Methods for improving disease diagnosis using measuring analytes

Abstract

To provide methods of improving accuracy of disease diagnosis and associated diagnostic tests involving a correlation of measured analytes with binary outcomes.SOLUTION: A computer-implemented method of diagnosing a disease is provided, comprising steps of: determining concentration of at least three biomarkers in a blind sample collected from a subject; selecting one or more meta-variables associated with the subject; transforming the concentration of the biomarkers as a function of one or more population distribution characteristics and the one or more meta-variables to compute pseudo-concentration representing each biomarker; comparing the pseudo-concentration with a training set model of pseudo-concentration determined for members of population who are known either to have or not have the disease; and determining whether the comparison indicates that the subject has the disease.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to methods for improving the accuracy of disease diagnosis and related diagnostic tests involving correlation of measured analytes with binary outcomes. [Background technology]

[0002] Correlation methods, which use three or more independent variables to correlate a binary outcome (e.g., the presence or absence of a given disease), are often used along with cluster or neighborhood search methods, regression methods, and wavelet methods. For disease prediction, correlations are attempted by measuring common components of blood or serum and using these concentrations as independent variables for predicting various disease states. For a given disease state in which the outcome is "disease" or "not disease," logistic regression methods are often used. Other techniques involve, for example, genetic algorithms. The predictive power of these methods depends heavily on the component analytes selected for the method. Those skilled in the art recognize that many analytes and parameters that appear to have predictive power will not actually improve diagnostic and analytical power.

[0003] Regression methods use the trend of independent variables to correlate with the outcome. Linear methods are based on linear trends, while logistic regression is based on logarithmic trends. Logistic regression is often used in biological disease prediction.

[0004] Group clustering methods look at the variable correlation topology for grouping of similar outcomes. Clustering methods have the advantage of being able to find correlations where the trend is not continuous but has a local reversal of topology in the trend. This method is more sensitive to outcomes with large nonlinearities and large local variability, but has smaller measurement errors and may be more predictive in biological use. In addition, either method can be combined in most cases with cluster methods applied on a smaller scale for full regressions.

[0005] However, some independent variables that theoretically would be correlated do not show predictive trends, and therefore approaches are needed to improve diagnostic accuracy by utilizing patient- and population-specific variables that have not previously provided useful information for diagnosing disease states.

[0006] Numerous studies have been conducted to find biomarkers that can predict disease states, alone or in combination, with sufficient reproducibility and predictive power for clinical use. These studies have met with limited or no success. High abundance proteins (HAPs) have been extensively explored to find single proteins that can make this prediction. Numerous examples have been found, but none with sufficiently low false negative levels to allow patients to be screened for disease by the marker. As a result, such single biomarkers are the only ones used for therapy monitoring, except for PSA, for prostate cancer. To reduce false negatives, the test is forced to severely distort the concentration that indicates a biopsy is appropriate, resulting in very high false positive levels. As many as 80% of men who are shown to need a biopsy are actually negative for prostate cancer.

[0007] DNA markers have also proven to be highly relevant for cancer subtyping in some cases, but again are not suitable for screening for the same reasons as HAP above.

[0008] Proteomic approaches are also being explored using a wide variety of proteins. This work is again focused on HAP or on high levels of effector proteins. This work is dominated by multiplexed protein measurement methods, such as immunoassays, chips and mass spectrophotometry. Very early work has had some success in ovarian cancer. However, the problem with all these methods is that many of the selected proteins do not show a strong correlation with progression from healthy to disease (and many have no known biological association with pathology, as is typical for mass spectrometry, for example). Furthermore, in mass spectrometry, the fact that whole serum samples are interrogated for protein levels by a spectrophotometer creates a significant oversampling problem, and therefore makes it difficult to train correlation algorithms. In the case of mass spectrometry, a whole serum sample may contain more than 200 proteins and have 10,000 mass spectrometry peaks.

[0009] There is also a need in the diagnostic field for techniques that utilize lower abundance proteins that are more useful for diagnostic purposes than HAP, and analytical techniques that allow for the analysis of low abundance biomarkers. Summary of the Invention [Problem to be solved by the invention]

[0010] The present invention and various embodiments are described in the claims that form part of this patent application.Without limiting the above, in a preferred embodiment of a preferred aspect, the present invention relates to improving the diagnostic power and diagnostic accuracy of methods for predicting disease states using multivariate (multivariate) correlation methods.These methods include proteomics, metabolomics and other techniques that involve determining the levels of various biomarkers found in body fluids and tissue samples.

[0011] Various embodiments contemplated by the inventors and discussed in this application include the use of meta-variables, in particular methods to adjust the impact of the measured biomarker analytes on the correlation score. Such meta-variables may be identified based on expert knowledge of immune system responses and knowledge of possible measurement errors. These methods are applicable to blinded samples during training set model building or diagnosis. [Means for solving the problem]

[0012] In one embodiment, the present invention relates to a method for diagnosing a disease, comprising the steps of: a) determining the concentrations of at least three predetermined analytes in blinded samples from a subject; b) selecting one or more meta-variables associated with the subject that vary in a population associated with the subject for members of the population known to have or not have the disease; c) transforming the concentrations of the analytes as a function of one or more population distribution characteristics and the one or more meta-variables to calculate pseudo-concentrations representing each analyte; d) comparing the pseudo-concentrations with a training set model of pseudo-concentrations determined for members of the population known to have or not have the disease; and e) determining whether the comparison indicates that the subject has the disease. It is contemplated that step (a) of determining the concentrations (or levels) of the predetermined analytes can be performed at a different time and place than the remaining steps of the method. Similarly, the other step(s) of the method can also be performed in whole or in part at different times and places. Thus, the inventors also consider as the present invention a method comprising fewer steps, in particular only steps (b) to (e).

[0013] In one embodiment of the present invention, the method uses at least three, at least four, at least five or at least six or more analytes measured in a biological sample taken from a subject or patient or their levels determined in said sample. In another embodiment, the method involves the assessment or prediction of the presence or absence of a given disease, such as solid tissue cancer, including but not limited to breast cancer, prostate cancer and lung cancer.

[0014] In some embodiments, the meta-variable is age. In certain embodiments, the meta-variable is selected from the group consisting of pre-, peri- and post-menopausal status, pubertal stage, weight, geographic location of sample source, body fat percentage, age, race or racial or ethnicity, species or time period (or range).

[0015] In another embodiment, the "comparing" step described herein involves the use of a correlation method selected from techniques including, but not limited to, clustering, neighborhood search, regression, or wavelet analysis. It may also include the use of a mismatched training set model, as appropriate. Such a mismatched training set module can be used with any of the methods of the present invention, for example, in conjunction with the transforming, comparing, and determining steps, which can be repeated using a second training set model, as needed. This second training set model can identify a non-disease state of a population of subjects that partially resembles the serum analyte changes in a disease state, but is not caused by the disease state, as opposed to the state or pathology of the disease itself. Thus, a related embodiment includes a second training set model, and the assessment and prediction of three states: non-disease state, non-disease state partially resembles a disease state, and disease state.

[0016] In another aspect of the invention, the method of the invention may be computer implemented using a microprocessor and further comprise the step of outputting the score in a form useful to medical personnel, e.g., a physician making a disease diagnosis.

[0017] Certain embodiments of the invention utilize mathematical methods to normalize concentrations and smooth irregular and discontinuous distributions of concentrations, including the logarithm of the ratio of measured concentrations and age-adjusted mean values ​​of concentrations of proteins for non-pathological and pathological conditions for which individual samples are predictive, and the ratio of concentrations of proteins for non-pathological and pathological conditions, so that the distribution of the resulting new independent variables used in the correlation is compressed to aid in the correlation calculation.

[0018] In another embodiment of the invention, the relationship between the independent variables and the meta-variables encompasses population distribution characteristics of the independent variables that are related to the degree of non-linearity of the relationship between diseased and non-disease states, one or more groups (normal or non-normal), group mean values, group means, group medians, and group dynamic range values.

[0019] Certain embodiments of the invention include tuning of the training set model to weight the influence of individual biomarkers based on prior (or expert) knowledge about the up- or down-regulation characteristics of the individual biomarkers, e.g., subgrouping or degree of non-linearity in the course of disease progression in typical subjects of the relevant population.

[0020] Certain other embodiments include tuning of the training set model to weight the influence of individual biomarkers based on prior (or expert) knowledge about bimarker plane topology instability in the bimarker plane caused by steep or deep peaks or valleys in the topology of each bimarker plane sufficient to significantly change the risk score or disease state prediction.

[0021] In other embodiments, the training set model is adjusted to weight the influence of individual biomarkers based on prior (or expert) knowledge about biomarker assay uncertainty, e.g., uncertainty that can occur at very low or very high levels in an assay result curve.

[0022] In another embodiment of the invention, unmatched training models are used to adjust or correct for individual blinded samples that exhibit instability in outcome prediction due to topological instability in the bimarker plane caused by steep or deep peaks or valleys in the topology of the bimarker plane sufficient to significantly change the risk score for a given blinded sample.

[0023] Another embodiment of the invention relates to a more personalized medicine approach to diagnosis and therapy in which baseline values ​​of individual protein (or other analyte, e.g., metabolite) concentrations for a disease for which prediction of the presence or diagnosis of the disease is desired, rather than population values, are determined for a subject over a period of time, including periods during which the subject is non-diseased.

[0024] Yet another aspect of the invention relates to the measurement of low abundance biomarkers, including signaling proteins, including at least one biomarker in each of at least three of the categories including immune system inflammatory markers, tumor anti-development markers, cell apoptosis markers, angiogenic protein-related markers, and tissue markers. In one embodiment of the invention, the low abundance biomarkers are very low abundance proteins having concentration levels of less than about 1 pg / ml in samples taken from at least about 20 percent of a relevant population of a given subject.

[0025] Another embodiment of the invention involves determining the concentrations in a biological sample of at least three biomarkers selected from the categories including immune system inflammatory markers, tumor anti-development markers, cell apoptosis markers, angiogenic protein markers and tissue markers, wherein any one or more of the at least three biomarkers other than the tissue markers are low abundance proteins having concentrations determined for at least about 20 percent of a relevant population of a given subject in a subpopulation having a disease for which diagnosis or prediction of the likelihood of disease is desired, of less than about 1 pg / ml.

[0026] In a preferred embodiment, the disease is cancer, more particularly a solid tumor.

[0027] In other embodiments, at least two or more, at least three or more, at least four or more, at least five or more, at least six or more, at least seven or more, at least eight or more, at least nine or more, or at least ten or more biomarkers are evaluated (or their levels are determined).

[0028] Yet another embodiment of the invention includes an evaluation or analysis in which at least one determined (or measured) analyte concentration value is below the LOD, and such analyte(s) concentration value is determined by a linear or other suitable standard curve fitting method between the LOD and the lowest reading of the analyte. Preferably, no analyte exhibits a zero or negative value, and no analyte exhibits a value lower than the approximate lowest acceptable value for that analyte in similar samples.

[0029] In other of its embodiments, the invention involves diagnostic kits comprising reagents for detecting one or more analytes, two or more analytes, three or more analytes, four or more analytes, five or more analytes, six or more analytes, seven or more analytes, eight or more analytes, nine or more analytes or ten or more analytes that are below the LOD for the analyte.

[0030] Another embodiment of the present invention involves computer systems and microprocessor-mediated devices and systems useful for carrying out any of the methods, diagnostic predictions and analyses described herein, including any one or more of the steps discussed herein.

[0031] The following figures, described below, which are incorporated in and form a part of this specification, illustrate exemplary embodiments according to the present disclosure and should not be considered as limiting the scope of the invention, which is therefore open to other equally effective embodiments: The figures are not necessarily to scale, and certain features and certain views of the figures may be exaggerated in scale and outline for clarity and conciseness. [Brief description of the drawings]

[0032] [Figure 1] 1 is a flow chart showing the process of building a training set model (or diagnostic model) and then generating a diagnostic score for blinded samples that assess the risk of having a pathological or non-pathological condition.

[0033] [Diagram 2] FIG. 1 shows a typical population distribution, in this case for the cytokine interleukin-6.

[0034] [Diagram 3] FIG. 1 shows a bimarker plane for one of ten such planes showing the pseudo-concentrations of two of the biomarkers used in the diagnostic method.

[0035] [Figure 4] FIG. 1 shows the bimarker plane containing the data points of the training set.

[0036] [Diagram 5] FIG. 13 shows the bimarker plane without any data points from the training set.

[0037] [Figure 6] FIG. 1 shows a bimarker plane with shaded areas that have reduced influence on immune system response.

[0038] [Figure 7]FIG. 13 shows a bimarker plane with a shaded area in which the impact of topology stability issues is reduced.

[0039] [Figure 8] FIG. 13 shows a bimarker plane with shaded areas in which the effect of known assay measurement uncertainty is reduced.

[0040] [Figure 9] FIG. 1 shows the results of a blinded study with two samples that did not pass the topology uncertainty test and were corrected using the discrepancy algorithm.

[0041] [Figure 10] FIG. 1 shows the results of a breast cancer clinical study, in this case training set cancer scores for training set Model I using 10 bimarker planes.

[0042] [Figure 11] FIG. 1 shows the results of a breast cancer clinical study, in which the training set cancer scores for training set Model II using 105 bimarker planes are shown.

[0043] [Figure 12] FIG. 1 shows the results with actual diagnoses of blinded samples tested in a clinical study.

[0044] [Figure 13] FIG. 1 shows a calibration curve for the protein TNFα with blinded sample data points.

[0045] [Figure 14] FIG. 1 shows error bars for TNFα protein assay for a cancer score error of 5%. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0046] Table 1 shows the stability calculations and figures of merit for Algorithm I from clinical studies.

[0047] Table 2 shows the results of a preclinical study of 868 women for breast cancer.

[0048] Table 3 shows the relative predictive power of various correlation methods for a breast cancer study of 868 women.

[0049] Table 4 shows the results of a preclinical study of 107 women with ovarian cancer.

[0050] Table 5 shows the results of a preclinical study of 259 men for prostate cancer.

[0051] The present invention is described below with reference to detailed exemplary embodiments. It is apparent that the present invention can be embodied in various forms, some of which may differ significantly from those of the disclosed embodiments. Therefore, the specific structural and functional details disclosed below are merely representative of the present invention and do not limit the scope of the present invention.

[0052] In a preferred embodiment, the present invention relates to a method for diagnosing a disease, which is described in more detail below. As an introduction to the following description, the method generally utilizes measured concentrations of a given analyte in blinded samples from patients for whom a predictive diagnosis of disease or not is desired. According to the method of the present invention, each analyte concentration is converted to a pseudo concentration based on the use of at least one selected meta-variable associated with the patient. This meta-variable also varies in a selected population associated with the patient. In the selected population, the concentration of the same given analyte is measured for members of the population known to have or not have the disease. For disease diagnosis, the pseudo concentrations are processed by the methods and algorithms described below. The processed pseudo concentration values ​​are compared to a diagnostic model (or training set model) of the determined and similarly processed pseudo concentrations for members of the population known to have or not have the disease. Finally, it is determined whether the evaluation of the sample taken from the patient indicates that the patient's condition is in a non-pathological or pathological population group. This determination can be seen as a result output from the computerized system for use by, for example, a health care provider.

[0053] This step-by-step process is depicted in the flow chart of Figure 1. A training set model is first constructed, the final result of which can provide a diagnostic outcome for unknown patient samples, referred to as blinded samples (because the exact diagnosis is not known at the time of analysis of these blinded samples). In general, the present invention provides a risk score to a health care provider, who then considers this score along with other patient factors to make a medical decision about the presence or absence of a given condition.

[0054] definition "Analytical sensitivity" is defined as 3 x standard deviations above the zero calibrator. Diagnostic indications are considered inaccurate for concentrations below this level. Thus, clinically relevant concentrations below this level are considered inaccurate and are not used for diagnostic purposes in clinical laboratories.

[0055] A "bimarker" is a set of two pseudo-concentrations that, when plotted in a two-axis graph (or grid), hereinafter referred to as the "bimarker plane", as shown, for example, in FIG. 3, is normalized with respect to the biological transition from non-disease to disease and is functionally related to the variation of a metavariable.

[0056] "Biological sample" means a tissue or bodily fluid, e.g., blood or plasma, obtained from a subject in which the concentration or level of an analyte (also called a marker or biomarker) that provides diagnostic information can be determined.

[0057] "Biomarker" or "marker" refers to a biological component of a subject's biological sample, typically a protein or metabolomic analyte measured in a bodily fluid, e.g., serum proteins. Examples include cytokines, tumor markers, etc.

[0058] A "blinded sample" is a biological sample taken from a subject who is not known to have a given disease diagnosis and about whom a prediction as to the presence or absence of that disease is desired.

[0059] The "limit of detection" (LOD) is defined as a concentration value that is 2 x standard deviations above the value of the "zero" concentration calibrator. Typically, the zero calibration is performed with 20 or more replicates to obtain an accurate representation of the standard deviation of the measurements. Concentration determinations below this level are considered to be zero or nonexistent, for example, for viral or bacterial detection. For purposes of the present invention, the use of 20 replicates is preferred, but 1.5 x standard deviation may be used if samples are run in duplicate. Diagnostic representations requiring concentration values ​​will not be below this level.

[0060] "Low abundance protein" is a protein that is present at very low levels in serum. This level is not clearly defined in the literature, but as used herein, this level will be less than about 1 picogram / milliliter in serum or plasma and other body fluids from which the sample is taken.

[0061] "Metavariable" refers to information characteristic of a given subject, other than the concentrations or levels of analytes and biomarkers, that is not necessarily individualized or specific to that subject. Examples of such metavariables include, but are not limited to, the subject's age, menopausal status (premenopausal, perimenopausal, and postmenopausal) and other conditions and characteristics, such as pubertal stage, weight, geographic location or region of patient residence, geographic source of the biological sample, body fat percentage, age, race or racial mix, or time of year.

[0062] "Population distribution" refers to the range of concentrations of a particular analyte in biological samples of a given subject population. A specific "population" refers to, but is not limited to, individuals selected from a geographic region, a particular race, or a particular gender. The population distribution characteristics selected for use as described in this application further contemplate the use of two distinct subpopulations within that larger defined population that are members of the population diagnosed with a given disease state (disease subpopulation) and without the disease state (non-disease subpopulation). The population may be any population for which disease prediction is desired. Furthermore, it is believed that a suitable population includes subjects with disease that has progressed to a particular clinical stage compared to other stages for a given disease.

[0063] "Population distribution characteristics" are determinable within a population distribution of a biomarker, for example, the arithmetic mean concentration of a particular analyte, or its median concentration, or the dynamic range of concentrations, or how the population distribution falls into groups that are discernible as distinct peaks as the degree of up- or down-regulation of various biomarkers or metavariables of interest is affected by disease onset or progression as a patient undergoes a biological transition or progression from a non-pathological to a pathological state.

[0064] "Predictive power" refers to the average value of the sensitivity and specificity of a diagnostic assay or test.

[0065] "Pseudo concentration" refers to a replacement or substitution value for the concentration of a measured biomarker, and is in effect a new independent variable that can be used in diagnostic correlation analysis. Pseudo concentrations are related to and calculated from the concentration of a measured biomarker analyte, where such analyte has inherent predictive power for a given disease state. The pseudo concentration is calculated using the population distribution characteristics of interest, adjusted by the meta-variables, to transform the actual concentration of the predictive biomarker for a given patient for whom a diagnosis is desired.

[0066] "Topological instability" is a region on the lattice of most or all of the bimarker planes, where all and most of the points of the region are located in the steep section of the topology. The topology is the shape of a multidimensional correlation calculation that takes into account all of the measured independent variables (i.e., the determined biomarker concentrations) and meta-variables. This topology for a single meta-variable value is at least five-dimensional (and can be more) for five biomarker measurements. The topology also changes shape as the values ​​of the meta-variables change. This multidimensional topology can be visually visualized in a decomposed state by taking ten biplanar slices through the topology. This creates a "risk" that the calculated disease score is incorrect due to measurement noise. This score can be obtained by weighting the individual bimarker plots in terms of their predictive power for diseased and non-pathological conditions, as well as by considering other factors, such as topology measurement instability and simple measurement error. The score range is arbitrary, as known to those skilled in the art, and the value represents the percent probability that the patient is in a diseased or non-pathological condition.

[0067] A "training set" is a group of patients (typically 200 or more to achieve statistical significance) with known biomarker concentrations, known metavariable values, and known diagnoses. The training set is used to determine the axis values, the "pseudo concentrations" of the "bimarker" plane, and the score grid points from the cluster analysis used to score the individual blinded samples.

[0068] A "training set model" is a group of algorithms constructed from a training set that allows the evaluation of blinded samples for predicted outcomes in terms of the probability that a subject (or patient) has or does not have a disease. The "training set model" is then used to calculate a score for the blinded samples for clinical and diagnostic purposes. For this purpose, it provides a score that indicates the percent likelihood of disease or non-disease, or any other readout that a health care provider developing a patient diagnosis may prefer, over any range.

[0069] The "unmatched training set model" (or "second-order algorithm") is a second-order training set model that uses a different phenomenological data reduction method such that individual points on the grid of the bimarker plane are unlikely to be unstable in both the first-order correlated training set model and this second-order algorithm.

[0070] Consideration Certain aspects of the invention, including defined terms, are discussed in more detail below as a guide to those of skill in the art in the practice of the invention.

[0071] Meta variables The claimed invention relates, in part, to improved diagnostic methods for correlation diagnostic assays using meta-variables. Such meta-variables may contribute predictive power for diagnosis when they represent a significant range of variation of "normal" values ​​in members of a population known to have or not have a given disease. As described herein, meta-variables are used to convert or transform measured analyte levels into "pseudo-concentrations." It is contemplated that the levels or concentrations of various analytes can be measured or determined by techniques known to those skilled in the diagnostic arts.

[0072] Even if a meta-variable is not particularly predictive in itself, it may be relatively more informative if it is physiologically or physicochemically related to the subject's biological status that reflects changes over time as the pathology develops. For example, body mass index (BMI) is an available meta-variable, and body weight itself affects the levels of various signaling proteins in heart disease. In the method of the present invention, BMI can be significantly more useful in predictive assays when used as a meta-variable rather than as an independent variable, such as another independent variable, e.g., the measured levels of various circulating blood proteins. The present invention is based, in part, on the discovery that body weight variation across a population associated with human subjects is further associated with a determinable population distribution pattern of measured serum protein levels. These protein (or biomarker) levels are the independent variables measured for diagnostic purposes as a given subject undergoes a biological transition (or progression) from a non-pathological to a pathological condition.

[0073] Similarly, the inventors have shown that the age of a subject is not clinically predictive in the diagnosis of disease, e.g., cancer, when used by itself as an independent variable along with measured analyte levels in traditional correlation methods, but when age is used as a meta-variable according to the methods of the present invention, its use improves diagnostic accuracy.

[0074] In general, a meta-variable as defined and described in this application will be predictive for diagnostic purposes if the population distribution characteristics of the measured analytes show significant functional separation (or variance) compared to non-diseased and diseased subjects, meaning that the relationship between the meta-variable (age) and the population characteristic of interest (the arithmetic mean value of the population) for diseased and non-diseased conditions is significantly different.

[0075] The graph in Figure 2 shows the population distribution of IL-6, which notably includes four distinct population subgroups, named I to IV, reflecting immune responses to cancer progression. The higher concentration groups are believed to be the result of stronger immune responses to immune stimuli, such as infection, wounds, allergies, and of course cancer, with the highest group (IV) showing a pronounced nonlinear response of IL6 to these stimuli. These population distribution characteristics and such groupings of measured analytes can be used to improve the predictive power of training set diagnostic models.

[0076] The metavariable diagnostic method of the present invention is based on a study involving the evaluation of approximately 868 patient samples in which the concentrations of five probative low-level signaling proteins (PSA, IL-6, IL-8, TNFα, and VEGF) were measured in subjects who were clinically diagnosed with or without breast cancer. Protein (or analyte) levels are classical independent variables for disease diagnostic tests.

[0077] The inventors also obtained age information for each of these subjects.Classical logistic regression analysis of five biomarkers achieved a predictive power of about 82%, and it was found that the use of age as the sixth independent variable in this analysis method brought about only a small improvement in predictive power.The data clustering method using only the determined biomarkers achieved a slightly higher predictive power of about 88%.Again, the use of age as an independent variable was not substantially more predictive.

[0078] Similarly, the use of classical cluster analysis in conjunction with log-transformed concentration values ​​achieved a predictive power of about 92%, while the use of age as an independent variable increased its predictive power by less than 0.5%. In this type of analysis, as is known, the log of the sample concentration was used because these analyte concentrations can span a dynamic range of four or more logs. It is also known that the blood levels of five analytes in cancer patients tend to increase to very high concentrations, but this is not always the case. Thus, this approach brings the points of the training set model closer together on the multidimensional cluster plot, avoiding the tendency to oversample cluster points at low concentrations. This log compression method is often used because it reduces interval bias.

[0079] The inventors have confirmed that, for example, the age of individual subjects can be used to create what is referred to herein as a meta-variable. This meta-variable is then used to create a new independent variable, referred to herein as a pseudo-concentration. The pseudo-concentration is calculated from the population distribution characteristics of the measured independent variable (in this case, protein concentration). By doing so, a predictive power of about 97% to more than 100% was obtained in two analytical models, as discussed in more detail herein. For the purposes of this application, a predictive power of more than 100% means that "pathological" and "non-pathological" are separated by a significant gap in correlation scores.

[0080] Population distribution characteristics: Examples of population distribution characteristics that the inventors found to be probative are the arithmetic mean concentration values ​​of each measured analyte for patients with or without disease, when age is selected as a metavariable; and the median concentration values ​​for the subgroupings of concentration values ​​for the separated age groups (or the population as a whole) for the diseased and non-diseased patient groups shown in Figure 2. The subgroupings shown in the population distribution plots mathematically adjust differently, i.e., compress the high levels of upregulated concentrations in the group above 5 pg / ml. These relationships create new age-based independent variables, which are then used directly in the cluster analysis, rather than the actual concentration levels of the biomarker analytes measured.

[0081] Arithmetic mean values ​​of concentrations are determined for age for patients diagnosed with non-diseased and diseased conditions. This is referred to as the age-adjusted arithmetic mean for each condition. Pseudo-concentration values ​​are calculated from these arithmetic mean values ​​and the actual concentrations for the patients using Equation 1 (described in the subsection on pseudo-concentrations).

[0082] As is common in multiple independent variable correlation analysis, the multiple independent variables are paired in a multidimensional plot that is the basis of classical cluster proximity analysis. Also, to be able to handle the relatively large dynamic range of the measured variables (i.e., analytes), these variables are often compressed. In this case, we describe a method in which the independent variable is a "pseudo-concentration" value based on age or other meta-variables and actual concentration. It is believed that a person skilled in the art can easily identify and select population distribution characteristics that improve predictive power.

[0083] Pseudo-Concentration: To derive predictive power for patient age, which may itself be treated as an independent variable in other diagnostic approaches, we used age as a meta-variable to calculate a "pseudo-concentration" that is then used as a substitute for the actual concentration of the determined analyte. To derive predictive power for all meta-variable values, e.g., age in the case of breast cancer, it is important that the meta-variable must separate between non-disease and disease states in a given population of interest. By "separation" we mean that the relationship observed between the population characteristics of subjects in non-disease and disease subpopulations is different for a particular meta-variable. This can be empirically determined by measurement and calculation. The population of interest is also determined by the nature of the desired predicted outcome group, which can be, but is not limited to, a geographic region, such as the continental United States; a particular racial or ethnic grouping; or a particular gender, e.g., women.

[0084] In the implementation of the method according to the invention for predicting breast cancer disease, preferably the concentrations of five or six very low signaling proteins are the independent variables of the diagnostic correlation. Each of the five biomarkers was measured and the variation with age was determined in a test population. The test population was women aged 35-80 years, 1 / 2 healthy and 1 / 2 diagnosed with breast cancer. Although subjects in a population are generally considered to have or not have disease, the biomarker level determined for a single subject will change as the subject's disease progresses, and not all subjects will have the same level of the biomarker throughout their various stages before and during the onset and progression of their disease. Thus, the population distribution characteristics reflect the normal variability of the biomarker in non-diseased subpopulations and the normal variability of the biomarker in diseased subpopulations.

[0085] The population distribution characteristics in this case were: 1) arithmetic mean values ​​for non-disease and disease (cancer in this case) versus age; and 2) the degree and non-linearity of biomarker upregulation moving from non-disease to disease. For example, the cytokine IL-6 level reflects the subject's response to the immune challenge caused by cancer, so in FIG. 2 we show the upregulation of interleukin 6. IL-6 is a known proinflammatory responder that is upregulated as a signal to the immune system to turn up its systemic response. The four separate groupings show different upregulation levels. These levels are taken into account when calculating the pseudo concentration. For example, the degree of data compression in moving from real concentration to pseudo concentration varies depending on the grouping position, and is very significant in the top group 4.

[0086] To do the above, a population with equal numbers of subjects with non-pathological condition and pathological condition must be measured.The size of this training set is optimally determined by the number of biomarkers used.The preferred size is one in which the predictive power of the training set model is within about 95% of accuracy for a similar or larger blinded population set.The age-adjusted arithmetic mean value of these two conditions can then be determined, and the effect of disease on the degree of upregulation or downregulation of biomarkers can be recognized.

[0087] Since the immune system upregulates proteins when triggered by the presence of cancer or other proinflammatory conditions, Figure 2 shows the population distribution characteristics of one biomarker, IL-6. The conversion from measured to pseudo concentration again involves normalization of the concentration to the age-adjusted arithmetic mean of non-diseased and diseased populations, and compression of the dynamic range of measured concentration values. For example, highly scattered outlying concentrations in the grouping above 5 pg / ml, extending to 100 pg / ml, are highly compressed. This improves predictive power. The result is a new unit-free, normalized independent variable, referred to in this application as pseudo concentration, which in a preferred embodiment reflects the age variation of the population distribution of the biomarker.

[0088] A relationship of the following form is used, involving age-adjusted arithmetic means for non-diseased and diseased and actual concentrations in patient samples: Equation 1: Pseudo-concentration α natural logarithm ((Ci / C (cまたはh) )-(C h / C c )) 2 [In the formula, C i = measured concentration of actual patient analyte C( cまたはh ) = patient age adjusted concentration of this patient analyte (this value is adjusted for whether the patient is non-diseased or diseased) C h = Patient age-adjusted arithmetic mean concentration of analyte in non-diseased patients C c = patient age-adjusted arithmetic mean concentration of diseased patient analyte].

[0089] This equation 1 is designed to adjust compression and expansion depending on the upregulation grouping. See, for example, the peaks in FIG. 2. The above equation for pseudo-concentration accomplishes this requirement, but it will be apparent to one skilled in the art that many other forms of this equation can be implemented. For example, C i , C h and C c σ could directly be the concentration, or the arithmetic mean, the concentration distance from the median, or the distance from the subgroup median or dynamic range edge, as discussed above.

[0090] Pseudo concentrations (unit-less, and therefore not concentrations or levels) are then used in correlation cluster multidimensional plots for analysis. All plots are also normalized to a common characteristic of population distributions; age arithmetic means, medians, or dynamic ranges of subgroupings of non-diseased and diseased (age-adjusted or not). These methods can result in an improvement in predictive power of 5 percentage points or more.

[0091] The case for personalized medicine is gaining in popularity and effectiveness. It is believed that the above disease prediction methods can be personalized by replacing the population distribution characteristics developed and described in these patents for non-disease states with the individual's non-disease baseline measurements. In other words, the C h Values ​​would be actual baseline values ​​for individual patients and not non-diseased population arithmetic mean values, and disease assessments would then be based accordingly on an individual's transition from these measured non-diseased characteristics to disease characteristics indicative of the general population.

[0092] Bi-marker plane: For an analysis using five biomarkers (and one metavariable), there are 10 such bimarker planes. The plot in Figure 3 shows disease grid points in red and non-disease grid points in yellow. The training set samples are characterized for age populations using the measured independent variable (concentration) and metavariable (age), the pseudoconcentration distances on the axes are calculated, and these pseudoconcentrations are applied to each of the 10 bimarker planes. The plot is divided into 2,000 grids on each axis for a total of 40,000 grid points.

[0093] The determination of whether a grid point is diseased or non-disease is computed by determining the distance from each grid point to the nearest measured data point for the training set samples. Figure 3 is an example, where the two biomarkers are IL-6 and VEGF, and the meta variable is age. Both the ordinate and abscissa are pseudo concentrations determined as described above. Thus, the meta variables and measured independent variables are incorporated into the pseudo concentrations on these plots. The grid points are evaluated as non-disease and disease, respectively, and given a corresponding numerical score (e.g., +1 and -1, although the actual numerical value is arbitrary). This score is determined by the calculated distance to the two training set data points of non-disease or disease. The shortest distance determines this score. The number of training set samples can be changed (e.g., from 4 to 6) to perform this distance determination. See Figure 3. A relatively small number of comparison samples may result in a relatively low predictive power. Similarly, increasing the number of comparison samples may reduce predictive power as the "reach" of the lattice points extends into distant topologically non-local regions. The best number is determined by experimental calculations.

[0094] The preferred number of comparison sample points is where the training set model best matches the actual diagnosis. Figure 3 shows the computational process for two biomarkers, IL-6 and VEFG. The unknown grid point (the small square on the plane at about 12.00 on the y-axis and about 4.00 on the x-axis) is assigned to non-disease or disease by determining its distance to the three closest training set data points for each non-disease and disease, in this case. These distances are added, and the grid point is then assigned to the appropriate state of non-disease or disease (computational score of +1 or -1, respectively). Any blinded samples that are diagnosed at some point in the future will be assigned a state score based on their position on this grid. Each blinded sample will also be scored for all biomarker planes. The total number of training set samples can be at least 200 or more.

[0095] The overall cancer score for the blinded samples is determined by using the individual grid point determinations for each sample from all bimarker planes multiplied by the overall predictive power for each individual bimarker plane. Each blinded sample grid point value (e.g., +1 or -1) is multiplied by the predictive power (or sensitivity) of the individual bimarker plane. All 10 planes are then summed together. A final overall score for all bimarker planes is obtained using typical linear and / or root sum of squares methods. The scores are normalized and shifted to yield a score between 0 and 200. This is the output used by the healthcare provider. This range is arbitrary.

[0096] A larger set of bimarker planes can be constructed from the same group of biomarkers by mathematically manipulating them. These larger bimarker sets will likely have greater predictive power, or may construct unmatched training set models (or secondary algorithms) to further improve predictive power. For example, using the ratio of concentrations of five biomarkers (rather than the concentrations themselves) to construct each pseudo concentration would create 10 pseudo concentration values ​​and 45 bimarker planes. The construction of pseudo concentrations and bimarker planes is likely more predictive, but may require a larger training set to accurately correlate with the general population. For example, the ratio of each concentration value divided by (1-another concentration value) could also be used. Those skilled in the art can easily determine whether these alternative methods for adjusting data for multidimensional cluster analysis have better predictive power by testing the method(s) for predictive power on blinded sample sets.

[0097] To further improve predictive power, these age- or grouping-adjusted concentrations are adjusted to normalize them, reducing or eliminating the interval bias of clustering across multidimensional grouped marker plots for cluster proximity analysis. See Figure 3, which shows the bimarker planes for IL-6 and VEGF. There are 10 of these planes for the 5-biomarker breast cancer testing panel. In this case, the calculated pseudo-concentration values ​​are normalized and shifted to yield any value between zero and 20, and outlier, highly upregulated concentrations are highly compressed.

[0098] Each of the bimarker projections of the multidimensional marker plane onto the same normalized interval across concentrations from the age / grouping analysis is compressed and normalized to the age-adjusted arithmetic mean and age (or total population) adjusted subgrouping.

[0099] Improving the predictive power of training set models using adjustable bimarker plane influence levels: Typically, bimarker plane is scored with binary numbers (e.g., +1 and -1) for non-disease and disease.The pseudo-concentration method described herein is suitable for further improving predictive power by selectively adjusting the influence level of these two binary numbers.The following method is developed in the training set model, and once set, it is fixed in the model.

[0100] Figures 4 and 5 below show a projection of one bimarker plane for five biomarkers used to predict the presence of a pathology (in this case breast cancer using five markers; IL-6, IL-8, TNFα, VEGF and PSA). Figure 4 shows a training set model with the data used to score the grid points on the plot by cluster search analysis. Figure 5 shows a training set model without the data. This constitutes the training set model. Since each of the 40,000 grid points is scored and the blinded samples are scored by their position on the grid, no training set data is required to create the model. The topology shows red positive for cancer and blue negative for cancer. When computing the total score in this case, the non-disease grid points are set to +1 and the disease (cancer) grid points are set to -1. Each bimarker in this five biomarker example is analyzed in five orthogonal spaces, of which Figure 5 is a projection of one in two dimensions. On this plot, the topology of the different subgroupings of the immune system response is shown. In this case, all grid spots (in this case 2000 x 2000 or 40,000) are scored in the usual way, assigning a value of -1 for pathology positive (breast cancer) and +1 for non-disease. This bimarker plane is normalized for the metavariable age above by the pseudoconcentration interval.

[0101] FIG. 6 shows the same bimarker model, and also immune response grouping inside the shaded area (see FIG. 2). The influence of the shaded areas is adjusted to reflect the fact that each shaded block has a slightly different influence on the probability that the patient is non-disease or diseased. This adjustment can be done by human estimation using training set validation (the adjustment produced a corrected training set result) or by rigorous computational multivariate incremental analysis. Two separate bimarker planes are created for the two outcomes, diseased and non-disease. In this case, the blinded data points in immune response group IV are much more likely to be diseased, and the influence (absolute value) will be increased slightly (e.g., by changing the score from -1 to -1.1). The actual amount of this increment will preferably be determined by computational analysis or possibly by rigorous manual methods. This method is feasible for the cluster search method of correlation analysis, but other means could be used in the same spirit. These methods of weighting the influence of disease associations can result in an improvement in predictive power of about 1%. With a predictive power of over 95%, this is highly significant.

[0102] Figure 7 also shows the same bimarker plane, with the circled shaded areas in the composite region being nonlinear rapidly changing disease vs. non-disease topology. Identification of such regions can be done by inserting noise-introduced test blind sample values ​​(e.g., + / - 10%) into the model and then introducing a measure of noise. Most of these blind points will not substantially change the disease (here, cancer) score. However, some grid points may be observed that change dramatically from non-disease to disease scores after this type of noise adjustment. These are regions with rapidly changing topology where most or all of the bimarker planes overlap with the entire multidimensional bimarker plane. By carefully reducing the influence in these regions, the weighting can be increased in a small number of relevant bimarker planes, resulting in noisy data being located in a broad plane without approaching the changing outcome boundary. This method has been shown to correct for mispredictions. In the above case, the influence of the red cancer region would shift downwards (in absolute value), for example, from -1.0 to -0.9. Alternatively, the blue non-disease region would be shifted downward from +1.0 to -0.9. The optimal shift level could be determined by rigorous computational analysis.

[0103] Assay noise can affect the accuracy of correlation analysis. This noise can be particularly problematic at or below the detection limit of the assay. This noise can also be mitigated by reducing the influence of measurement points for individual biomarkers that are in these unstable zones. Figure 8 also shows the PSA and IL-6 bimarker planes for a breast cancer panel. All areas within the shaded rectangular area at the bottom left of this figure are below the traditional limit of detection (LOD) of the assay. Traditionally, the LOD is defined as two standard deviations of 20 zero calibrators + the average value of the 20 zero calibrators. The statistical certainty of values ​​at this level is 95% within two standard deviations, and of course, the measurement certainty decreases when the measurement sample falls below the LOD. The data still has useful information, but should be applied to the analysis with less influence. In this case, the influence of the blind sample data points within the shaded area is reduced, for example, from +1.0 to -0.9, with respect to the grid points of the training set model within the shaded area. This increases the influence of data points of this test sample above the limit of detection on those other biomarker surfaces.

[0104] The above methods are complementary and can be implemented in parallel by combining influence shifts.

[0105] Methods for improving predictive power by testing blinded samples for instability: Once the training set model is complete and fixed, it is used to calculate the Cancer Score for blinded patient samples. We use two preferred methods to create the Cancer Score. The first method, called the Linear Method (CSl), takes the topological location scores (+1 or -1) multiplied by their predictive power for that bimarker plane. These are then summed, scaled, and shifted to result in a score between 0 and 200. The second score, called the q score (CSq), is calculated by using a root-sum-of-squares method on these same values. This second method highlights the differences between individual bimarker scores and is useful in the overall physician's final diagnosis.

[0106] Due to the highly nonlinear nature of the correlation clustering method, topological instabilities still remain in the bimarker plane and cannot be completely eliminated. The location of these instabilities can be found by extensive and rigorous evaluation of the computational cancer scores of each grid point with incremental pseudo-concentration values ​​across each bimarker plane. This involves a large number of computational values, 40,000 grid points x 10 bimarker planes x number of biomarkers (2,000,000 computations for 5 biomarkers). Unstable regions will be revealed by large variations in the cancer scores at neighboring grid points. This can also be done less rigorously by visual overlay of all 10 bimarker planes, looking for regions of close transition from healthy to disease (e.g., cancer) on about 5 or more planes. It has been found that the visually found regions can then be verified by a smaller number of computational validation calculations.

[0107] According to another embodiment of the present invention, the technique involving stability testing and noise introduction can be applied to blinded data sets. Also, the unmatched training set model can be used to adjust or correct the cancer score. In this embodiment of the present invention, a fixed level of noise is introduced for each blinded patient data set (e.g., plus or minus 10%). If the blinded sample set is about 100 patients, the actual training set model is run in silico for 300 samples each in triplicate (raw data ± noise). The resulting triplicate data sets are tested for stability (a is -10%, b is +10%, and c point is raw data). Table 1 shows the results of the stability test on data from a clinical study. Note that three samples show very high instability in the cancer score. Samples 138, 207, 34 and 29 all show very high figures of merit. The figure of merit (lower is better) should encompass both the degree of score shift and, in particular, whether or not the score predicting health is shifted to a score predicting cancer or vice versa. These data sets from blinded samples are at high risk of being inaccurate in predicting diagnosis.

[0108] The mismatched training model can be used to adjust the sample data set of patients who are "at risk" of failing the merit noise test. These points are at risk because of the inevitable random or systematic measurement noise, coupled with the extreme topological instability caused by the fact that most, if not all, of the blinded sample data points are located on the very steep slope of the bimarker plane, so that small disturbances lead to large amplitudes in the score. Table 1 shows samples with noise introduced. Each sample has three values: 1) plus noise, 2) minus noise, and 3) raw data without noise. These samples show a cancer score that changes from diseased to non-disease and vice versa with the introduction of ±10% noise. These sample data can be judged as unstable in this case. The instability level cannot be precisely defined, but adjustments can be made for various noise introduction levels. In this case, they are corrected for ±10% noise and a stability score greater than 200 (note that stability score and cancer score are two obviously different numbers with different meanings).

[0109] Measurement noise can be adjusted by this discrepancy second algorithm. The discrepancy algorithm used for adjustment improves the chances that the points are accurate, so that it can be used to correct these "at risk" patient samples, even if it has a slightly lower predictive power compared to the main algorithm. In this case, two were corrected (see FIG. 9); sample 138 had a non-disease score of 85, which was corrected to 195 with the discrepancy algorithm (this point was stable by algorithm I, and sample 34 had a score of 102 (linear method), which was also corrected to 198 using algorithm II). Samples 29 and 207 were not changed by the discrepancy algorithm.

[0110] The mismatched training set model (algorithm II) used 105 bimarker planes. These same samples show stability in the algorithm II stability test, so the mismatched training set model (algorithm II) is mismatched to the primary training set model (algorithm I). The mismatched training set model is tested in exactly the same way as the primary training set model. Note that logistic regression could have been used to calculate these scores as well. Algorithm II was used because it has high predictive power. The training set model for tuning can be used even if its predictive power is lower than the primary algorithm (although preferably the predictive power is 50% or more), as long as it produces results that are likely to be accurate without instability. Note that the correction is dramatic for the blinded samples in question that failed the noise test. In fact, all of these samples were cancers with high scores. Eight of the 10 bimarker planes for these uses were on topologies with very high instability lattice points. Thus the scores were at risk and indeed inaccurate (one inaccurate and one uncertain, giving a score of 100 / 120). In this case, one sample was corrected, improving the predictive power from 97% to 98%, with a very significant reduction in error (50%). One sample was changed to uncertain but cancer, which was also corrected.

[0111] Methods for improving the predictive power of disease-correlated binary outcomes by eliminating independent states that partially resemble one of the outcome states of the primary disease analysis: Cluster analysis generally uses three or more independent variables, often serum protein concentrations of the patient. Correlation algorithms may only operate on a binary outcome of non-disease or disease, but result in a continuous score that is more closely related to the probability of the actual outcome being the two binary states. In some cases, there are other conditions that are nominally classified as non-disease that partially resemble the pathological condition within the population distribution of the biomarkers used. In some of these cases, this non-disease "MIMIC" condition may cause a false positive outcome of the correlation analysis. The solution to resolve this type of false positive result is to create a further new correlation analysis that is completely separate from the non-disease or disease analysis. This new correlation analysis preferably uses the exact same biomarker measurement data as for the non-disease or disease correlation, or may use some or all different biomarkers. This new correlation analysis produces a "non-disease MIMIC" or "disease" result, or at least a score that allows a judgment to be made about the true state of the patient. Uncertain or close transition scores on non-disease or disease analyses, in conjunction with very low or high scores on non-disease MIMIC or disease correlations, can help medical practitioners improve disease determination and reduce false positive scores.

[0112] An example of this situation where a non-disease state resembles a disease state is benign prostatic hyperplasia (BPH), a non-malignant condition. This condition typically exhibits elevated levels of at least one biomarker used to diagnose prostate cancer. For example, the biomarker prostate-specific antigen will be elevated in men with BHP and also in men with prostate cancer. Table 5 shows that this further correlation analysis method can distinguish between monogamous men with BHP and men with prostate cancer, as well as between men who are presumed to be non-disease and those who are diseased and have confirmed prostate cancer, using the same biomarkers but different training set models. In only a few men did the non-disease vs. cancer training set model produce false positives that would be distinguished by the BHP vs. cancer training set model. In these cases, two scores, one presumed non-disease vs. cancer and one BHP vs. cancer, would help the physician or other medical personnel decide on the next diagnostic steps. For example, for a total score of 0-200 for both models (for CS1 or CSq), a score of 110 for non-prostate cancer or prostate cancer would indicate a weak score for being positive for cancer, whereas a second score of 30, considered as BPH or cancer, would indicate to the medical practitioner a high probability of BPH but not cancer. The medical practitioner would use this added information along with other medical information and the patient history to determine the next steps in diagnosis.

[0113] Several methods have been described herein to improve the predictive power of traditional proteomic correlation methods for diagnosing disease. These include 1) using metavariables and pseudoconcentrations in the correlation, and 2) using expert knowledge of topological stability and assay measurement properties to adjust the influence of bimarker planes in the training set model. Methods have also been described to use mismatched training set models to find and correct blinded sample stability problems specific to a particular training set model. In addition, methods have been described to find and correct non-disease conditions that partially resemble the training set model for a given pathology. All of these methods are complementary and can be used simultaneously. For example, adjusting the training set model for likely regions of instability does not completely remove this problem from the blinded sample prediction calculation, and therefore both methods can be used to improve predictive power. We have found that by combining these methods, we can obtain a predictive power of more than 95%, and in the breast cancer study discussed in Example 1, we obtained a predictive power of more than 98% (sensitivity 100%, specificity 97.5%). EXAMPLES

[0114] Clinical Research - Evaluation of Breast Cancer Blood Test The performance of the OTraces BC Sera Dx test kit and the OTraces CDx immunochemistry instrument system (www.otraces.com) was evaluated in an experiment to assess the risk of the presence of breast cancer. The test kit measures the concentration of five very low level cytokines and tissue markers and uses the training set model developed as described above to calculate scores, namely CSl and CSq, to ​​assess the risk of breast cancer. The proteins measured were IL-6, IL-8, VEGF, TNFα, and PSA. The experiment consisted of the measurement of approximately 300 patient samples roughly divided into 50% of patients with biopsy-diagnosed breast cancer and 50% of patients presumably non-diseased (i.e., free of breast cancer in this case). Of this group, the biopsy results of 200 samples were precisely divided into 50% non-diseased and 50% with breast cancer disease, and each group was further subdivided into specific age groups.

[0115] The sample analysis results were used to develop a training set model to predict disease state. The remaining samples (approximately 110) were then run through the training set model as blinded samples, resulting in numerical cancer risk scores. These scores were disclosed to the host clinical center. These blinded sample scores were subsequently analyzed by the clinical center to assess the clinical accuracy of the results.

[0116] Two diagnostic models were developed for this experiment, which are referred to herein as Algorithm I (or Training Set Model I) and Algorithm II (or Training Set Model II), as discussed above. Neighborhood cluster analysis was used for both algorithms. Subject age was used not as an independent variable, but as a meta-variable to convert the measured concentration into a new independent variable, referred to herein as pseudo-concentration. This pseudo-concentration was directly used in correlation analysis. The difference between Algorithm I and Algorithm II is the number of new independent variables used in correlation. Algorithm I uses five pseudo-concentration variables in a 10-dimensional cluster space. This space can be seen by the human eye through a projection of this multidimensional space or a slice across it to see a two-dimensional bimarker plane. Algorithm I has 10 such planes.

[0117] Algorithm II generates and uses 10 times more independent variables, so there are about 100 bimarker planes. 200 samples are expected to be sufficient for the training set model, resulting in a model that is reasonably close to the general population. A secondary or mismatched training set model was developed from the same 200 sample training data set. This training set model is the primary scoring method used to describe the results herein. The mismatched training set model is used to adjust the cancer scores calculated by the primary training set model that are considered unstable, i.e., scores that are in the area of ​​topological instability. Although the mismatched training set model is somewhat less accurate for blinded samples, it can still adjust the primary training set model, thus improving predictive power.

[0118] The above clustering method analysis has a major advantage over logistic regression analysis in that it can accommodate highly nonlinear trends in the independent variables used to generate the calculated outcome. The outcome is either disease or non-disease (cancer or no cancer in this case), which is based on the pseudo-concentration for the training set model calculation. The drawback of this method is that highly nonlinear regions can be associated with very steep topological gradients. Thus, unknown (or blinded) samples may be in steep peaks or deep steep valleys that act to amplify small errors in the computer-calculated pseudo-concentration. We evaluated the stability of the calculated scores in a proprietary stability test, and then used algorithm II to adjust algorithm I if algorithm II indicated stability for the sample.

[0119] Figures 10, 11, and 12 show the results of the Algorithm I training set. The model itself consists of 10 bimarker planes of 40000 topological points, each scored by a clustering method for non-disease and disease (here, breast cancer). The ability of the model to separate the two sets, non-cancer and cancer, is shown in these figures. The model must be built from two outcome states that are very close to 50% vs. 50%, preferably exactly 50% vs. 50%. The method also uses age as a transformation metavariable. The training set samples had samples distributed across the age groups of interest. The model for Algorithm 1 (Figure 10) was built from 100 healthy women and 98 women with breast cancer.

[0120] The summary table in Figure 10 shows the numerical results, where the number of samples is N = 198. CI are correctly predicted samples, FI are incorrectly predicted samples, and 4 samples were marked as uncertain.

[0121] A secondary training set model was developed to distinguish the four uncertain samples that resulted from using the primary training set model. This model is the unmatched training set model. This secondary model uses the same training set data as the primary model.

[0122] Figure 11 shows the results of the unmatched training set model calculation. Algorithm II shows 100% separation with a separation of more than 60 points.

[0123] Results of blinded breast cancer study: Figure 12 shows the results of the blinded samples evaluated in the clinical study. The results show a sensitivity of 100% and a specificity of 97.5%. The diagnostic transition values ​​were set by an oncologist at the clinical research center to ensure that all breast cancer positive samples were correctly identified. Thus, two non-disease samples were predicted to be cancer positive. This is medically plausible because all samples that test positive undergo the next diagnostic step, breast imaging. Many women do not undergo breast imaging because they do not live close enough to a facility that has the medical equipment to do so. However, blood can be drawn from these women and sent on ice to a laboratory in a large city, even if it is far from a clinical laboratory. EXAMPLES

[0124] Use of the metavariable “age” to improve diagnostic accuracy. Table 2 shows a table of results from a breast cancer clinical study consisting of a sample of 868 subjects. Table 3 shows a comparison of various methods for correlation calculation. The standard method (logistic regression analysis) showed only a predictive power of 82%. Standard neighborhood cluster analysis improved this, resulting in a predictive power of 88%. Using a metavariable and weighting approach, conditioning on topological stability, grouping by immune system response, and conditioning on weights for assay performance - combined with blinded sample instability testing and discrepancy algorithm correction - the method described herein resulted in a predictive power of over 97%. EXAMPLES

[0125] Using the metavariable "age" to improve diagnostic accuracy in ovarian cancer research Table 4 shows the results of a study of 107 women with and without ovarian cancer using the metavariable methods described herein. Although this study did not use all of the predictive power improvements described herein, it still achieved a relatively good predictive power of 95%. EXAMPLES

[0126] Use of the metavariable “age” to improve diagnostic accuracy in prostate cancer. Table 5 shows the results of a study of 259 men with prostate cancer or benign prostatic hyperplasia (BPH) using the metavariable method described herein. This study also did not use all of the predictive power improvements described herein, but still achieved a relatively good predictive power of 94%. It should be noted that BPH is by far the most common condition that causes false positive results in current PSA tests for prostate cancer. In traditional diagnosis of prostate cancer, men with BPH are about 4 out of 5 positive, and therefore most prostate cancer biopsies are negative. The metavariable method can correct these inaccurate diagnoses discussed above.

[0127] The above results in Examples 3 and 4 (for ovarian and prostate cancer, respectively) did not use meta-variables, effect adjustment methods (LOD, subpopulation grouping and instability), or blinded sample stability methods, as these had not yet been discovered by the inventors when this data was measured.

[0128] II. Preferred Analyte Categories and Diagnostic Methods Using Analytes Measured Below Conventional Limits of Detection The present invention is also based on the discovery that certain immune system proteins allow for the diagnosis of risk for a given disease at measured concentrations significantly lower than those currently utilized in commercial diagnostic tests. These include cytokines, which, primarily, but not exclusively, function as signaling proteins, belonging to several specific groups: immune system inflammatory markers, tumor anti-development, cell apoptosis, and tumor angiogenesis markers, as well as known tumor tissue markers.

[0129] The present inventors have shown that by selecting several ultra-low abundance protein LAPs and using non-traditional methods to perform correlation analysis and to determine concentrations from immunoassay analysis, predictive power is greatly improved. These low-level proteins, called signaling proteins (i.e., proteins that perform one or more of several types of functions in signaling networks), act from a direct immune system response to the presence of tumors or by tumors that instruct the organism to provide the required physiological response that tumors require to progress. Furthermore, by selecting several LAPs (preferably 6 or less), the challenges of sampling and training set size determination are solved. The fact that these proteins are below the traditionally defined detection limit has previously prevented the study of their usefulness.

[0130] The present invention is based on the surprising discovery that concentration values ​​below the conventionally determined detection level provide important information in the correlation risk assessment of disease. Such information has not traditionally been used in clinical diagnosis because of their imprecision. However, the inventors have found it useful to curve fit a line from the LOD calibration point to the lowest signal value in the test run and utilize those values. Doing so provides a smooth Gaussian distribution for population distribution analysis and, surprisingly, also provides accurate cancer score prediction. The diagnostic method should not report any readings below the LOD if they are lower than what is normally seen in large population evaluations of this marker. For example, if the signal level (or measured concentration) is below the LOD, the use of the techniques described herein is appropriate up to the lowest level found in normal serum. Thus, if the LOD for IL-6 is about 250 fg / ml, but values ​​in serum as low as 10 fg / ml have also been reported, this level should be the lowest value used in the assay according to the present invention. Also, the value cannot be zero or negative. This approach works with a variety of conventional standard curve construction strategies.

[0131] The inventors have surprisingly found that cytokines, immune system proteins whose functions belong to several specific groups of immune system inflammation, tumor anti-development, cell apoptosis, and tumor angiogenesis markers, as well as known tumor tissue markers, can be used to achieve predictive power such that correlation models have better than 95% false negative performance and better than 95% false positive performance. These proteins require methods to extract useful concentration information at levels much lower than 1 pg / ml for some markers. For example, a specific proteomic test panel for breast cancer using PSA for tissue markers, IL-6 for inflammatory response, IL-8 for inflammation and angiogenesis, VEGF for angiogenesis, and TNFα for anti-tumor development has been provided with a predictive power of over 98%. Some of these markers have significant population distributions below 1 pg / ml (down to below 50 fg / ml). For this reason, researchers have not explored the use of these proteins in clinical diagnostic methods.

[0132] The present invention contemplates including diagnostic tests that predict breast cancer using, for example, immune system inflammatory (IL-6, IL-8), angiogenesis (IL-8, VEGF), anti-tumorigenic (TNFα) proteins, and tissue markers (PSA). Other markers in these categories can also be used, such as tissue marker CA19.9 instead of PSA; or IL-1 instead of or in addition to IL-6. For prostate cancer, it is contemplated that predictive assays are utilized that include immune system inflammatory markers (IL-6, IL-18), angiogenesis markers (IL-8, VEGF), anti-tumorigenic (TNFα) proteins, and tissue markers (PSA). Other markers in these categories can also be used, such as IL-1 instead of or in addition to IL-6. Contemplated ovarian cancer status predictions are made using immune system inflammatory markers (IL-6, IL-18), angiogenesis (IL-8, VEGF), anti-tumorigenic (IL-12) proteins, and tissue markers (CA125). Other markers within these categories can also be used (eg, IL-12 instead of TNFα).

[0133] The inventors have also found that correlation analysis requires that 100% of the population have practically accurate measurements, otherwise predictive power is compromised.Practically accurate does not mean that the accuracy of these measurements must be equivalent to those expected for clinical diagnostic methods currently used in clinical laboratories.In clinical laboratories, when multiple concentrations are required for diagnosis, the points of the assay calibration curve must be higher than the analytical sensitivity.This means that 99.7% certainty in the result is within 3 x standard deviation of the actual value.

[0134] Any sample with an indeterminate or zero value for a marker will not anchor the correlation calculation to this sample as completely inaccurate. The combination of several LAP signaling proteins and very low level concentration measurement extraction methods greatly improves the results. Even at very low levels below the conventional assay detection limit, acceptable accuracy is obtained by simply using a straight line from the LOD to the sample with the lowest signal and using the lowest physiological level found in serum as the concentration at this point in the test run for many samples. Sample points between the LOD and the lowest reading are presumed to be on this straight line. Other standard curve fitting methods can also be used. These improvements are large enough that the measurement panel is useful for cancer screening (providing a predictive power of 98% or more).

[0135] The inventors have found that cytokines, immune system proteins that belong to several specific groups whose functions are immune system inflammation, tumor anti-development, cell apoptosis, and tumor angiogenesis markers, as well as known tumor tissue markers, can be used to achieve predictive power such that the false negative performance of correlation models is better than 95% and the false positive performance is also better than 95%. These proteins require methods to derive useful concentration information at levels much lower than 1 pg / ml for some markers.

[0136] For example, a specific proteomic test panel for breast cancer using PSA for tissue markers, IL-6 for inflammatory response, IL-8 for inflammation and angiogenesis, VEGF for angiogenesis, and TNFα for anti-tumor development has yielded a predictive power of over 95%. All of these markers have significant population distributions below 1 pg / ml (down to below 100 fg / ml). Correlation analysis requires that a very high percentage (100%) of the population have practically accurate measurements, otherwise the correlation is lost. Any sample with an indeterminate or zero value for a marker will not anchor the correlation calculation, which may make this sample completely inaccurate. The combination of several LAP signaling proteins and very low level densitometric extraction methods greatly improves the results. This is large enough that the measurement panel may be useful for screening for cancer.

[0137] Figure 13 shows a typical ELISA calibration curve for TNFα. In this case, the limit of detection (LOD) is approximately 1.0 pg / ml. This is about the best that can be achieved with current rapid screening assay technology. Note that a significant proportion of the population is below the LOD. Surprisingly, these data points are very useful and are in fact necessary to achieve a highly predictive correlation.

[0138] In clinical laboratories, it is customary to define the limit of detection as 2 x standard deviations above the zero calibrator, with the standard deviation usually calculated using 20 zero standards. Measurement results below this level are usually not reported, or if reported, are classified as below the LOD. When using a single analyte assay for clinical diagnostic purposes, the assay must have an accurate freestanding value above the LOD to adequately provide diagnostic information to medical personnel, and conventional methods are required.

[0139] As described herein, for diagnostic assays using low levels of signaling proteins, measurements aimed at scoring the probability of disease vs. non-disease with high confidence use measurements below these LODs. Figure 14 shows the data from Figure 13, with error bars showing the amount of error that can be tolerated to keep the calculated cancer score error below 5%. At extremely low levels of detection below LOD, a relatively small amount of error can be tolerated. The same is true for very high concentrations.

[0140] A very important region for the calibration curve of a diagnostic assay is where the diagnostic prediction score lies in the range from a score clearly indicating a non-disease state to a score clearly indicating a disease state. Also, note that if data below the LOD are excluded or reported as 0, the scoring of that patient may simply result in extreme errors, with a low score of healthy being a high score of cancer. This is because the algorithm must be able to "anchor" all five markers at some general level at the extremes and be accurate at the transitions. If one marker is not anchored, the calculation is based only on the remaining four. There may well be situations where a high result of these four does not indicate cancer, but all five are needed, and all five must be high to indicate cancer and generate a score accurately. Each individual marker has a limited impact on the individual cancer score, which is weighted by the true impact of the marker and the level of noise that arises based on its position in the calibration curve.

[0141] For example, the breast cancer testing panels discussed herein that include PSA in the panel show only a 3-4% improvement in overall training set model predictive power for a combined analysis of a large training set (200 samples). However, excluding it from one blinded sample deviates the scoring from the training set model enough to shift this sample from healthy or cancer to other (e.g., from a score of 50 to 180 on a scale of 0-200). These blinded samples may be rare. Only 1 in 100 will produce a drop in predictive power of 1%. With these methods, we have achieved a predictive power of 98%. Thus, a drop of 1% is very significant. Blinded sample data points for PSA, estimated at very low levels, for example, around 10 fg / ml, may be 20-fold apart without causing a significant error in the cancer score. On the other hand, if PSA is omitted or zeroed out, the cancer score becomes "unanchored" by the low PSA level, and may shift to the point where the cancer score becomes completely inaccurate if not included.

[0142] It is important to understand that there is uncertainty involved at these low measurement levels. For duplicate measurement samples, the confidence level is 95% at 1.5 standard deviations from the nominal measurement point. At 0.75 standard deviations from the nominal measurement point, the confidence level is 67%. This level of precision is more than adequate at the very low levels of these methods; having no numerical values ​​at all would make the method useful.

[0143] The method for evaluating the values ​​of measurement points for a correlation evaluation according to the invention requires: 1) A traditional calibration curve should extend from the LOD of the assay being evaluated during development, through the dynamic range of the population, and as high as possible using a typical immunoassay method, e.g., ELISA. 2) Below the LOD, all points are considered to have significant noise, and therefore the actual measured point may be less than the lowest calibrator. In these cases, the reported results should be as follows: a. Above zero. Negative concentrations are not possible, and a zero value from the correlation algorithm adversely affects the accuracy of the results to the same extent as no value at all. b. Rather than simply having values ​​assigned to all patient sample point concentrations undetermined, it is determined which would skew the population distribution that is necessary for building the subgroup training set model. c. Values ​​lower than those typically seen in large population evaluations of this marker are not reported.

[0144] The inventors have surprisingly found that it is appropriate to simply use a linear curve fitting from the LOD calibration point to the lowest signal value in the test run. This results in a smooth Gaussian distribution for population distribution analysis and accurate cancer score calculation. Using this approach, any number of different standard curve generation strategies will work, provided that the above rules are followed.

[0145] It is contemplated that the disclosed techniques for utilizing analyte concentrations below conventional LODs can be used in any assay that utilizes, for example, the measured concentration of a marker in a patient sample.Preferred embodiments include diagnostic assays for various diseases, such as solid tumors, including prostate cancer, lung cancer, breast cancer, and ovarian cancer.These techniques can be optionally, but preferably, combined with other data analysis and diagnostic techniques described elsewhere herein.

[0146] Those skilled in the art will also understand that analytical errors caused by insufficient training set size, such as instability of the correlation method used or elsewhere herein, such as instability of the bimarker plane topology, are caused by completely different phenomena and require completely different methods for mitigation. For example, certain disclosed errors and correction methods are appropriate for extremely nonlinear (or very steep) gradients where test sample points are on the correlation bimarker topology. And such errors can be found by injecting artificial noise and adjustments in the mismatched training set model. Although the noise in this case is inherent to the assay measurements, resulting from experimental error, it is very important to understand the impact of this noise on the resulting cancer score.

[0147] III. Implementation via a Computer System The implementation of the various techniques described herein can be implemented in digital electronic circuitry, or in computer hardware, firmware, software, or combinations thereof. The implementation can be implemented as a computer program product, e.g., a computer program tangibly embodied in an information carrier, e.g., a machine-readable storage device or a propagated signal, for execution by a data processing device or for controlling the operation of a data processing device, e.g., a programmable processor, a computer, or multiple computers. Computer programs, such as those described above, can be written in any form of programming language, including compiled or interpreted languages, and can be arranged in any form, including as a stand-alone program or as a module, component, subroutine, or other unit suitable for use in a computing environment. Computer programs can be arranged to be executed on one computer or on multiple computers at one site, or distributed across multiple sites and interconnected by a communication network.

[0148] The method steps may be performed by one or more programmable processors executing a computer program which performs functions by manipulating input data and producing output. The method steps may also be performed by, and an apparatus may be implemented as, special purpose logic circuitry, such as an FPGA (field programmable gate array) or an ASIC (application specific integrated circuit).

[0149] Processors suitable for executing computer programs include, for example, both general-purpose and special-purpose microprocessors, and any one or more processors of any kind of digital computer. Generally, a processor will receive instructions and data from a read-only memory, a random access memory, or both. The elements of a computer can include at least one processor for executing instructions and one or more storage devices for storing instructions and data. Generally, a computer can include one or more mass storage devices for storing data, such as magnetic, magneto-optical, or optical disks, or can be operatively coupled to receive data, transfer data, or both. Information carriers suitable for incorporating computer program instructions and data include all forms of non-volatile memory, including, for example, semiconductor storage devices, such as EPROM, EEPROM, and flash storage devices; magnetic disks, such as internal hard disks or removable disks; magneto-optical disks; and CD-ROM and DVD ROM disks. The processor and memory can be supplemented by, or incorporated in, special-purpose logic circuitry.

[0150] To provide for interaction with a user, the implementations can be implemented in a computer having a display device, e.g., a cathode ray tube (CRT) or liquid crystal display (LCD) monitor, for presenting information to the user, and a keyboard and pointing device, e.g., a mouse or trackball, by which the user can provide input to the computer. Other types of devices can also be used to provide for interaction with a user. For example, feedback provided to the user can be any form of sensory feedback, e.g., visual feedback, auditory feedback, or tactile feedback, and input from the user can be received in any form, including acoustic, speech, or tactile input.

[0151] An implementation may be implemented in a computing system including a back-end component, e.g., as a data server, or a middleware component, e.g., including an application server, or a front-end component through which a user may interact with the implementation, e.g., a client computer having a graphical user interface or a web browser, or any combination of such back-end, middleware, or front-end components. The components may be interconnected by any form or medium of digital data communication, e.g., a communications network. Examples of communications networks include local area networks (LANs) and wide area networks (WANs), e.g., the Internet.

[0152] The invention described herein generally relates to a method for improving the diagnostic accuracy or predictive power of proteomic and metabolomic correlation methods for predicting disease states. Although certain exemplary embodiments have been described above and illustrated in the accompanying drawings, it should be understood that such embodiments are merely illustrative and do not limit the broad invention. In particular, it should be recognized that the teachings of the present invention apply to a variety of diseases. Also, although the preferred embodiment of the present invention relates to the diagnosis of human diseases in subjects (or patients) for whom disease diagnosis is desired, the methods and systems disclosed herein are useful for diagnostic purposes in non-human species, particularly primates and other mammals, and as such are also contemplated as part of the present invention.

[0153] While certain features of the described implementations have been illustrated as set forth herein, many variations, substitutions, modifications, and equivalents will occur to those skilled in the art. It is therefore understood that the invention is not limited to the particular embodiments or configurations disclosed, but rather is intended to cover any modifications, adaptations, or variations that are within the spirit and scope of the invention as defined by the appended claims.

[0154] References: All journal articles and all other publications, patents, and textbooks mentioned herein are incorporated by reference in their entirety, including: (1) Drukier, et al., “High-Sensitivity Blood-Based Detection of Breast Cancer by Multi Photon Detection Diagnostic Proteomics,” Journal of Proteome Research 2006, 5:1908,1915. (2) Lokshin et al., “Multimarker assay for early diagnosis of ovarian cancer,” American Association for Cancer Research, Amer Assoc Cancer Res 2006, 47:653. CME: Disclosure. (3) Drukier, et al., Ultra-Sensitive Immunoassays Using Multi Photon Detection in Diagnostic Proteomics of Blood,” Journal of Proteome Research 2005, 4:2375-2378. (4) Drukier, “Supersensitive Immunoassays,” U.S. Patent No. 7,604,956 (2009).

Claims

1. 1. A computer implemented method for diagnosing cancer, comprising: a) a computer processor determining concentrations of at least three predetermined biomarkers indicative of a cancer or a pro-inflammatory condition in a blinded sample from the subject; b) the computer processor selecting one or more meta-variables to associate with the subject; c) the computer processor transforms the concentrations of the biomarkers as a function of one or more population distribution characteristics and the one or more meta-variables to calculate pseudo-concentrations representative of each biomarker, wherein the one or more population distribution characteristics separate populations into distinct groups. d) said computer processor comparing said pseudo-concentration to a training set model of pseudo-concentrations determined for members of a population known to have or not have said cancer; and e) said computer processor determining whether said comparison indicates that the subject has said cancer or pro-inflammatory condition. A method comprising:

2. 2. The method of claim 1, wherein the population distribution characteristic is calculated as a mean value of one or more concentrations of the biomarkers, or as a median value of one or more concentrations of the biomarkers, or as a dynamic range of one or more concentrations of the biomarkers, or based on the degree of upregulation or downregulation of one or more of the biomarkers.

3. 3. The method of claim 1 or 2, wherein the one or more biomarkers comprise a low abundance protein.

4. The method of any of claims 1 to 3, wherein the biomarkers include at least one of PSA, IL-6, IL-8, TNFα, or VEGF.

5. 5. The method of any of claims 1 to 4, wherein said meta-variables are selected from the group consisting of pre-, peri- and post-menopausal status, pubertal stage, body weight, geographic location of sample source, body fat percentage, age, race or racial / ethnicity, species or time of period.

6. The method according to any of claims 1 to 5, wherein the transformation comprises normalizing the concentrations and smoothing irregular or discontinuous distributions of the concentrations.

7. The method of any of claims 1 to 6, wherein the training set model comprises at least 200 individuals.

8. The method of any of claims 1 to 7, wherein the cancer is selected from the group consisting of breast cancer, prostate cancer, ovarian cancer, and lung cancer.

9. 1. A system for diagnosing cancer or a pro-inflammatory condition, comprising: receiving concentrations of at least three predetermined biomarkers indicative of said cancer or pro-inflammatory condition in a blinded sample from a subject; selecting one or more meta-variables associated with said subject; transforming the concentrations of the biomarkers as a function of one or more population distribution characteristics and the one or more meta-variables to calculate pseudo-concentrations representative of each biomarker, wherein the one or more population distribution characteristics separate populations into distinct groups; comparing said pseudo-concentrations to a training set model of pseudo-concentrations determined for members of a population known to have or not have said cancer or pro-inflammatory condition; and determining whether the comparison indicates that the subject has the cancer or pro-inflammatory condition; A system, including a computer.

10. 10. The system of claim 9, wherein the population distribution characteristic is calculated as a mean value of one or more concentrations of the biomarkers, or as a median value of one or more concentrations of the biomarkers, or as a dynamic range of one or more concentrations of the biomarkers, or based on the degree of upregulation or downregulation of one or more of the biomarkers.

11. The system of claim 9 or 10, wherein the one or more biomarkers comprise a low abundance protein.

12. The system of any one of claims 9 to 11, wherein the biomarkers include at least one of PSA, IL-6, IL-8, TNFα, or VEGF.

13. 13. The system of any of claims 9 to 12, wherein the meta-variables are selected from the group consisting of pre-, peri- and post-menopausal status, pubertal stage, weight, geographic location of sample source, body fat percentage, age, race or racial / ethnicity, species or time of period.

14. The system of any of claims 1 to 13, wherein the transformation comprises normalizing the concentrations and smoothing irregular or discontinuous distributions of the concentrations.

15. The system of any one of claims 9 to 14, wherein the training set model comprises at least 200 individuals.

Images