Recent rapid advances in molecular biology have created more demand for high volume testing based on the need to screen ever larger compound libraries, validate ever increasing numbers of genetic markers and test ever more diversified patient populations. This has led to the development of new array formats, particularly for nucleic acid and protein-protein interaction analysis, which permit parallel processing by performing requisite assays in a “multiplexed” format.
Conventionally, such assays are performed by producing arrays of nucleic acids and antibodies by way of “spotting” or “printing” of aliquot solutions on filter paper, blotting paper or other substrates. However, notwithstanding their widespread current use in academic research for gene expression analysis and protein profiling, arrays produced by spotting have shortcomings, particularly in applications placing high demands on accuracy and reliability and calling for the analysis of a large volume of samples at high rates of throughput In another more recently developed technique, spatially encoded probe arrays are produced by way of in-situ photochemical oligonucleotide synthesis. However, this technology is limited in practice to producing short oligonucleotide probes—as a result, alternative technologies are required for the production of cDNA and protein arrays. In addition, in-situ synthersis precludes rapid probe array customization given the time and cost involved in the requisite redesign of the photochemical synthesis process.
In addition to these inherent difficulties relating to assay performance, spatially encoded arrays produced by spotting or in-situ synthesis conventional methods generally require specialized optical scanning instrumentation to extract data of useable quality. Commercial systems available for this purpose rely on confocal laser scanning—a slow process—even at the typically modest spatial resolution of ˜5:m—which must be repeated for each signal color.
In order to resolve many of the problems associated with diagnostic and analytical uses of “spotted arrays” of oligonucleotides and proteins (as outlined in “Multianalyte Molecular Analysis Using Application-Specific Random Particle Arrays,” U.S. application Ser. No. 10/204,799, filed on Aug. 23, 2002; WO 01/98765), arrays of oligonucleotides or proteins arrays can be formed by displaying these capture moieties on chemically encoded microparticles (“beads”) which are then assembled into planar arrays composed of such encoded functionalized carriers. See U.S. patent application Ser. No. 10/271,602 “Multiplexed Analysis of Polymorphic Loci by Concurrent Interrogation and Enzyme-Mediated Detection,” filed Oct. 15, 2002, and Ser. No. 10/204,799 supra.
Microparticle arrays displaying oligonucleotides or proteins of interest can be assembled by light-controlled electrokinetic assembly (see, e.g., U.S. Pat. Nos. 6,468,811; 6,514,771; 6,251,691) or by a direct disposition assembly method (previously described in “Arrays of Microparticles and Methods of Preparation Thereof,” filed Jul. 9, 2002; Ser. No. 10/192,352).
To perform nucleic acid or protein analysis, such encoded carrier arrays are placed in contact with samples anticipated to contain target polynucleotides or protein ligands of interest. Capture of target or ligand to particular capture agents displayed on carriers of corresponding type as identified by a color code produces an optical signature such as a fluorescence signal, either directly or indirectly by way of subsequent decoration, in accordance with one of several known methods. The identity of capture agents including probes or protein receptors (referred to herein sometimes also collectively as “receptors”) generating a positive assay signal can be determined by decoding carriers within the array. See U.S. patent application “Multianalyte Molecular Analysis Using Application-Specific Random Particle Arrays” Ser. No. 10/204,799.
These microparticle (“bead”) arrays generally contain a number of spectrally distinguishable types of beads within an area small enough to be viewed by a standard optical microscope. The small footprint and high signal contrast permit “instant” direct (“snapshot”) multicolor imaging of the entire array under a microscope, thereby obviating the need for confocal laser scanning.
Analyzing the images recorded from such random encoded bead arrays involves several steps and issues, as described in U.S. patent application “ANALYSIS, SECURE ACCESS TO, AND TRANSMISSION OF ARRAY IMAGES” Ser. No. 10/714,203. To identify individual receptors scoring positive (and negative) in a given assay, the assay image comprising a pattern of signal intensities recorded from individual beads within the array is compared to a decoding image, taken before (or after) the assay. As described in detail in Ser. No. 10/714,203, this process is performed in an automated fashion by constructing a decoding map comprising groups (“clusters”) of beads of the same type (displaying the same type of receptor). A rapid process for constructing the decoding map (also referred to as the “cluster map”) is disclosed herein which invokes an adaptive template.
Determining the intensities of the individual beads in the assay image allows quantitation of results, for example, in gene expression or protein biomarker analysis. To determine bead signal intensities, the image is divided into segments by constructing a grid such that each field in the grid contains at most one bead—the intensities of the grid field intensities are then recorded. Because the grid finding step is performed on an image that is aligned within the viewing field, a first processing step is that of image rotation.
While the grid finding step can be performed on a bright field image (as described in Ser. No. 10/714,203), it is advantageous to eliminate the recording of a bright field image by instead finding the grid directly for decoding and assay images, typically comprising fluorescence images. In addition, the fluorescence image typically has a substantially higher contrast, that is, the ratio of signal to background. However, several problems must be overcome. First, beads displaying receptors which do not produce an assay signal will not be visible in the fluorescence image, and this generally will lead to the corruption of grid edges, thereby affecting the step of image rotation which relies on the accurate determination of the original image orientation by measuring the orientation of edges. Second, signal sources such as fluorescent beads may be randomly displaced from the center of each grid field, a fact which can affect the intensity recorded from that field.
Disclosed are methods of automating the process of analysis in a manner that increases the confidence in the results of the analysis of assay data recorded in the form of images of arrays of signals representing the strength of receptor-ligand interactions; and more particularly, where different receptors are associated with different encoded microparticles (“beads”), and signals from “positive” beads represent the amount of ligand captured by the corresponding receptors displayed on such beads. The methods herein allow:
The methods and algorithms set forth below address the problem of segmenting an image of an array of discrete signal sources which generally will be misaligned and will be composed of randomly interspersed “bright” and “dark” sources.
Image Misalignment—As illustrated in
Randomly Placed “Dark” Sources—In the processing of images of signals from arrays, a critical step is to generate a grid having a number of fields, where a field contains a distinct signal source. Fluorescence images such as those shown in the examples herein are recorded using color filters and generally have high contrast which significantly exceeds that of a “brightfield image,” which, as described in Ser. No. 10/714,203, is used for decoding.
However, the problem in drawing a grid with fields around individual signal sources in an array of signal sources is that, in general, not all signal sources will be “on” when the grid is being drawn. For example, where the signals are fluorescence signals indicating a positive assay score (i.e., a receptor-ligand interaction), and signal sources are distinct beads in an array of such beads, only those bead scoring positive will be visible, the remainder, randomly distributed throughput the array, will be invisible (see
I.1 Segmentation of Images of Source Arrays of Hexagonal Configuration—Certain embodiments of random encoded arrays use a set of mechanical traps to confine individual beads. In such embodiments, a hexagonal bead layout (
If an array with horizontal lines is precisely aligned with respect to the horizontal axis of the frame (that is, with an alignment angle of 0°, as in
The main steps of finding a grid in an image of an array of fluorescence signal sources are as follows:
The goals of image processing are:
Reduce noise sources by blurring the image;
Find the edges of the image;
Thin the edges by erosion;
Enhance the edges; and
Convert the image into a binary image (i.e., each signal source position is either “on” or “off”).
The input parameters for finding the array orientation angle are:
The binary image itself;
The signal source (bead) size;
Distance between two neighbor beads, or Bead Distance;
Distance between rows, or Row Distance.
An array can be placed in a grid using one set of lines along the row direction and two sets of lines along the column direction (in order to account for beads which are offset from the lines). Given small variations in chip manufacturing and array imaging, actual distances between beads and distances between lines of beads generally will vary slightly, i.e., one cannot simply insert grid lines using “design” dimensions and these lines must be constructed explicitly from the image itself.
Using the binary image, one searches for positive signals, and then determines if a “blob” represents a signal source (such as a bead). A blob is designated as a source if its dimensions in both the x and y direction are close to the expected bead dimensions. When a bead is found, the position of its center is determined and stored in a list. As described herein further below, center positions can be refined as necessary in order to enhance the accuracy of the intensity value recorded from the grid field containing any given bead. From the list of bead center positions, one builds a list of all hexagon lines in the image by examining the relationship between every bead and the existing hexagon lines.
There are four types of possible relationships between beads and hexagon lines:
After constructing all hexagon lines, they are partitioned into groups according to misalignment angle, that is, the angle from the horizontal (and/or from the vertical) using a preset variance of each group (corresponding to acceptable variance in the orientation of the array) of +/−15°. There should be at least three groups of lines. This is so because, if all hexagon lines are included, the lines within each group can be oriented, respectively, at or near 0°, 60° and 120°. From the mean orientation angles of the set of lines in each of the three groups, one can determine the offsets from 0°, 60°, and 120°, respectively, and from this in turn the alignment of the array. To enhance efficiency and accuracy, the algorithm provides that any two hexagon lines that lie along the same line, be connected and merged into one longer line.
The partitioning function developed as part of this algorithm takes a set of data (the points that make up a line) and a preset variance for each group is determined, and the deviation in alignment of any line from the mean must be within the variance in order for that line to be included in the appropriate group. Below is a pseudocode description of this function.
After sorting groups based on their size, the first three (largest) groups are designated as major line groups. Among those three major line groups, one should find one group of lines either close to 90° or close to 0°. That group is designated as a row group. The row group type determines whether the array has vertical lines of beads or horizontal lines of beads (see
The bead occupancy in a fluorescent image may vary, and if the occupancy is too low, it may be difficult to construct the grid. One can simulate the effect of reducing occupancy, by starting with an image with sufficient bead occupancy to conduct an orientation calculation. The simulation then throws out the beads sequentially in a random manner and re-calculates orientation. The orientation angle is used to determine the robustness of the algorithm since it is a critical step in the algorithm.
As illustrated in
For safety and precision, there should be at least 1000 beads for the successful segmentation of a grid when using a fluorescent image. In other words, the occupancy of image should be at least 25% to make algorithm work. Requiring 25% occupancy demonstrates that the method disclosed herein is indeed robust.
To improve the accuracy of the process of locating array boundaries, the method disclosed here invokes a boundary box surrounding an array and then computes its orientation angle again and finds the center of the array.
In finding the box, one first looks for the first and the last rows in the array.
The steps of finding them are:
1. Project all bead centers along the direction of rows;
Partition results of all projected (X, Y) positions into data groups; If the rows are vertical, partition all projected x values into groups. If rows are horizontal, partition all projected y values into groups. Each group has a variance size of +/−(Projected Row Distance/3), where Projected Row Distance=Row Distance/cos (Orientation Angle).
Sort partitioned data groups, and find two boundary groups;
The conditions for a boundary group are:
(i) The group is at one end and has at least 4 bead centers;
(ii) Its distance to the next group is close to the Projected Row Distance. These two conditions prevent one from selecting, by mistake, groups of noise as boundary groups.
On then checks the rest of the groups in projected rows between the two boundary groups and retains groups whose distance to their neighbors is close to the Row Distance.
To illustrate the above steps, an array is shown in
After finding row boundaries, one separates bead centers into two groups. One group contains beads from rows of odd number counted in an ascending order of row positions, overlaid by dashed lines in
After the process, one has 4 boundary groups, each perpendicular to the direction of the rows. Within those 4 boundary groups, one retains the two groups that are column boundaries for all columns, as illustrated by the two vertical lines on the perimeter in
Transforming Array Orientation into Alignment with External Frame
The next step is to transform the array so that the orientation angle is 0° to the horizontal. One first composes a rectangle around the image with two row boundaries and two column boundaries. Given the orientation angle (calculated above) and the center of the boundary rectangle (which is determined) one generates a transformation matrix to rotate the fluorescent image by the desired amount.
To find grid lines which represent the boundaries of the fields between beads, one uses the bead centers determined in previous steps. But first, one transforms all bead centers inside the boundary rectangle by the same transformation used for rotating the array, to place them where desired.
Using transformed bead centers, one performs similar steps to those used in finding boundaries of the array before transformation. The steps are:
After finding all row positions and their grid lines, one finds two sets of column grid lines. One first separates transformed bead centers into two groups. One group contains all beads from rows of odd numbers, counted in an ascending order of row positions, while the other group contains all beads from rows of even number.
For each group, one applies similar steps as those for finding grid lines for rows, but with data representing the different dimensions and distances between columns. At the end of the process, one has a set of row grid lines and two sets of column grid lines (the two sets of lines are offset from each other by half a grid width because of a staggered layout of the columns, (as shown in
Once the grid is established for a fluorescent image, it should be aligned with fluorescent images from other color channels, to avoid having to establish a grid de novo for each fluorescent image.
Fluorescent images often reside in several color channels. For instance, an array on a single substrate can include green, orange and UV fluorescent images. Because of optical variations, there are shifts in the fluorescent images between different color channels. The shifts in the images are in the x, y directions, and are the shifts do not appear when the images are rotated. The images can be aligned with each other to eliminate these shifts.
The algorithm below can be used to find the grid for one fluorescent image, and use the grid information for alignment with the other fluorescent images. The conventional algorithm used in grid-finding of a fluorescent image is based on finding the intensity profile for an array of fluorescent images. As noted above, and as illustrated in
The edge profile of the array is determined by examining the difference between pixels near the edges of the array (the signal from each bead is made up of several pixels). No matter whether dark beads or lit beads are viewed, a peak always is present at the edge (see
The pseudocode for the alignment algorithm is as follows:
Once the grid has been found and fluorescence images have been aligned, one can extract the intensity values from the different color channels to form a multi-dimensional scatter plot of signal source intensities collected from two or more color channels. The scatter plot is converted into a signal source map, one example being a decoding map, by an automated clustering algorithm described below.
A signal source (e.g., a bead), when it is smaller than the grid field, can be displaced from the center of the field Such random lateral shifts in bead position within the grid fields can be corrected for in order to enhance the accuracy of the analysis.
One can view the intensity distribution across a single source such as bead within a given grid field through a window (“mask”) that is placed onto each grid field, the mask generally chosen to be smaller than the grid field. If the bead is not located in the center of the grid field, viewing the intensity through the window will reduce the mean intensity computed from the set of discrete intensity readings for each mask pixel. As a result, the assay image or decoding image, and the corresponding source maps, will display regions of “false” reductions in intensity.
The grid field with a window is illustrated in
Note that the boundary region of the grid field in
There is another pre-condition in using the window. If the bead intensity in a grid field is weak, the mean intensity at the center of the grid field will be used as the actual intensity. The reason for this adjustment is that weak beads are subject to interference from noise, especially near the grid field boundaries, and the center region of a grid field provides more consistent results.
A bead is determined to be weak if its intensity is below a dynamic threshold. The chip background intensity is first calculated by measuring mean intensity in the region outside the bead array. If the bead intensity is determined to be less than the background intensity by a specified percentage, the bead will be treated as a weak bead, and its intensity will only be measured at the center of the grid field.
A critical step in the decoding of a signal source map in the form of an assay image is the construction of a decoding map by clustering of beads of identical type. A decoding map (also cluster map or bead map) is a multi-dimensional representation of groups of intensities representing the level of several encoding dyes within each bead; these intensities are recorded in two or more color channels.
Given a scatter plot of dots, each such dot representing an encoding intensity, a clustering operation groups dots located close to one another within the scatter plot into clusters which are assigned a number label. A decoding map contains, therefore, for each cluster, a position (e.g., that of the centroid) and a label.
To achieve high throughput in the decoding step, automatic clustering is essential. In automatic clustering one starts with a conventional decoding map and from that, generates a project decoding map template, with clusters in the same positions as the original map, but without labeling of clusters.
Converting a Decoding Map into an Image Format for Automatic Clustering
In automatic clustering, one starts with a decoding map and converts it into an image format, and applies a watershed algorithm to that image in order to group beads by increasingly higher encoding intensities within the image. Within the image format, one can also apply image processing or filtering methods, for example to remove dots in the scatter plot which are regarded as “outliers.”
Steps in the conversion process are:
Watershed Algorithm: Clusters are Assigned a Number so that Their Location is Known
The watershed method of finding intensity in an image is well known in image processing and described in, e.g., “Digital Image Processing,” K. R. Castleman, Prentice hall press. It is used to find intensity levels by treating segments of different intensity levels of an image as basins of different depth. It then simulates rainfall on those basins until all basins are filled. The results of the watershed method are affected primarily by the starting position in each segment, before simulating rainfall. In this application, the starting position of a cluster shall be at or close to the central and the densest position of that cluster, in order to avoid erroneous results.
A bead map can be enclosed in a perimeter, e.g., box rectangle shown in
To enclose the bead map of
In
2. Transform One Decoding Map into Another: Adaptive Templates
After finding bounding rectangles for both the template bead map and the project bead map, one can establish a transformation between the two bounding rectangles so that one can transform a point in the template bead map to a relative position in the project bead map. Assume here that one knows the bounding rectangle for a template bead map by its lower left corner (TLBx, TLBy) its width (TemplateW) and its height (TemplateH). And assume that one also knows the bounding rectangle in a project bead map by its lower left corner (PLBx, PLBy) its width (ProjectW) and its height (ProjectH). To transform a point (x, y) in template to a position (u, v) in project bead map, one uses the following formula:
u=PLBx+(x−TLBx)/TemplateW*ProjectW;
v=PLBy+(y−TLBy)/TemplateH*ProjectH.
One can also transform a point (u, v) from project bead map to a position (x, y) in a template with the following formula:
x=TLBx+(u−PLBx)/ProjectW*TemplateW;
y=TLBy+(v−PLBy)/ProjectH*TemplateH.
3. Convert Project Decoding Map into Image Format
After obtaining the project bead map, it can be converted to an image format using the same techniques described above for converting a bead map and into an image format.
Here, each cluster in a template bead map is considered to be enclosed in a perimeter (here, a rectangular box). To define a cluster's bounding rectangle, one uses the position range of the dots which form the cluster. For each cluster in a template bead map, one transforms its bounding rectangle to an image format of a project bead map.
For each transformed bounding rectangle for each cluster, convert the interior region into a binary signal using a threshold of average pixel values in that region. Then find the biggest “blob” of image in the binary region. The center of the biggest blob is the starting point for that cluster.
Apply the watershed algorithm with the starting point of each cluster in the image format to find the intensity of the cluster. At the end of this process, each pixel either belongs to a specific cluster or to no cluster.
7. Convert Results from Image Format Back to Project Bead Map (Cluster Assignment of Each Dot in the Project Bead Map)
For each dot the image format, one must find the cluster it belongs to in the project bead map. To do so, one finds which pixel each dot belongs to in the image format. Then copy the cluster assignment of that pixel to that dot in the project bead map.
It should be understood that the terms, expressions and examples herein are exemplary and not limiting, and that the invention is defined only in the claims which follow, and includes all equivalents of the subject matter of the claims. The method steps in the claims are not necessarily in order, and unless specified in the claim, may be carried out in any order, including that specified in the claims.
Number | Date | Country | |
---|---|---|---|
Parent | 11439597 | May 2006 | US |
Child | 12049288 | US | |
Parent | 10893007 | Jul 2004 | US |
Child | 11439597 | US |