Ai system and method to predict ovarian cancer propensity based on metabolic fingerprints in blood samples

Abstract

An exemplary machine learning-based system and method are disclosed for a clinical tool that can provide a comprehensive prediction of ovarian cancer propensity using metabolite and metabolomic profiles acquired from ultra-performance liquid chromatography, high-resolution mass spectrometry (UPLC-MS) analysis. Each serum sample may be independently processed through two different columns (HILIC and C18 reversed phase). Different ionization modes (negative and positive) may be used. The clinical tool can validated to provide successful prediction of ovarian cancer by over 90%. Ultra-performance liquid chromatography coupled with tandem mass spectrometry (UPLC-MS / MS) and the like can be used.

Description

Al SYSTEM AND METHOD TO PREDICT OVARIAN CANCER PROPENSITYBASED ON METABOLIC FINGERPRINTS IN BLOOD SAMPLESCROSS REFERENCE TO RELATED APPLICATIONS

[0001] This application claims the benefit of priority to U.S. Provisional Application No. 63 / 519,831, filed August 15, 2023, which is incorporated by reference herein in its entirety.BACKGROUND

[0002] Early cancer diagnosis is one of the most important contributing factors to the successful treatment of the disease. Early diagnosis is especially challenging for cancers like ovarian cancer (OC) that can progress rapidly and yet display little to no clinical symptoms early in their development.

[0003] The ideal cancer diagnostic should not only be highly accurate but also non-invasive and low-cost to be widely available to the general public. Despite efforts to develop such cancer diagnostics over the last several decades, the solution remains elusive.

[0004] There exists a need and a benefit to improving ovarian cancer diagnosis.SUMMARY

[0005] An exemplary machine learning-based system and method are disclosed for a clinical tool that can provide a comprehensive prediction of ovarian cancer propensity using metabolite and metabolomic profiles acquired from ultra-performance liquid chromatography, high-resolution mass spectrometry (UPLC-MS) analysis. Each serum sample may be independently processed through two different columns (HILIC and Cl 8 reversed phase). Different ionization modes (negative and positive) may be used. The clinical tool can validated to provide successful prediction of ovarian cancer by over 90%. Ultra-performance liquid chromatography coupled with tandem mass spectrometry (UPLC-MS / MS) and the like can be used.

[0006] In an aspect, a method is disclosed comprising: providing, by a processor, a multi- omics data set including a plurality of metabolomic profile data acquired from a liquid chromatography-mass spectrometry analysis of an individual (e.g., ultra-performance liquid chromatography, high-resolution mass spectrometry (UPLC-MS) analysis or ultraperformance liquid chromatography coupled with tandem mass spectrometry (UPLC- MS / MS)); determining, by one or more trained Al models, using the set of metabolomic profile data, one or more probability values that the individual does have ovarian cancer (and a probability value that the individual does not have ovarian cancer), wherein the one or moretrained Al models can be configured to provide a comprehensive assessment of a presence of ovarian cancer including serous papillary; and outputting, by the processor, via a report or a graphical user interface, the one or more probability values, or a value derived therefrom (e.g., consensus classifier output) that an individual does have ovarian cancer, wherein the output provided in the report or graphical user interface can be subsequently used by a clinician to (i) determine a diagnosis of ovarian cancer and / or (ii) direct treatment or screening of ovarian cancer.

[0007] In some embodiments, the liquid chromatography-mass spectrometry analysis can include features from three or more of a HILIC positive analysis, a HILIC negative analysis, a C18 reversed phase negative analysis, and a Cis reversed phase positive analysis.

[0008] In some embodiments, the one or more trained Al models can include at least one trained Al model for each of the three or more of the HILIC positive analysis, the HILIC negative analysis, the Cl 8 reversed phase negative analysis, and the Cl 8 reversed phase positive analysis.

[0009] In some embodiments, the one or more trained Al models, collectively, can be configured to use at least 10 features, wherein at least 50% of the 10 features can be from FEATURE ID NOs: 1-102.

[0010] In some embodiments, the one or more trained Al models, collectively, can be configured to use at least 100 features, wherein 50% of the 100 features can be from FEATURE ID NOs: 1-102.

[0011] In some embodiments, the one or more trained Al models, collectively, can be configured to use at least 1,000 features, wherein the at least 1,000 features can include at least 50% of FEATURE ID NOs: 1-102.

[0012] In some embodiments, the one or more trained Al models, collectively, can be configured to use at least 10,000 features, wherein the at least 10,000 features can include at least 50% of FEATURE ID NOs: 1-102.

[0013] In some embodiments, the one or more trained Al models, collectively, can be configured to provide a comprehensive assessment of ovarian cancer, including an early indicator of pre-metastatic ovarian cancer and an indicator of post-metastatic ovarian cancer.

[0014] In some embodiments, the one or more trained Al models can be configured to provide a comprehensive assessment of ovarian cancer including an early indicator of pre- metastatic ovarian cancer.

[0015] In some embodiments, the method can further include: determining, by the trained Al model, a second probability value that the individual does not have ovarian cancer; andoutputting, by the processor, via the report or the graphical user interface, the second probability value that the individual does not have ovarian cancer, wherein the probability value and the second probability value can both be used in the diagnosis of ovarian cancer.

[0016] In some embodiments, the method can further include: generating a binary output from the probability value by applying the probability value to a threshold value.

[0017] In some embodiments, the one or more trained Al models can include at least a first trained Al model and a second trained Al model, wherein the first trained Al model and the second Al model respectively can include at least one of logistic regression classifier (LRC), random forest classifier (RFC), support vector machine (SVM), k-nearest neighbor (KNN), and adaptive boosting (ADA).

[0018] In some embodiments, the trained Al model can include a consensus classifier that combines probability outputs of the first trained Al model and the second trained Al model to generate the one or more probability values.

[0019] In some embodiments, the probability output assigned by the consensus classifier can provide a background distribution of probabilities that a given sample was cancer or normal.

[0020] In some embodiments, the method can further include: determining a range of a set of ranges for the one or more probability values, or the value derived therefrom, wherein the set of ranges can include (i) a highly unlikely range, (ii) a second range in which a number of samples in the score range has a comparable number of cancerous and non-cancerous patients and (iii) a majority of training patients has been diagnosed with cancer, wherein the range can inform a guideline for subsequent testing and testing frequency (e.g., yearly monitoring, immediate evaluation).

[0021] In some embodiments, the method can further include: presenting, via the report or the graphical user interface, a range of a set of ranges for the one or more probability values, or the value derived therefrom, wherein the set of ranges can include (i) a highly unlikely range, (ii) a second range in which a number of samples in the score range has a comparable number of cancerous and non-cancerous patients and (iii) a majority of training patients has been diagnosed with cancer, wherein the range can inform a guideline for subsequent testing and testing frequency (e.g., yearly monitoring, immediate evaluation).

[0022] In some embodiments, the liquid chromatography-mass spectrometry analysis can include features from at least one of a HILIC positive analysis, a HILIC negative analysis, a Cl 8 reversed phase negative analysis, and a C18 reversed phase positive analysis.

[0023] In another aspect, a system is disclosed comprising: a processor; and a memory having instructions stored thereon, wherein execution of the instructions by the processor causes the processor to execute any one of the disclosed methods.

[0024] In another aspect, a non-transitory computer-readable medium is disclosed having instructions stored thereon, wherein execution of the instructions by a processor can cause the processor to execute any one of the disclosed methods.

[0025] Other systems, methods, features and / or advantages will be or may become apparent to one with skill in the art upon examination of the following drawings and detailed description. It is intended that all such additional systems, methods, features and / or advantages be included within this description and be protected by the accompanying claims.BRIEF DESCRIPTION OF DRAWINGS

[0026] FIGS. 1A and IB each show an exemplary machine learning -based system 100 (shown as 100a, 100b) for a clinical tool that can provide a comprehensive prediction of ovarian cancer propensity using metabolite and metabolomic profiles acquired from ultra-performance liquid chromatography, high-resolution mass spectrometry (UPLC-MS) analysis in accordance with an illustrative embodiment.

[0027] FIG. 2 shows an example method of operating a clinical tool to provide a comprehensive prediction of ovarian cancer propensity using metabolite and metabolomic profiles acquired from ultra-performance liquid chromatography, high-resolution mass spectrometry (UPLC-MS) analysis.

[0028] FIG. 3 depicts the relative abundance of select features before and after SERRF normalization. In the raw datasets, noticeable unwanted variation is observed between batches A and B in the QC samples, with significant drift within each batch. This drift can be attributed to instrumental drift and batch effects, which can introduce systematic errors. However, SERRF normalization effectively addresses most of these systematic errors, minimizing the impact of instrumental drift and batch effect on the data (R = Cis reverse phase columns; H = HILIC columns; N = negative ionization mode; P = positive ionization mode).

[0029] FIGS. 4-5 depict feature quality control and filtering. FIG. 4 shows a comparison of nonparametric residual standard deviation (nRSD) between raw and SERRF normalized datasets. Prior to applying the normalization, most features in the QC samples are observed in the red area (nRSD > 10%), which indicates high variability of the features. FIG. 5 shows the distribution of dispersion ratios (D-ratios) observed in the normalized datasets. Features that are to be removed due to not meeting the threshold (D-ratio > 10%) are in the red area as theyindicate higher variability in the QC samples than in the biological samples (R = Cis reversed- phase columns; H = HILIC columns; N = negative ionization mode; P = positive ionization mode).

[0030] FIGURES 6A-6D depict frequencies and Gini importance values of the features obtained from RFECV. FIGS. 6A-6B show frequency plots that demonstrate that, across the four datasets, a substantial proportion of the features are consistently observed across 100 iterations of RFECV (using stratified K-fold CV), rather than being limited to a select few features. FIGS. 6C-6D show that the Gini importance values for the four datasets reveal a contrasting pattern. Features ranked below the top 100 highest Gini importance values exhibit relatively smaller importance values. Notably, in RP, a discernible pattern emerges with the presence of a few distinguishing features, while HN shows little difference between the importance of top-ranked features and those in the lower ranks.

[0031] FIGURES 7A-7B depict the effect of alpha on weighted mean squared error (MSE) and evaluation metrics. FIG. 7A is a line plot that illustrates the relationship between alpha and weighted MSE. As alpha increases, the contribution of feature frequencies in the weight calculation also increases. As the alpha value increases, the reconstruction error from the autoencoder also increases. Therefore, prioritizing Gini importance appears more appropriate to retain the structure and information present in the original datasets. FIG. 7B is a scatter plot that shows small differences in various evaluation metrics across different alpha values (Fl = F 1 -score; PPV = positive predictive value; B A = balanced accuracy; NP V = negative predictive value; MCC = Matthew’s correlation coefficient).

[0032] FIGURES 8A-8B depict HMDB super classes of the putatively annotated features. FIG. 8A is a bar plot that illustrates the percentage of features with putative annotations and their assigned HMDB super class. While a small percentage of the features have putative annotations, lipids, and lipid-like molecules are the most observed super class compared to other super classes, particularly in higher ranks based on weights. FIG. 8B shows pie charts that provide a more detailed examination of the putatively annotated features in terms of their classes. Consistent with the findings in FIG. 8A, lipids and lipid-like molecules are the most observed, followed by organic acids and derivatives. Other super classes are also displayed for comparison.

[0033] FIGS. 9A-9B depict an assessment of feature extraction using an autoencoder. FIG. 9A shows that weighted MSE was examined for different numbers of nodes in the bottleneck layer during feature extraction. For all four datasets, the curve plateaus around 50 nodes, indicating that adding more nodes beyond this point adds no substantial improvement duringreconstruction. FIG. 9B shows that principal component analysis (PCA) was performed on the original input data, extracted features from the autoencoder, and reconstructed data from the extracted features. The scatterplot illustrates the representation of cancer and normal samples in the first two principal components from PCA. While the positions of the points change in the extracted and reconstructed data, the preservation of the original structure is evident from the density plots shown at the top and right sides of the scatterplots, where the distribution of both cancer and normal samples remains similar.

[0034] FIGS. 10A-10B depict a workflow diagram illustrating the data acquisition and preparation process. FIG. 10A shows the workflow diagram. Serum samples from ovarian cancer patients and non-cancerous individuals are collected from multiple geolocations. They are analyzed using UPLC-MS / MS in an untargeted workflow to characterize the metabolome of ovarian cancer patients. Normalization and filtering of the features are performed following the best practices to obtain the preprocessed metabolomic profiles for downstream analyses. FIG. 10B shows scatter plots of principal component analysis performed on the preprocessed data after accounting for systematic and random errors. QC samples (orange) which are shown to mostly cluster together with no clear separation between the two batches, indicating unwanted experimental variation has been eliminated.

[0035] FIGURE 11 depicts a bar plot that visualizes the number of early (yellow), late (red), and normal (blue) samples within each age group. The numbers on top of the bars indicate the actual number of samples in the age group, whereas the y-axis displays the percentage of each stage within the age group. For example, the age group ‘20s’ has 2 samples each from early, late, and normal (hence -33% per stage).

[0036] FIGURE 12 depicts the frequency and Gini importance values of features in each dataset. The x-axis and y-axis correspond to feature frequency and importance value, respectively. The weights computed by combining frequencies and importance values were represented by the sizes and opacity of the points. The analysis revealed that across all datasets, many features had high frequency but relatively lower levels of importance. The HN dataset exhibited the smallest range of importance values, while most features were observed frequently. The RN and HP datasets showed a similar pattern, with the HP dataset being particularly noteworthy due to a subset of features displaying lower frequencies but higher importance values. The RP dataset displayed the largest number of features with high levels of both frequency and importance values.

[0037] FIGS. 13A-13B depict a comparison of consensus classifier performance. FIG. 13A shows the performance characteristics of the models graphically represented throughprecision-recall (PR) curves. Besides the HP dataset, the models for the remaining datasets showed similar levels of performance. FIG. 13B shows that cumulative confusion matrices, also compiled from repeated CV, further reinforced these observations despite the false positives (FP) and false negatives (FN).

[0038] FIGS. 14A-14B depict evidence for the classifier's ability to clearly distinguish between cancerous and non-cancerous samples. FIG. 14A shows the distributions of voting scores that have been converted from class probabilities and ranged from -2 to 2 to improve visual clarity. The scores represent averages obtained from repeated stratified k-fold cross- validation (CV) for each sample. Clear differentiation can be observed between the scores of cancer (red) and normal (blue) samples across all four datasets. The peaks in the distributions indicate the most frequently occurring score range for the samples. FIG. 14B illustrates the best average score across the four datasets, demonstrating a notable improvement in classifying both cancer and normal classes. FIG. 14C illustrates the distribution of samples across various score ranges. This revealed that early- and late-stage samples clearly distinguish themselves from the normal samples. FIG. 14D shows that selecting the best score improves the final score for the scores at the stage-level.

[0039] FIG. 15 depicts box plots that visualize the distribution of the mean scores observed for patients 50 years old or younger (blue) and those that are older than 50 years (yellow). Based on the distribution, there is a negligible difference between the two age groups.

[0040] FIG. 16 depicts a diagram visualizing the potential adoption of the proposed workflow in a clinical setting. Given the absence of approved screening methods for ovarian cancer, this approach enables women to undergo serum profiling at a clinic to predict their cancer status. This could result in three possible scenarios: an individual's serum profile (X) falls within a score range where misdiagnosis is unlikely, enabling a confident ruling out of a cancer diagnosis. An individual's score (Y) falls within a range where 94% of others with this score have been diagnosed with cancer. Lastly, determining the cancer status of an individual (Z) may be challenging, as there are only a few samples within this score range, and it is in the middle of the distribution.DETAILED DESCRIPTION

[0041] It is appreciated that certain features of the disclosure, which are, for clarity, described in the context of separate aspects, can also be provided in combination with a single aspect. Conversely, various features of the disclosure, which are, for brevity, described in the context of a single aspect, can also be provided separately or in any suitable subcombination.Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art. Methods and materials similar or equivalent to those described herein can be used in the practice or testing of the present disclosure.

[0042] Some references, which may include various patents, patent applications, and publications, are cited in a reference list and discussed in the disclosure provided herein. The citation and / or discussion of such references is provided merely to clarify the description of the disclosed technology and is not an admission that any such reference is “prior art” to any aspects of the disclosed technology described herein. In terms of notation, “[n]” corresponds to the nth reference in the list. For example, [1] refers to the first reference in the list. All references cited and discussed in this specification are incorporated herein by reference in their entirety and to the same extent as if each reference was individually incorporated by reference.

[0043] DEFINITIONS

[0044] In this specification and in the claims that follow, reference will be made to a number of terms, which shall be defined to have the following meanings:

[0045] As used herein, “comprising” is to be interpreted as specifying the presence of the stated features, integers, steps, or components as referred to, but does not preclude the presence or addition of one or more features, integers, steps, or components, or groups thereof. Moreover, each of the terms “by”, “comprising,” “comprises”, “comprised of,” “including,” “includes,” “included,” “involving,” “involves,” “involved,” and “such as” are used in their open, non-limiting sense and may be used interchangeably. Further, the term “comprising” is intended to include examples and aspects encompassed by the terms “consisting essentially of’ and “consisting of.” Similarly, the term “consisting essentially of’ is intended to include examples encompassed by the term “consisting of.

[0046] As used in the specification and the appended claims, the singular forms “a,” “an,” and “the” include plural referents unless the context clearly dictates otherwise. Thus, for example, reference to “a compound,” “a composition,” or “a cancer” includes, but is not limited to, two or more such compounds, compositions, or cancers, and the like.

[0047] It should be noted that ratios, concentrations, amounts, and other numerical data can be expressed herein in a range format. It can be further understood that the endpoints of each of the ranges are significant both in relation to the other endpoint and independently of the other endpoint. It is also understood that there are a number of values disclosed herein and that each value is also herein disclosed as “about” that particular value in addition to the value itself. For example, if the value “10” is disclosed, then “about 10” is also disclosed. Ranges can beexpressed herein as from “about” one particular value and / or to “about” another particular value. Similarly, when values are expressed as approximations, by use of the antecedent “about,” it can be understood that the particular value forms a further aspect. For example, if the value “about 10” is disclosed, then “10” is also disclosed.

[0048] When a range is expressed, a further aspect includes one particular value and / or the other particular value. For example, where the stated range includes one or both of the limits, ranges excluding either or both of those included limits are also included in the disclosure, e.g., the phrase “x to y” includes the range from ‘x’ to ‘y’ as well as the range greater than ‘x’ and less than ‘y’. The range can also be expressed as an upper limit, e.g., ‘about x, y, z, or less’ and should be interpreted to include the specific ranges of ‘about x,’ ‘about y’, and ‘about z’ as well as the ranges of Tess than x,’ less than y’, and Tess than z.’ Likewise, the phrase ‘about x, y, z, or greater’ should be interpreted to include the specific ranges of ‘about x,’ ‘about y’, and ‘about z’ as well as the ranges of ‘greater than x,’ greater than y’, and ‘greater than z.’ In addition, the phrase “about ‘x’ to ‘y’,” where ‘x’ and ‘y’ are numerical values, includes “about ‘x’ to about ‘y’.”

[0049] It is to be understood that such a range format is used for convenience and brevity and, thus, should be interpreted in a flexible manner to include not only the numerical values explicitly recited as the limits of the range but also all the individual numerical values or subranges encompassed within that range as if each numerical value and sub-range is explicitly recited. To illustrate, a numerical range of “about 0.1% to 5%” should be interpreted to include not only the explicitly recited values of about 0.1% to about 5% but also include individual values (e.g., about 1%, about 2%, about 3%, and about 4%) and the sub-ranges (e.g., about 0.5% to about 1.1%; about 5% to about 2.4%; about 0.5% to about 3.2%, and about 0.5% to about 4.4%, and other possible sub-ranges) within the indicated range.

[0050] As used herein, the terms “about,” “approximate,” “at or about,” and “substantially” mean that the amount or value in question can be the exact value or a value that provides equivalent results or effects as recited in the claims or taught herein. That is, it is understood that amounts, sizes, formulations, parameters, and other quantities and characteristics are not and need not be exact but may be approximate and / or larger or smaller, as desired, reflecting tolerances, conversion factors, rounding off, measurement error and the like, and other factors known to those of skill in the art such that equivalent results or effects are obtained. In some circumstances, the value that provides equivalent results or effects cannot be reasonably determined. In such cases, it is generally understood, as used herein, that “about” and “at or about” mean the nominal value indicated ±10% variation unless otherwise indicated orinferred. In general, an amount, size, formulation, parameter, or other quantity or characteristic is “about,” “approximate,” or “at or about,” whether or not expressly stated to be such. It is understood that where “about,” “approximate,” or “at or about” is used before a quantitative value, the parameter also includes the specific quantitative value itself unless specifically stated otherwise.

[0051] EXAMPLE SYSTEM

[0052] FIGS. 1A and IB each show an exemplary machine learning -based system 100 (shown as 100a, 100b) for a clinical tool that can provide a comprehensive prediction of ovarian cancer propensity using metabolite and metabolomic profiles acquired from ultra-performance liquid chromatography, high-resolution mass spectrometry (UPLC-MS) analysis in accordance with an illustrative embodiment. In the example shown in FIGS. 1A and IB, the systems 100a, 100b each include an analysis system 102 (shown as 102a, 102b) comprising a trained Al model comprising a classifier configured to provide a comprehensive assessment of a presence of ovarian cancer, including serous papillary using measurements 104 of metabolites (shown as “HN,” “HP,” “RN,” and “RP” 104a in FIG. 1A and “HN,” “RN,” and “RP” 104b in FIG. IB) from ultra-performance liquid chromatography, high-resolution mass spectrometry 106. The output 108 is a probability value that the individual does have ovarian cancer. System 100 includes hardware and software components that are designed to work together in combination to facilitate the analysis and presentation of an estimation score using the algorithm to allow a physician to use that score, e.g., to assess for the presence or non-presence of ovarian cancer.

[0053] The terms “HN”, “HP”, “RN”, and “RP”, respectively, refer to metabolic profiles acquired from ‘liquid chromatography-mass spectrometry analysis including metabolic profiles from a HILIC positive analysis (HP), a HILIC negative analysis (HN), a Cl 8 reversed phase negative analysis (RN), and a Cl 8 reversed phase positive analysis (RP). The metabolic profiles, and thus features in the classifier, are peaks in the HN, HP, RN, and RP analysis that corresponds to specific molecular weight. Other features corresponding to other metabolites can be used. It is observed that Liquid chromatography-mass spectrometry provides high precision measurements that can provide consistent and comprehensitive estimation of the likelihood of ovarian cancer being present in the sample, inclduing early onset of ovarian cancer. Liquid chromatography-mass spectrometry is an analytical chemistry technique that combines the physical separation capabilities of liquid chromatography with the mass analysis capabilities of mass spectrometry.

[0054] In FIG. 1A, features corresponding to metabolites measured from four metabolic profiles from a HILIC positive analysis (HP), a HILIC negative analysis (HN), a Cl 8 reversed phase negative analysis (RN), and a Cl 8 reversed phase positive analysis (RP) are used. In Fig. IB, features corresponding to metabolites measured from three metabolic profiles from a HILIC positive analysis (HP), a C18 reversed phase negative analysis (RN), and a Cl 8 reversed phase positive analysis (RP) are used. The features of FIG. IB was further analyzed by excluding certain features, e.g., via recursive feature reduction, to provide a feature set that provides 100% specificity, a sensitivity of 98% for early ovarian cancer (OV), a sensitivity of 89% for late-stage OV, and an overall sensitivity of 91.9%. On the molecular level, cancer is a highly heterogeneous disease not only between different types of cancer but even among individuals with the same cancer type. As a consequence, finding a single molecular biomarker or set of biomarkers that are universally shared among individuals with even the same type of cancer is extremely difficult. To this extent, a larger number of features corresponding to a larger number of metabolites can provide a more accurate test, while a similar analysis using a smaller number of features can provide reasonable estimations that are quicker to analyze. Analogous to fingerprint analysis, a large number of features in a fingerprint can be used to provide a high degree of accuracy, while too many features can result in an overfit. Similarly, pre-screening operations can be performed to trigger a more comprehensive test. FEATURE ID NOs: 1-102 are key features that can be used in the analysis for ovarian cancer as a signature for the disease or a subset thereof.

[0055] In some embodiments, the trained Al models, collectively, are configured to use at least 10 features, wherein at least 50% of the 10 features are from FEATURE ID NOs: 1-102 derived from the HILIC positive analysis (HP), the HILIC negative analysis (HN), the Cl 8 reversed phase negative analysis (RN), and the Cl 8 reversed phase positive analysis (RP).

[0056] In some embodiments, the trained Al models, collectively, are configured to use at least 100 features, wherein 50% of the 100 features are from FEATURE ID NOs: 1-102.

[0057] In some embodiments, the one or more trained Al models, collectively, are configured to use at least 1,000 features, wherein the at least 1,000 features include at least 50% of FEATURE ID NOs: 1-102.

[0058] In some embodiments, the rained Al models, collectively, are configured to use at least 10,000 features, wherein the at least 10,000 features include at least 50% of FEATURE ID NOs: 1-102.

[0059] Usage of the 10 features, 100 features, 1000 features, or 10000 features from the FEATURE ID NOs: 1-102 can provide a successful estimation of the probaility of the presenceof ovarian cancer being present in a subject or patient with varying degrees of time, computing resource, and accuracy. Other features corresponding to other metabolites can be used.

[0060] In some embodiments, a prescreen operation is performed using, e.g., 10 or 100 features to which a full operation can be performed using 1000 features, or 10000 features.

[0061] Data store 110 (shown as 110a), which can be cloud-based, stores data from the ultra-performance liquid chromatography, high-resolution mass spectrometry 106. The analysis system 102 can generate the output 108 and store the output in a data store 100 (shown as 110b). The output 108 can be incorporated into a report that can be outputted as a physical report or provided to a graphical user interface for presentation to a clinician. In some embodiment, the report 104 is displayed via a secure web-based portal designed to provide healthcare providers access to their patient’s reports. The report can be put into a secure database and may securely interface and communicate with other systems, such as a hospital or physician-hosted, remotely hosted, or remote electronic health records systems.

[0062] The analysis systems 100a, 100b can each perform assessments for acquired metabolic data, e.g., to confirm validity. The assessment may employ control to determine if the control exceeded a predefined threshold. Upon a failed assessment, the analysis system 100a, 100b can present the data from being further analyzed. In some embodiments, the analysis systems 100a, 100b can generate a notification to inform the operator of a failed assessment.

[0063] If the processing is performed as a batch operation, the analysis systems 100a, 100b may continue the assessment of other data sets without stopping following a detected error for one of the data.

[0064] Feature Value Computation. The analysis systems 100b, 100b can perform feature extraction and computation to calculate feature output values from the HP, RN, and RP data (and, in some embodiments, from the HN data). The raw data from the scans may include spectra that represent the intensity of the ions at different m / z where each peak in the spectra corresponds to a different ion detected. In the context of metabolomics analysis, these ions are referred to as “features”, which represent the detected metabolites or analytes. The peak intensities may be outputted for a given metabolite which is used in the classifier. In some embodiments, multiple classifiers are used to which the outputs are combined, e.g., in an ensemble. In other embodiments, the features are consolidated into a single classifier. The output is a probability value of the presence of ovarian cancer. The output can also include a second probability value of a non-presence of ovarian cancer. The output of the classifier models in the ensemble is shown in Equation 1.(Eq. 1)

[0065] In the example of the ovarian cancer detection algorithm, five machine-learned classifier models are each calculated using the calculated feature outputs. The five classifier models include logistic regression classifier (LRC), random forest classifier (RFC), support vector machine (SVM), k-nearest neighbor (KNN), and adaptive boosting (ADA). Other classifiers can be used, including those additional described or references herein.

[0066] Physician Portal Visualization. The patient’s report may include a visualization of the acquired patient metabolite data and the results of the ovarian cancer prediction analysis. A healthcare provider, e.g., a physician, can review the report and interpret it to provide a diagnosis of ovarian cancer, direct subsequent diagnostic / testing (e.g., biopsy), or direct a treatment plan.

[0067] The analysis system (e.g., 100a, 100b) can be implemented via a processing unit (e.g., processor) configured to execute computer-readable instructions. In a configuration, the processing unit includes at least one processing circuit (e.g., core) and system memory. Depending on the exact configuration and type of computing device, system memory may be volatile (such as random-access memory (RAM)), non-volatile (such as read-only memory (ROM), flash memory, etc.), or some combination of the two. The processing unit may be a standard programmable processor that performs arithmetic and logic operations necessary for the operation of the computing device. As used herein, processing unit and processor refers to a physical hardware device that executes encoded instructions for performing functions on inputs and creating outputs, including, for example, but not limited to, microprocessors (MCUs), microcontrollers, graphical processing units (GPUs), and application-specific circuits (ASICs).

[0068] While instructions may be discussed as executed by a processor, the instructions may be executed simultaneously, serially, or otherwise executed by one or multiple processors. The processing unit may also include a bus or other communication mechanism for communicating information among various components of the computing device. In some embodiments, the processing unit is configured with co-processors (e.g., FPGA, ASIC) or AI- processors.

[0069] EXAMPLE METHOD

[0070] FIG. 2 shows an example method 200 of operating a clinical tool to provide a comprehensive prediction of ovarian cancer propensity using metabolite and metabolomicprofiles acquired from ultra-performance liquid chromatography, high-resolution mass spectrometry (UPLC-MS) analysis. In the example shown in FIG. 2, Method 200 includes providing (202), by a processor, a multi-omics data set comprising a plurality of metabolomic profile data acquired from a liquid chromatography-mass spectrometry analysis of an individual (e.g., ultra-performance liquid chromatography, high-resolution mass spectrometry (UPLC-MS) analysis or ultra-performance liquid chromatography coupled with tandem mass spectrometry (UPLC-MS / MS)).

[0071] Method 200 then includes determining (204), by one or more trained Al models, using the set of metabolomic profile data, one or more probability values that the individual does have ovarian cancer (and a probability value that the individual does not have ovarian cancer), wherein the one or more trained Al models are configured to provide a comprehensive assessment of a presence of ovarian cancer including serous papillary.

[0072] Method 200 then includes outputting (206), by the processor, via a report or a graphical user interface, the one or more probability values, or a value derived therefrom (e.g., consensus classifier output) that an individual does have ovarian cancer, wherein the output provided in the report or graphical user interface is subsequently used by a clinician to (i) determine a diagnosis of ovarian cancer and / or (ii) direct treatment or screening of ovarian cancer.

[0073] In some embodiments, the liquid chromatography-mass spectrometry analysis includes features from three or more of a HILIC positive analysis, a HILIC negative analysis, a Cl 8 reversed phase negative analysis, and a C18 reversed phase positive analysis.

[0074] In some embodiments, the one or more trained Al models include at least one trained Al model for each of the three or more of the HILIC positive analysis, the HILIC negative analysis, the Cl 8 reversed phase negative analysis, and the Cl 8 reversed phase positive analysis.

[0075] In some embodiments, the one or more trained Al models, collectively, are configured to use at least 10 features, wherein at least 50% of the 10 features are from FEATURE ID NOs: 1-102.

[0076] In some embodiments, the one or more trained Al models, collectively, are configured to use at least 100 features, wherein 50% of the 100 features are from FEATURE ID NOs: 1-102.

[0077] In some embodiments, the one or more trained Al models, collectively, are configured to use at least 1,000 features, wherein the at least 1,000 features include at least 50% of FEATURE ID NOs: 1-102.

[0078] In some embodiments, the one or more trained Al models, collectively, are configured to use at least 10,000 features, wherein the at least 10,000 features include at least 50% of FEATURE ID NOs: 1-102.

[0079] In some embodiments, the one or more trained Al models, collectively, are configured to provide a comprehensive assessment of ovarian cancer, including an early indicator of pre-metastatic ovarian cancer and an indicator of post-metastatic ovarian cancer, including an early indicator of pre-metastatic ovarian cancer.

[0080] In some embodiments, the method 200 further includes determining, by the trained Al model, a second probability value that the individual does not have ovarian cancer; and outputting, by the processor, via the report or the graphical user interface, the second probability value that the individual does not have ovarian cancer, wherein the probability value and the second probability value are both used in the diagnosis of ovarian cancer.

[0081] In some embodiments, the method 200 includes generating a binary output from the probability value by applying the probability value to a threshold value.

[0082] In some embodiments, The one or more trained Al models include at least a first trained Al model and a second trained Al model, wherein the first trained Al model and the second Al model respectively include at least one of logistic regression classifier (LRC), random forest classifier (RFC), support vector machine (SVM), k-nearest neighbor (KNN), and adaptive boosting (ADA).

[0083] In some embodiments, the trained Al model includes a consensus classifier that combines probability outputs of the first trained Al model and the second trained Al model to generate the one or more probability values.

[0084] In some embodiments, the probability output assigned by the consensus classifier provides a background distribution of probabilities that a given sample was cancer or normal.

[0085] In some embodiments, the method 200 includes determining a range of a set of ranges for the one or more probability values, or the value derived therefrom, wherein the set of ranges includes (i) a highly unlikely range, (ii) a second range in which a number of samples in the score range has a comparable number of cancerous and non-cancerous patients and (iii) a majority of training patients has been diagnosed with cancer, wherein the range informs a guideline for subsequent testing and testing frequency (e.g., yearly monitoring, immediate evaluation).

[0086] In some embodiments, the method 200 includes presenting, via the report or the graphical user interface, a range of a set of ranges for the one or more probability values, or the value derived therefrom, wherein the set of ranges includes (i) a highly unlikely range, (ii) asecond range in which a number of samples in the score range has a comparable number of cancerous and non-cancerous patients and (iii) a majority of training patients has been diagnosed with cancer, wherein the range informs a guideline for subsequent testing and testing frequency (e.g., yearly monitoring, immediate evaluation).

[0087] METHOD OF TRAINING

[0088] In addition to the machine learning features described above, the various analysis system can be implemented using one or more artificial intelligence and machine learning operations. The term “artificial intelligence” can include any technique that enables one or more computing devices or comping systems (i.e., a machine) to mimic human intelligence. Artificial intelligence (Al) includes but is not limited to knowledge bases, machine learning, representation learning, and deep learning. The term “machine learning” is defined herein to be a subset of Al that enables a machine to acquire knowledge by extracting patterns from raw data. Machine learning techniques include, but are not limited to, logistic regression, support vector machines (SVMs), decision trees, Naive Bayes classifiers, and artificial neural networks. The term “representation learning” is defined herein to be a subset of machine learning that enables a machine to automatically discover representations needed for feature detection, prediction, or classification from raw data. Representation learning techniques include, but are not limited to, autoencoders and embeddings. The term “deep learning” is defined herein to be a subset of machine learning that enables a machine to automatically discover representations needed for feature detection, prediction, classification, etc., using layers of processing. Deep learning techniques include but are not limited to artificial neural networks or multilayer perceptron (MLP).

[0089] Machine learning models include supervised, semi-supervised, and unsupervised learning models. In a supervised learning model, the model learns a function that maps an input (also known as feature or features) to an output (also known as target) during training with a labeled data set (or dataset). In an unsupervised learning model, the algorithm discovers patterns among data. In a semi-supervised model, the model learns a function that maps an input (also known as a feature or features) to an output (also known as a target) during training with both labeled and unlabeled data.

[0090] Neural Networks. An artificial neural network (ANN) is a computing system including a plurality of interconnected neurons (e.g., also referred to as “nodes”). This disclosure contemplates that the nodes can be implemented using a computing device (e.g., a processing unit and memory as described herein). The nodes can be arranged in a plurality of layers such as an input layer, an output layer, and optionally one or more hidden layers withdifferent activation functions. An ANN having hidden layers can be referred to as a deep neural network or multilayer perceptron (MLP). Each node is connected to one or more other nodes in the ANN. For example, each layer is made of a plurality of nodes, where each node is connected to all nodes in the previous layer. The nodes in a given layer are not interconnected with one another, i.e., the nodes in a given layer function independently of one another. As used herein, nodes in the input layer receive data from outside of the ANN, nodes in the hidden layer(s) modify the data between the input and output layers, and nodes in the output layer provide the results. Each node is configured to receive an input, implement an activation function (e.g., binary step, linear, sigmoid, tanh, or rectified linear unit (ReLU), and provide an output in accordance with the activation function. Additionally, each node is associated with a respective weight. ANNs are trained with a dataset to maximize or minimize an objective function. In some implementations, the objective function is a cost function, which is a measure of the ANN’S performance (e.g., error such as LI or L2 loss) during training, and the training algorithm tunes the node weights and / or bias to minimize the cost function. This disclosure contemplates that any algorithm that finds the maximum or minimum of the objective function can be used for training the ANN. Training algorithms for ANNs include but are not limited to backpropagation. It should be understood that an ANN is provided only as an example machine learning model. This disclosure contemplates that the machine learning model can be any supervised learning model, semi-supervised learning model, or unsupervised learning model. Optionally, the machine learning model is a deep learning model. Machine learning models are known in the art and are therefore not described in further detail herein.

[0091] A convolutional neural network (CNN) is a type of deep neural network that has been applied, for example, to image analysis applications. Unlike traditional neural networks, each layer in a CNN has a plurality of nodes arranged in three dimensions (width, height, depth). CNNs can include different types of layers, e.g., convolutional, pooling, and fully- connected (also referred to herein as “dense”) layers. A convolutional layer includes a set of filters and performs the bulk of the computations. A pooling layer is optionally inserted between convolutional layers to reduce the computational power and / or control overfitting (e.g., by downsampling). A fully-connected layer includes neurons, where each neuron is connected to all of the neurons in the previous layer. The layers are stacked similarly to traditional neural networks. GCNNs are CNNs that have been adapted to work on structured datasets such as graphs.

[0092] As used herein, the term “encoder-decoder Al model” describes an architecture of a machine learning system in which input data are encoded or coded in order then to be decodedagain immediately afterward. In the middle between the encoder and the decoder the necessary data are present as a type of feature vector. During decoding, depending on the training of the machine learning model, specific features in the input data can then be specially highlighted.

[0093] As used herein, the term “U-nef ’ describes an architecture of a machine learning system which is based on a convolutional network architecture. This architecture is particularly well suited to a fast and accurate segmentation of digital images in the biological / medical field. A further advantage of such a machine learning system is that it manages with fewer training data and allows a comparatively accurate segmentation.

[0094] Other Supervised Learning Models. A logistic regression (LR) classifier is a supervised classification model that uses the logistic function to predict the probability of a target, which can be used for classification. LR classifiers are trained with a data set (also referred to herein as a “dataset”) to maximize or minimize an objective function, for example, a measure of the LR classifier’s performance (e.g., error such as LI or L2 loss), during training. This disclosure contemplates that any algorithm that finds the minimum of the cost function can be used. LR classifiers are known in the art and are therefore not described in further detail herein.

[0095] A Naive Bayes’ (NB) classifier is a supervised classification model that is based on Bayes’ Theorem, which assumes independence among features (i.e., the presence of one feature in a class is unrelated to the presence of any other features). NB classifiers are trained with a data set by computing the conditional probability distribution of each feature given a label and applying Bayes’ Theorem to compute the conditional probability distribution of a label given an observation. NB classifiers are known in the art and are therefore not described in further detail herein.

[0096] A k-NN classifier is an unsupervised classification model that classifies new data points based on similarity measures (e.g., distance functions). The k-NN classifiers are trained with a data set (also referred to herein as a “dataset”) to maximize or minimize a measure of the k-NN classifier’s performance during training. This disclosure contemplates any algorithm that finds the maximum or minimum. The k-NN classifiers are known in the art and are therefore not described in further detail herein.

[0097] The exemplary system and method can be based on any type of machine learning or artificial intelligence system, including, and not limited to, convolutional neural networks, autoencoders, recombinant neural networks, and long-term short-term memory (LSTM) networks.

[0098] EXPERIMENTAL RESULTS AND ADDITIONAL EXAMPLES

[0099] A study was conducted that explores the development of an ML-based approach for the early detection of OC using metabolomic profiles in blood. Analyses were carried out on serum samples collected from 431 OC patients and 133 normal (donors with no known medical pathologies) at four geographic locations in the United States and Canada. The utility of a consensus classifier was evaluated using four independent sets of metabolomic profiles. Combining the best predictions from each profile using the consensus classifier resulted in a final set of predictions that can distinguish cancer from control samples with high accuracy (PPV 93%). The study illustrated how the frequency distribution of individual patient scores can be used to develop a useful clinical tool that may be used to assign a likelihood that an individual does or does not have OC.

[0100] Study Motivation. In recent years, various computational methods, including machine learning (ML), have been applied in efforts to identify patterns embedded within large omics datasets (e.g., genomic / proteomic / metabolomic) that may constitute an accurate diagnostic of cancer [5,6] and other diseases [7], For example, perturbations of metabolic levels in the blood and / or other body fluids have long been considered promising indicators of cancer and other diseases [8,9] because metabolites constitute end points of many, if not most, of the molecular processes underlying biological functions. As such, metabolic profiles have been proposed as a molecular phenotype of biological systems, reflective of collective information encoded at the genome level and realized at the transcriptome and proteome levels

[0010] ,

[0101] Despite the inherent advantages of metabolic patterns as biomarkers of cancer and other diseases, extreme care is required in both the selection and analysis of metabolomic datasets. For example, potential technical inconsistencies in data acquisition (e.g., variation in sensitivity between instruments / laboratories and / or analytic drift associated with the same instrument over time) can easily compromise the reliability of acquired datasets unless frequent standardization with control samples is employed throughout the analytic process. In addition, extra precaution is needed in both the computational analysis of metabolic and other omics datasets and the interpretation of results. For example, there are a variety of ML approaches to the analysis of omics data, and each is associated with individual strengths and weaknesses

[0011] ,

[0012] , Despite these challenges, the use of metabolomic and other omics profiles as early indicators of cancer is not insurmountable and may provide clinicians with a powerful and highly accurate tool for personalized cancer diagnosis when properly addressed.

[0102] Materials and Methods

[0103] Sample collection: Serum samples from ovarian cancer (serous papillary) patients and non-cancer controls (donors with no known medical pathologies) were obtained from 1) a first clinical location in Atlanta, GA (having 10 early- and 142 late-stage cancer samples); 2) a second clinical location in Philadelphia, PA (having 51 early- and 68 late-stage cancer samples, 133 control samples); 3) a third clinical location in Chapel Hill, NC (having 17 early-stage cancer samples); 4) a fourth clinical location in Alberta, Canada (having 23 early- and 120 late- stage cancer samples). Serum samples were aliquoted (500ul) into separate tubes and stored at -80 °C.

[0104] Sample preparation: Serum samples were sent to Creative Proteomics (New York) for analysis. 100 pL of each serum sample was transferred into a 1.5 mL Eppendorf centrifuge tube with 300 pL of 80% methanol added into each tube before the mixture was vortexed for 30 seconds. Subsequently, the samples were kept at -40 °C for 1 hour. Samples were vortexed again for 30 seconds and then stored once more at -40 °C for 30 minutes. Following this, the samples were centrifuged at 12,000 rpm and 4 °C for 15 minutes. The supernatant was carefully transferred to clean tubes and kept at -40 °C for an additional hour. After this last incubation, the samples were centrifuged again under the same conditions for another 15 minutes. Finally, 120 pL of the supernatant and 3 pL of DL-o-Chlorophenylalanine (100 pg / mL concentration) were transferred to a vial, ready for UPLC-MS / MS analysis. In addition to the individual patient samples, pooled quality control (QC) samples were created by combining small aliquots of the study samples in order to monitor the performance and consistency of the analysis.

[0105] UPLC-MS / MS analysis of serum samples: Separation of the molecules was performed using Cis (reversed phase, “R”) and HILIC (hydrophilic interaction liquid chromatography “H”) columns with the UltiMate™ 3000LC combined with Q Exactive MS (ThermoFisher). For Cis, the LC system includes an ACQUITY UPLC BEH Cis (100><2. Immx 1.7 pm). The mobile phase contains solvent A (0.05% formic acid-water) and solvent B (acetonitrile) with a gradient elution (0-1.0min, 95% A; 1.0-12.0 min, 95%-5% A; 12.0-13.5 min, 5% A; 13.5-13.6 min, 5%-95% A, 13.6-16 min, 95% A). The flow rate of the mobile phase was 0.3 mL / min. For HILIC, the LC system is comprised of an XBridge BEH Amide Column (150x4.6 mmx3.5 pm) with mobile phase utilizing solvent A (5% acetonitrile water, containing 10 mM ammonium formate) and solvent B (95% acetonitrile-water, containing 10 mM ammonium formate) with a gradient elution (0-0.5 min, 5% B; 0.5-9.5 min, 5%-35% B; 9.5-11.5 min, 35%-60% B; 11.5-13.5 min, 60% B; 13.5-13.6 min, 60%-5% B; 13.6-16.0 min, 5% B). The flow rate of the mobile phase was 0.6mL / min. The temperature for both columns was maintained at 40 °C with the sample manager temperature set at 4 °C. Forevery analysis, the injection volume of each sample was 10 pL, with a QC sample being analyzed every ten patient samples.

[0106] The following mass spectrometry (MS) parameters were used during positive (“P”) and negative (“N”) ionization modes using electrospray ionization (ESI): heater temperature of 300 °C, sheath gas flow rate of 45 arbitrary units (arb), aux gas flow rate of 15 arb, sweep gas flow rate of 1 arb, and capillary temperature of 350 °C. The spray voltage and S-Lens RF level for ESI+ and ESI- modes are 3.0KV and 30%, and 3.2KV and 60%, respectively.

[0107] Data generation: Two different MS data acquisition methods were used during this study: full scan and data-dep endent scanning (ddMS2). For full scan data acquisition, the resolution was set at 70,000 full-width half max (fwhm), the automatic gain control (AGC) target was 3xl06, the maximum injection time was 100 ms, the scan range was 70-1050 mass- to-charge ratio (m / z), the polarity was set to negative or positive, and the spectrum data type was centroid. For ddMS2, the resolution was set at 17,500 fwhm, the AGC target was IxlO5, the maximum injection time was 50 ms, the isolation window was 1.7 m / z, the loop count was 10, the normalized collision energy (NCE) for fragmentation of selected precursor ions based on the specified isolation window was 15 / 30 / 45 eV, and the spectrum data type was centroid.

[0108] The raw data from the scans includes spectra that represent the intensity of the ions at different m / z where each peak in the spectra corresponds to a different ion detected. In the context of metabolomics analysis, these ions are referred to as “features”, which represent the detected metabolites or analytes. Compound Discoverer Software (version 3.1) from ThermoFisher was used to process the raw data using the following parameters. For processing spectra data, the minimum precursor mass was 0 dalton (Da), the maximum precursor mass was 5,000 Da, and the signal-to-noise threshold was 1.5. For the alignment model used to correct variations in retention times across different runs, the adaptive curve was employed with a maximum shift of 2 minutes and mass tolerance of 5 parts per million (ppm). Putative annotations of the features were assigned for preliminary identification from database searches using Compound Discoverer.

[0109] Data preprocessing'. The metabolomic datasets acquired through UPLC-MS / MS analysis of patient serum samples required multiple preprocessing stages to minimize inherent analytical errors. These errors are generally divided into two primary categories: systematic and random errors.

[0110] Systematic errors, also referred to as repeatable inaccuracies, are primarily introduced as batch effects and longitudinal drifts during sequential sample processing. To address these, the study employed a strategy known as Systematic Error Removal usingRandom Forest (SERRF)

[0038] , This approach leverages pooled QC samples to approximate and correct batch and injection order effects (FIG. 3). To evaluate the effectiveness of this correction, non-parametric residual standard deviations (nRSDs) were calculated both pre- and post-correction.

[0111] Alongside systematic errors, data may also be influenced by random errors — unpredictable factors that are difficult to identify and quantify. Although these random errors present a unique challenge, it is possible to alleviate their impact by filtering features based on defined dispersion ratios (D-ratios) that measure the balance between technical and biological variation, with a lower D-ratio indicating higher quality and reliability. Filtering for D-ratio and nRSDs based on a given threshold (< 0.1 for this study) ensures the inclusion of only reliable and consistent features in downstream analyses (leaving only features that satisfy both nRSD and D-ratio thresholds). The number of features satisfying the said thresholds, as well as the distribution of nRSDs and D-ratios can be found in TABLE 1 and FIG. 4, respectively.TABLE 1. Number of features satisfying nRSD and D-ratio thresholds (< 10%).

[0112] Workflow: The workflow outlined in this study includes three main steps. The first involves examining frequencies and importance values of the features in each of the four datasets using recursive elimination with cross-validation (RFECV). This specific task serves to identify stable “drivers” that would be important when making downstream predictions. To this end, a combined metric, “weight,” is computed for each feature, allowing for rankings among the features within each dataset.

[0113] With these rankings, putatively annotated features were grouped based on their characteristics and functions. These rankings are taken into consideration during feature extraction via autoencoder (see below). This reduces the dimensionality of the original datasetsprior to evaluating the consensus classifier that uses five individual classifiers to make predictions.

[0114] Finally, the probabilities assigned to individual samples in the study during the classifier evaluation step were further examined. This additional assessment serves not only to gauge the confidence of the predictions, but also to create a background distribution of the class (Cancer / Normal) probabilities that may potentially have a clinical utility in determining an individual’s likelihood of having cancer. The workflow was specifically designed for binary classification, involving Cancer and Normal classes.

[0115] Recursive feature elimination with cross-validation (RFECV)'. As outlined above, RFECV was used to obtain information about the features in the dataset. This process involves selecting the most important features in a model, starting with the model training on all features. Then, at each iteration, the least informative features are pruned from the current set of features until the algorithm reaches the specified number of features. Coupling RFE with CV ensures a more robust selection of the least informative features for removal. In evaluating the model during RFECV, the Fl -score was used as the performance metric. For the calculation of performance metrics like the Fl score, a threshold of 0.5 was set for cancer probability. However, depending on the specific application setting that adopts this approach, it may be beneficial to explore whether sensitivity (reduce false negatives) or specificity (reduce false positives) should be prioritized. Area under the receiver operating characteristic (AUROC) and precision-recall area under the curve (PR AUC) are provided in TABLE 2.TABLE 2. Area under the receiver operating characteristic (AUROC) and precision-recall area under the curve (PR AUC).

[0116] The RFECV implementation employed in this study utilized a tree-based ensemble method known as “Extremely Randomized Trees”

[0039] , Extremely Randomized Trees (ERT) differ from standard random forests by selecting split points entirely at random. This added an additional layer of randomness, potentially increasing model robustness. In the methodology, the study first used cross-validation (CV) with ERT to identify the most effective hyperparameters (e.g., depth of the tree). Following this, CV was applied again to select the best set of features, ensuring a comprehensive and systematic approach to model optimization.

[0117] Although integrating CV with ERT requires significant computational resources, it is vital for thorough model evaluation and hyperparameter tuning. While this process is not without challenges, parallel processing can be utilized with larger feature sets / datasets to mitigate computational demands. Each CV fold serves as a distinct training and validation set for ERT, enabling an exhaustive evaluation across various data subsets. This contributes to the creation of a model that is more likely to generalize effectively.

[0118] As with other ensemble methods in this category, it starts with multiple decision trees that are responsible for dividing the samples based on the values of the features in the dataset. While other methods aim to choose the best feature and its value that can minimize the mix of classes after the split, the extremely randomized trees method randomly selects a feature and its value to perform the split. Each tree then computes a decrease in impurity using the Gini index, reflective of the degree of heterogeneity resulting from the split. Gini impurity was calculated using Equation 2, and ranges from 0 (maximum purity - all samples belong to a single class) to 0.5 (specific to binary classification and represents an equal number of samples belonging to each class). In the equation, n represents the number of classes and p is the proportion of samples in the node belonging to class i.(Eq. 2)

[0119] The feature importance values, calculated by averaging the Gini index observed across all the trees in the ensemble, are utilized during RFECV to rank and prune features. This approach has demonstrated the ability to not only reduce over fitting but also maintainpredictive performance. The implementation of RFECV, as described in this section, was done using ExtraTreesClassifier and RFECV with stratified K-fold cross-validation from scikit-learn library

[0040] , Repeated stratified K-fold cross-validation involves performing K-fold cross- validation multiple times with different random data splits. In each iteration, the dataset is divided into 'K' folds, with each fold serving as the validation set once, ensuring every data point is used for validation. This process is repeated 'N' times with new random splits. The results from all repetitions are collected (e.g., all true positives, false positives, true negatives, and false negatives) to compute cumulative performance metrics. In the study, this method was applied consistently for both model evaluation and other processes like feature selection and training consensus classifiers, integrating these steps into the overall cross-validation framework to prevent overfitting and ensure comprehensive model assessment.

[0120] Feature assessment: To assess the features of the datasets, RFECV was repeated 100 times with K=5 to examine both frequencies and mean importance values. Frequencies (F() refer to the number of times a given feature shows up throughout the iterations with the importance value (G / ) being the average Gini importance controlled for the frequency (FIG. 6). These two metrics were merged using weighted average to compute weights for the features. Then, features were ranked using the weights of the features. In addition to giving the mean importance more relative influence in the final weight (Wj) to be assigned, the decision of setting to 0.3 was also influenced by its downstream impact on reconstruction error after the feature extraction step (FIG. 7A) as well as the classifier evaluation (FIG. 7B).

[0121] Super class of a metabolite refers to a high-level categorization of compounds based on their chemical properties, functions, or biological activities. To assign such classes to the features, names and chemical formulas of the putative annotations were queried using the metabolite data (XML format) from Human Metabolome Database (HMDB)

[0041] to find a feature’s super class (FIG. 8, TABLEs 3A and 3B).TABLE 3A. FEATURE_ID NOs: 1-102: HMDB superclasses of the features with putative annotation.

[0122] Feature extraction using Autoencoder: An autoencoder is an artificial neural network often used for feature extraction where the goal is to learn a compressed representation of the original input data

[0042] ,

[0043] ,

[0044] , It included two components: an encoder and a decoder. The encoder outputs a lower-dimensional, or “latent”, representation that captures the most important features of the input data. In contrast, the decoder evaluates the reconstruction error after decompressing the encoded data, providing a measure of how effectively the autoencoder preserved the original input. When implementing an autoencoder, there are components that require specification: the architecture of the autoencoder including the number of layers and nodes per layer, activation functions, and optimizer.

[0123] In this study, an autoencoder with a single layer was implemented using the TensorFlow library

[0045] , This single layer, known as the bottleneck, performs compression to provide latent representation of the input data. The number of nodes in this layer was set to the square root of the number of features in the original data, which follows a heuristic that balances retaining information and dimensionality reduction

[0046] ,

[0047] , Increasing the number of nodes beyond this did not significantly improve the weighted mean squared error for the four datasets (FIG. 9).

[0124] Following the architecture setup, the exponential linear unit (ELU)

[0048] was selected as the activation functions for its ability to transform input values into a consistentrange of real values post-compression. These functions are essential components of the autoencoder as they allow the autoencoder to learn and represent complex patterns found in the input data.

[0125] Finally, Adam optimization

[0049] was employed to adjust the autoencoder’s parameters, aiming to minimize the reconstruction error between the autoencoder’s output (Xi) and the original input data (X£). The loss function used for this optimization is the weighted mean squared error (MSE). This function incorporates the weights (w() of the features, which are assigned based on the feature assessment phase, to emphasize the importance of higher- ranked features during reconstruction. In essence, these weights prioritize the fidelity of reconstruction for features considered more crucial, guiding the autoencoder to focus on reducing the error more significantly for these features.Weighted MSE = ln(- = lnw£* Xt— X2

[0126] In training the autoencoder, the default parameters from TensorFlow were used. These parameters include the batch size, the number of training epochs, and the learning rate.

[0127] Consensus classifier'. The implementation of consensus classifiers involves aggregating predictions from multiple classifiers to arrive at a final decision. Recognizing that different classifiers have different biases, strengths, and weaknesses, the consensus approach utilizes diversity to enhance its robustness

[0050] ,

[0128] In this study, five types of classifiers - logistic regression classifier (LRC)

[0051] , random forest classifier (RFC)

[0046] , support vector machine (SVM)

[0052] , k-nearest neighbor (KNN)

[0053] , and adaptive boosting (ADA)

[0054] - were employed using the implementations from scikit-learn library

[0040] , Hyperparameters for each classifier were fine-tuned using a grid search to ensure optimal performance prior to being incorporated into the consensus classifier (TABLE 4). The probabilities generated from individual classifiers were calibrated using Platt scaling

[0055] to convert the output from individual classifiers into the likelihood of prediction. Once ready, a soft-voting strategy was used to average the class probabilities predicted by each classifier to make the collective predictions.TABLE 4. Final hyperparameters used to train the individual classifiers, dct max depth = decision tree w / max depth of N; LRC = logistic regression classifier; RFC = random forest classifier; SVC = support vector machine; KNN = k-nearest neighbor; ADA = adaptive boosting.

[0129] To evaluate the performance of the individual and consensus classifiers in the presence of class imbalance, balanced accuracy, Fl -score, average precision, and Matthew’s correlation coefficient, sensitivity, specificity, negative predictive value (NPV), and positive predictive value (PPV) were used in conjunction with repeated cross-validation using stratified K-fold ( C=5) strategy (TABLE 5, TABLE 6). Moreover, cumulative confusion matrices were computed using the number of prediction outcomes collected from the repeated cross- validations. Finally, the performance of the consensus classifier was analyzed by shuffling the class labels to verify whether the classifier was truly learning meaningful patterns and relationships between the input features and original class labels rather than relying on random noise. The observed decrease in performance after shuffling the labels confirms the importance of using correct class labels during the classification (TABLE 7).TABLE 5. Individual classifier evaluation metrics for training and validating set. ADA: Adaboost; KNN: K-nearest neighbor; LRC: logistic regression; RFC: random forest; SVC: support vector machine; BA: balanced accuracy; Fl : fl-score; MCC: Matthew’s Correlation Coeffcient; NPV: negative predictive value; PPV: positive predictive value.TABLE 6. Consensus classifier evaluation metrics for training and validating set. BA: balanced accuracy; Fl : fl -score; MCC: Matthew's Correlation Coeffcient; NPV: negative predictive value; PPV: positive predictive value.TABLE 7. Observed Fl scores from shuffling class labels (cancer / normal).

[0130] The probabilities assigned to individuals in the study during repeated stratified k- fold cross-validation were utilized to create a background distribution of the class probabilities for Cancer and Normal classes. The background distribution represents the spectrum of probabilities observed among individuals in the study cohort. Although this distribution was not used to directly assign a final probability to any sample, it serves as a valuable reference. It provides a basis for relative comparison against scores or probabilities that may be assigned to future patients. For visualization purposes, these probabilities were converted to a “voting score” scaled to fit within a range of -2 to 2.

[0131] Results

[0132] Data acquisition'. The data acquisition process for this study is summarized in FIG. 10A. Serum samples were collected from 431 OC patients and 133 non-cancerous / normal individuals (FIG. 11). Samples were characterized using ultra-performance liquid chromatography coupled with tandem mass spectrometry (UPLC-MS / MS). Each serum sample was independently processed through two different columns (HILIC and Cis reversed phase) and analyzed using two different ionization modes (negative and positive), resulting in four distinct datasets. Because of the large number of samples, metabolomic analyses were conducted over two separate batches. To detect and correct instrument drift within and between runs, a pooled quality control (QC) sample was run following analysis of every ten patient samples. A scatter plot of principle component analyses performed on the preprocessed data confirmed that no significant experimental variation was detected between batches after quality control of the data (FIG. 10B).

[0133] Accessing the stability of metabolomic features'. The overall goal of the study is the identification / development of an ML classifier that utilizes metabolic profiles to accurately distinguish individuals with or without OC. Toward this end, the study independently examined the predictive accuracy of five ML classifiers for each of the four datasets: RFC, SVC, ADA, KNN, and LRC.

[0134] Prior to the independent evaluation of each of these classifiers, the study identified reliable features (metabolites) with recursive feature elimination (RFE)

[0013] coupled with repeated cross-validation (CV). The output from these processing steps for each of the four datasets was an assignment of a relative ranking of features reflective of the relative frequencies of the features after repeated CV iterations, as well as their relative contribution levels as determined by the Gini importance scores.

[0135] Across all four datasets, a moderate positive correlation (HN: R = 0.26, p < 0.001; HP: R = 0.56, p < 0.001; RN: R = 0.39, p < 0.001; RP: R = 0.50, p < 0.001) was observed between the relative frequency of features and their importance (FIG. 12). This trend is most apparent in the RP dataset where the vast majority of features of high importance were in high frequency. In contrast, the HP dataset displayed a number of lower frequency features of high Gini importance.

[0136] Features were assigned weights, a combined metric of both relative frequency and importance, then ranked and grouped into rank groups. Features were then classified with respect to putative functions using the human metabolome database (HMDB). Lipids and lipid- like molecules were found to be widely distributed across rank groups while most other putatively annotated classes of metabolites were predominantly associated with lower rankfeatures (FIG. 8). The vast majority of the highly ranked features remain unannotated. Indeed, only ~7% of the complete set of features identified in this study was associated with metabolite information from HMDB.

[0137] Evaluation of classifier performance'. Prior to the evaluation of the classifier performance, a neural network-based autoencoder was used to reduce the dimensionality of the datasets while preserving informative representation of the original (FIG. 9). Using the compressed dataset, the ability of each of the five classifiers (RFC, SVC, ADA, KNN, LRC) to correctly identify cancer samples and non-cancer controls was independently evaluated using four metrics: 1) Positive predictive value (PPV, a.k.a. precision), 2) negative predictive value (NPV), 3) Fl-score (Fl), and 4) Matthew's correlation coefficient (MCC). PPV is the number of true positives divided by the number of true positives plus false positives (potential range: 0-100%), while NPV is the number of true negatives divided by the number of true negatives plus false negatives (potential range: 0-100%). The Fl -score, which symmetrically represents both precision and recall in a single metric, is the harmonic mean of precision and recall (a.k.a. sensitivity; potential range: 0-100%). MCC reflects the correlation between the observed and predicted binary classifications (potential range: -1 to +1). An MCC of +1 represents a perfect prediction, 0 no better than a random prediction, and -1 indicates total disagreement between predictions and observations. MCC considers true and false positives and negatives and is generally regarded as a balanced measure of predictive accuracy, even if the classes are of very different sizes

[0014] , The performance of each of the five classifiers and the consensus classifier based on repeated cross-validation is presented in TABLE 8 in order of descending PPV performance value for each dataset.TABLE 8. Performance evaluation metrics for individual and consensus classifier.

[0138] While the performance of the individual and consensus classifiers varied across different datasets, the differences were minor. The HP dataset displayed a slightly lower performance relative to the HN, RN, and RP datasets. However, the overall performance was consistently high across the four datasets (PPV > 93%; NPV > 87%; Fl > 92%; MCC > 0.78; FIG. 13A). The cumulative confusion matrix from the consensus classifier (FIG. 13B) is generally consistent with these results demonstrating a relatively low misclassification rate of false negatives (~2 to 3%) and slightly higher rate of false positives (~11 tol7%).

[0139] Utility of class probabilities as background distributions'. The results of the repeated stratified k-fold cross-validation scores can be used to assign a mean “voting score” (adjusted to fall within a - 2 to 2 range) that signifies the certainty of either a cancerous or non- cancerous classification. These probabilities were averaged for each sample, and the distributions for each of the four datasets are displayed in FIG. 14A. The results highlight the classifier's ability to clearly distinguish between cancerous and non-cancerous samples in all four datasets.

[0140] By combining results from the four datasets and selecting the best average score among them, a notable improvement in classifying both cancer and normal samples wasobserved. This underscores that each dataset brings its unique contribution to the accurate prediction of cancer or non-cancer status (FIG. 14B). A striking 97% of the cancer samples scored within the 1.0-2.0 range, with no (0%) misclassification of non-cancerous samples (TABLE 9). In contrast, 83% of the non-cancerous / normal samples were found to fall within the -2.0 to -1.0 score range, indicating that the consensus classifier is better at predicting cancer than non-cancer.TABLE 9. Distribution of samples belonging to score ranges.

[0141] To determine if age had an impact on classification performance, the study independently computed scores for individuals greater and lesser than 50 years of age The results demonstrate no significant impact of age (FIG. 15).

[0142] When binary classification results for the cancer samples are subdivided into early - stage (Stage I / II) and late-stage (Stage III / IV) cancer groups, the classifier still demonstrates high accuracy. It not only identifies late-stage cancer samples effectively but also classifies early-stage samples accurately. This holds true when considering both individual scores (FIG. 14C) and best scores (FIG. 14D). Using the best scores, the classifier's predictive accuracyreaches 98% for early-stage samples and 92.7% for late-stage cancers for score range 1.5-2 (TABLE 9)

[0143] Adoption of the proposed workflow in a clinical setting would enable women to undergo serum profiling at a clinic to predict their cancer status. For example, consider the scenarios presented in FIG. 16. Scenario (X) represents a situation in which an individual's serum profile falls within a score range that makes cancer highly unlikely. In such a case, the individual may only require yearly monitoring. In scenario (Z), the detection is more problematic due to a relatively small number of samples in this score range and a comparable number of cancerous and non-cancerous patients. In such cases, a patient may be referred for more additional and / or more frequent screening. Scenario (Y) depicts a situation where an individual's score lies in a range where a majority (94%) of patients has been diagnosed with cancer. In such a case, the patient would likely be referred for immediate advanced screening / treatment.

[0144] Discussion

[0145] Although the incidence of OC is relatively low (2.5% of all malignancies in women

[0015] ), it is among the most lethal of all cancers due to its high mortality rate. The reason for this is largely attributable to the fact that the disease is not typically diagnosed until the advanced (post-metastatic) stages of development (Stage III / IV), when effective treatment is difficult. For example, the most common sub-type of OC, serous papillary (65% of OC patients), is typically not diagnosed until Stage III / IV when the 5-year survival rate is only 31%. In contrast, if the disease is identified and treated early at less advanced stages of development (Stage I / II), the 5-year survival rate is 93%. These statistics dramatically underscore the dire need for an early diagnostic test for OC and other cancers where early-stage clinical symptoms are virtually non-existent.

[0146] The traditional approach for the identification of non-invasive biomarkers of cancer has been the screening of blood (or other body fluids) in search of significant changes in the presence / levels of molecules (typically proteins) associated with the disease

[0016] , A well- known example of such a diagnostic is the PSA (prostate-specific antigen) biomarker for prostate cancer

[0017] ,

[0147] The OC biomarker candidate, CA125 (a.k.a., mucin 16 / MUC16)

[0018] was first introduced in 1996

[0019] , Although an elevated level of CA125 is detected in ~90% of latestage (III / IV) OC patients, it is elevated in only ~50% of early-stage patients, making it a poor biomarker of early-stage disease with a PPV of only ~30%

[0020] , In 2003, a second candidatebiomarker for OC, HE4 (human epididymis protein), was introduced

[0021] , While HE4 was an improvement over CA125 in having a reported PPV of ~58%, it is still not sufficiently accurate to serve as a diagnostic test. Combining the results of the HE4 and CA125 together did not significantly improve PPV. However, the combination did lead to the development of a logistic regression model called ROMA (risk of malignancy algorithm) that was approved by the FDA in 2011 as a method to classify patients with a pelvic mass into those with high vs. low risk of having OC

[0020] ,

[0148] By the early 2000s, it was becoming progressively clear that, on the molecular level at least, cancer was a much more complex disease than originally envisioned [4], This realization was supported by findings indicating the existence of a multitude of disrupted molecular pathways (and underlying mutations) capable of leading to even the same cancer type. Such molecular level heterogeneity among individual cancer patients made the likelihood of identifying one or two biomarkers capable of accurately diagnosing all individuals with even the same type of cancer highly unlikely. As a consequence, the search for more accurate ways to diagnose cancer became focused on exploring larger combinations of biomarkers that might better capture the molecular heterogeneity underlying the disease

[0022] ,

[0023] ,

[0024] ,

[0025] ,

[0149] In the case of OC, there were a number of multi-biomarker diagnostic tests developed in the early 2000s

[0026] ,

[0027] ,

[0028] , However, none of these early efforts were sufficiently validated to acquire FDA approval. In 2009, an assay (trade name OVA1) was proposed that incorporated levels of five serum proteins combined with proprietary software to generate a high or low probability that an ovarian mass was a malignant tumor

[0029] , While the test was approved by the FDA as a clinical aid in determining if a patient should be referred for further analysis, the test's low PPV (31%)

[0030] eliminated it from consideration as an effective OC diagnostic. A more recent version of the OVA1 test (initially known as OVA2 but now trademarked as OVERA) uses a slightly different set of proteins upon which to generate its predictions. Although an improvement over OVA1, OVERA continues to be associated with a relatively low PPV (~40%)

[0031] , thereby again excluding it as a reliable OC diagnostic.

[0150] With the expanded availability of omics technologies and associated datasets (e.g., genomic, transcriptomic, proteomic, metabolomic) in recent years, a new approach to diagnostics began to emerge

[0032] , The application of various Al (artificial intelligence) approaches most notably machine learning, to the analysis of large omics datasets of diseased and non-diseased individuals opened the possibility of the identification of patterns by whichthese categories could be distinguished. Predictive models built upon such classifications might then constitute a new generation of diagnostic tests.

[0151] While this basic concept is straightforward, its application is certainly not. There are multiple approaches to ML, and each is associated with individual strengths and weaknesses

[0011]

[0012] , In addition, the output from ML analyses of omics datasets is heavily dependent upon both the quality and type of data being analyzed. For example, classifiers that are based exclusively on ML analysis of DNA sequence datasets may be appropriate if the onset and progression of the disease in question are exclusively attributable to genetic mutations. Certainly, there is a significant genetic component to cancer, but other environmental (e.g., diet, lifestyle, microbiome, etc.) and molecular (e.g., epigenetic / gene expression changes, gene-gene / protein-protein interactions, etc.) factors are also known to play significant roles. Indeed, it has been proposed that the accurate characterization of a complex disease like cancer will ultimately require simultaneous analyses of multi-omics datasets

[0033] , While this may well be the case, the development of computational methodologies sufficiently complex to accurately characterize multi-omics datasets is only in its infancy

[0034] ,

[0035] ,

[0152] In lieu of an approach that simultaneously analyzes multi-omics datasets, the study chose a currently available alternative, i.e., working with a dataset that reflects biological changes occurring on multiple levels. Metabolic profiles are widely viewed as a molecular phenotype reflective of underlying collective information encoded at the genome level and realized at the transcriptome and proteome levels. As such, metabolic profiles have long been considered promising indicators of cancer and other complex diseases [8], [9],

[0153] To help ensure the quality of the metabolic data, individual normal and OC patient samples were collected from four geographically divergent locations and analyzed using ultraperformance liquid chromatography coupled with tandem mass spectrometry (UPLC-MS / MS- positive and negative modes and each sample independently pre-processed through two columns) generating four distinct datasets (HN: HILIC negative; HP: HILIC positive; RN: Cis reversed phase negative; RP: Cis reversed phase positive). To guard against instrumental drift between runs, the same control samples were analyzed following every ten biological samples. Principle component analyses of data generated from the MS analyses across different batches and times demonstrated little experimental variation between runs.

[0154] Each of the four datasets was analyzed separately to determine if any particular dataset contained more relevant information than any other. The study found little difference in the accuracy of predictions computed using each dataset individually. When the study combined the best average score from each of the four datasets, an improvement in theclassification of both cancer and normal samples was observed. This suggests that each of the four datasets contribute uniquely to the accurate prediction of cancer status.

[0155] Computationally, the study evaluated the performance of five independent ML classifiers. A consensus classifier that generates average predictive probabilities from the probabilities of each of the individual classifiers gave the best overall performance with a PPV of 93%. While predictive accuracy of the consensus classifier was not significantly impacted by age, accuracy was found to be slightly better for early- relative to late-stage patients. Latestage patients displayed greater heterogeneity in molecular profiles than early-stage patients. While the reason for this dichotomy is currently unknown, the preliminary findings suggest that OCs may become more metabolically heterogeneous as they progress / metastasize.

[0156] In this regard, it may be relevant to note that the terms “early stage” vs. “late stage” when applied to OC do not necessarily imply a temporal progression. A more accurate terminology might be “pre-metastatic” vs. “post-metastatic”. Indeed, it is now widely accepted that ovarian cancers should properly be classified into those with low metastatic potential (Type 1) vs. those with high metastatic potential (Type 2)

[0036] , Thus, while OCs currently classified as “late-stage” are, by definition, “post-metastatic”, it is possible that at least some “late-stage” OCs may have arisen and progressed to an advanced stage of development more rapidly than at least some Type 1 “early-stage” samples. It is because of this ambiguity in classification that the study chose to develop the models using a mixture of early- and late-stage samples.

[0157] The model's accuracy in predicting women with OC is slightly greater than its accuracy in predicting women without the disease. The reason for this is currently unknown but may, at least in part, be due to the fact that the model may be detecting disease in women prior to clinical symptoms and clinical diagnosis. Time course studies are currently being instituted to test this hypothesis.

[0158] The high PPV (93%) associated with the consensus classifier supports the notion that ML analysis of omics data, and of metabolomic data in particular, is an extremely promising approach for the future diagnosis of ovarian and possibly other cancers as well. Such analyses will likely lead to a more probabilistic approach to cancer diagnosis that will serve to personalize the process much as genomic profiling of individual patient tumors is personalizing cancer treatment (i.e., precision cancer medicine).

[0159] Despite these highly favorable prospects, it is important to keep in mind the limitations of ML analyses of omics data. For example, the PPV associated with even the same ML based predictive model can be highly sensitive to the size and composition of the datasets employed in building and testing the models. For example, in an earlier pilot study of themetabolic profiles of a relatively small number (46) of OC patient samples collected from one of the same areas sampled in the current study (Northside Hospital, Atlanta), the authors generated a predictive model with a putative accuracy of 100%

[0037] , The relative reduction in accuracy associated with the current model relative to this earlier study coupled with the fact that none of the top ranked features in the earlier study ranked within the top 100 features in the current study underscores the impact of datasets on ML / metabolomic based predictive models. Future refinements in the development of metabolomic (and likely all omics) based ML models will need to address the issue of how many samples over what geographic area are needed to reflect the full spectrum of diversity in OC (and other cancer types).

[0160] To exemplify how the type of results generated in the study might, in the future, translate into a clinically useful tool, the study grouped the quantity and percentage of the samples into score ranges. This study envisions a clinical tool in which the scores of individual patients can be mapped across such a distribution providing a likelihood that an individual patient does or does not have cancer (FIG. 16). Such information could serve as a significant aid in determining the need for treatment or continued monitoring.

[0161] In summary, these results confirm the overall potential of an integrative approach using metabolomic profiles and ML-based classifiers for the detection of OC. The accuracy of these classifiers is highly dependent upon both the quality and quantity of the data upon which models are built. The study found little difference in the accuracy of predictions generated using alternative ML classifiers, although the consensus classifier generated the most accurate predictions. Application of results generated from the consensus classifier illustrated how the frequency distribution of individual patient scores can be used to develop a useful clinical tool that assigns a likelihood that an individual does or does not have OC. It is believed that this personalized / probabilistic approach to cancer diagnostics is more robust and clinically informative than the more traditional binary (yes / no) tests and may represent a promising new direction in the early detection of OC and perhaps other cancer types as well.

[0162] Example #2

[0163] A study also conducted the boosted tree regression method with features derived from mass spectrometry of four methods: UPLC-MS / MS in reversed (Cis) and HILIC (hil) phases using positive and negative ionization modes. The study used the maximal intensity value of all training samples (these values will be the same for all testing samples) of each method normalized to the intensity of each sample.

[0164] The latest version of the ovarian cancer diagnostic system (OCDX-2.0’s) include excluding possible FDA-approved drugs and the HILIC negative set (that does not perform aswell as other sets); partition the 412 set randomly into a training (212 samples) set and testing set (200 samples) other than the previous leave one out testing. These result in at 100% specificity, a sensitivity of 48 / 49 (98%) for early ovarian cancer (OV), a sensitivity of 77 / 87 (89%) for late stage OV, and an overall sensitivity of 125 / 136(91.9%).

[0165] The following patents, applications and publications as listed below and throughout this document are hereby incorporated by reference in their entirety herein.Reference List[1] D. Crosby et al., Early detection of cancer, Science. 375 (2022) eaay9040.[2] C. Stewart et al., Ovarian cancer: an integrated review, Semin. Oncol. Nurs. 35 (2019) 151—156.[3] J.D. Brooks, Translational genomics: the challenge of developing cancer biomarkers,Genome Res. 22 (2012) 183-187.[4] R.A. Burrell et al., The causes and consequences of genetic heterogeneity in cancer evolution, Nature. 501 (2013) 338-345.[5] J.T. Shreve et al., Artificial intelligence in oncology: current capabilities, future opportunities, and ethical considerations, Amer. Soc. Clin. Oncol. Edu. Book. 42 (2022) 842-851.[6] J.F. McDonald, Back to the future - the integration of big data with machine learning is reestablishing the importance of predictive correlations in ovarian cancer diagnostics and therapeutics, Gynecol. Oncol. 149 (2018) 230-231.[7] Y. Kumar et al., Artificial intelligence in disease diagnosis: a systematic literature review, synthesizing framework and future research agenda, J. Ambient. Intell. Humaniz. Comput. 14 (2023) 8459-8486.[8] C. Beuchel et al., Whole blood metabolite profiles reflect changes in energy metabolism in heart failure, Metabolites. 12 (2022) 216.[9] V.H. Telle-Hansen et al., A comprehensive metabolic profiling of the metabolically healthy obesity phenotype, Lipids Health Dis. 19 (2020) 90.

[0010] O.P. Trifonova et al., Metabolic profiling of human blood, Biochem. Moscow Suppl. Ser.B. 7 (2013) 179-186.

[0011] P.S. Reel et al., Using machine learning approaches for multi-omics data analysis: a review, Biotechnol. Adv. 49 (2021), 107739.

[0012] S. Albaradei et al., Machine learning and deep learning methods that use omics data for metastasis prediction, Comput. Struct. Biotechnol. J. 19 (2021) 5008-5018.

[0013] I. Guyon et al., Gene selection for cancer classification using support vector machines,Mach. Learn. 46 (2002) 389-422.

[0014] D. Chicco et al., The Matthews correlation coefficient (MCC) is more reliable than balanced accuracy, bookmaker informedness, and markedness in twoclass confusion matrix evaluation, BioData Min. 14 (2021) 13.

[0015] L.A. Torre, et al., Ovarian cancer statistics, 2018, CA Cancer J. Clin. 68 (2018) 284-296.

[0016] S.K. Chatterjee et al., Cancer biomarkers: knowing the present and predicting the future,Future Oncol. 1 (2005) 37-50.

[0017] A.R. Rao et al., The discovery of prostate-specific antigen, BJU Int. 101 (2008) 5-10.

[0018] E.P. Diamandis et al., Reflection on the discovery of carcinoembryonic antigen, prostatespecific antigen, and cancer antigens CA125 and CA 19-9, Clin. Chem. 59 (2013) 22- 31.

[0019] K. Nustad et al., Specificity and affinity of 26 monoclonal antibodies against the CA 125 antigen: first report from the ISOBM TD-1 workshop, Tumor Biol. 17 (2009) 196-219.

[0020] K. Holcomb et al., Human epididymis protein 4 offers superior specificity in the differentiation of benign and malignant adnexal masses in premenopausal women, Am. J. Obstet. Gynecol. 205 (358) (2011) el-e6.

[0021] I. Hellstrbm et al., The HE4 (WFDC2) protein is a biomarker for ovarian carcinoma,Cancer Res. 63 (2003) 3695-3700.

[0022] S.F. Shariat et al., Multiple biomarkers improve prediction of bladder cancer recurrence and mortality in patients undergoing cystectomy, Cancer. 112 (2008) 315-325.

[0023] K.A. Landers et al., Use of multiple biomarkers for a molecular diagnosis of prostate cancer, Int. J. Cancer 114 (2005) 950-956.

[0024] D.P. Malinowski, Multiple biomarkers in molecular oncology. I. Molecular diagnostics applications in cervical cancer detection, Expert. Rev. Mol. Diagn. 7 (2007) 117-131.

[0025] K.N. Kang et al., Multiple biomarkers are more accurate than a combination of carbohydrate antigen 125 and human epididymis protein 4 for ovarian cancer screening, Obstet. Gynecol. Sci. 65 (2022) 346-354.

[0026] E.F. Petricoin et al., Use of proteomic patterns in serum to identify ovarian cancer, Lancet359 (2002) 572-577.

[0027] G. Mor et al., Serum protein markers for early detection of ovarian cancer, Proc. Natl.Acad. Sci. U. S. A. 102 (2005) 7677-7682.

[0028] I. Visintin et al., Diagnostic markers for early detection of ovarian cancer, Clin. CancerRes. 14 (2008) 1065-1072.

[0029] F.R. Ueland et al., Effectiveness of a multivariate index assay in the preoperative assessment of ovarian tumors, Obstet. Gynecol. 117 (2011) 1289.

[0030] R.E. Bristow et al., Ovarian malignancy risk stratification of the adnexal mass using a multivariate index assay, Gynecol. Oncol. 128 (2013) 252-259.

[0031] R.R. Urban et al., Combined symptom index and second-generation multivariate biomarker test for prediction of ovarian cancer in patients with an adnexal mass, Gynecol. Oncol. 150 (2018) 318-323.

[0032] V.A. Hristova et al., Cancer biomarker discovery and translation: proteomics and beyond,Exp. Rev. Proteo. 16 (2019) 93-103.

[0033] M. Lu et al., The crucial role of multiomic approach in cancer research and clinically relevant outcomes, EPMA J. 9 (2018) 77-102.

[0034] G. Nicora et al., Integrated multi-omics analyses in oncology: a review of machine learning methods and tools, Front. Oncol. 10 (2020).

[0035] A. Dhillon et al., A systematic review on biomarker identification for cancer diagnosis and prognosis in multi-omics: from computational needs to machine learning and deep learning, Arch Computat. Methods Eng. 30 (2023) 917-949.

[0036] R.J. Kurman et al., The dualistic model of ovarian carcinogenesis revisited, revised, and expanded, Am. J. Pathol. 186 (2016) 733-747.

[0037] D.A. Gaul et al., Highly accurate metabolomic detection of early-stage ovarian cancer,Sci. Rep. 5 (2015) 16351.

[0038] S. Fan et al., Systematic error removal using random forest for normalizing large-scale untargeted lipidomics data. Anal. Chem. 91, 3590-3596 (2019).

[0039] P. Geurts et al., Extremely randomized trees. Mach. Learn. 63, 3-42 (2006).

[0040] F. Pedregosa et al., Scikit-leam: Machine Learning in Python. J. Mach. Learn. Res. 12,2825-2830 (2011).

[0041] D.S. Wishart et al., HMDB 5.0: The human metabolome database for 2022. Nucleic AcidsRes. 50, D622-D631 (2022).

[0042] D. Bank, Autoencoders, arXiv:2003.05991v2 Apr. 03, 2021.

[0043] Y.N. Kunang et al., Automatic features extraction using autoencoder in intrusion detection system. International Conference on Electrical Engineering and Computer Science. 2018, 219-224 (2018).

[0044] L. Macias-Garcia et al., Autoencoded DNA methylation data to predict breast cancer recurrence: Machine learning models and gene-weight significance. Artif. Intell. Med. 110, 101976 (2020).

[0045] M. Abadi et al., TensorFlow: Large-Scale machine learning on heterogeneous distributed systems, Zenodo, May 30, 2023.

[0046] L. Breiman, Random Forests. Mach. Learn. 45, 5-32 (2001).

[0047] T. Hastie et al., The Elements of Statistical Learning: Data Mining, Inference, andPrediction, Second ed., Springer-Verlag, 2009.

[0048] D.-A. Clevert et al., Fast and Accurate Deep Network Learning by Exponential LinearUnits (ELUs). arXiv, Feb. 22, 2016.

[0049] D.P. Kingma et al., Adam: A method for stochastic optimization. arXiv, Jan. 29, 2017.

[0050] R. Polikar, Ensemble based systems in decision making. IEEE Circuits and Systems Mag.6, 21-45 (2006).

[0051] R.-E. Fan et al., LIBLINEAR: A library for large linear classification. J. Mach. Learn.Res. 9, 1871-1874 (2008).

[0051] C.-C. Chang et al., LIBSVM: A library for support vector machines. ACM Trans. Intell. Syst. Technol. 2, 1-27 (2011).

[0053] J. Goldberger et al., Neighbourhood components analysis, in advances in neural information processing systems, MIT Press (2004).

[0054] Y. Freund et al., A decision-theoretic generalization of on-line learning and an application to boosting. J. Comput. Syst. Sci. 55, 119-139 (1997).

[0055] J. Platt, Probabilistic Outputs for Support Vector Machines and Comparisons toRegularized Likelihood Methods, A. J. Smola, P. Bartlett, B. Scholkopf, D. Schuurmans, Eds. In: Advances in Large Margin Classifiers, (MIT Press, Boston, 1999), pp. 61-74.

[0056] U.S. Patent US11,555,818

[0057] U.S. Patent US10393745

[0057] PCT WO 2009 / 151967A1

[0058] EP Application EP 4343775A1

[0059] US. Published Patent Application No. US20220228102A1

Claims

What is claimed is:

1. A method comprising: providing, by a processor, a multi-omics data set comprising a plurality of metabolomic profile data acquired from a liquid chromatography-mass spectrometry analysis of an individual (e.g., ultra-performance liquid chromatography, high-resolution mass spectrometry (UPLC-MS) analysis or ultra-performance liquid chromatography coupled with tandem mass spectrometry (UPLC-MS / MS)); determining, by one or more trained Al models, using the set of metabolomic profile data, one or more probability values that the individual does have ovarian cancer (and a probability value that the individual does not have ovarian cancer), wherein the one or more trained Al models are configured to provide a comprehensive assessment of a presence of ovarian cancer including serous papillary; and outputting, by the processor, via a report or a graphical user interface, the one or more probability values, or a value derived therefrom that an individual does have ovarian cancer, wherein the output provided in the report or graphical user interface is subsequently used by a clinician to (i) determine a diagnosis of ovarian cancer and / or (ii) direct treatment or screening of ovarian cancer.

2. The method of claim 1, wherein the liquid chromatography -mass spectrometry analysis includes features from three or more of a HILIC positive analysis, a HILIC negative analysis, a Cl 8 reversed phase negative analysis, and a Cl 8 reversed phase positive analysis.

3. The method of claim 2, wherein the one or more trained Al models include at least one trained Al model for each of the three or more of the HILIC positive analysis, the HILIC negative analysis, the Cl 8 reversed phase negative analysis, and the Cl 8 reversed phase positive analysis.

4. The method of any one of claims 1-3, wherein the one or more trained Al models, collectively, are configured to use at least 10 features, wherein at least 50% of the 10 features are from FEATURE ID NOs: 1-102.

5. The method of any one of claims 1-3, wherein the one or more trained Al models, collectively, are configured to use at least 100 features, wherein 50% of the 100 features are from FEATURE ID NOs: 1-102.

6. The method of any one of claims 1-3, wherein the one or more trained Al models, collectively, are configured to use at least 1,000 features, wherein the at least 1,000 features include at least 50% of FEATURE ID NOs: 1-102.

7. The method of any one of claims 1-3, wherein the one or more trained Al models, collectively, are configured to use at least 10,000 features, wherein the at least 10,000 features include at least 50% of FEATURE ID NOs: 1-102.

8. The method of any one of claims 1-7, wherein the one or more trained Al models, collectively, are configured to provide a comprehensive assessment of ovarian cancer including an early indicator of pre-metastatic ovarian cancer and an indicator of post- metastatic ovarian cancer.

9. The method of any one of claims 1-7, wherein the one or more trained Al models are configured to provide a comprehensive assessment of ovarian cancer including an early indicator of pre-metastatic ovarian cancer.

10. The method of any one of claims 1-9 further comprising: determining, by the trained Al model, a second probability value that the individual does not have ovarian cancer; and outputting, by the processor, via the report or the graphical user interface, the second probability value that the individual does not have ovarian cancer, wherein the probability value and the second probability value are both used in the diagnosis of ovarian cancer.

11. The method of any one of claims 1-10 further comprising: generating a binary output from the probability value by applying the probability value to a threshold value.

12. The method of any one of claims 1-11, wherein the one or more trained Al models include at least a first trained Al model and a second trained Al model, wherein the first trained Al model and the second Al model respectively include at least one of logistic regression classifier (LRC), random forest classifier (RFC), support vector machine (SVM), k-nearest neighbor (KNN), and adaptive boosting (ADA).

13. The method of claim 12, wherein the trained Al model includes a consensus classifier that combines probability outputs of the first trained Al model and the second trained Al model to generate the one or more probability values.

14. The method of claim 13, wherein the probability output assigned by the consensus classifier provides a background distribution of probabilities that a given sample was cancer or normal.

15. The method of any one of claims 1 - 14, comprising: determining a range of a set of ranges for the one or more probability values, or the value derived therefrom, wherein the set of ranges includes (i) a highly unlikely range, (ii) a second range in which a number of samples in the score range has a comparable number of cancerous and non-cancerous patients and (iii) a majority of training patients has been diagnosed with cancer, wherein the range informs a guideline for subsequent testing and testing frequency.

16. The method of any one of claims 1 - 14, comprising: presenting, via the report or the graphical user interface, a range of a set of ranges for the one or more probability values, or the value derived therefrom, wherein the set of ranges includes (i) a highly unlikely range, (ii) a second range in which a number of samples in the score range has a comparable number of cancerous and non-cancerous patients and (iii) a majority of training patients has been diagnosed with cancer, wherein the range informs a guideline for subsequent testing and testing frequency.

17. The method of claim 1, wherein the liquid chromatography -mass spectrometry analysis includes features from at least one of a HILIC positive analysis, a HILIC negative analysis, a Cl 8 reversed phase negative analysis, and a Cl 8 reversed phase positive analysis.

18. A system comprising: a processor; and a memory having instructions stored thereon, wherein execution of the instructions by the processor causes the processor to execute any one of the methods of claims 1-17.

19. A non-transitory computer-readable medium having instructions stored thereon, wherein execution of the instructions by a processor causes the processor to execute any one of the methods of claims 1-17.

Images