US 20060293859 A1
An analytic apparatus and method is provided for diagnosis, prognosis and biomarker discovery using transcriptome data such as mRNA expression levels from microarrays, proteomic data, and metabolomic data. The invention provides for model-based analysis, especially using kernel-based models, and more particularly similarity-based models. Model-derived residuals advantageously provide a unique new tool for insights into disease mechanisms. Localization of models provides for improved model efficacy. The invention is capable of extracting useful information heretofore unavailable by other methods, relating to dynamics in cellular gene regulation, regulatory networks, biological pathways and metabolism.
1. A method for analyzing biological specimens comprising the steps of:
receiving multivariate observations each comprising measurements of the amounts of a plurality of constituents in biological specimens;
generating expected estimates of said amount measurements in each said observation as a function of each said observation;
generating residuals for said measurements from the difference between said estimates and said measurements for each said observation; and
testing said residuals to determine if a constituent is deviating from expected amounts in said biological specimens.
2. A method according to
3. A method according to
4. A method according to
5. A method according to
6. A method according to
7. A method according to
8. A method according to
9. A method according to
10. A method according to
11. A method according to
12. An apparatus for analyzing a biological specimen comprising:
a memory for storing a set of reference observations each comprised of measurements of a plurality of constituents from a biological specimen, each such observation being associated with a biological specimen, and each specimen being of a first selected class;
means for receiving input observations each comprised of measurements of the plurality of constituents from a biological specimen;
a processor for generating estimates of at least some of the measurements of the plurality of constituents in each said input observation, as a function of each input observation and as a function of said reference observations, said processor further disposed to generate residuals for at least some of those constituents for which estimates are generated, by differencing the estimates and the measurements of the input observations, said processor further disposed to test at least one of said residuals to determine if the associated constituent is deviating from expected behavior; and
means for displaying the results of the test performed by said processor.
13. An apparatus according to
14. An apparatus according to
15. An apparatus according to
16. An apparatus according to
17. An apparatus according to
18. An apparatus according to
19. An apparatus according to
20. An apparatus according to
This application claims the benefit of priority under 35 U.S.C. § 119(e) to U.S. Provisional application Ser. No. 60/670,950 filed Apr. 13, 2005.
The invention relates to biomarker discovery and disease diagnosis and prognosis, especially based on transcriptomic data such as gene expression levels and proteomics.
Recent advances in the biological sciences have made it possible to measure thousands of gene expression levels in cells using microarrays, and to quantify the cellular content of thousands of types of proteins using e.g., mass spectrometry, in single experiments with one sample of tissue or serum. Gene expression molecules, i.e., messenger RNA, and the proteins encoded by the mRNA comprise the “transcriptome” of the cell, the components of the cell that are the direct products of the transcription (the genome) and translation (the proteome) of DNA. Analysis of the transcriptome offers the potential to identify “bad actors” in the genetic expression process that are responsible for disease, and may offer both a signature for disease diagnosis, as well as targets for disease amelioration, whether through drugs or other means. By measuring thousands of components at once, the pace of biomarker discovery can be greatly accelerated, and the interactions between multiple components can be analyzed for complex disease mechanisms.
Genomic expression data representing semi-quantitatively the amount of mRNA present in a cell by on a gene-by-gene basis, may reveal markers for disease as well as patterns of expression that effectively differentiate or classify disease states and progression. Similarly, quantification of the count of proteins in the cell by protein molecule type, provides further insight into the overall regulation of the cell from DNA to protein product, and provides key insight into “post-transcriptional modifications” whereby protein structures are modified after translation from RNA. Pattern matching unknown patient samples against known microarray expression level patterns of disease state, may provide diagnosis or prognosis. Disease differentiation is key to applying the appropriate treatment. Cancers may manifest as seemingly similar phenotypes in identical tissue, yet have very different prognoses and treatment outcomes, because the biological pathway of disease is in fact different at the genetic level. It is essential to be able to identify what form of disease a patient has. Another useful application utilizing this kind of data is classifying patients as to whether they are candidates or not for the use of a drug which may, in a small subpopulation, have adverse effects related to the genetic makeup of the individual.
This new wellspring of genomic and proteomic data presents many challenges to discovery of the important information buried therein. One of the challenges of this data is the high dimensionality and the relative scarcity of patient observations. Often, thousands of gene expression levels are measured by a single microarray applied to patient tissue or a cell culture, but the total count of patients may be less than 20. Each such patient represents an observation of a vector comprising these thousands of expression levels. Another challenge is that the data can be extremely noisy, confounding complex analysis approaches that are sensitive to bad input. Yet another challenge is that the new measurement technologies are very sensitive to methodology, resulting in vastly different results when prepared by different researchers. Microarrays themselves have yet to be standardized, and microarrays from different manufacturers often have different oligonucleotide probes for the same gene, introducing variability due to oligo length or affinity. Yet another challenge is that the data often represents a mere snapshot of transcriptomic content, which is in a constant state of flux, as cellular genetic expression changes to adjust to the cell's needs and environment.
A number of methods are known for analysis of genomic and proteomic data. According to one known method, genetic expression levels for genes are compared across samples of normal tissue and diseased tissue, and where differential expression is sufficiently large, the gene is identified as a possible biomarker of the disease. This method is intended to identify gross amplification of a gene, or the complete silencing of a gene. According to known clustering methods, pair-wise correlations of the expression levels of pairs of genes are calculated, and genes are arranged in clusters according to a ranking of their pair-wise correlation across a number of samples. According to an approach that uses Support Vector Machines (SVMs), an optimal separation boundary with maximum margin (or maximum soft margin) is generated (possibly in higher dimensional space) for multivariate expression levels from two classes (normal and diseased), thus providing a classifier for future patterns.
However, these methods have their shortcomings. Gross differential expression analysis assumes static expression levels, and is easily confounded by expression dynamics. Furthermore, many diseases may be the result of more subtle deviations in expression than is typically tolerated by differential expression studies. Clustering only takes advantage of correlative information between pairs of genes, and thus misses more complex interactions and multi-gene dynamics. SVMs can be very computationally intensive, and produce optimal solutions that may overfit the typically small sample size, and as a consequence do not generalize well.
What are needed are improved analytic methods capable of tapping into the dynamic information in multivariate transcriptomic data, to reveal more subtle signatures for classification and data mining of biomarkers. Further, what are needed are faster methods that can be used interactively by discovery personnel, to provide leverage to extensive expertise in the exploration and discovery process. Further, methods are needed that are robust to the noisy nature of the transcriptome data, so that information can be derived from data from today's level of accuracy in microarray and mass spec measurement technologies.
The present invention provides a unique and advantageous model-based approach to analysis of genomic, proteomic, and general transcriptomic data, for diagnostics, prognostics and biomarker discovery. Accordingly, each patient sample is treated as an input vector, and gene expression levels are treated as variables. A set of patient data for known disease states is used to train a model, ideally a similarity-based model; thereafter, the model can be used to classify disease state, differentiate cancer types, determine candidacy for use of a drug or therapy, etc.; or can be used to model data from another class of tissue or serum sample, to provide insight into which components deviate and are thus potential markers for the disease mechanism.
Unlike conventional methods, which generally rely on statistical approaches and can only detect gross changes on the assumption of relatively static genetic expression, the present invention is capable of detecting both static expression level changes as well as deviations in dynamic behavior. Living cells are mostly dynamic systems, expressing and then turning off genes in complex regulation networks and biological pathways involving tens or hundreds of factors, in response to the metabolic needs of the cell. Disease most often occurs when these delicately balanced networks are upset and mis-regulated. Such an event can occur, for example, when a mutation occurs in a gene that produces a protein for regulating other, downstream genes. Since the gene product—the regulating protein—may no longer function, downstream regulation is impacted, with cascading effects. The present invention is capable of detecting these upset dynamics.
The invention takes the form of software for analysis of data. The software can run on any conventional platform, and can even be deployed in a remote, served environment, such as over the internet, because the needed data files can be sent to the processor for processing, and the results returned at a later time. However, a particular advantage of the present invention is that it has a small computing footprint and processes data quickly, making it ideal for an interactive tool, or for embedding in a distributed product for diagnostics (e.g., a CDROM deployed with a diagnostic microarray).
In a first embodiment of the invention, a model is trained with multivariate transcriptome profiles representative of normal health, and the model is then used to detect deviations in transcriptome patterns and dynamics representative of a diseased state, which deviations point to underlying disease mechanisms. An autoassociative model of gene expression data for normal tissue is trained from normal tissue data. A new input vector representing expression level data from diseased tissue is modeled with this model to provide an autoassociative estimate of the expression levels, which is then differenced with the actual measurements to provide residuals. A residual threshold test detects gene expression levels that are abnormal, identifying these genes as potential markers. Furthermore, the residual pattern can be pattern matched against stored patterns for known disease types to classify the input vector vis-à-vis disease.
In a second embodiment of the invention, a diagnostic classification is made for a dynamic multivariate genomic or proteomic profile. An inferential model is trained to recognize normal and diseased state transcriptome profiles, and new samples are analyzed and classified accordingly.
The classification capability of the invention can advantageously be extended to prognostics, or assessing the progression of a disease. Thus, in a third embodiment of the analytic method, especially useful for multi-class classification though also useful for binary classification, an autoassociative model is built for each class represented, containing only samples/observations from that class (and none that are not from that class). A new input observation to be classified is modeled by all models in the system, and an autoassociative estimate is produced from each. Each estimate is then pattern matched to the input vector to provide a “global similarity” score: The estimate vector is compared to the input vector using the similarity operator, and the vector similarity score represents the global similarity. The model that generates the estimate with the highest global similarity score represents the class of the input vector.
An important aspect of the present invention is the use of a kernel-based model, and more particularly a similarity-based model. Such models are capable of learning complex dynamic relationships from data, with a small and fast computing footprint, and are robust to noise in the input. Even more particularly, the model can be a localized model, which is reconstituted with each new input sample, thus filtering out less relevant learned profiles.
The present invention can be used to analyze data from the transcriptome, e.g., genomic expression data or proteomic data, to provide diagnostic and prognostic information, or to provide biomarker discovery information, among other things. Genetic expression level data can be obtained using a number of methods known in the art, including microarrays, and generally can comprise raw expression level information or processed comparative expression level data, e.g., from two-color experiments. In brief, a microarray comprises a plurality of probes fixed at isolated locations on a substrate. Each probe is specific to a nucleotide sequence, usually known to correspond to at least a part of a known gene. Each probe location has a different probe on it, and the probe location itself may comprise thousands to hundreds of millions of identical oligonucleotides anchored to the substrate. When DNA from a sample (e.g., processed biopsy tissue or blood serum) is exposed to the microarray, sample DNA molecules having complementary genetic sequences to the oligonucleotides will specifically attach, or hybridize, with the matching probe. Labeling of the sample DNA allows for fluorescent imaging of the washed microarray, and reveals the amount of sample DNA that hybridized with probes at each location. This image information can be converted to expression level values. Assuming the amount of DNA obtained in the processing of the sample tissue or serum corresponds to the actual amount in the cells of that sample, and assuming that hybridization occurs with statistically equal likelihood for all oligonucleotide probes, the microarray data should provide a reasonable estimate of the expression levels for each gene in the cells of the sample, that are targeted in the microarray. A microarray can have any number of genes represented on it, which may or may not be comprehensive of the genes active in the cells of the sample.
Alternatively, quantification of protein content of cells or extracellular fluids from tissue and serum samples can be obtained from known methods such as mass spectrometry. In mass spec, the sample's protein contents are separated through any of a number of possible methods, introduced into the mass spec, and a mass-to-charge pattern is obtained which can be identified as specific proteins. The strength of the signature can be equated to protein levels in the sample.
Other cellular or extracellular content levels are also subject to measurement techniques known in the art, such as measurement of “metabolomic” components such as sugars, ions, and the like, and all such data can be used in the present invention, either separately or in combination. For example, a disease mechanism may be revealed by utilizing data from both protein content and gene expression levels in combination in the present invention. As another example, a disease diagnostic signature may best be differentiated utilizing metabolomic and proteomic data in the present invention. For purposes of the description of the present invention, “expression level” will be used, but should be taken to mean quantified data regarding cellular or serum components of any kind, not just genomic expression, unless otherwise specifically indicated.
The present invention advantageously utilizes empirical modeling to provide information from genomic, proteomic, and metabolomic data. An empirical model is “trained” using data gathered at a model development stage, and then the model is used to process a new sample to provide a diagnosis or prognosis, or alternatively, the model is used to process a different class or classes of samples to reveal mechanisms responsible for the differences between the class samples, e.g., normal from disease, or cancer of one molecular mechanism from cancer of a second molecular mechanism. The present invention uniquely provides a model that itself need not be derived from scientific information of biological pathways or mechanisms, but rather is purely data driven. Furthermore, the use of a model provides insights not afforded with conventional statistical techniques that may not be able to make use of dynamic data effectively.
In step 105, expression level data from samples of the same class or a second class, e.g., diseased or cancer type 2, is provided. This data should correspond to the same components measured and modeled in step 100. Each sample represents equivalently a vector of readings of expression level, one vector element per component measured; furthermore, each sample typically will be taken from a unique tissue section or patient, although time series samples from a single patient are also contemplated hereunder.
In optional step 108, the input sample data can be culled down to a subset of data based on filtering or feature selection. For example, though expression level data for 10,000 genes may be provided in step 105, this may be reduced to 100 or 500 or 1000 select feature genes in step 108. A number of techniques are known in the art for selecting genes as “features” of a larger set of data, including selecting genes based on their signal to noise ratio. In the signal-to-noise ratio test, the expression level readings for a gene for one class of samples (e.g., normal) provides a mean expression level and a standard deviation in expression level. The same mean and standard deviation are calculated for data for that gene from samples from a second class of data. The gene's signal-to-noise (S2N) ratio is equal to the difference of the means divided by the sum of the standard deviations:
In optional step 112, the input sample data may be normalized. Normalization of genetic expression or protein content data typically means normalizing across all genes or proteins for a given sample. This may be necessitated by the fact that each microarray experiment or spectrometer often has its own baseline of level measurement, which is not easily compared from microarray to microarray or spectrometer run to spectrometer run. It is desirable then to normalize each gene expression level vis-à-vis the expression levels of the other genes from that same sample. One method for normalizing is to calculate the mean expression level and standard deviation across all the genes (or constituents) of a sample; then scale the expression levels with the mean set to zero, and the standard deviation set to some constant, e.g., one. Whatever normalization method is used on the training data from which the model was created in step 100 should be applied to the data in this step 112.
In step 116, the model is used to generate autoassociative estimates of the expression level data in response to inputting the input sample. In autoassociative estimation, each input variable is also an output. The input in this case is the vector of all expression levels for a sample; the output is an estimate of what the model says the value should be, based on the input, and may differ from the input values. Accordingly, a wide variety of modeling techniques can be used, and are discussed below in detail. For each sample vector input, there should be a sample estimate.
In the alternative, an inferential model may also be used for this step 116, whereby additional estimates are generated in the model that are not present as inputs. For example, the input vector of constituent expression levels or quantities can comprise some of the constituents of a biological regulatory network, and the model can estimate the expression levels or quantities of other constituents of the biological regulatory network there from. In order for the model to be capable of this, the model is trained with reference data comprising both the variables that will serve as inputs, as well as the variables that will be inferred as estimates.
In step 120, the autoassociative estimates are differenced with the actual input values to produce “residuals”. If an inferential model is used, while the actual measurements for constituents or components for which inferential estimates are generated are not inputted to the model, they are used at this point to form a difference with the model estimates. In step 124 one or more tests are performed on the residuals to provide information as to which genomic or proteomic components are deviating, thus identifying potential biomarkers of disease or intervention targets. The steps described above can be performed for all samples in each of the two classes, including the class that the model was trained on, to provide comparative residual information for step 124. Hence, the model is trained on data from class 1, and further provides estimates and residuals for data from class 1 (for example, with test data from that class, or on a leave-one-out modeling basis); then that same model is used to provide estimates and residuals for class 2 data.
Accordingly, one test that may be applied to the residuals is to calculate the mean of the residual for each genetic/proteomic component of the input, and compare the shift in means between one class and another. Typically, if there is no deviation in the behavior of a component, the residual mean will be close to zero, whereas it will be biased if behavior has changed. Components in which the residual mean has shifted substantially are identified as potential biomarkers, because their residuals have become pronounced when testing with data from a second class in a model built from the first class.
An extension of this mean shift test is to calculate a residual signal-to-noise ratio, as described above for raw signals: The difference in residual means between the two classes is divided by the sum of the residual standard deviations of the two classes. This conditions the shift in the mean by the apparent noisiness of the residuals.
Yet another test is to examine the standard deviation of the residuals of the two classes, and look for significant changes. This is useful when the mean of a residual may remain around zero, but a change in expression behavior is evidenced by an increase in standard deviation.
Given that microarray data can be very noisy, and the dynamics of the samples quite volatile, yet another residual test, called normalized cumulative residual, which can be applied is to sum the absolute value of the residuals for each component, and divide by the range of the expression level data for that gene.
Results of the estimation process can also be viewed interactively by a researcher to see where deviations occur between actual expression levels and estimated expression levels. A software tool can be provided that shows these differences graphically, and allows a researcher to compare different sets of samples, or different selected features. Turning to
A simulated system comprising 15 constituents was developed whereby the 15 constituents related in their dynamic behavior to one another with varying degrees of linkage, emulating a regulatory network in a metabolic system. A set of reference data for this system, comprising observations of the 15 variables throughout various states of the dynamics of the system, was used to train a kernel-based model. Then, sets of normal and diseased observations, respectively, were generated, wherein one of the 15 constituents was perturbed to be slightly lower than it should be, regardless of its raw value. Turning to
Further according to the invention, with reference to
Further according to the invention, the diagnostics exemplified in
Turning now to the model, a number of techniques may be used to capture and model the relationships between genetic components to the extent they exist as biological pathways and networks of gene regulatory products. The model technique according to the invention must provide for multivariate autoassociative estimation, and must be capable of working without time/sequence information, since the sample inputs are typically from distinct patient samples and have no known time element. Preferably, the modeling technique is empirical. One class of models appropriate for the invention is the class of kernel-based models, exemplified by the equation:
In a preferred embodiment, the kernel-based model is a similarity-based model (SBM), which in its autoassociative form is given by:
The similarity operator is a type of kernel, which compares the input vector (e.g., Xin above) to another vector (e.g., one of the learned vectors Xi) and generates a scalar score from the comparison. Generally, in terms of a similarity score, the result should range between two boundary values (e.g., from zero to one), and when the input vectors are identical, the value of one of the boundary values should be returned, and as the vectors become increasingly dissimilar, the similarity score should approach the other boundary. One example is the Gaussian operator given by:
In a preferred embodiment of the invention, localization is used to recreate the model with each new sample. Accordingly, turning to
A variety of criteria can be used to constitute the localized D matrix membership, including the similarity operator itself. Hence, according to one embodiment of the invention, the input vector can be compared to the library H of stored samples, and a similarity s can be calculated for each comparison. This provides a ranking of the stored samples, and a certain top fraction of them can be included in the D matrix. In a preferred embodiment of this localization aspect, vectors in the ranked list for H are included to the extent they provide a value for one of the transcriptome components that “brackets” the corresponding value in the input vector. This search down the ranked list is performed until either all values in the input vector are bracketed on both the low and high side, or until a maximum limit of vectors to include in D is reached. Other slight modifications in determining the membership of localized D are within the scope of the invention.
Turning now to
In optional step 607, the expression level data is normalized, to remove experiment-to-experiment variation. As mentioned above, the data can be normalized using scaling according to the mean and standard deviation of the expression levels in a given patient sample. It may be preferable to perform the normalization prior to the filtering step 613, because it will better reflect the overall baseline of the sample input experiment to use all the genomic components in the normalization, rather than just selected/filtered features. Hence, in this
In optional step 613, the data may be filtered, for example to choose certain feature genes as a subset of the total number of expression levels available in the input sample. (However, it should be noted that in a clinical setting, it is likely an appropriately down-sized microarray will have been tailored to the clinical test, and will have only the feature genes on it that are useful for the particular classification. Hence, the input sample may have expression level data for only a dozen or two dozen probes.) The signal-to-noise ratio score is a commonly used ranking for genomic feature selection and can be used in the present invention to filter a subset of expression levels for classification. The signal-to-noise scores are determined from the training data that was used to train the model, and is shown above.
In step 620, the input sample is used by the model to generate an inferred variable, the class, of the input sample. The model is “inferential” in the sense that the class variable is not one of the inputs, as would be the case in an “autoassociative” model. Whereas the class variable in the training data is typically a discrete value (e.g., “1” or “−1”), the estimate directly from the model itself will be a continuous variable over the span of the discrete class variable range.
In step 623, the estimate of the class variable is tested. Since the class estimate is a continuous variable, there is a need to “discretize” the outcome. In the simplest form, the estimated class can be thresholded, and is assigned the discrete class that it is closer to. With class variables “1” and “−1” this would be equivalent to a threshold at zero. Alternatively, the threshold can be something other than the midway point between the discrete values, if this is for example warranted by the statistical distribution of class estimates made on the training data as a test. Accordingly, the distribution of estimates for the training data can be used to establish a threshold that optimally separates the two distributions. A further possibility is to calculate a likelihood of membership function for each distribution, and test the input sample class estimate with these membership functions.
In step 630, the classification of the input sample is output to the clinician or user. The output may be accompanied by a confidence assigned from the membership functions of the two distributions.
The model used in classification is again preferably an empirical model as described above, and more preferably a kernel-based model. In contrast to the autoassociative case, the model for classification must take an inferential form (this form is also used for residual generation with inferential models, described hereinabove). Generally, the kernel-based methods will take the form:
Most preferably, a similarity-based model is used. Similarity-based modeling has a different form of the weights c as described above, and in the inferential form is used as:
The output can be an inferred variable y (Dout comprises the output classification yi associated with the training vectors), or an inferred multivariate vector y (again, Dout comprises the output vectors yi associated with the training vectors). According to the classification embodiment of the invention summarized above, the model is an inferential model that predicts the class variable y, which may take on two values in the training data, and the estimate of which will range between those two values. (In the above-described generation of residuals using an inferential model, it is inferred constituent estimate(s) that are generated instead of the classification estimate).
According to the invention, localization can be used in the inferential modeling process. The D matrix can be reconstituted on-the-fly based on the input observation. This is described above.
The global similarity scores from a set of models can be compared in step 820 to determine the actual classification of the input. For example, the highest global similarity score may designate the classification of the input vector.
However, the global similarity from each model is preferably normalized so that the global similarities can be meaningfully compared. For example, one way is to normalize the global similarity from a model according to its general global similarity mean and variance for known classification in its class. This general mean and variance of the global similarity can be provided by doing leave-one-out modeling of the training set, for example. One class model may have typical global similarities at 0.97 mean and little variance, while the next maybe at 0.84 and large variance. The actual measured global similarity can be scaled to this mean and variance to provide a normalized score. For example, the score can be the difference between the actual measured value and the mean, divided by the variance. Other calculations for normalizing are well known and can be used. For example, the distribution of general global similarity scores can be modeled with a kernel regression and the actual measured global similarity converted to a probability score related to how likely it is the measured similarity belongs to the distribution associated with that model.
The present invention can take the form of a software product, either loaded on a workstation for processing local data, or on an application server on the web for processing data sent to it (“ASP” mode). In the workstation mode, the software may provide an environment for analyzing data produced from another software package, such as gene chip analysis software from vendors of microarrays, such as Affymetrix and Agilent, whether by loading files exported by those systems, or being called from those systems and receiving data directly from them. The processing environment is similar to a batch analysis workbench, for loading data from files and running alternative analyses on the data, and reporting and graphics functions. In the ASP mode, there may additionally be a large data repository on the back end of the system that stores a large amount of data from which a wide variety of disease-condition models have been trained. A facility for receiving new data to be analyzed or classified serves the input to the appropriate model (user selectable, or selectable by an automated algorithm) to provide the classification output or residual-based disease diagnosis or prognosis. A number of algorithms might be used to determine which models to employ against the incoming data in the absence of user selection. Often, however, this is dictated by the gene expression targets selected in the input microarray experiment or test.
In a clinical setting, the software may be loaded on a workstation as part of an apparatus for large scale patient sample testing. Tailored microarrays fit for testing particular disease conditions may be used to provide expression levels for just a handful of previously identified components of the disease signature, and these expression levels are processed in the workstation software to determine diagnosis or prognosis as described above in a number of alternative embodiments. Alternatively, large microarrays may be used for broad panel testing for disease, in which case the software may run the multi-classification modeling against the data. In the context of a mass spectrometry device, quantification of constituents is provided to the workstation software from the mass spectrometry control system or via exported files, which can then generate model-based residuals for biomarker identification, or for disease diagnosis/prognosis, using models either stored locally, or downloaded or otherwise accessed via an Internet connection. For example, a diagnostic lab may use the present invention in software to analyze the constituent quantities determined through mass spectrometry from a patient sample for a plurality of constituents, whereby a set of constituent residuals is generated for at least some of the measurements for purposes of diagnosis. The model that is used to generate the residuals can be a model specific to the disease test being run, in terms of the constituents modeled together or in terms of the type of reference data used to train the model. Disease diagnosis can be performed based on the residuals, e.g., a residual pattern. Residual patterns indicative of the specific disease can be stored locally, or can be accessed over the Internet from a large repository of such patterns. Further, the results of the patient test can be uploaded (for example, later upon further disease confirmation) to improve the accuracy of the online repository, and further tune the model and/or the residual patterns indicative of disease. Diagnostic results are reported from the software to screen and/or to a printed report, and can be served to a web page or to a fax number at a physician's office as is known in the art.
It will be appreciated by those skilled in the art, that modifications to the foregoing preferred embodiments may be made in various aspects. The present invention is set forth with particularity in the appended claims. It is deemed that the spirit and scope of that invention encompasses such modifications and alterations to the preferred embodiment as would be apparent to one of ordinary skill in the art and familiar with the teachings of the present application.