Method for identifying biomarkers using Fractal Genomics Modeling

Abstract

This present invention relates to methods of manipulation, storage, modeling, visualization and quantification of datasets. One application of the present invention is related to developing FGM models of datasets represented by the various points in a multi-dimensional map. The invention can be adapted to genomic analysis by Fractal Genomics Modeling (FGM) which can be used to identify biomarkers to develop treatments, diagnoses or prognoses of disease by exploiting the map of interactions and causality-pathway conjecture-rendered by this technology.

Description

FEDERALLY SPONSORED RESEARCH OR DEVELOPMENT

Not Applicable.

BACKGROUND OF THE INVENTION

1. Technical Field

This present invention relates to methods of manipulation, storage, modeling, visualization and quantification of datasets. One application of the present invention is related to developing point-models of datasets represented by the various points in a multi-dimensional map. The invention can be adapted to genomic analysis by Fractal Genomics Modeling (FGM) which can be used to identify biomarkers to develop treatments, diagnoses or prognoses of disease by exploiting the map of interactions and causality-pathway conjecture-rendered by this technology.

2. Background Art

The standard techniques currently employed to analyze large datasets are Cluster Analysis and Self-Organizing Maps. These approaches can be effective in identifying broad groupings of genes connected with well understood phenotypes but fall short in identifying more complex gene interactions and phenotypes, which are less well defined. They do not allow for the fingerprinting and visualization of an entire dataset, and missing values are not easily accommodated. The computational requirements are high for these techniques, and the mapping time increases exponentially with the size of the dataset. Furthermore, the current data must be reanalyzed when new datasets are added to the analysis, and vastly different results can occur for each new dataset or group of datasets added.

In order to take full advantage of the information in multiple, large sets of data, we need new, innovative tools. There is a need for methods that more easily enable identification and visualization of potentially significant similarities and differences between multiple datasets in their entirety. There is also a need for methods to intelligently store and model large datasets.

Recent studies have revealed genome-wide gene expression patterns in relation to many diseases and physiological processes. These patterns indicate a complex network interaction involving many genes and gene pathways, over varying periods of times. On a parallel track, recent studies involving mathematical models and biophysical analysis have shown evidence of an efficient, robust, network structure for information transmission when these networks are examined as large-scale gene groups. The problem comes in producing analysis of information transmission and network structure on the scale of individual genes and genetic pathways. Fractal Genomics Modeling (FGM) solves this problem by taking advantage of universal principles of organization. From the Internet, to social relations, to biochemical pathways, the fundamental patterns are similar. The natural relationship among many different types of networks, when mathematically represented, enables the extrapolation of vast quantities of data, capable of computerized analysis. FGM is computationally efficient because the method is performed incrementally, is almost perfectly parallel, and is substantially linear. Consequently, there is no scaling problem with FGM. Furthermore, of significant interest, FGM can be used to identify biomarkers and develop systems for diagnoses or prognoses of disease by exploiting the map of interactions and causality-pathway conjecture-rendered by this technology.

BRIEF DESCRIPTION OF THE DRAWINGS

These and other aspects and attributes of the present invention will be discussed with reference to the following drawings and accompanying specification.

FIGS. 1A and 1B are a flow chart of the operational steps for manipulation, storage, modeling, visualization and quantification of datasets;

FIG. 2 is a flow chart of the operational steps for an iterative algorithm and processing which provides a comparison string;

FIG. 3 is a model showing an efficient, robust, network structure for information transmission of the kind that has been found in many complex networks, including gene regulatory networks;

FIG. 4 is a model showing clinical expression of acute lymphoblastic leukemia (ALL) based on gene expression patterns in the ALL genetic network;

FIG. 5 is a log-log probability distribution for lined FGM models derived from an arbitrarily chosen gene expression data from a sample of Control/Normal subject and Down's Syndrome subject. Dashed line=scale-free fit;

FIG. 6 is a log-log probability distribution for lined FGM models derived from all gene expression data of Control/Normal subject and Down's Syndrome subject in the study. Dashed line=scale-free fit;

FIG. 7 is a chart showing gene group models used for a 7-gene diagnostic test in Example 3;

FIG. 8 is a chart showing gene group models used for 5-gene diagnostic tests in Example 3;

FIG. 9 is a chart showing gene group models used for 10-gene diagnostic tests in Example 3;

FIG. 10 is a flow chart showing a downstream causality from two diagnostic 7-gene groups used in the 7-gene test in Example 3;

FIG. 11 shows the differentiation of subjects by gene group correlation C. The large squares represent the HIV negative subjects; and

FIG. 12 represents a proposed network diagram derived from the biomarkers found in Example 4.

DETAILED DESCRIPTION OF THE INVENTION

The present invention is susceptible to embodiments in many different forms. Preferred embodiments of the invention are disclosed with the understanding that the present disclosure is to be considered as exemplifications of the principles of the invention and are not intended to limit the broad aspects of the invention to the embodiments illustrated.

Generation of Point-Models of Datasets in a Multi-Dimensional Map

This present invention relates to methods of manipulation, storage, modeling, visualization and quantification of datasets. FIGS. 1A and 1 show a flow chart of the method of the present invention for generating a multi-dimensional map of one or more target strings in which the target strings can be represented by marked points in the map. The target strings correspond to datasets to be analyzed. Each point marked in the map serves as a point-model for one or more target strings. The method can be used in the manipulation, storage, modeling, visualization and quantification of datasets in the target strings. The dataset in each target string consists of a sequence of numbers of length N*. One example of a dataset to be analyzed and its corresponding target string is the yearly income of a population, the target string being each person's income listed in a sequence. Another example is the body temperature readings of a group of patients in a hospital ward, with the target string being those readings listed in a sequence. A further example is a DNA sequence, such that each different type of base (A, C, T, G) is labeled with a number (0, 1, 2, 3), producing a target string with a corresponding numerical sequence. A further example is a protein sequence, such that each type of amino acid in the protein chain is labeled with a different number, producing a target string with a corresponding numerical sequence.

For FIGS. 1A and 1B, suppose each dataset to be analyzed is a string of measurements resulting from an experiment involving several thousand genes. Further suppose that there is a number connected with the experimental result from each gene. Such a number could be the gene expression ratio, which represents the differences in fluorescence calculated from the gene combined with some other chemical on a biochip or on a slide. This calculation is not a part of the present invention but provides the numbers in the example target strings. The number of numbers in the example target strings is N*.

Starting with FIG. 1A, the method starts (step 101) by providing a set of M such target strings of length N* (step 103). A region, R, is selected (step 104) such that each point in the region can serve as the domain of an iterative function. The iterative algorithm calculates the comparison string from a point, p, in some region, R. Preferably, the region, R, is in the complex plane corresponding to the area in and around the Mandelbrot Set. Although the Mandelbrot Set is used in the preferred embodiment of the present invention, other sets, such as Julia Sets, may also be used. Using this iterative method, every point within the Mandelbrot Set can be made to correspond to a data sequence of arbitrary length. Because the Mandelbrot Set is made up of an infinite number of points, the method allows any number of datasets containing any number of values to be compared by mapping the datasets to points in or near the Mandelbrot Set.

The Mandelbrot Set is an extremely complex fractal. The term fractal is used to describe non-regular geometric shapes that have the same degree of non-regularity on all scales. It is this property of a self-similarity that allows pictures of artificial systems built from fractals to resemble complex natural systems.

A comparison string of length N is also provided (step 107). The comparison string is generated from a point, p, in the Region, R, by using an iterative algorithm N times to generate the comparison string having a length of N. The comparison string is also a data string and may be of any length relative to the target string. FIG. 2 shows an example of the steps involved in an iterative algorithm to generate a comparison string of length N provided in step 107 of FIG. 1A. The algorithm of FIG. 2 for the Mandelbrot Set is an example of an algorithm that can be used. If a set of points from a different iterative domain is used in this method instead of the Mandelbrot Set, a different algorithmic function would instead be used for that different set of points. The algorithm starts (201), and a counter, n, is initialized to zero (step 221). A variable to be used in the algorithm, z₀, is initialized to zero (step 227). A point, p, is chosen from region R, preferably the region corresponding to the area in and around the Mandelbrot Set (step 231). An example of choosing such a point might be to overlay a grid upon the Mandelbrot Set and, then, choose one of the points in the grid.

Determine if N numbers, which constitute the comparison string, have been calculated (step 241). In other words, check if n =N. If all the numbers of the comparison string have not yet been calculated (step 241), then the point, p, is used as input to the iterative algorithm z_n+1=z_n²+p (step 251). For example, the first iteration based on a point, p, is z₁=z₀²+p, or z₁0+p, z₁=p, or z₁=p. Since p is a complex number of form a +bi when decomposed into its real and imaginary parts, z₂ takes the form z₂=(a²+2i*a*b−b²)+a+bi or (a²−b²+a)+i(b *(2a+1)).

If the absolute value of Z_n+1 is greater than 2.0, or |z_n+1|>2.0 (step 261), the iteration is stopped because it is unbounded, and the z_n+1 will become infinitely large. Thus, point, p, is no longer under consideration. Instead, n is initialized to zero (step 221), z₀ is initialized to zero (step 227), and another point is instead chosen from the region R (step 231), preferably in and/or near the Mandelbrot Set. This prematurely stopped string, however, may be used as a comparison string with a length of less than N.

If the absolute value of Z_n+1 is equal to 2.0 or less, increment n by one (step 271) and check if N numbers have been calculated which constitute the comparison string (step 241). In other words, the algorithm iterates until n =N. If n <N, then perform the next iteration on point p (step 251). This next iteration will calculate the next number in the string of numbers comprising the comparison string. The process iterates until a string of variables, z₁, through Z_N can be produced that is of length N, so long as |z_n+1|≦2.0.

If n=N (step 241), or when the iteration is stopped because the absolute value of z_n+1 is greater than 2.0, or |z_n+1|>2.0 (step 261), then the comparison string has been generated. However, the numbers in the comparison string may need to be transformed to have values within a value set of interest (step 281). Suppose the numbers in the example target string representing gene expression ratios are real numbers between 0 and 10. If we wish to explore the similarities between the comparison string and the target string the value set of interest would be the real numbers between 0 and 10. The numbers of the comparison string may need to undergo some transformation to produce real numbers in this range. One way to produce such a real number is the function r=10.0 * b/|z_n|. This will produce real numbers r falling in the range between 0 and 10 for Z_n=a+bi. Provide the comparison string (step 291), and the algorithm ends (step 299).

Referring to FIG. 1A, an optional step is to determine if certain properties of the comparison string should be marked (step 109). Examples of properties that might be marked are the mean value of the comparison string or the Shannon entropy. If certain properties of the comparison string should be marked (step 109), mark the properties of the comparison string (step 111). Optionally, the comparison string can also be checked to determine if it meets pre-scoring criteria (step 113), regardless of whether the properties of the comparison string are marked. This step involves preliminary testing of the comparison string's properties alone as criteria to initiate scoring. Examples of pre-scoring criteria are measuring the mean value of the comparison string to see if it is higher or lower than desired and determining if the Shannon entropy of the comparison string is too low or too high. When marking prior to scoring, it may be determined that an entire subregion of the region has a large number of points that do not meet the pre-scoring criteria. For example, this subregion may be part of a grid. It may be determined that the rest of the points in that subregion will not be considered, even though the original intent was to consider all points in the region.

If the comparison string is pre-scored as described above and it does not meet the pre-scoring criteria (step 113), then the current comparison string is no longer under consideration. Another comparison string is instead provided (step 107). The new comparison string is generated using the exemplary iterative algorithm of FIG. 2 on a new point, p, from region R.

If the comparison string is pre-scored and it meets the pre-scoring criteria (step 113), then scoring of the comparison string is performed (step 121). Scoring refers to some test of the comparison string using the target string. Scoring of the comparison string can also be performed without marking the properties of the comparison string or pre-scoring the comparison string. In the example of real numbers r falling in the range between 0 and 10 described above, the score could be the correlation coefficient between the comparison string consisting of numbers r and the target string. A simple example of scoring might be counting the number of one-to-one matches between the comparison string and the target string over some length L where L<=N*, where N* is the length of the target string. Alternatively, a one-to-one comparison between numbers in the comparison and target strings may be performed for a non-contiguous number L of the numbers. For example, compare the second, fourth, and sixteenth numbers for a number L=3.

Determine if the point, p, corresponding to the comparison string should be marked depending on the score or other properties (step 123). If it is determined that the point should be marked (step 123), mark the point, p, in the region, R (step 127). The marked point is a point-model in the region, R, to represent the target string, M. The comparison string generated from this marked point with the iterative algorithm represents the target string, M. Marking can be used in an environment where a pixel or character corresponds to point p on a visual display or marking can refer to annotating the coordinates of point p in some memory, a database or a table. The point is marked by changing some graphical property of the corresponding pixel, such as color, or changing the corresponding character. The point may also be marked by annotating the coordinates of point p in some memory, a database or a table based on the score. Optionally, point p can be marked, either additionally or solely, according to quantification of properties of the comparison string, without regard to the score. Such properties can be general, such as using some color or annotation to reflect the mean value of the string being in a certain range or markings reflecting the number of 3's in the string, or the value of the Shannon entropy. Such marking can be used as an aid in searching for preliminary criteria for scoring. When marking point p, it may be determined that an entire subregion of the region has a large number of points that do not meet the scoring criteria or other properties. For example, this subregion may be part of a grid. It may be determined that the rest of the points in that subregion will not be considered, even though the original intent was to consider all points in the region.

If it is determined that the point should not be marked (step 123), determine if a sufficient number of the M target strings have been checked for the comparison string derived from point p (step 129). For instance, in our gene expression example, there may be several experiments or datasets that are being scored against each comparison string. If more of the M target strings should be checked, the comparison string is scored against another of the M target strings (step 121). The comparison string can be used to compare to all M target strings. Not all of the target strings may exhibit similarity to a comparison string, and, therefore, not all target strings may be marked. Also, more than one target string may demonstrate some homology with a comparison string. Moreover, target strings may be marked multiple times, exhibiting correlative relationships to multiple comparison strings.

If a sufficient number or all of the M target strings have been checked (step 129), determine if a sufficient number of points corresponding to comparison strings have been checked (step 133). If more of the points corresponding to comparison strings should be checked, provide another comparison string (step 107). The new comparison string is generated using the same iterative algorithm as used in generating the previous comparison string, such as the one detailed illustratively in FIG. 2, on a new point, p, from region R. Any number of the same M target strings will then be used to score the new comparison string.

If a sufficient number of points corresponding to comparison strings has been checked (step 133), the scoring process stops. In the case of determining the points, p, from a grid, this could be the number of points in the grid. The highest scoring point or points are then mapped (step 137). Mapping refers to placing the coordinates of highest scoring point or points in memory, a database or a table. The target string or strings, represented by the coordinates, may also be visually marked on a visual display.

Target strings may be analyzed and/or compared by examining, either visually or mathematically, their relative locations and/or absolute locations within the region, R. When scoring similarity measures between the comparison strings and the target strings, target strings with greater similarity are generally mapped closer to each other based on Euclidean distance on the map. This is because comparison strings with greater similarity are generally closer to each other on the map. However, this is not always true because the metrics involved are more complicated. For example, shading of points corresponding to comparison strings with high scores for a given target string represents a metric which shows similarity between this target string and others mapped in this shaded region. The target strings in this case, however, may not appear close together on the map or display but can be identified as being similar.

Continuing to FIG. 1B, determine whether points in region R should be marked (in a similar manner as previously described) based on their relative scores or properties compared to other points in region R (step 139). If it is determined that the points should be marked (step 139), mark the points (step 141). For example, one might wish to mark all the points whose score falls within 10% of the highest score of a chosen target string, or mark points whose comparison strings have the lowest or highest Shannon entropy for the region. When marking points, it may be determined that an entire subregion of the region has a large number of points that do not meet the relative score criteria or other properties. For example, this subregion may be part of a grid. This may be used to determine whether this subregion is of interest or not.

In one embodiment of the present invention, once the decision has been made as to whether such points should be marked (step 139), determine if a subregion of R is of interest (step 143). If a subregion of R is of interest (step 143), then this subregion is examined with higher resolution, called zooming (step 147). The subregion of R replaces the previous region R. (step 104 of FIG. 1A). Comparison strings will be generated from the new subregion of R and will be scored against any number of the same set of M target strings originally provided. Points in a subregion of interest, which were previously unchecked, will be examined because the new region, R, is a higher resolution version of the subregion of interest. The points in the subregion will tend to produce a greater percentage of similar comparison strings to those previously examined in region R. If the subregion of interest is a high scoring region this will, in general, produce a greater percentage of high scores and some differences will emerge to produce higher scores or properties which are closer to some desired criteria.

After zooming (step 147) and before examining the subregion, the target strings and comparison strings may optionally be transformed to attempt to improve the precision and resolution of the mapping and marking in the method. Suppose in the gene expression example, the target strings values, instead of real numbers from 0 to 10, were binned into 10 contiguous intervals, such that the first bin corresponds to real number values from 0 to 1, the second bin to real number values from 1 to 2, etc. Suppose these bins were labeled 0 to 9. The target string would then be a string of integers with values from 0 through 9. Suppose that a similar transformation was done on the transformed comparison strings. Suppose the method is performed and after zooming (step 147), the gene expression ratios and comparison strings are split into 20 such bins from 0 to 0.5, 0.5 to 1.0, etc. Thus, the target and comparison strings will be re-scaled before repeating the process in the new subregion (104 of FIG. 1A).

This re-scaling can improve the precision and accuracy of the mapping and marking in the method. There are several well-studied methodologies that can be used to approach such a re-scaling to improve the precision and resolution of the mapping and marking process as zooming is performed. These include, but are not limited to, methodologies such as Simulated Annealing, Hill Climbing Algorithms, Genetic Algorithms, or Evolutionary Programming Methods.

If no other subregions of R are of interest (step 143), the method of FIGS. 1A and 1B ends (step 199). This generally results when there is no improvement in the score after some number of zooms.

It should be apparent to one skilled in the art that this technique can be used to study the behavior of any (scoring) function that uses the target strings and the comparison strings as variables. Attempting to find the highest value of the similarity measure scoring function is a particular case of this. As such, this method could be used to attempt to optimize any scoring function, using a target string or multiple target strings and comparison strings as variables, to find the functions minima and maxima. In addition, each comparison string can simply be used alone as input into the variables of a scoring function for such a purpose.

It should be apparent to one skilled in the art that this method can be used for data compression. If the model of the target string represented by a comparison string is sufficiently similar to the target string, and the coordinates of the point p corresponding to that comparison string can be represented in a more compact way than the target string, then the target string can be replaced with its more compact representation in the form of the coordinates of point p. This is because the comparison string generation algorithm can then be used to recreate a sufficiently similar representation of target string from point p.

This method has special applicability to multiple large datasets. Uses for this method include analysis of DNA sequence data, protein sequence data, and gene expression datasets. The method can also be used with demographic data, statistical data, and clinical (patient) data. The uses for this method are not limited to these datasets, however, and may be applied to any type of data or heterogeneous mixtures of different data types within datasets. Some of the steps of this method can involve determinations and interventions made by a user of the method or they can be automated.

Fractal Genomics Modeling (FGM)

The previously described method can be adapted for use in a new data analysis technique, Fractal Genomics Modeling (FGM), to explore the structure of genetic networks. It is possible to produce hypotheses for unknown gene interactions, for proposed pathways, and for pathway interconnections of large-scale gene expression through Fractal Genomics Modeling (FGM). By virtue of its correlational power, FGM inherently results in the discovery of putative biomarkers that can classify disease. Such disease indicators are discovered by the rendering and ordering of the underlying genetic elements that engender the illness, as it progresses and changes over time. Three distinct disease models ensue, each exemplifying the predictive capability of FGM: Down's Syndrome, Human Immunodeficiency Virus (HIV) infection, and leukemia.

The conventional approach to analyze large-scale gene expression has been cluster analysis and self-organizing maps. This approach can be effective in identifying broad groupings of genes connected with well understood phenotypes, but falls short in identifying more complex gene interactions and phenotypes which are less well defined.

When applying cluster analysis to microarrays, typically a function is applied to every gene expression value in such a way that similar values cluster in similar locations on (usually) a two dimensional surface. With FGM, a special modeling method is used so that every point on a surface uses its own function to represent a cluster model of gene expression values, effectively “clustering the clusters.” This allows for much greater insight into gene expression patterns and the similarities between them. By using FGM, the analysis moves from conventional approaches of examining gene expression values to examining gene expression patterns.

FIG. 3 illustrates an efficient, robust, network structure for information transmission of the kind that has been found in many complex networks, including gene regulatory networks. The points represent what are called nodes in the networks, and the lines represent what are called links. The nature of the type of network shown in FIG. 3 is that there are few nodes with many direct links, forming hubs, and many nodes with only a small number of direct links. These types of networks are often called scale-free or power-law networks. They are characterized by the fact that the number of genes with a given number of links falls off as a power law. For example, there may be twice as many nodes with 2 links as with 4 links. The robust nature of this type of network comes from the fact that if one removes or disables one of the nodes, it is more likely to be one with only a few links and cause less harm to the network as a whole.

Suppose the World Wide Web is organized this way. The points around the center would be web sites like Yahoo or Google, the points slightly further from the center might be web sites like Amazon.com or Expedia, and the outside points might be personal web sites (obviously this requires a much larger picture to show this accurately!). The flow of information tends to go from the inside out. For example, information flows easily from Yahoo to the rest of the network because it has so many direct links. Information flow from a personal web site to the rest of the web is possible, but less likely. One can see the robust nature of the web in the fact that sites and servers go offline all the time without effecting the network. Of course the occasional times when an “inner” site such a Yahoo goes offline can have a very large impact!

Each node in FIG. 3 can also represent a gene and the lines can represent correlated behavior to other genes in a genetic network. The most “connected” genes, or genes with the most direct links, would be ones in the center of this picture.

As an example, FIG. 4 represents how acute lymphoblasitc leukemia (ALL) might express itself in the genetic network of an actual patient. In this model, the flow of information, in this case biochemical interactions, begins far upstream, in what could be called “super-regulatory genes” (SRG) due to their importance. This area is labeled SRG/Normal. In this patient and in most people, these SRG behave as they would in a normal, healthy individual.

As the biochemical gene expression patterns propagate out of the center through downstream links, however, something occurs which causes a divergence from a normal, healthy pattern. Due to mutation, biochemical or environmental factors, or chance, a group of genes residing somewhere in the ringed area labeled SRG/Carcinogenic begins a cascade through the network that propagates into a clinical expression of cancer. Further downstream are nodes in the network that define the clinical outcome as a specific type of cancer, illustrated by the group of genes labeled leukemia and, still further downstream, as a subclass of leukemia, ALL (and extending out to genes not seen). It should be noted that this is not a simple cascade from the center outward. Many interconnecting pathways are involved with both promoter and inhibitory links between genes.

FGM is a hybrid technique that blends some of the concepts of wavelet analysis with cluster analysis. FGM ”wavelets” are a series of real-valued numbers derived from complex logistic maps, such as Julia sets, generated from iterations of a single point in the complex plane.

FGM searches points on the complex plane for the model that gives the greatest Pearson correlation with the actual localized data, using a minimum cutoff correlation whose absolute value is >0.95. The similarity metric between point-models on the complex plane found in this way is very intricate but, in general, similar models tend to cluster in similar areas on the surface. This is particularly true if the point-models fall within a given “threshold” determined by Euclidean measure.

Since a genome-wide pattern is mirrored in a small number of genes due to underlying fractal structure, FGM can be used to model the gene expression of small groups of genes, each having n number of genes (for example, n is 7 or 14 genes) from a much larger gene pool. The larger gene pool can be a sample of an organism's genome or of an organism's entire genome, such as the entire human genome. Illustratively, the genes in the gene pool can be arranged randomly in microarrays of commercial gene chips (e.g., Affymetrix Human Genome U95A chips consisting of about 12,000 genes) to measure the gene expression levels of the genes. Significantly, at least one small characterizing group of genes must exist.

Since FGM models are usually scored based on their Pearson correlation, the overall magnitude of gene expression within these small groups does not matter in probing for similar patterns throughout the array, only the relative expression patterns within the groups. Other mathematical relations may be used other than Pearson's correlation. When comparing patterns of gene expression between these groups, we sometimes worked with only the models of these gene groups (in “model” space), and we sometimes worked with the actual gene expression values. Unless noted, we usually compare model values and not actual gene expression values although they are often similar.

Choosing gene expression values from small groups of arbitrarily chosen genes in a network is the same as a series of short, random walks of random step-size on a modular structure. By analogy, one should see a comparative distribution of gene expression values between such “walks” much different than if genes were randomly linked within the genome or acting largely independently. Similarities between the gene expression patterns in these groups should reveal information about the genetic network structure with correlations between gene groups skewed around gene groups chosen that align with the inherent modularity. Clusters on the FGM surface can serve to identify and to analyze such a skewed distribution.

The FGM method for modeling gene expression of a small group of genes in a genetic network of a subject comprises the following steps: (a) providing a dataset of gene expression values of the small group of genes from the subject; (b) providing a surface wherein each point on the surface can serve as a domain for an iterative algorithm; (c) selecting a point on the surface; (d) generating a comparison string from the selected point using the iterative algorithm; (e) scoring the comparison string against the gene expression values in the dataset; (f) determining if the score of the comparison string meets a pre-determined condition or property; and (g) marking the point if the score meets the pre-determined condition or property to generate a fractal genomics modeling (FGM) model of the target string on the surface. The method above can include the additional steps of repeating the steps (c) through (g) for a plurality of gene expression values from a plurality of small groups of gene in the genetic network to generate a plurality of FGM models on the surface.

Identifying Biomarkers

Within the point-models (also known as FGM models) on the FGM surface, clusters are found containing models of the same gene groups corresponding to only one of the phenotypes. If such a gene group is found, it is then individually tested across all datasets to verify that between these n-gene patterns the Pearson correlation, or any other suitable correlations, is markedly different depending on the phenotype from which the dataset is drawn. If such a gene group is found, further testing is done to choose the n-gene group from the sample within the cluster that produces the most marked difference. Such a gene group or its FGM model then becomes a candidate biomarker for the particular phenotype being studied and provides insight into the biochemical pathways linked to the phenotype present. The biomarker can then be used to develop treatments, diagnoses or prognoses of diseases. A diagnostic test can also be designed to diagnose a disease of a test subject by comparing the gene expression values of the phenotype of the test subject against the biomarker.

In a preferred embodiment, the method for identifying the biomarker for a phenotype includes the steps of: (a) identifying clusters containing FGM models of the small group of genes corresponding to the phenotype; (b) individually testing each of the small group of genes across all datasets to verify that the pre-determined condition or property between the small groups of genes is markedly different with regard to the phenotype; and (c) selecting the small group of genes that produces the most marked difference in the pre-determined condition or property as a biomarker for the particular phenotype.

In another preferred embodiment, the method for identifying the biomarker for a phenotype, such as the phenotype of a disease, includes the steps of: (a) providing a plurality of datasets of gene expression values wherein each dataset is from a small group of genes, and the plurality of datasets is from one or more subjects having the phenotype; (b) providing a surface wherein each point on the surface can be served as a domain for an iterative algorithm; (c) selecting a point on the surface; (d) generating a comparison string from the selected point using the iterative algorithm; (e) scoring the comparison string against the gene expression values in the dataset; (f) determining if the score of the comparison string meets a pre-determined Pearson correlation value; (g) marking the point if the score meets the pre-determined Pearson correlation value to generate a FGM model of the target string on the surface; (h) repeating steps (c) through (g) for a plurality of the datasets to generate FGM models for said plurality of datasets; (i) identifying clusters containing FGM models of the same small group of genes corresponding to the phenotype; (j) individually testing each of the small group of genes across all datasets to verify that the Pearson correlation between the small groups of genes is markedly different with regard to the phenotype; and (k) selecting the small group of genes that produces the most marked difference in the Pearson correlation as a biomarker for the particular phenotype.

“Shotgun Approach” with a Weighted Process in Studying Genetic Networks and Identifying Biomarkers

FGM can be also used to study the genetic networks in a population of subjects with regard to a specific phenotype, as well as the correlative behavior between the nodes within the networks, using a “shotgun” approach. In a preferred embodiment, the phenotype is a disease, such as HIV infection.

In a typical “shotgun” approach, the gene expression values are measured from a large gene pool from subjects within a population, the subjects having a positive group with the phenotype and a negative group without the phenotype. In a preferred embodiment, the number of genes in the gene pool is about 10,000 or more. The large gene pool may consist of the entire genome of the subject. The gene expression values can be measured from a tissue from the subject by any method known in the art. One common method to measure the gene expression values is using a gene chip such as an Affymetrix gene chip (e.g., the Affymetrix Human Genome U95A chip consisting of 12,558 genes).

The gene expression values from this large gene pool are then randomly broken into overlapping or non-overlapping subgroups each containing N gene expression values, for example, from 3 to 20 genes, and preferably from 5 to 20 genes. The number of genes N in each subgroup is chosen using the criteria of picking a relatively small number in the range, for example, 3 to 20, or 5 to 20, which goes evenly into the number of genes in the large gene pool. For example, in a large gene pool of 12,558 genes, the small number can be 7 or 14. FGM, which is described previously, then maps each of these subgroups to a single point (also known as the “point-model,” or the “FGM model,” or simply the “model”) on a multi-dimensional surface, where metrics can be applied to evaluate correlative behavior between the groups based on their proximity on the surface. The surface can be derived from a Julia set or from within or near the boundary of a Mandelbrot set. The similarity metric between FGM point models on the multi-dimensional surface is very intricate but, in general, similar models tend to cluster in similar areas on the surface. This is particularly true if the points (models) fall within a given “threshold” determined by Euclidean measure. In a preferred embodiment, the multi-dimensional surface is a two-dimensional surface. The FGM is conceptually similar to cluster analysis and self-organizing maps, except that one is clustering groups of gene expression values rather than single expression values.

If clusters on the FGM surface are found that contain the same gene group for a significant number of either the positive or negative subjects, then the correlative patterns between these subgroups are examined in all subjects selected to see if the pattern of expression in this subgroup effectively differentiates positive from negative subjects. What is meant by “a significant number” is any number equal to or greater than two. In this way, candidates for genetic markers of the phenotype are found. The gene groups found having this property are deemed candidates for Biomarker Evaluation (BE). Gene groups are also deemed candidates for BE if the only valid models produce were two or more of the same positive or negative phenotype.

The genes and pattern of expression within the subgroups can be further evaluated for biochemical relationships and to see if they are part of known pathways or networks. Importantly, since no consideration is given for the ordering of the original gene expression values from the large gene pool, the values are randomly scrambled according to the same key for all subjects and FGM is repeated again to repeat the process to find another pathway, piece of a network, or an overlap with a previous piece found. The process is analogous to shotgun approaches to DNA sequencing, where a network is shredded into component pieces and analyzed rather than a DNA sequence.

The candidates for Biomarker Evaluation (BE) obtained as described above can be further examined to see if they successfully allow classification of all the subjects in the study. If such a gene group is found, it is deemed to have passed BE and is listed as a biomarker. The candidates are considered to have successfully classified all subjects in the study based on two criteria:

• • (1) The candidate (gene group) only produces valid models (for example, the absolute value of the Pearson correlation |C|>0.95) for all phenotype positive or all phenotype negative subjects.
  • (2) The candidate's expression values from a single subject produce a ranked differentiation by phenotype when correlated with the other subjects.

Once a gene group is labeled a biomarker, a weighted procedure is performed to determine which genes in the group are deemed most important in distinguishing phenotype using the following procedure:

• • (1) The average expression value for each of the N genes is calculated for the positive and negative subjects with regard to the phenotype.
  • (2) The correlation value, such as the Pearson correlation value, is evaluated between the two N average gene expression values from the phenotype positive subjects and the phenotype negative subjects.
  • (3) For each of the N genes, the average phenotype negative expression value is made equivalent to the phenotype negative value and the correlation, such as the Pearson correlation, between all N values is calculated. The absolute value of the difference from the correlation calculated in step (2) is evaluated. (4) The N values generated in step (3) are ranked from a highest to lowest and the weight of 1.0 is given to the highest value. The ratios of the remaining values over the highest value are calculated to produce the weighting of the other values.

Biomarkers on the list are checked against known pathways and networks and in some cases conjectures are made based on biochemical considerations. Overlaps in such findings are given extra significance.

The results produced so far using the method described are intriguing, and were produced by relatively low-powered hardware in a partially automated setting that can be fully automated. The specific implementation of FGM described in the present can be expanded in a multitude of ways. FGM can be applied to other processes such as metabolic and protein interactions.

EXAMPLES

Example 1

Evidence of Scale-free Genetic Network and Identification Biomarkers in Down's Syndrome

This example demonstrates the use of FGM both to provide evidence of scale-free genetic network in Down's Syndrome and to identify specific small gene groupings, consisting of 7 genes, that can serve as biomarkers relating to Down's Syndrome.

In this study, FGM was used to model small groups of 7 genes from much larger microarrays (Affymetrix Human Genome U95A chips) consisting of 12,558 genes. The data was derived from fibroblasts of 4 subjects with and 4 subjects without Down's Syndrome-totaling 8 subjects. The number of genes within the groups, in this case 7, was decided using the criteria of picking a relatively small number-in the range of 5-20-that when divided into 12,558 yields a real number without a remainder. Thus, arbitrarily choosing the gene groups by grouping the genes as they appeared on the gene chip, 1,794 7-gene groups were established. Consequently, 14,352 (1,794 gene groups x 8 subjects) target strings, M, each with 7 gene expression values, were provided for FGM analysis.

Comparison strings were generated from points in the multi-dimensional map or complex plane for each target string and were scored against each of the target string. These comparison strings served as potential FGM models for the target strings. These FGM models were scored based on their overall Pearson correlation, using a minimum cutoff correlation of absolute value >0.95. Among the point-models (also known as the FGM models) on the FGM surface, clusters were found containing models of the same gene groups corresponding to only one of the phenotypes.

In order to test a genetic network for the threshold requirements of scale-free and modular behavior, a log-log plot of k vs. P(k) of gene expression data from a Control/Normal sample and a Down's Syndrome subject is graphed. P(k) is the probability of finding a 7-gene group with k links to another 7-gene group. A group is considered linked to another group if it falls within the same FGM cluster of a given size.

FIG. 5 is the log-log plot of k vs. P(k) of gene expression data for an arbitrarily chosen Downs and Control sample. The resulting plot is linear, demonstrating both modular and scale-free characteristics. The network organization appears to be hierarchical in nature for smaller clusters but deviates from linearity for larger clusters. This could be due to an effect called saturation that limits how large a cluster can get in real-world networks, due to physical constraints and stability.

FIG. 6 is the same plot derived from clusters for all samples on the same FGM map. This is, in effect, a picture of the combined genome for all the data. The picture conveyed from FIGS. 5 and 6 together brings to light further notions of universal constructs within such complex networks.

Using the method described above, a 7-gene group was discovered that corresponded only to subjects with Down's syndrome. The corresponding results are shown in Tables 1 and 2.

TABLE 1
Ranked absolute values of the Pearson Correlation for 7-gene FGM
models with the 7-gene biomarker candidate model (left) and the
corresponding correlations with actual expression values (right). Down's
subject marked with “D” (The model/actual values of the genes from
subject marked with * were used.)
	Subject	Pearson	Subject	Pearson
	6-194D*		6-194D	1
	4213-34D	1	42135-34D	0.97
	5-186D	1	5-186D	0.97
	7-197D	0.92	10-A01C	0.89
	8-367C	0.87	7-197D	0.88
	10-A01C	0.83	3648FC	0.85
	9-367C	0.62	8-367C	0.84
	3648FC	No model found	9-367C	0.71

TABLE 2
The 7-gene Down's Syndrome biomarker candidate.
Model and actual gene expression values for subject 6-194D (produced
highest correlation in Table 1) and description of genes in the group.
*Denotes the fact that this model is negatively correlated to the actual
values (absolute Pearson used in model scoring).
FGM (Model) Values* Actual Values
57.5                200.9
112.3               22.5
70.9                170.7
106.9               7.9
103.3               8.2
99.7                14.5
112.9               4.7
Cluster Incl. D11466: Homo sapiens mRNA for PIG-A protein
Cluster Incl. D10925: Human mRNA for HM145
Cluster Incl. U13395: Human oxidoreductase
Human scavenger receptor cysteine rich Sp alpha mRNA
Homo sapiens properdin (PFC) gene
H. sapiens mRNA for BMPR-II
Human apoptotic cysteine protease

The 7-gene Down's Syndrome biomarker candidate found was located within some of the larger clusters (which did not contain any control samples of the same gene group) on the FGM surface. This could be significant when exploring linkages to larger gene groups.

To test for artifacts from the FGM surface, a “random” U-95A mock sample, 25 produced from 12,558 uniformly distributed random numbers from 0-10,000, was analyzed as 7-gene groups. Only one cluster of three genes and 23 pair-clusters were found in the entire sample.

Example 2

Identification of Biomarkers in Human Immunodeficiency Virus (HIV) Infection

In this example, FGM was used to model small groups of 14 genes from much larger microarrays (Affymetrix Human Genome U95A chips) consisting of 12,558 genes. The data was derived from the brain tissue of 5 HIV-1 negative and 4 HIV-1 infected subjects-totaling 9 subjects. The number of genes within the groups, in this case 14, was decided using the criteria of picking a relatively small number-in the range of 5-20-that goes evenly into 12,558. Thus, arbitrarily choosing the gene groups by grouping the genes as they appeared on the gene chip, 897 14-gene groups were established. Consequently, 8,073 (897 gene groups * 9 subjects) target strings, M, each with 14 gene expression values, were provided for FGM analysis.

Comparison strings were generated for each target string, as previously described. These FGM models were scored based on their overall Pearson correlation, using a minimum cutoff correlation of absolute value >0.95. Therefore, the overall magnitude of gene expression with in these small groups did not matter in probing for similar patterns throughout the array, only the relative expression patterns within the groups.

When comparing gene expression between the gene groups, the models of comparison strings were most often used, though sometimes the actual gene expression values were used.

Among the point-models (also known as FGM models) on the FGM surface, clusters were found containing models of the same gene groups corresponding to only one of the phenotypes. One 14-gene group was discovered that corresponded only to HIV-1 infected subjects. This 14-gene group was then individually tested across all data for each subject in order to verify that between these n-gene patterns (n=14 in this case) the Pearson correlation was noticeably different depending on the phenotype from which the data sample was drawn. The 14-gene group from the sample within the cluster that produced the most noticeable difference was identified as a putative biomarker. The correlation values with this particular gene group and the corresponding gene groups, across all samples, are shown in Table 3. The left side of Table 3 uses the FGM model values, both ranked from highest to lowest correlation.

Table 3: Ranked absolute values of the Pearson Correlation for 14-gene FGM models with the 14-gene biomarker candidate model (left) and the corresponding correlations with actual expression values (right). HIV- 1

TABLE 3
Ranked absolute values of the Pearson Correlation for 14-gene FGM
models with the 14-gene biomarker candidate model (left) and the
corresponding correlations with actual expression values (right). HIV-1
positive marked with “+”. (The model/actual values for the genes from
subject marked with * were used.)
	Subject	Pearson	Subject	Pearson
	G0036+*		G0036+*
	G0017+	0.98	D97 2916−	0.97
	H0011+	0.94	G0017+	0.96
	H0002+	0.91	H0011+	0.94
	G0010+	0.86	G0010+	0.92
	BTB 3455−	No model found	H0002+	0.89
	BTB 3648−	No model found	BTB 3648−	0.88
	BTB 3749−	No model found	BTB 3455−	0.72
	D97 2916−	No model found	BTB 3749−	0.71

The actual marker genes and the model and actual expression values of the

TABLE 4
HIV-1 brain biomarker candidate.
Model and actual gene expression values for subject G0036+ (produced
highest correlation in Table 3) and description of genes in the group.
FGM (Model) Actual
Values      Values
310.8       180.7
126.3       55.6
298.8       158.4
264.5       51.9
274.4       174.7
585.9       912.1
233.4       264.3
248         245.6
478.6       572.3
144.4       55.2
363         218.5
328.3       312.4
457.5       626.5
1074        1593.2
Cluster Incl. U39067: Homo sapiens translation initiation factor eIF3 p36
Cluster Incl. AL050106: Homo sapiens mRNA; cDNA DKFZp586I1319
Cluster Incl. AF047181: Homo sapiens NADH-ubiquinone oxidoreductase
Cluster Incl. AF007872: Homo sapiens torsinB (DQ1)
Cluster Incl. AF007871: Homo sapiens torsinA
Cluster Incl. AB011116: Homo sapiens mRNA for KIAA0544
Cluster Incl. AF032456: Homo sapiens ubiquitin conjugating enzyme G2
Cluster Incl. D87454: Human mRNA for KIAA0265 gene
Cluster Incl. AF001383: Homo sapiens amphiphysin II mRNA
Cluster Incl. U69263: Human matrilin-2 precursor mRNA
Cluster Incl. D31889: Human mRNA for KIAA0072 gene
Cluster Incl. AL050265: Homo sapiens mRNA
Cluster Incl. AL038340: DKFZp566K192_s1
Cluster Incl. AL038340: DKFZp566K192_s1 (duplicate description)

Example 3

Genetic Network and Biomarkers in Leukemia

Input data from the study produced by Golub et al. (Golub T. R., et al., Science, Vol. 286, pp. 531-536, 1999) are used in this example in order to further demonstrate the utility of the present invention. The data in the Golub study contained Affymetrix gene expression data for 7070 genes acquired from patients diagnosed with either acute lymphoblastic leukemia (ALL) or acute myeloid leukemia (AML). The data was composed of a training set of data from 27 ALL patients and 11 AML patients to develop diagnostic approaches based on the Affymetrix data and an independent set of 34 patients for testing.

Genetic Network in the Clinical Expression of Leukemia

In order to determine what kind of genetic network is involved in the clinical expression of leukemia, the more than 7,000 gene expression values in the Golub data were broken into groupings of 5, 7, and 10 genes based only on the order in which the genes were arranged on the Affymetrix chip. FGM was used to create point-models (also known as FGM models) of the gene expression patterns in these small groups and looked for correlations, or clustering between the 5, 7, and 10 gene models in each of the 38 patients in the Golub training set.

The number of ways to arrange to arrange 7 genes out of 7,000 is ˜10²⁷. Unless there is coordinated behavior between a large number of these 7,000 genes, there would be almost no chance of finding correlations between (effectively) arbitrary 7-gene groupings, even when clustering a thousand of them. On the other hand, if there is a genetic network of the scale-free type described above, there should be a large number of genes whose behavior is correlated to only a few genes.

For the 7-gene grouping, our analysis found that there were significant model clusters in every patient. The largest cluster had an average size of approximately ten 7-genes models. Pearson correlations of >|0.95| between the models confirmed the similarities within these clusters. This provides statistical evidence that there are at least a few genes whose behavior is connected with well over 1,000 other genes. This also agrees with an earlier gene expression study based on time-based gene expression data.

The clusters that contained 7-gene groups from only the patients with ALL were then scrutinized. Two 7-gene group models correlated to the largest number of corresponding models in ALL patients but with no AML patients. The two 7-gene groups are listed in FIG. 7 with their respective gene model values as well as the actual gene expression values.

The first group of the 7-gene group contains the following genes:

• • 1. GATA2 GATA-binding protein 2
  • 2. Alcohol dehydrogenase 6 gene
  • 3. GB DEF=Protein-tyrosine phosphatase mRNA
  • 4. Globin gene
  • 5. Pre-mRNA splicing factor SF2, P32 subunit precursor
  • 6. Major histocompatibility complex enhancer-binding protein
  • 7. MSN Moesin.

The second group of the 7-gene group contains the following genes:

• • 1. Onconeural ventral antigen-1 (Nova-1) mRNA
  • 2. Inil mRNA
  • 3. RORA RAR-related orphan receptor A
  • 4. FUSE biding protein mRNA
  • 5. Rar protein mRNA
  • 6. Fetal ALZ-50-reactive clone 1 (FACI) mRNA
  • 7. MB-1 gene.

These two 7-gene group models were used for a 7-gene diagnostic test. The two 7-gene group model values from two patients in the training set (above) were used to characterize ALL in the independent set. The test was an OR test, where if the corresponding 7-gene models in the independent set patients had a Pearson correlation with either of these 7-gene model values such that the absolute value was >0.95, the patient was classified as ALL. The results for the 7-gene grouping are as follows:

Overall Accuracy 0.853
ALL only         0.95
AML only         0.714

Pathways related to this result comprise the Ras-Independent pathway in NK cell-mediated cytotoxicity. The gene of special interest from this result is MB-I gene.

In addition, it was found that the second 7-gene group above allows for the differentiation of patients with ALL into those who have the T-cell ALL from B-cell ALL. The test using this 7-gene group model was accurate to 100% in the test set in classifying B-cell vs. T-Cell(See Table 7). The gene segments used are summarized in Table 8.

TABLE 7
Summary of using the second 7-gene group to predict B-cell and
T-cell ALL
	Absolute Pearson cor-
	relation between patient
Gene-chip	gene segment model	Predicted
(Patient)	and classifier model	(>.95 = B-Cell)	Actual	Correct
39	0.9997	B-Cell	B-Cell	Yes
40	0.9509	B-Cell	B-Cell	Yes
41	0.9954	B-Cell	B-Cell	Yes
42	1	B-Cell	B-Cell	Yes
43	0.9974	B-Cell	B-Cell	Yes
44	0.9995	B-Cell	B-Cell	Yes
45	1	B-Cell	B-Cell	Yes
46	0.9995	B-Cell	B-Cell	Yes
47	0.9996	B-Cell	B-Cell	Yes
48	0.9999	B-Cell	B-Cell	Yes
49	0.9999	B-Cell	B-Cell	Yes
55	0.9792	B-Cell	B-Cell	Yes
56	1	B-Cell	B-Cell	Yes
59	0.9616	B-Cell	B-Cell	Yes
67	0.6753	T-Cell	T-Cell	Yes
68	1	B-Cell	B-Cell	Yes
69	0.9996	B-Cell	B-Cell	Yes
70	0.9998	B-Cell	B-Cell	Yes
71	1	B-Cell	B-Cell	Yes
72	0.9998	B-Cell	B-Cell	Yes

TABLE 8
Gene Segments Used to predict B-cell and T-cell ALL
	Gene Model
	(Classifier)	Actual
Gene Segment (Classifier) Used	Values	Values
Onconeural ventral antigen-1 (Nova-1) mRNA	U04840_at	828.711548	93
Ini1 mRNA	U04847_at	758.938538	123
RORA RAR-related orphan receptor A	U04898_at	237.028641	−60
FUSE binding protein mRNA	U05040_at	1345.72998	891
Rar protein mRNA	U05227_at	958.517456	−38
Fetal Alz-50-reactive clone 1 (FAC1) mRNA	U05237_at	1616.58411	635
MB-1 gene	U05259_rna1_at	5244.02344	5314

Clusters were also found in the 5 and 10 gene grouping runs. These clusters were generally smaller but the analysis of these groups also gave indications of large-scale correlation between many genes.

The five gene-grouping runs resulted in several 5-gene groups. FIG. 8 is a chart showing gene group models used for 5-gene diagnostic tests. Five different gene model value sets consisting of four 5-gene groups each (20 genes total) were used to create five

Results from 5-gene test 1:
Overall Accuracy 0.824
ALL only         0.8
AML only         0.857
Results from 5-gene test 2:
Overall Accuracy 0.735
ALL only         0.8
AML only         0.643
Results from 5-gene test 3:
Overall Accuracy 0.824
ALL only         0.8
AML only         0.857
Results from 5-gene test 4:
Overall Accuracy 0.765
ALL only         0.8
AML only         0.714
Results from 5-gene test 5:
Overall Accuracy 0.735
ALL only         0.75
AML only         0.714

Pathways related to this result comprise:

• • Regulation of hematopoiesis by cytokines,
  • IL-2 Receptor Beta Chain in T cell Activation,
  • Tumor Suppressor Arf Inhibits Ribosomal Biogenesis,
  • Neuropeptides VIP and PACAP inhibit the apoptosis of activated T cells,
  • FAS signaling pathway (CD95 ),
  • HIV-I Nef: negative effector of Fas and TNF,
  • Fc Epsilon Receptor I Signaling in Mast Cell,
  • p38 MAPK Signaling Pathway, and
  • Induction of apoptosis through DR3 and DR4/5 Death Receptors.

FIG. 9 is a chart showing gene group models used for 10-gene diagnostic tests. Two different gene model values sets consisting of two 10-gene groups each (50 genes total) were used to create two different 10-gene diagnostic tests. The gene group models used are listed in FIG. 9.

Results from 10-gene test 1
Overall Accuracy 0.735
ALL only         0.65
AML only         0.857
Results from 10-gene test 2
Overall Accuracy 0.676
ALL only         0.55
AML only         0.857

Pathways related to this result comprise

• • Free Radical Induced Apoptosis,
  • PDGF Signaling Pathway,
  • Rac 1 cell motility signaling pathway, and
  • Selective expression of chemokine receptors during T-cell polarization.

Genes of special interest from this result are SODI, Sm protein F, Sm protein G, and HOXA9.

Transmission Pattern within the Network of ALL

In order to determine if a particular transmission pattern within this network (gene expression pattern) can be identified with acute lymphoblastic leukemia (ALL), point models (also known as FGM models) from all 7-gene groups for all 38 patients were clustered. Clusters were examined that contained only 7-gene groups from the patients with ALL. Two 7-gene group model patterns were identified, which correlated with the largest number of corresponding models in other ALL patients and with none of the AML patients. To test how accurately these two patterns classified ALL patients, correlations were also tested with this diagnostic/classification method on the Golub independent data. This method identified ALL patients form AML patients to ˜85% accuracy. (See the Results section) This gives credence to this method both as a diagnostic technique and lends significance to the gene models used. The chance of these two gene group model patterns producing an 85% result by chance is roughly 1 in 50,000. Similarly, tests were performed on the 5 and 10-gene groups. The diagnostic accuracy varied from 67.6 to 82.4%. Many pathways and genes were identified as being significant in the course of this test. Several of these appeared to mesh with current knowledge in the field (See Results section).

The test cited above identified a particular group of genes and a gene expression pattern within them that appears to identify ALL. This does not necessarily mean, however, that this group of genes is in the hypothetical ALL ring within a network of the kind illustrated in FIG. 2. To produce evidence of this type of large-scale transmission a test was produced which compared all 7-gene models to all corresponding models between patients in the independent set and a randomly chosen ALL and AML patient from the training set. All model correlations were calculated and averaged for both the ALL and AML patients chosen. The diagnostic decision was based on which comparison had the higher average correlation. This test produced a diagnostic accuracy of 82.4%. More importantly, this result is a statistically significant indication of gene expression pattern reflecting a clinical expression of ALL throughout the 7,000+gene set. The same test was also performed with the 10-gene models to also produce a statistically significant result (See Results section).

The results of the7-gene grouping all models to all models diagnostic test (based on average correlation with randomly chosen ALL and AML patient from the training set) are as follows:

Overall Accuracy 0.824
ALL only         0.85
AML only         0.786

The results of the 10-gene grouping all models to all models diagnostic test (based on average correlation with randomly chosen ALL and AML patient from the training set) are as follows:

Overall Accuracy 0.735
ALL only         0.7
AML only         0.786

The results for all 7-gene models to 7-gene group 1 model pattern diagnostic test (based on average correlation with randomly chosen ALL and AML patient from the training set) are as follows:

Overall Accuracy 0.765
ALL only         0.9
AML only         0.571

Upstream and Downstream Pathways in ALL Genetic Network

It can be further determined if this transmission pattern be traced upstream in the network. Starting with the two specific 7-gene model patterns used to diagnosis ALL, an attempt was made to find correlations between these patterns and all 7-gene models in both ALL and AML patients in the training set.

The assumption was that finding this expression pattern in an area closer inside than the “ALL ring” in FIG. 4 would constitute finding an upstream gene grouping. In this area ALL and AML have yet to reach genes which will determine their specific clinical expression.

There was one 7-gene grouping whose models correlated with one of the ALL diagnostic patterns in all patients, both ALL and AML. There were also two other 7-gene groups that met this condition in almost all patients in the training set. All three of the gene groups are listed under the heading “Most Common Upstream Gene Groups correlated to 7-gene Model Patterns Used in Diagnostic Test” in the Results section.

To strengthen the assumption that this pattern was being transmitted through a large section of the network, we performed the following test. We correlated the single 7-gene diagnostic pattern cited above against all the 7-gene models in each of the AML patients in the training set. The highest average correlation was found. The same correlation test was performed across all the independent patients. A patient was identified as ALL if the average correlation was greater than the highest average AML correlation from the training set. This test identified ALL to 76% accuracy. The diagnostic score is somewhat low, but the probably of chance occurrence is roughly 1 in a 1,000. This provides statistical evidence that not only can large-scale gene expression be seen in ALL patients, a single pattern can be seen as being transmitted through a large section of a genetic network involved in the clinical expression of ALL.

Most common upstream gene Groups correlated to 7-gene model patterns which can be used in a diagnostic test are:

Group 1

• • GAA gene extracted from Human lysosomal alpha-glucosidase gene exon 1
  • AGA Aspartylglucosaminidase
  • 2-19 gene (2-19 protein) extracted from H.sapiens G6PD gene for glucose-6-phosphate dehydrogenase
  • CYCLIC-AMP-DEPENDENT TRANSCRIPTION FACTOR ATF- 1
  • Usf mRNA for late upstream transcription factor
  • PRTN3 Proteinase 3 (serine proteinase, neutrophil, Wegener granulomatosis auto antigen)
  • RPS3 Ribosomal protein S3

Group 2

• • XP-C repair complementing protein (p58/HHR23B)
  • KIAA0031 gene
  • Estrogen responsive finger protein
  • C3G protein
  • CDH 11 Cadherin 11 (OB-cadherin)
  • 60S RIBOSOMAL PROTEIN L23
  • SM22-ALPHA HOMOLOG

Group 3

• • CD1D CD1D antigen, d polypeptide
  • 5,10-methenyltetrahydrofolate synthetase mRNA
  • PTPRD Protein tyrosine phosphatase, receptor type, delta polypeptide
  • GT197 partial ORF mRNA, 3′ end of cds
  • The longest open reading frame predicts a protein of 202 amino acids, with fair Kozak consensus at the initial ATG codon; an in-frame TGA codon is seen at nucleotide 8; ORF; putative gene extracted from Homo sapiens
  • GT198 mRNA, complete ORF
  • GT212 mRNA
  • RPL37 Ribosomal protein L37

Pathways related to upstream gene groups comprise:

• • Oxidative reactions of the pentose phosphate pathway,
  • TNF/Stress Related Signaling, fMLP induced chemokine gene expression in HMC-1 cells,
  • Proepithelin Conversion to Epithelin and Wound Repair Control,
  • Rac1 cell motility signaling pathway, and
  • Catabolic pathway for asparagine and asparate.

FIG. 10 shows a preliminary diagram of downstream causality from two diagnostic 7-gene groups used in the 7-gene test. Pathways related to downstream causality groups comprise ALK in cardiac myocytes, WNT Signaling Pathway, BCR Signaling Pathway, Fc Epsilon Receptor I Signaling in Mast Cell, Neuropeptides VIP and PACAP inhibit the apoptosis of activated T cells, Regulation of hematopoiesis by cytokines, Cytokines and Inflammatory Response, Integrin Signaling Pathway, AKT Signaling Pathway, Regulation of transcriptional activity by PML, mTOR Signaling Pathway, and Regulation of eIF4e and p70 S6 Kinase.

Genes of special interest from this result include: FEZI, EIF4A

Causal Picture of the Network

In order to determine if a transmission pattern can be used to create a causal picture of the network, a partial picture of causality going downstream from the 7-gene diagnostic groups was constructed using a combination of correlations with the actual diagnostic patterns and correlations with the actual 7-gene diagnostic group models for each patient. A 7-gene group was considered a candidate for a downstream link if the gene model did not correlate with the corresponding model in any of the ALL patients and its 7-gene model correlated with one of the two diagnostic patterns. Downstream causality was considered found when the last condition only occurred when there was a correlation between its 7-gene model and the diagnostic group 7-gene models. The assumption is that this 7-gene group's expression (as part of an ALL network) was apparently “switched on” by the diagnostic 7-gene group correlation upstream. The results of this preliminary causal analysis are in the Results section.

In summary, this example describes a method of pathway conjecture and diagnosis using fractal genomics modeling (FGM). The 7-gene group results were focused on, but many interesting pathway and gene inferences seems to come out of the 5 and 10 gene tests. Within the related pathways listed, there is a great deal of overlap between the pathways connected with the downstream links and the 5-gene groups. This is intriguing because in a scale-free network of the kind shown in FIG. 2, the genes with 5 links would tend to be both downstream of genes with 7-links and also more prevalent. This could provide a framework for building the interconnected downstream pathways actually represented in these groups. This would also lend credence to the idea that the 10-gene models tend to reflect pathways upstream of the 7-gene groups. Together these two notions could perhaps be used to map the biochemistry within the “ALL ring” in FIG. 4. This also might explain why the 5-gene and 10-gene results were results less accurate, since they were dealing with pathways slightly removed from the “critical point” in ALL clinical expression. There could also be other biophysical reasons for this. Statistical evidence was produced toward validation of the model of clinical expression shown in the genetic network in FIG. 4. In this process of arriving at this evidence, new tools and approaches have been identified for extracting a great deal of information about the structure and function of such a network. New diagnostic methods have also been identified. The diagnostic results, although statistically significant, were still somewhat low compared to other methods. This could well be due to problems with the Golub methodology which were accurately portrayed in a false diagnosis by FGM. We will apply FGM to more up-to-date and accurate gene expression studies to further validate, improve, and extend the diagnostic approaches and pathway information of this invention. In the process, we will continue to translate the biophysics of gene expression models into the pathways and targets of interest to researchers in the medical field. Since FGM is data independent, we hope to apply these approaches to proteomic and even clinical data as well.

Example 4

Identification of Biomarkers and Sudying of Genetic Network of HIV Infection using “Shotgun” FGM Approach

In this example, the gene expression values were measured from 12,558 genes from the brain tissue for each of 9 human subjects, 5 of which were positive for HIV infection and 4 were negative for HIV. The gene expression values were measured using the Affymetrix Human Genome U95A chips. The 12558 values were randomly broken into 897 non-overlapping subgroups of 14 genes. The number of genes in each subgroup (14) was chosen using the criteria of picking a relatively small number in the range, for example, 5 to 20, which went evenly into 12,558. In the initial run, the ordering of the genes for random grouping into groups of 14 was the ordering of the genes on the Affymetrix chip. The raw data were preprocessed using Affymetrix Microarray Suite version 4.0 (MAS 4.0).

FGM then mapped each of these subgroups to a single point (also known as the “point-model,” or the “FGM model,” or simply the “model”) on a two-dimensional surface. Clusters of the point-models on the surface were found which contained corresponding to the same gene group for a significant number of either the positive or negative subjects. For example, a cluster containing the 47^th 14-gene group on the U95A chip for 4 of the HIV-1 infected subjects and only 1 of the HIV negative subjects, or all of the HIV positives and none of the negatives. The gene groups found having this property were deemed candidates for Biomarker Evaluation (BE) Gene groups were also deemed candidates for BE if the only valid models produced were two or more of the same positive or negative phenotype.

After the initial run using the Affymetrix ordering on the chip, the ordering of the gene expression results were randomly scrambled and run again. This was done a total of 17 times in this example. This produced 9 14-gene groups that passed Biomarker Evaluation. The genes within those 9 groups were then weighted as to their importance within each group toward differentiating between positive and negative subjects.

The candidates for Biomarker Evaluation (BE) were each examined to see if they successfully allowed classification of all 9 subjects in the study. If such a gene group was found, it was deemed to have passed BE and was listed as a biomarker. In this example, the candidates were considered to have successfully classified all 9 subjects based on two criteria:

• • (1) The candidate (gene group) only produces valid models (the absolute value of the Pearson correlation |C|>0.95) for all HIV positive or all HIV negative subjects.
  • (2) The candidate's expression values from a single subject produced a ranked differentiation by phenotype when correlated with the other 8 subjects. For example, if the 14 gene expression values in group 47 from HIV positive patient 1 produced higher Pearson values when correlated against the 14 values from group 47 in the 4 other HIV positive subjects than those produced by the correlation with group 47 in the HIV negative subjects.

An example of the differentiation described by criteria (2) above is shown in FIG. 11. Once a gene group was labeled a biomarker, a weighting procedure was performed to determine which genes in the group were deemed most important in distinguishing phenotype. This was done using the following procedure:

• • (1) The average expression value for each of the 14 genes was calculated for the HIV positive and HIV negative subjects.
  • (2) The Pearson correlation value was evaluated between the two 14 average gene expression values from the HIV positive subjects and the HIV negative subjects.
  • (3) For each of the 14 genes the average HIV negative expression value was made equivalent to the HIV positive value and the Pearson correlation between all 14 values was calculated. The absolute value of the difference from the correlation calculated in step (2) was evaluated.
  • (4) The 14 values generated in step (3) were ranked from a highest to lowest and the weight of 1.0 was given to the highest value. The ratios of the remaining values over the highest value were calculated to produce the weighting of the other 13 values.

Nine biomarkers were found in the FGM analysis based on 17 FGM runs. The genes in each group with weighting of 0.02 or greater are listed in Table 9.

TABLE 9
List of 9 Biomarkers from 17 FGM Runs
	Weight of
Biomarker	gene in the
No.	biomarker	Regulation	Probe ID	Description of gene
1	1	UP	32243_g_at	Cluster Incl. AL038340:DKFZp566K192_s1 Homo
				sapiens cDNA, 3 end clone=DKFZp566K192 /
				clone_end=3 /gb=AL038340
	0.12	UP	32238_at	Cluster Incl. AF001383:Homo sapiens
				amphiphysin II mRNA, complete cds /
				cds=(171,1619) /gb=AF001383 /gi=2199534
	0.11	UP	32242_at	Cluster Incl. AL038340:DKFZp566K192_s1 Homo
				sapiens cDNA, 3 end /clone=DKFZp566K192 /
				clone_end=3 /gb=AL038340
	0.1	DOWN	32235_at	Cluster Incl. AB011116:Homo sapiens mRNA for
				KIAA0544 protein, partial cds /cds=(0,1751) /
				gb=AB011116 /gi=3043611
	0.03	DOWN	32241_at	Cluster Incl. AL050265:Homo sapiens mRNA;
				cDNA DKFZp564O1716 (from clone
				DKFZp564O1716) /cds=(88,1332) /gb=
	0.03	UP	32236_at	Cluster Incl. AF032456:Homo sapiens ubiquitin
				conjugating enzyme G2 (UBE2G2) mRNA,
				complete cds /cds=(55,552) /gb=
	0.02	UP	32231_at	Cluster Incl. AL050106:Homo sapiens mRNA;
				cDNA DKFZp586I1319 (from clone
				DKFZp586I1319) /cds=UNKNOWN /gb=
	0.02	UP	32234_at	Cluster Incl. AF007871:Homo sapiens torsinA
				(DYT1) mRNA, complete cds /cds=(42,1040) /
				gb=AF007871 /gi=2358278 /ug=
	0.02	DOWN	32240_at	Cluster Incl. D31889:Human mRNA for KIAA0072
				gene, partial cds /cds=(0,1513) /gb=D31889 /
				gi=505105 /ug=Hs.193725
	0.02	UP	32239_at	Cluster Incl. U69263:Human matrilin-2 precursor
				mRNA, partial cds /cds=(0,941) /gb=U69263 /
				gi=2072789 /ug=Hs.19368
2	1	UP	36892_at	Cluster Incl. AF032108:Homo sapiens integrin
				alpha-7 mRNA, complete cds /cds=(161,3574) /
				gb=AF032108 /gi=2897115
	0.71	DOWN	34400_at	Cluster Incl. AI540957:PEC1.2_15_G03.r Homo
				sapiens cDNA, 5 end /clone_end=5 /gb=AI540957 /
				gi=4458330 /ug=Hs.
	0.26	DOWN	36169_at	Cluster Incl. N47307:yy87a10.s1 Homo sapiens
				cDNA, 3 end /clone=IMAGE-280506 /
				clone_end=3 /gb=N47307 /gi=1188473
	0.04	DOWN	37778_at	Cluster Incl. AJ005273:Homo sapiens mRNA for
				Kin17 protein /cds=(65,1246) /gb=AJ005273 /
				gi=3850703 /ug=Hs.123647
	0.02	DOWN	38536_at	Cluster Incl. D87717:Human mRNA for KIAA0013
				gene, complete cds /cds=(721,3792) /gb=D87717 /
				gi=1663709 /ug=Hs.
3	1	UP	39026_r_at	Cluster Incl. AF052114:Homo sapiens clone
				23887 mRNA sequence /cds=UNKNOWN /
				gb=AF052114 /gi=3360421 /ug=Hs.
	0.19	DOWN	35773_i_at	Cluster Incl. AA527880:nh86h10.s1 Homo
				sapiens cDNA, 3 end /clone=IMAGE-965443 /
				clone_end=3 /gb=AA527880 /gi=
	0.09	DOWN	34858_at	Cluster Incl. D79998:Human mRNA for KIAA0176
				gene, partial cds /cds=(0,797) /gb=D79998 /
				gi=1136411 /ug=Hs.4935 /len=
	0.08	DOWN	1169_at	D88799 /FEATURE= /DEFINITION=D88799
				Homo sapiens mRNA for cadherin, partial cds
	0.05	UP	39161_at	Cluster Incl. AF052093:Homo sapiens clone
				23685 mRNA sequence /cds=UNKNOWN /
				gb=AF052093 /gi=3360399 /ug=Hs.
	0.02	DOWN	40897_at	Cluster Incl. M26061:Human cGMP
				phosphodiesterase alpha subunit (CGPR-A)
				mRNA, complete cds /cds=(120,2702) /gb=
	0.02	UP	41634_at	Cluster Incl. D87445:Human mRNA for KIAA0256
				gene, complete cds /cds=(1424,3331) /
				gb=D87445 /gi=1665778 /ug=Hs.
4	1	UP	41745_at	Cluster Incl. X57352:Human 1-8U gene from
				interferon-inducible gene family /cds=(237,638) /
				gb=X57352 /gi=311374 /ug=
	0.56	DOWN	33712_at	Cluster Incl. N63574:yy63f05.s1 Homo sapiens
			1211403	cDNA, 3 end /clone=IMAGE-278241 /
				clone_end=3 /gb=N63574 /gi=
5	1	DOWN	36591_at	Cluster Incl. X06956:Human HALPHA44 gene for
				alpha-tubulin, exons 1-3 /cds=(0,1343) /
				gb=X06956 /gi=32014 /ug=Hs.75318
	0.89	DOWN	1009_at	U51004 /FEATURE= /DEFINITION=HSU51004
				Homo sapiens protein kinase C inhibitor (PKCI-1)
				mRNA, complete cds
	0.2	DOWN	36080_at	Cluster Incl. AB002332:Human mRNA for
				KIAA0334 gene, complete cds /cds(251,2791) /
				gb=AB002332 /gi=2224608 /ug=
	0.18	DOWN	39407_at	Cluster Incl. M22488:Human bone morphogenetic
				protein 1 (BMP-1) mRNA /cds=(29,2221) /
				gb=M22488 /gi=179499 /ug=
	0.11	DOWN	38258_at	Cluster Incl. U79290:Human clone 23908 mRNA
				sequence /cds=UNKNOWN /gb=U79290 /
				gi=1710270 /ug=Hs.90449 /len=
	0.09	UP	41531_at	Cluster Incl. AI445461:tj34g07.x1 Homo sapiens
				cDNA, 3 end /clone=IMAGE-2143452 /
				clone_end=3 /gb=AI445461 /gi=4288374
	0.04	DOWN	40598_at	Cluster Incl. W20138:zb40d12.r1 Homo sapiens
				cDNA, 5 end /clone=IMAGE-306071 /
				clone_end=5 /gb=W20138 /gi=1296008
	0.03	UP	33452_at	Cluster Incl. M15518:Human tissue-type
				plasminogen activator (t-PA) mRNA, complete
				cds /cds=(76,1764) /gb=M15518 /
	0.02	UP	32684_at	Cluster Incl. AF038174:Homo sapiens clone
				23579 mRNA sequence /cds=UNKNOWN /
				gb=AF038174 /gi=2795893 /ug=Hs.
	0.02	DOWN	35992_at	Cluster Incl. AF087036:Homo sapiens musculin
				mRNA, partial cds /cds=(0,606) /gb=AF087036 /
				gi=3599520 /ug=Hs.42474
	0.02	UP	39977_at	Cluster Incl. U69274:Human zinc finger protein
				mRNA, complete cds /cds=(161,3322) /
				gb=U69274 /gi=4097821 /ug=Hs.47371
	0.02	UP	40690_at.	Cluster Incl. X54942:H.sapiens ckshs2 mRNA for
				Cks1 protein homologue /cds=(95,334) /
				gb=X54942 /gi=29978 /ug=Hs
	0.01	UP	35676_at	Cluster Incl. AF006386:Homo sapiens axonemal
				dynein light chain (hp28) mRNA, complete cds /
				cds=(56,829) /gb=AF006386
6	1	DOWN	38480_s_at	Cluster Incl. U66867:Human ubiquitin conjugating
				enzyme 9 (hUBC9) mRNA, complete cds /
				cds=(806,1282) /gb=U66867
	0.65	UP	36451_at	Cluster Incl. AI743299:wg91b04.x1 Homo sapiens
				cDNA, 3 end /clone=IMAGE-2372431 /
				clone_end=3 /gb=AI743299 /gi=
	0.11	DOWN	41134_at	Cluster Incl. AB023181:Homo sapiens mRNA for
				KIAA0964 protein, complete cds /cds=(435,3404) /
				gb=AB023181 /gi=4589571
	0.05	UP	38972_at.	Cluster Incl. AF052169:Homo sapiens clone
				24775 mRNA sequence /cds=UNKNOWN /
				gb=AF052169 /gi=3360480 /ug=Hs
	0.02	UP	32167_at	Cluster Incl. AF054182:Homo sapiens
				mitochondrial processing peptidase beta-subunit
				mRNA, complete cds /cds=(13,1482)
	0.01	DOWN	37925_r_at	Cluster Incl. AJ245434:Homo sapiens mRNA for
				G3a protein (G3a gene, located in the class III
				region of the major histocompatibility
7	1	DOWN	41602_at	Cluster Incl. AB015202:Homo sapiens gene for
				hippocalcin /cds=(235,816) /gb=AB015202 /
				gi=4417205 /ug=Hs.114215 /
	0.35	UP	32744_at	Cluster Incl. AI526078:DU3.2-7.G08.r Homo
				sapiens cDNA, 5 end /clone_end=5 /gb=AI526078 /
				gi=4440196 /ug=Hs.1948
8	1	UP	39045_at	Cluster Incl. W26655:34c9 Homo sapiens cDNA /
				gb=W26655 /gi=1307498 /ug=Hs.11641 /
				len=1007
	0.1	DOWN	1089_i_at	M64936 /FEATURE= /DEFINITION=HUMRIRT
				Homo sapiens retinoic acid-inducible endogenous
				retroviral DNA
	0.02	DOWN	37284_at	Cluster Incl. U60800:Human semaphorin (CD100)
				mRNA, complete cds /cds=(87,2675) /gb=U60800 /
				gi=1663566 /ug=Hs.
9	1	DOWN	34906_g_at	Cluster Incl. AA977136:oq24f02.s1 Homo sapiens
				cDNA, 3 end /clone=IMAGE-1587291 /
				clone_end=3 /gb=AA977136 /gi=
	0.2	UP	1562_g_at	U27193 /FEATURE= /DEFINITION=HSU27193
				Human protein-tyrosine phosphatase mRNA,
				complete cds
	0.17	UP	38855_s_at	Cluster Incl. D82343:Homo sapiens mRNA for
				AMY, complete cds /cds=(285,692) /gb=D82343 /
				gi=1841335 /ug=Hs.18551
	0.13	DOWN	33211_at	Cluster Incl. AW051889:wz04f05.x1 Homo
				sapiens cDNA, 3 end /clone=IMAGE-2557089 /
				clone_end=3 /gb=AW051889 /gi=
	0.03	UP	39681_at	Cluster Incl. AF060568:Homo sapiens
				promyelocytic leukemia zinc finger protein (PLZF)
				gene, complete cds /cds=(90,2111)
	0.02	UP	32117_at	Cluster Incl. U51698:HSU51698 Homo sapiens
				cDNA /gb=U51698 /gi=1255268 /ug=Hs.16178 /
				len=800

Their apparent regulation, up or down, for HIV positive subjects is also shown. FIG. 12 represents a proposed network diagram derived from the biomarkers found.

The results of this example validate the FGM “shotgun” approach toward genetic network analysis and identification of biomarkers. It is contemplated that the same approach can be applied to differentiate other phenotypes or diseases.

While specific embodiments have been illustrated and described, numerous modifications cone to mind without departing from the spirit of the invention and the scope of protection is only limited by the scope of the accompanying claims.

Claims

1. A method for identifying biomarkers for a phenotype or determining a genetic pathway of a larger gene pool regarding the phenotype in a population, the method comprising: a) providing a plurality of subjects from the population, the plurality of subjects having a positive group with the phenotype and a negative group without the phenotype; b) obtaining gene expression values of the genes in the larger gene pool from each subject; c) randomly breaking the gene expression values into subgroups for each subject, each subgroup containing N gene expression values from a small group of N genes; d) developing a single point-model for each subgroup of gene expression values using a fractal genomics modeling (FGM) method generated from a multi-dimensional FGM surface; e) mapping the FGM point models of each subgroup of genes for all subjects on the FGM surface; f) clustering the point models based on their proximity on the surface; g) identifying a small group of genes in a cluster containing the same group of genes for a significant number of either the positive or negative subjects; h) determining the correlative patterns between the small group of genes for all subjects in the plurality of subjects to determine if the subgroup effectively differentiates the positive from the negative subjects; and i) identifying the small group of genes in step (h) that positively differentiates the positive from the negative subjects as a candidate for a biomarker for the phenotype in the population.

2. The method of claim 1, further comprising repeating steps (c) through (i) wherein the gene expression values are scrambled to generate a different ordering of the genes in the larger gene pool to find another biomarker or another pathway.

3. The method of claim 1, further comprising evaluating the candidate for biomarker wherein the candidate is listed as a biomarker for the phenotype if it successfully allows classification of all the subjects tested.

4. The method of claim 3, wherein the candidate successfully allows classification of all the subjects tested when: (a) the candidate only produces valid models for all phenotype positive or all phenotype negative subjects; and (b) the candidate's expression values from a single subject produce a ranked differentiation by phenotype when correlated with the other subjects.

5. The method of claim 3, wherein a weighted procedure is performed to determine which genes in the biomarker are deemed most important in distinguishing phenotype, the procedure comprising: (a) calculating the average expression value for each of the N genes for the positive and the negative subjects with regard to the phenotype; (b) evaluating the correlation value between the two N average gene expression values from the phenotype positive subjects and the phenotype negative subjects; (c) making each of the N genes the average phenotype negative expression value equivalent to the phenotype positive value, calculating the correlation between all N values and evaluating the absolute value of the difference from the correlation calculated in step (b); (d) ranking the N values generated in step (c) from highest to lowest and giving the weight of 1.0 to the highest value; and (e) taking the ratios of the remaining values other the highest ranked value to produce the weighting of the other values.

6. The method of claim 5 further comprises checking the biomarkers identified against known pathways and networks.

7. The method of claim 1 wherein the larger gene pool consists of more than about 10,000 genes.

8. The method of claim 1 wherein the larger gene pool consists of the entire genome of the subject in the population.

9. The method of claim 1 wherein gene expression values are obtained by using a gene chip.

10. The method of claim 1 wherein the number of genes N from the small gene group is from 5 to 20.

11. The method of claim 1 wherein the subject is a human being.

12. The method of claim 1 wherein the phenotype is a disease.

13. The method of claim 12, wherein the disease is cancer.

14. The method of claim 12, wherein the disease is human immunodeficiency virus (HIV) infection.

15. The method of claim 12, wherein the disease is Down's Syndrome.

16. The method of claim 1, wherein the correlative pattern is a Pearson correlation.

17. The method of claim 1, wherein the significant number in step (g) is equal to two or greater.

18. The method of claim 1, wherein the multi-dimensional surface is a two dimensional surface.

19. The method of claim 1, wherein the surface is derived from a Julia set.

20. The method of claim 1, wherein the surface is from or near the boundary of a Mandlebrot set.

21. The method of claim 4, wherein the models are valid when the absolute value of Pearson correlation is greater than 0.95.

22. The method of claim 5, wherein the correlation is a Pearson correlation.

23. The method of claim 5, further comprising checking the biomarkers against known pathways and networks to give extra significance to the biomarkers based on biochemical considerations.

24. The method of claim 1 wherein the method is automated.

25. The method of claim 2, wherein the method is automated.

26. The method of claim 4, wherein the method is automated.

27. The method of claim 5, wherein the method is automated.