be estimated robustly from the appropriate quantiles. As a result, the influence of the problem can be eliminated by estimating the width of the index distribution according to the cell distribution. Systematic noise and hybridization problems can change the compensation factor 0.25 to somewhat larger values. In this description, the constant 0.31 obtained from the actual measurement mode value is used, which is smaller than many other values that would have been affected by many problems (FIG. 1). All indicators are adjusted or standardized by dividing by this constant.

The threshold is derived from the test level determined by analysis prior to manipulation, as is the screening level in other statistical tests. The parametric properties of data handling make it possible to estimate how many indicators are more (less) from 500,000 results. The program asks the operator how many windows to predict. If the array is problem free, the predicted number of windows is recognized by a random neighborhood of biological responses on the chip. In practice, the affected indicator will not follow a normal distribution and will take a value above the threshold.

B-2 Program Parametric scanning method program is provided in the form of function R. The function requires the library “affy” available from BioC (http://www.bioconductor.org/). Outsourcing services are available in the data standardization part (http://www.super-norm.com).

B-3 Data Source and Data Processing Arabidopsis GeneChip data was obtained from TAIR (http://www.arabidopsis.org/index.jsp). The leaf data of the two research groups used to compare the ideal criteria for hybridization are the 15 arrays of Rosetta leaves (Schmid, M., Davison, TS, Henz,
SR, Pape, UJ, Demar, M., Vingron, M., Scholkopf, B., Weigel, D., and
Lohmann, J., A gene expression map of Arabidopsis development.Nature Genetics
37: 501-506, 2005) and 18 arrays of control plants (http://www.arabidopsis.org/index.jsp) 0.5-5 days after infection experiment of Dr. F. Ausubel group . Human data was obtained from the public domain resource RCAST, University of Tokyo (http://www.genome.rcast.u-tokyo.ac.jp/normal/). The PM data of the array is a three-parameter method (Konishi, T., Three-parameter lognormal distribution ubiquitously found
in cDNA microarray data and its application to parametric data treatment.BMC
Bioinformatics, 5: 5, 2004).

C Results C-1 Hypothesis Verification C-1-1 Stability of Hybridization Criteria This approach compares each data with the ideal criteria for hybridization that should represent a stable pattern of sample tissue. If the pattern is really stable, the pattern will match other reference patterns determined using different data sets for the same tissue. To confirm this agreement, we compared criteria obtained using data from two study groups. Both groups determine the leaf transcriptome, one is part of the plant atlas and the other is the control for the infection experiment. The criterion was obtained as a trimmed average of logarithmic data median-normalized using the median. The results were compared on a scatter plot for 1000 corresponding cell data (FIG. 2). Agreement between laboratories was confirmed. Similar agreement was presented in comparisons between several other laboratories and within laboratories. Such a match is not a coincidence. For example, criteria obtained from different tissues have different tendencies and appear as wider scatter in the plot. Such a trend indicates the organization's dependence of the standards, which should be noted in the actual use of the program.

C-1-2 The proposed method of normality of the difference between the array data and the reference assumes that the difference between each data and the ideal reference for hybridization will generally be normally distributed. This assumption was confirmed by the QQ plot of the data distribution. The distribution has a long tail that will reflect systematic additive noise. However, all distributions are in agreement with the theoretical values from -1.5 to 1.5 (FIGS. 3A, 3B, 3C), indicating that over 85% of the data follows a normal distribution. Since problems and noise affect the distribution, hybridizations with many problems are more closely matched as seen in FIG. 3C (ATGE 14C).

C-1-3 Normality of Index Distribution The present invention also assumes that the index derived from the moving window median will be normally distributed in the absence of major problems. This assumption was confirmed by the QQ plot (FIGS. 4A, 4B, 4C). The observed distribution is almost normal as predicted by the central limit theorem. A standard deviation of 0.31 determined from many hybridizations (FIG. 1) can better compensate for the width of the distribution and the slope of the plot (FIGS. 4A, 4B). As expected, the width of the distribution increased with the severity of the problem (FIG. 4C ATGE 14C).

C-2 Confirmation of this method C-2-1 Improvement of reproducibility in repeated experiments If parametric scanning effectively removes problems from data, the variation seen in repeated experiments should be reduced. This effect was checked using a set of repeated experiments based on Arabidopsis leaves (http://www.arabidopsis.org/index.jsp). Before and after data cancellation, PM data was standardized using a SuperNORM® algorithm based on a three parameter approach. The resulting z-scores were compared on the scatter plot (FIG. 5), indicating that the proposed method can reduce the diffusion seen in the plot (FIG. 5 left) and obtain the expected reproducibility (FIG. 5). Center).

In other statistical tests, parametric scanning has also shown cancellation of clean, defect-free data. In a sense, this is the cost required to find something by means of statistical testing. However, in this algorithm, the number of canceled clean data is not large. The characteristics of the canceled data were checked from the reproducibility of the experiment (right of FIG. 5). The number of data on the plot increases as the quality of hybridization degrades. The canceled data did not concentrate narrowly with respect to the y = x line, but was dispersed instead (right in FIG. 5). A match is seen only when a lot of cell data is canceled (lower right in FIG. 5), and the data concentrated on the y = x line is only a limited part of the canceled data.

Some of the experimental examples shown in FIG. 5 have large variations. These experimental examples are no exception in many tests. FIG. 6 compares cancellation data under different expected values. The data sources in FIG. 6 are as follows: □ (Schmid, M., Davison, TS, Henz, S.
R., Pape, UJ, Demar, M., Vingron, M., Scholkopf, B., Weigel, D., and
Lohmann, J., A gene expression map of Arabidopsis development. Nature
Genetics 37: 501-506, 2005); ○ (http://www.arabidopsis.org/index.jsp); △ (Ge, X., Yamamoto, S., Tsutsumi,
S., Midorikawa, Y., Ihara S., Wang S., Aburatani H., Interpreting expression
profiles of cancers by genome-wide survey of breadth of expression in normal
tissues, Genomics. 86: 127-141, 2005.) It can be seen that extreme experimental examples are not taken from outliers.

The improvement in reproducibility was checked by reducing the standard deviation of the difference in z-scores of the corresponding PM cells in the pair hybridization. In order to minimize the effects of additive noise and measurement saturation, standard deviations were calculated using standardized values (0-1). The effect was checked on a scatter plot (FIG. 7) and parametric scanning was shown to reduce the standard deviation in the difference between the obtained z-scores.

C-2-2 Comparison with other algorithms This method was evaluated by processing the same set of arrays rather than new experimental data using other methods in the dChip package. All spikes and outliers recognized by dChip were canceled using the PM-only model, and the data was standardized in the same way. As shown in FIG. 8, dChip shows lower reproducibility (left side of FIG. 8) and low detection capability. This does not mean that dChip retains the defective data, but cancels the entire set of cells for some genes (0.004-6.4% of the total) None of the genes were completely canceled in the parametric scan. In such genes, no information is retained for analysis. Table 1 shows the number of genes in which the entire cell was canceled (unit:%).

C-2-3 Sensitivity of Threshold Parameter The number of data actually canceled in each hybridization did not clearly depend on the test level threshold parameter determined by the operator. The number of rejected data is greater than the expected number estimated from the threshold parameter, even if 50 cells in two windows are expected, a quarter of the total number of cells (tens of thousands) Reached up to one. However, the number of cells that were canceled did not increase by a factor of 10, even though it was expected to increase from 2 to 20. The relationship between the expected number and the actually canceled number becomes weaker as the number of canceled data increases. Processing data acquired from three different laboratories suggests a stable relationship between windows canceled with two expected numbers (FIG. 7). It is noted that the expected number appearing at (1.7, 2.7) in the plot satisfies the inferred relationship (FIG. 7).

The number of data canceled can depend on the quality of hybridization. When a major problem is found, more cancellations are observed (Figure 5). Canceled windows often form clusters at the chip, suggesting that there is a single cause within the cluster (FIG. 9). Such a cluster is seen regardless of the value of the expected parameter. The frequency and area of cancellation varied between data obtained from different laboratories (FIG. 6). The data at one particular laboratory (Δ in the figure) is clearly larger than the data from other laboratories. Many clusters represent polishing or non-uniform hybridization of the chip surface. Differences in problem frequency tend to be due to protocols and skills in wet experiments that will vary with laboratory and preparation time. These problems are accentuated by high index values and produce many canceled windows in the case of severe defective cells even when the expected number is smaller.

The above results show that parametric scanning is less sensitive to the value of the expected value parameter, i.e., the proposed method is accurate with respect to problem detection. Such insensitivity means the objectivity of this algorithm. This is because the threshold is the only parameter that is affected by the operator's choice.

Based on the above view, the proposed method is encouraged for actual use in all GeneChip expression data prior to standardization. The assumptions in this approach have been verified through analysis of data distribution, indicating that only arbitrary parameters can have a limited impact on the results. Furthermore, through many additional experiments (not shown), it has been found that parameter scanning techniques are effective in eliminating hybridization problems. The validity of this approach can be checked in every analysis by providing the necessary data for the checking process provided by the software. The number of canceled data is always higher than expected, suggesting that most hybridizations have some problem.

The detected problems have a pattern that indicates surface polishing, non-uniform hybridization, and errors in the fabricated cell structure. A symmetrical cluster (lower right of FIG. 9) surrounding the center of the chip is identified as polishing artifacts. In such cases, the signal in the affected area is always low and is therefore insensitive to the expected value. When the degree of surface polishing progresses, a common donut-like cluster pattern is formed. In contrast, amorphous clusters tend to show heterogeneous hybridization. Within the cluster, the data will tend to increase or decrease, producing a scatter plot with experimental reproducibility (Figure 5). Such inhomogeneities are thought to be derived from several causes, and some unique areas are insensitive to the expected values, but others are not (Figure 9 ATGE_14_C). The difference in sensitivity corresponds to the difference in defect size. Defects detected as smaller clusters or isolated windows can be those formed by dust. Again, some features are specific, others are not. Errors in the chip structure are perceived as repeated clusters in the same part of multiple chips, often forming a regular shape surrounded by straight lines. Many of these defects are not a problem, they can be caused by problems, but they control the cells that are designed and placed on the chip and appear on all chips in the same batch number (ie, the same production lot) . Such problems can be caused by product errors that are not detected in quality control and can lead to serious problems. In the case shown in FIG. 5, the large upward diffusion is due to such defects (lower left of FIG. 9).

The proposed method reduces false positives in microarray data analysis. Such errors are not unique to microarray analysis, and multiple tests performed using microarrays increase the severity of the error. Multiplicity has been recognized over a wide range of microarrays and other post-genome analyzes that produce characteristically different analytical targets compared to conventional methods that measure only a limited number of gene products. In excessive multiple comparisons, a large number of false positives hinder analysis and create observational discrepancies between and within laboratories. For example, if a 1% probability type-I error is allowed, 500,000 double-sided tests will generate 10,000 errors. Ignoring the hybridization problem will greatly improve this expectation (FIG. 5). Such issues will also affect data normalization and gene summary data. Therefore, hybridization problems need to be detected and eliminated before standardization. The proposed method can extract clean data from an area free from hybridization defects, standardize the data remaining after cancellation, and use it for further analysis. The remaining data set showed good agreement with the corresponding pair in the reproducibility experiment (middle of FIG. 5). The overall experimental cost can be reduced as compared to the ad hoc technique of canceling the genes in the array and / or the entire array.

R programs can be influenced by tissues in finding the ideal criteria for hybridization. That is, the criteria can vary according to cell differences in the sample. Such an effect can occur when processing a small number of arrays with a large number of arrays in different tissues. Also, it is not recommended to process data using fewer than four arrays. This is because the standard is not considered stable. The stability of the criteria can be checked using the technique shown in FIG. 2, and the effect of the organization is noticed by a significant increase in the number of cancellations without generating the canceled window cluster seen in FIG. be able to. Such a problem can be prevented by finding a reference independent of the recognition process. In practice, two alternative approaches can be employed to find the ideal criteria. Use a randomly selected sample from a variety of tissues in many arrays, or find a tissue-specific criterion and use it in the corresponding array.

The present invention can be used for microarray analysis for detection of nucleotide hybridization including, for example, measurement of mRNA levels and discovery of SNPs.

Claims (17)

Providing target DNA microarray data comprising a set of cell values obtained from a DNA microarray;
Providing reference data comprising a set of reference values, each reference value corresponding to each cell value of the DNA microarray data;
Obtaining a difference value between each cell value of the DNA microarray data and each reference value of the reference data;
Replacing each cell value of the DNA microarray data with each difference value to obtain a pseudo image;
Calculating a value representative of a small area corresponding to a predetermined number of cells in the pseudo-image based on a difference value of the predetermined number of cells, wherein Obtaining a set of values representing a small region by repeating the calculation while moving one cell at a time on the image;
Detecting one or more subregions containing outlier representative values based on a comparison between an expected normal distribution of the representative value set and a distribution of the representative value set, wherein the detected 1 One or more subregions contain defective cell values;
A method for detecting defects in DNA microarray data.   The method according to claim 1, wherein the target DNA microarray data and the reference data are standardized.   The method according to claim 1, wherein the cell value and the reference value are logarithmic values.   The method according to claim 1, wherein the reference value is a value representative of each cell value acquired from a plurality of standardized DNA microarray data.   5. The method of claim 4, wherein the plurality of standardized DNA microarray data is obtained from the same type of DNA microarray data as the target DNA microarray data.   The method according to claim 4, wherein the plurality of standardized DNA microarray data for the reference value is acquired based on the same tissue.   The method according to claim 4, wherein the plurality of standardized DNA microarray data for the reference value is acquired based on a plurality of different tissues.   The method according to claim 4, wherein the value representative of each cell is a representative value (measure of central tendency).   9. The method of claim 8, wherein the measure of central tendency is a median, trimmed average, or weighted average.   The method according to any one of claims 1 to 9, wherein a size of the small region is 3 cells x 3 cells to 10 cells x 10 cells.   The method according to any one of claims 1 to 10, wherein the value representative of the small region is a representative value (measure of central tendency).   The method of claim 11, wherein the measure of central tendency is a median, trimmed average, or weighted average. The detecting step includes
Standardizing a set of values representative of the small region to obtain a set of indicators;
Detecting one or more subregions with an index that exceeds a predetermined rejection threshold based on an expected normal distribution of the set of indices;
The method of claim 1 comprising:   The method of claim 13, wherein the index and the rejection limit are z scores.   The method according to any one of claims 1 to 14, further comprising the step of rejecting cell values of a plurality of cells belonging to the detected one or more subregions.   A computer program for causing a computer to execute the method according to claim 1. Providing target DNA microarray data comprising a set of cell values obtained from a DNA microarray;
Providing reference data comprising a set of reference values, each reference value corresponding to each cell value of the DNA microarray data;
Obtaining a difference value between each cell value of the DNA microarray data and each reference value of the reference data;
Replacing each cell value of the DNA microarray data with each difference value to obtain a pseudo image;
Calculating a value representative of a small area corresponding to a predetermined number of cells in the pseudo-image based on a difference value of the predetermined number of cells, wherein Obtaining a set of values representing a small region by repeating the calculation while moving one cell at a time on the image;
Detecting one or more subregions containing outlier representative values based on a comparison between an expected normal distribution of the representative value set and a distribution of the representative value set, wherein the detected 1 One or more subregions contain defective cell values;
Discard cell values of all cells belonging to the detected one or more sub-regions;
A method for detecting and removing defects in DNA microarray data.

Images