Patent Application
20130217057
Lung conditions and particularly lung cancer present significant diagnostic challenges. In many asymptomatic patients, radiological screens such as computed tomography (CT) scanning are a first step in the diagnostic paradigm. Pulmonary nodules (PNs) or indeterminate nodules are located in the lung and are often discovered during screening of both high risk patients or incidentally. The number of PNs identified is expected to rise due to increased numbers of patients with access to health care, the rapid adoption of screening techniques and an aging population. It is estimated that over 3 million PNs are identified annually in the US. Although the majority of PNs are benign, some are malignant leading to additional interventions. For patients considered low risk for malignant nodules, current medical practice dictates scans every three to six months for at least two years to monitor for lung cancer. The time period between identification of a PN and diagnosis is a time of medical surveillance or “watchful waiting” and may induce stress on the patient and lead to significant risk and expense due to repeated imaging studies. If a biopsy is performed on a patient who is found to have a benign nodule, the costs and potential for harm to the patient increase unnecessarily. Major surgery is indicated in order to excise a specimen for tissue biopsy and diagnosis. All of these procedures are associated with risk to the patient including: illness, injury and death as well as high economic costs.
Frequently, PNs cannot be biopsied to determine if they are benign or malignant due to their size and/or location in the lung. However, PNs are connected to the circulatory system, and so if malignant, protein markers of cancer can enter the blood and provide a signal for determining if a PN is malignant or not.
Diagnostic methods that can replace or complement current diagnostic methods for patients presenting with PNs are needed to improve diagnostics, reduce costs and minimize invasive procedures and complications to patients. The present invention provides novel compositions, methods and kits for identifying protein markers to identify, diagnose, classify and monitor lung conditions, and particularly lung cancer. The present invention uses a blood-based multiplexed assay to distinguish benign pulmonary nodules from malignant pulmonary nodules to classify patients with or without lung cancer. The present invention may be used in patients who present with symptoms of lung cancer, but do not have pulmonary nodules.
The present invention provides a method of determining the likelihood that a lung condition in a subject is cancer by measuring an abundance of a panel of proteins in a sample obtained from the subject; calculating a probability of cancer score based on the protein measurements and ruling out cancer for the subject if the score) is lower than a pre-determined score, wherein When cancer is ruled out the subject does not receive a treatment protocol. Treatment protocols include for example pulmonary function test (PFT), pulmonary imaging, a biopsy, a surgery, a chemotherapy, a radiotherapy, or any combination thereof. In some embodiments, the imaging is an x-ray, a chest computed tomography (CT) scan, or a positron emission tomography (PET) scan.
The present invention further provides a method of ruling in the likelihood of cancer for a subject by measuring an abundance of panel of proteins in a sample obtained from the subject, calculating a probability of cancer score based on the protein measurements and ruling in the likelihood of cancer for the subject if the score in step is higher than a pre-determined score
In another aspect, the invention further provides a method of determining the likelihood of the presence of a lung condition in a subject by measuring an abundance of panel of proteins in a sample obtained from the subject, calculating a probability of cancer score based on the protein measurements and concluding the presence of said lung condition if the score is equal or greater than a pre-determined score. The lung condition is lung cancer such as for example, non-small cell lung cancer (NSCLC). The subject at risk of developing lung cancer
The panel includes at least 4 proteins selected from ALDOA, FRIL, LG3BP, IBP3, LRP1, ISLR, TSP COIA1, GRP78, TETN, PRXD1 and CD14. Optionally, the panel further includes at least one protein selected from BGH3, COIA1, TETN, GRP78, PRDX, FIBA and GSLG1.
The subject has or is suspected of having a pulmonary nodule. The pulmonary nodule has a diameter of less than or equal to 3 cm. In one embodiment, the pulmonary nodule has a diameter of about 0.8 cm to 2.0 cm.
The score is calculated from a logistic regression model applied to the protein measurements. For example, the score is determined as Ps=1/[1+exp(−α−Σi=1Nβi*{hacek over (I)}i,s)], where {hacek over (I)}i,s is logarithmically transformed and normalized intensity of transition i in said sample (s), βi is the corresponding logistic regression coefficient, α was a panel-specific constant, and N was the total number of transitions in said panel.
In various embodiments, the method of the present invention further comprises normalizing the protein measurements. For example, the protein measurements are normalized by one or more proteins selected from PEDF, MASP1, GELS, LUM, C163A and PTPRJ.
The biological sample such as for example tissue, blood, plasma, serum, whole blood, urine, saliva, genital secretion, cerebrospinal fluid, sweat and excreta.
In one aspect, the determining the likelihood of cancer is determined by the sensitivity, specificity, negative predictive value or positive predictive value associated with the score. The score determined has a negative predictive value (NPV) is at least about 80%.
The measuring step is performed by selected reaction monitoring mass spectrometry, using a compound that specifically binds the protein being detected or a peptide transition. In one embodiment, the compound that specifically binds to the protein being measured is an antibody or an aptamer.
The disclosed invention derives from the surprising discovery, that in patients presenting with pulmonary nodule(s), protein markers in the blood exist that specifically identify and classify lung cancer. Accordingly the invention provides unique advantages to the patient associated with early detection of lung cancer in a patient, including increased life span, decreased morbidity and mortality, decreased exposure to radiation during screening and repeat screenings and a minimally invasive diagnostic model. Importantly, the methods of the invention allow for a patient to avoid invasive procedures.
The routine clinical use of chest computed tomography (CT) scans identifies millions of pulmonary nodules annually, of which only a small minority are malignant but contribute to the dismal 15% five-year survival rate for patients diagnosed with non-small cell lung cancer (NSCLC). The early diagnosis of lung cancer in patients with pulmonary nodules is a top priority, as decision-making based on clinical presentation, in conjunction with current non-invasive diagnostic options such as chest CT and positron emission tomography (PET) scans, and other invasive alternatives, has not altered the clinical outcomes of patients with Stage I NSCLC. The subgroup of pulmonary nodules between 8 mm and 20 mm in size is increasingly recognized as being “intermediate” relative to the lower rate of malignancies below 8 mm and the higher rate of malignancies above 20 mm [9]. Invasive sampling of the lung nodule by biopsy using transthoracic needle aspiration or bronchoscopy may provide a cytopathologic diagnosis of NSCLC, but are also associated with both false-negative and non-diagnostic results. In summary, a key unmet clinical need for the management of pulmonary nodules is a non-invasive diagnostic test that discriminates between malignant and benign processes in patients with indeterminate pulmonary nodules (IPNs), especially between 8 mm and 20 mm in size.
The clinical decision to be more or less aggressive in treatment is based on risk factors, primarily nodule size, smoking history and age [9] in addition to imaging. As these are not conclusive, there is a great need for a molecular-based blood test that would be both non-invasive and provide complementary information to risk factors and imaging.
Accordingly, these and related embodiments will find uses in screening methods for lung conditions, and particularly lung cancer diagnostics. More importantly, the invention finds use in determining the clinical manangement of a patient. That is, the method of invention are useful in ruling in or ruling out a particular treatment protocol for an individual subject.
Cancer biology requires a molecular strategy to address the unmet medical need for an assessment of lung cancer risk. The field of diagnostic medicine has evolved with technology and assays that provide sensitive mechanisms for detection of changes in proteins. The methods described herein use a LC-SRM-MS technology for measuring the concentration of blood plasma proteins that are collectively changed in patients with a malignant PN. This protein signature is indicative of lung cancer. LC-SRM-MS is one method that provides for both quantification and identification of circulating proteins in plasma. Changes in protein expression levels, such as but not limited to signaling factors, growth factors, cleaved surface proteins and secreted proteins, can be detected using such a sensitive technology to assay cancer. Presented herein is a blood-based classification test to determine the likelihood that a patient presenting with a pulmonary nodule has a nodule that is benign or malignant. The present invention presents a classification algorithm that predicts the relative likelihood of the PN being benign or malignant.
More broadly, it is demonstrated that there are many variations on this invention that are also diagnostic tests for the likelihood that a PN is benign or malignant. These are variations on the panel of proteins, protein standards, measurement methodology and/or classification algorithm.
As disclosed herein, archival plasma samples from subjects presenting with PNs were analyzed for differential protein expression by mass spectrometry and the results were used to identify biomarker proteins and panels of biomarker proteins that are differentially expressed in conjunction with various lung conditions (cancer vs. non-cancer).
In one aspect of the invention, one hundred and sixty three panels were discovered that allow for the classification of PN as being benign or malignant. These panels include those listed on Table 1. In some embodiments the panel according to the invention includes measuring 1, 2, 3, 4, 5 or more proteins selected from ISLR, ALDOA, KIT, GRP78, AIFM1, CD14, COIA1, IBP3, TSP1, BGH3, TETN, FR1, LG3BP, GGH, PRDX1 or LRP1. In other embodiments the panel includes any panel or protein exemplified on Table 1. For, example the panel includes ALDOA, GRP78, CD14, COIA1, IBP3, FRIL, LG3BP, and LRP1
The one hundred best random panels of proteins out of the million generated are shown in Table 2.
Preferred panels for ruling in treatment for a subject include the panels listed on Table 3 and 4. In various other embodiments, the panels according to the invention include measuring at least 2, 3, 4, 5, 6, 7, or more of the proteins listed on Tables 2 and 3.
A preferred normalizer panel is listed in Table 5.
The term “pulmonary nodules” (PNs) refers to lung lesions that can be visualized by radiographic techniques. A pulmonary nodule is any nodules less than or equal to three centimeters in diameter. In one example a pulmonary nodule has a diameter of about 0.8 cm to 2 cm.
The term “masses” or “pulmonary masses” refers to lung nodules that are greater than three centimeters maximal diameter.
The term “blood biopsy” refers to a diagnostic study of the blood to determine whether a patient presenting with a nodule has a condition that may be classified as either benign or malignant.
The term “acceptance criteria” refers to the set of criteria to which an assay, test, diagnostic or product should conform to be considered acceptable for its intended use. As used herein, acceptance criteria are a list of tests, references to analytical procedures, and appropriate measures, which are defined for an assay or product that will be used in a diagnostic. For example, the acceptance criteria for the classifier refers to a set of predetermined ranges of coefficients.
The term “average maximal AUC” refers to the methodology of calculating performance. For the present invention, in the process of defining the set of proteins that should be in a panel by forward or backwards selection proteins are removed or added one at a time. A plot can be generated with performance (AUC or partial AUC score on the Y axis and proteins on the X axis) the point which maximizes performance indicates the number and set of proteins the gives the best result.
The term “partial AUC factor or pAUC factor” is greater than expected by random prediction. At sensitivity=0.90 the pAUC factor is the trapezoidal area under the ROC curve from 0.9 to 1.0 Specificity/(0.1*0.1/2).
The term “incremental information” refers to information that may be used with other diagnostic information to enhance diagnostic accuracy. Incremental information is independent of clinical factors such as including nodule size, age, or gender.
The term “score” or “scoring” refers to the refers to calculating a probability likelihood for a sample. For the present invention, values closer to 1.0 are used to represent the likelihood that a sample is cancer, values closer to 0.0 represent the likelihood that a sample is benign.
The term “robust” refers to a test or procedure that is not seriously disturbed by violations of the assumptions on which it is based. For the present invention, a robust test is a test wherein the proteins or transitions of the mass spectrometry chromatograms have been manually reviewed and are “generally” free of interfering signals
The term “coefficients” refers to the weight assigned to each protein used to in the logistic regression equation to score a sample.
In certain embodiments of the invention, it is contemplated that in terms of the logistic regression model of MC CV, the model coefficient and the coefficient of variation (CV) of each protein's model coefficient may increase or decrease, dependent upon the method (or model) of measurement of the protein classifier. For each of the listed proteins in the panels, there is about, at least, at least about, or at most about a 2-, 3-, 4-, 5-, 6-, 7-, 8-, 9-, or 10-, -fold or any range derivable therein for each of the coefficient and CV. Alternatively, it is contemplated that quantitative embodiments of the invention may be discussed in terms of as about, at least, at least about, or at most about 10, 20, 30, 40, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99% or more, or any range derivable therein.
The term “best team players” refers to the proteins that rank the best in the random panel selection algorithm, i.e., perform well on panels. When combined into a classifier these proteins can segregate cancer from benign samples. “Best team player” proteins is synonymous with “cooperative proteins”. The term “cooperative proteins” refers proteins that appear more frequently on high performing panels of proteins than expected by chance. This gives rise to a protein's cooperative score which measures how (in)frequently it appears on high performing panels. For example, a protein with a cooperative score of 1.5 appears on high performing panels 1.5× more than would be expected by chance alone.
The term “classifying” as used herein with regard to a lung condition refers to the act of compiling and analyzing expression data for using statistical techniques to provide a classification to aid in diagnosis of a lung condition, particularly lung cancer.
The term “classifier” as used herein refers to an algorithm that discriminates between disease states with a predetermined level of statistical significance. A two-class classifier is an algorithm that uses data points from measurements from a sample and classifies the data into one of two groups. In certain embodiments, the data used in the classifier is the relative expression of proteins in a biological sample. Protein expression levels in a subject can be compared to levels in patients previously diagnosed as disease free or with a specified condition.
The “classifier” maximizes the probability of distinguishing a randomly selected cancer sample from a randomly selected benign sample, i.e., the AUC of ROC curve.
In addition to the classifier's constituent proteins with differential expression, it may also include proteins with minimal or no biologic variation to enable assessment of variability, or the lack thereof, within or between clinical specimens; these proteins may be termed endogenous proteins and serve as internal controls for the other classifier proteins.
The term “normalization” or “normalizer” as used herein refers to the expression of a differential value in terms of a standard value to adjust for effects which arise from technical variation due to sample handling, sample preparation and mass spectrometry measurement rather than biological variation of protein concentration in a sample. For example, when measuring the expression of a differentially expressed protein, the absolute value for the expression of the protein can be expressed in terms of an absolute value for the expression of a standard protein that is substantially constant in expression. This prevents the technical variation of sample preparation and mass spectrometry measurement from impeding the measurement of protein concentration levels in the sample.
The term “condition” as used herein refers generally to a disease, event, or change in health status.
The term “treatment protocol” as used herein including further diagnostic testing typically performed to determine whether a pulmonary nodule is benign or malignant. Treatment protocols include diagnostic tests typically used to diagnose pulmonary nodules or masses such as for example, CT scan, positron emission tomography (PET) scan, bronchoscopy or tissue biopsy. Treatment protocol as used herein is also meant to include therapeutic treatments typically used to treat malignant pulmonary nodules and/or lung cancer such as for example, chemotherapy, radiation or surgery.
The terms “diagnosis” and “diagnostics” also encompass the terms “prognosis” and “prognostics”, respectively, as well as the applications of such procedures over two or more time points to monitor the diagnosis and/or prognosis over time, and statistical modeling based thereupon. Furthermore the term diagnosis includes: a. prediction (determining if a patient will likely develop a hyperproliferative disease) b. prognosis (predicting whether a patient will likely have a better or worse outcome at a pre-selected time in the future) c. therapy selection d. therapeutic drug monitoring e. relapse monitoring.
In some embodiments, for example, classification of a biological sample as being derived from a subject with a lung condition may refer to the results and related reports generated by a laboratory, while diagnosis may refer to the act of a medical professional in using the classification to identify or verify the lung condition.
The term “providing” as used herein with regard to a biological sample refers to directly or indirectly obtaining the biological sample from a subject. For example, “providing” may refer to the act of directly obtaining the biological sample from a subject (e.g., by a blood draw, tissue biopsy, lavage and the like). Likewise, “providing” may refer to the act of indirectly obtaining the biological sample. For example, providing may refer to the act of a laboratory receiving the sample from the party that directly obtained the sample, or to the act of obtaining the sample from an archive.
As used herein, “lung cancer” preferably refers to cancers of the lung, but may include any disease or other disorder of the respiratory system of a human or other mammal. Respiratory neoplastic disorders include, for example small cell carcinoma or small cell lung cancer (SCLC), non-small cell carcinoma or non-small cell lung cancer (NSCLC), squamous cell carcinoma, adenocarcinoma, broncho-alveolar carcinoma, mixed pulmonary carcinoma, malignant pleural mesothelioma, undifferentiated large cell carcinoma, giant cell carcinoma, synchronous tumors, large cell neuroendocrine carcinoma, adenosquamous carcinoma, undifferentiated carcinoma; and small cell carcinoma, including oat cell cancer, mixed small cell/large cell carcinoma, and combined small cell carcinoma; as well as adenoid cystic carcinoma, hamartomas, mucoepidermoid tumors, typical carcinoid lung tumors, atypical carcinoid lung tumors, peripheral carcinoid lung tumors, central carcinoid lung tumors, pleural mesotheliomas, and undifferentiated pulmonary carcinoma and cancers that originate outside the lungs such as secondary cancers that have metastasized to the lungs from other parts of the body. Lung cancers may be of any stage or grade. Preferably the term may be used to refer collectively to any dysplasia, hyperplasia, neoplasia, or metastasis in which the protein biomarkers expressed above normal levels as may be determined, for example, by comparison to adjacent healthy tissue.
Examples of non-cancerous lung condition include chronic obstructive pulmonary disease (COPD), benign tumors or masses of cells (e.g., hamartoma, fibroma, neurofibroma), granuloma, sarcoidosis, and infections caused by bacterial (e.g., tuberculosis) or fungal (e.g. histoplasmosis) pathogens. In certain embodiments, a lung condition may be associated with the appearance of radiographic PNs.
As used herein, “lung tissue”, and “lung cancer” refer to tissue or cancer, respectively, of the lungs themselves, as well as the tissue adjacent to and/or within the strata underlying the lungs and supporting structures such as the pleura, intercostal muscles, ribs, and other elements of the respiratory system. The respiratory system itself is taken in this context as representing nasal cavity, sinuses, pharynx, larynx, trachea, bronchi, lungs, lung lobes, aveoli, aveolar ducts, aveolar sacs, aveolar capillaries, bronchioles, respiratory bronchioles, visceral pleura, parietal pleura, pleural cavity, diaphragm, epiglottis, adenoids, tonsils, mouth and tongue, and the like. The tissue or cancer may be from a mammal and is preferably from a human, although monkeys, apes, cats, dogs, cows, horses and rabbits are within the scope of the present invention. The term “lung condition” as used herein refers to a disease, event, or change in health status relating to the lung, including for example lung cancer and various non-cancerous conditions.
“Accuracy” refers to the degree of conformity of a measured or calculated quantity (a test reported value) to its actual (or true) value. Clinical accuracy relates to the proportion of true outcomes (true positives (TP) or true negatives (TN) versus misclassified outcomes (false positives (FP) or false negatives (FN)), and may be stated as a sensitivity, specificity, positive predictive values (PPV) or negative predictive values (NPV), or as a likelihood, odds ratio, among other measures.
The term “biological sample” as used herein refers to any sample of biological origin potentially containing one or more biomarker proteins. Examples of biological samples include tissue, organs, or bodily fluids such as whole blood, plasma, serum, tissue, lavage or any other specimen used for detection of disease.
The term “subject” as used herein refers to a mammal, preferably a human.
The term “biomarker protein” as used herein refers to a polypeptide in a biological sample from a subject with a lung condition versus a biological sample from a control subject. A biomarker protein includes not only the polypeptide itself, but also minor variations thereof, including for example one or more amino acid substitutions or modifications such as glycosylation or phosphorylation.
The term “biomarker protein panel” as used herein refers to a plurality of biomarker proteins. In certain embodiments, the expression levels of the proteins in the panels can be correlated with the existence of a lung condition in a subject. In certain embodiments, biomarker protein panels comprise 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 60, 70, 80, 90 or 100 proteins. In certain embodiments, the biomarker proteins panels comprise from 100-125 proteins, 125-150 proteins, 150-200 proteins or more.
“Treating” or “treatment” as used herein with regard to a condition may refer to preventing the condition, slowing the onset or rate of development of the condition, reducing the risk of developing the condition, preventing or delaying the development of symptoms associated with the condition, reducing or ending symptoms associated with the condition, generating a complete or partial regression of the condition, or some combination thereof.
The term “ruling out” as used herein is meant that the subject is selected not to receive a treatment protocol.
The term “ruling-in” as used herein is meant that the subject is selected to receive a treatment protocol.
Biomarker levels may change due to treatment of the disease. The changes in biomarker levels may be measured by the present invention. Changes in biomarker levels may be used to monitor the progression of disease or therapy.
“Altered”, “changed” or “significantly different” refer to a detectable change or difference from a reasonably comparable state, profile, measurement, or the like. One skilled in the art should be able to determine a reasonable measurable change. Such changes may be all or none. They may be incremental and need not be linear. They may be by orders of magnitude. A change may be an increase or decrease by 1%, 5%, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, 95%, 99%, 100%, or more, or any value in between 0% and 100%. Alternatively the change may be 1-fold, 1.5-fold 2-fold, 3-fold, 4-fold, 5-fold or more, or any values in between 1-fold and five-fold. The change may be statistically significant with a p value of 0.1, 0.05, 0.001, or 0.0001.
Using the methods of the current invention, a clinical assessment of a patient is first performed. If there exists is a higher likelihood for cancer, the clinician may rule in the disease which will require the pursuit of diagnostic testing options yielding data which increase and/or substantiate the likelihood of the diagnosis. “Rule in” of a disease requires a test with a high specificity.
“FN” is false negative, which for a disease state test means classifying a disease subject incorrectly as non-disease or normal.
“FP” is false positive, which for a disease state test means classifying a normal subject incorrectly as having disease.
The term “rule in” refers to a diagnostic test with high specificity that coupled with a clinical assessment indicates a higher likelihood for cancer. If the clinical assessment is a lower likelihood for cancer, the clinician may adopt a stance to rule out the disease, which will require diagnostic tests which yield data that decrease the likelihood of the diagnosis. “Rule out” requires a test with a high sensitivity.
The term “rule out” refers to a diagnostic test with high sensitivity that coupled with a clinical assessment indicates a lower likelihood for cancer.
The term “sensitivity of a test” refers to the probability that a patient with the disease will have a positive test result. This is derived from the number of patients with the disease who have a positive test result (true positive) divided by the total number of patients with the disease, including those with true positive results and those patients with the disease who have a negative result, i.e. false negative.
The term “specificity of a test” refers to the probability that a patient without the disease will have a negative test result. This is derived from the number of patients without the disease who have a negative test result (true negative) divided by all patients without the disease, including those with a true negative result and those patients without the disease who have a positive test result, e.g. false positive. While the sensitivity, specificity, true or false positive rate, and true or false negative rate of a test provide an indication of a test's performance, e.g. relative to other tests, to make a clinical decision for an individual patient based on the test's result, the clinician requires performance parameters of the test with respect to a given population.
The term “positive predictive value” (PPV) refers to the probability that a positive result correctly identifies a patient who has the disease, which is the number of true positives divided by the sum of true positives and false positives.
The term “negative predictive value” or “NPV” is calculated by TN/(TN+FN) or the true negative fraction of all negative test results. It also is inherently impacted by the prevalence of the disease and pre-test probability of the population intended to be tested.
The term “disease prevalence” refers to the number of all new and old cases of a disease or occurrences of an event during a particular period. Prevalence is expressed as a ratio in which the number of events is the numerator and the population at risk is the denominator.
The term disease incidence refers to a measure of the risk of developing some new condition within a specified period of time; the number of new cases during some time period, it is better expressed as a proportion or a rate with a denominator.
Lung cancer risk according to the “National Lung Screening Trial” is classified by age and smoking history. High risk—age≧55 and ≧30 pack-years smoking history; Moderate risk—age≧50 and ≧20 pack-years smoking history; Low risk—≦age 50 or ≦20 pack-years smoking history.
The term “negative predictive value” (NPV) refers to the probability that a negative test correctly identifies a patient without the disease, which is the number of true negatives divided by the sum of true negatives and false negatives. A positive result from a test with a sufficient PPV can be used to rule in the disease for a patient, while a negative result from a test with a sufficient NPV can be used to rule out the disease, if the disease prevalence for the given population, of which the patient can be considered a part, is known.
The clinician must decide on using a diagnostic test based on its intrinsic performance parameters, including sensitivity and specificity, and on its extrinsic performance parameters, such as positive predictive value and negative predictive value, which depend upon the disease's prevalence in a given population.
Additional parameters which may influence clinical assessment of disease likelihood include the prior frequency and closeness of a patient to a known agent, e.g. exposure risk, that directly or indirectly is associated with disease causation, e.g. second hand smoke, radiation, etc., and also the radiographic appearance or characterization of the pulmonary nodule exclusive of size. A nodule's description may include solid, semi-solid or ground glass which characterizes it based on the spectrum of relative gray scale density employed by the CT scan technology.
“Mass spectrometry” refers to a method comprising employing an ionization source to generate gas phase ions from an analyte presented on a sample presenting surface of a probe and detecting the gas phase ions with a mass spectrometer.
The technology liquid chromatography selected reaction monitoring mass spectrometry (LC-SRM-MS) was used to assay the expression levels of a cohort of 388 proteins in the blood to identify differences for individual proteins which may correlate with the absence or presence of the disease. The individual proteins have not only been implicated in lung cancer biology, but are also likely to be present in plasma based on their expression as membrane-anchored or secreted proteins. An analysis of epithelial and endothelial membranes of resected lung cancer tissues (including the subtypes of adenocarcinoma, squamous, and large cell) identified 217 tissue proteins. A review of the scientific literature with search terms relevant to lung cancer biology identified 319 proteins. There was an overlap of 148 proteins between proteins identified by cancer tissue analysis or literature review, yielding a total of 388 unique proteins as candidates. The majority of candidate proteins included in the multiplex LC-SRM-MS assay were discovered following proteomics analysis of secretory vesicle contents from fresh NSCLC resections and from adjacent non-malignant tissue. The secretory proteins reproducibly upregulated in the tumor tissue were identified and prioritized for inclusion in the LC-SRM-MS assay using extensive bioinformatic and literature annotation. An additional set of proteins that were present in relevant literature was also added to the assay. In total, 388 proteins associated with lung cancer were prioritized for SRM assay development. Of these, 371 candidate protein biomarkers were ultimately included in the assay. These are listed in Table 6, below.
190 of these candidate protein biomarkers were shown to be measured reproducibly in blood. A moderately powered multisite and unbiased study of 242 blood samples from patients with PN was designed to determine whether a statistically significant subpanel of proteins could be identified to distinguish benign and malignant nodules of sizes under 2 cm. The three sites contributing samples and clinical data to this study were the University of Laval, University of Pennsylvania and New York University.
In an embodiment of the invention, a panel of 15 proteins effectively distinguished between samples derived from patients with benign and malignant nodules less than 2 cm diameter.
Bioinformatic and biostatistical analyses were used first to identify individual proteins with statistically significant differential expression, and then using these proteins to derive one or more combinations of proteins or panels of proteins, which collectively demonstrated superior discriminatory performance compared to any individual protein. Bioinformatic and biostatistical methods are used to derive coefficients (C) for each individual protein in the panel that reflects its relative expression level, i.e. increased or decreased, and its weight or importance with respect to the panel's net discriminatory ability, relative to the other proteins. The quantitative discriminatory ability of the panel can be expressed as a mathematical algorithm with a term for each of its constituent proteins being the product of its coefficient and the protein's plasma expression level (P) (as measured by LC-SRM-MS), e.g. C×P, with an algorithm consisting of n proteins described as: C1×P1+C2×P2+C3×P3+ . . . +Cn×Pn. An algorithm that discriminates between disease states with a predetermined level of statistical significance may be refers to a “disease classifier”. In addition to the classifier's constituent proteins with differential expression, it may also include proteins with minimal or no biologic variation to enable assessment of variability, or the lack thereof, within or between clinical specimens; these proteins may be termed typical native proteins and serve as internal controls for the other classifier proteins.
In certain embodiments, expression levels are measured by MS. MS analyzes the mass spectrum produced by an ion after its production by the vaporization of its parent protein and its separation from other ions based on its mass-to-charge ratio. The most common modes of acquiring MS data are 1) full scan acquisition resulting in the typical total ion current plot (TIC), 2) selected ion monitoring (SIM), and 3) selected reaction monitoring (SRM).
In certain embodiments of the methods provided herein, biomarker protein expression levels are measured by LC-SRM-MS. LC-SRM-MS is a highly selective method of tandem mass spectrometry which has the potential to effectively filter out all molecules and contaminants except the desired analyte(s). This is particularly beneficial if the analysis sample is a complex mixture which may comprise several isobaric species within a defined analytical window. LC-SRM-MS methods may utilize a triple quadrupole mass spectrometer which, as is known in the art, includes three quadrupole rod sets. A first stage of mass selection is performed in the first quadrupole rod set, and the selectively transmitted ions are fragmented in the second quadrupole rod set. The resultant transition (product) ions are conveyed to the third quadrupole rod set, which performs a second stage of mass selection. The product ions transmitted through the third quadrupole rod set are measured by a detector, which generates a signal representative of the numbers of selectively transmitted product ions. The RF and DC potentials applied to the first and third quadrupoles are tuned to select (respectively) precursor and product ions that have m/z values lying within narrow specified ranges. By specifying the appropriate transitions (m/z values of precursor and product ions), a peptide corresponding to a targeted protein may be measured with high degrees of sensitivity and selectivity. Signal-to-noise ratio is superior to conventional tandem mass spectrometry (MS/MS) experiments, which select one mass window in the first quadrupole and then measure all generated transitions in the ion detector. LC-SRM-MS.
In certain embodiments, an SRM-MS assay for use in diagnosing or monitoring lung cancer as disclosed herein may utilize one or more peptides and/or peptide transitions derived from the proteins set forth in Table 6. In certain embodiments, the assay may utilize peptides and/or peptide transitions from 100 or more, 150 or more, 200 or more, 250 or more, 300 or more, 345 or more, or 371 or more biomarker proteins. In certain embodiments, two or more peptides may be utilized per biomarker proteins, and in certain of these embodiments three or more of four or more peptides may be utilized. Similarly, in certain embodiments two or more transitions may be utilized per peptide, and in certain of these embodiments three or more; four or more; or five or more transitions may be utilized per peptide. In one embodiment, an LC-SRM-MS assay for use in diagnosing lung cancer may measure the intensity of five transitions that correspond to selected peptides associated with each biomarker protein. The achievable limit of quantification (LOQ) may be estimated for each peptide according to the observed signal intensities during this analysis. For examples, for sets of target proteins associated with lung cancer see Table 12.
The expression level of a biomarker protein can be measured using any suitable method known in the art, including but not limited to mass spectrometry (MS), reverse transcriptase-polymerase chain reaction (RT-PCR), microarray, serial analysis of gene expression (SAGE), gene expression analysis by massively parallel signature sequencing (MPSS), immunoassays (e.g., ELISA), immunohistochemistry (RIC), transcriptomics, and proteomics.
To evaluate the diagnostic performance of a particular set of peptide transitions, a ROC curve is generated for each significant transition.
An “ROC curve” as used herein refers to a plot of the true positive rate (sensitivity) against the false positive rate (specificity) for a binary classifier system as its discrimination threshold is varied. A ROC curve can be represented equivalently by plotting the fraction of true positives out of the positives (TPR=true positive rate) versus the fraction of false positives out of the negatives (FPR=false positive rate). Each point on the ROC curve represents a sensitivity/specificity pair corresponding to a particular decision threshold.
AUC represents the area under the ROC curve. The AUC is an overall indication of the diagnostic accuracy of 1) a biomarker or a panel of biomarkers and 2) a ROC curve. AUC is determined by the “trapezoidal rule.” For a given curve, the data points are connected by straight line segments, perpendiculars are erected from the abscissa to each data point, and the sum of the areas of the triangles and trapezoids so constructed is computed. In certain embodiments of the methods provided herein, a biomarker protein has an AUC in the range of about 0.75 to 1.0. In certain of these embodiments, the AUC is in the range of about 0.8 to 0.8, 0.9 to 0.95, or 0.95 to 1.0.
The methods provided herein are minimally invasive and pose little or no risk of adverse effects. As such, they may be used to diagnose, monitor and provide clinical management of subjects who do not exhibit any symptoms of a lung condition and subjects classified as low risk for developing a lung condition. For example, the methods disclosed herein may be used to diagnose lung cancer in a subject who does not present with a PN and/or has not presented with a PN in the past, but who nonetheless deemed at risk of developing a PN and/or a lung condition. Similarly, the methods disclosed herein may be used as a strictly precautionary measure to diagnose healthy subjects who are classified as low risk for developing a lung condition.
The present invention provides a method of determining the likelihood that a lung condition in a subject is cancer by measuring an abundance of a panel of proteins in a sample obtained from the subject; calculating a probability of cancer score based on the protein measurements and ruling out cancer for the subject if the score) is lower than a pre-determined score, wherein when cancer is ruled out the subject does not receive a treatment protocol. Treatment protocols include for example pulmonary function test (PFT), pulmonary imaging, a biopsy, a surgery, a chemotherapy, a radiotherapy, or any combination thereof. In some embodiments, the imaging is an x-ray, a chest computed tomography (CT) scan, or a positron emission tomography (PET) scan.
The present invention further provides a method of ruling in the likelihood of cancer for a subject by measuring an abundance of panel of proteins in a sample obtained from the subject, calculating a probability of cancer score based on the protein measurements and ruling in the likelihood of cancer for the subject if the score in step is higher than a pre-determined score
In another aspect the invention further provides a method of determining the likelihood of the presence of a lung condition in a subject by measuring an abundance of panel of proteins in a sample obtained from the subject, calculating a probability of cancer score based on the protein measurements and concluding the presence of said lung condition if the score is equal or greater than a pre-determined score. The lung condition is lung cancer such as for example, non-small cell lung cancer (NSCLC). The subject at risk of developing lung cancer
The panel includes at least 4 proteins selected from ALDOA, FRIL, LG3BP, IBP3, LRP1, ISLR, TSP COIA1, GRP78, TETN, PRXD1 and CD14. Optionally, the panel further includes at least one protein selected from BGH3, COIA1, TETN, GRP78, PRDX, FIBA and GSLG1.
The subject has or is suspected of having a pulmonary nodule. The pulmonary nodule has a diameter of less than or equal to 3 cm. In one embodiment, the pulmonary nodule has a diameter of about 0.8 cm to 2.0 cm.
The score is calculated from a logistic regression model applied to the protein measurements. For example, the score is determined as Ps=1/[1+exp(−α−Σi=1Nβi*{hacek over (I)}i,s)], where {hacek over (I)}i,s is logarithmically transformed and normalized intensity of transition i in said sample (s), βi is the corresponding logistic regression coefficient, a was a panel-specific constant, and N was the total number of transitions in said panel.
In various embodiments, the method of the present invention further comprises normalizing the protein measurements. For example, the protein measurements are normalized by one or more proteins selected from PEDF, MASP1, GELS, LUM, C163A and PTPRJ.
The biological sample such as for example tissue, blood, plasma, serum, whole blood, urine, saliva, genital secretion, cerebrospinal fluid, sweat and excreta.
In one aspect, the determining the likelihood of cancer is determined by the sensitivity, specificity, negative predictive value or positive predictive value associated with the score. The score determined has a negative predictive value (NPV) is at least about 80%.
The measuring step is performed by selected reaction monitoring mass spectrometry, using a compound that specifically binds the protein being detected or a peptide transition. In one embodiment, the compound that specifically binds to the protein being measured is an antibody or an aptamer.
In specific embodiments, the diagnostic methods disclosed herein are used to rule out a treatment protocol for a subject, measuring the abundance of a panel of proteins in a sample obtained from the subject, calculating a probability of cancer score based on the protein measurements and ruling out the treatment protocol for the subject if the score determined in the sample is lower than a pre-determined score. In some embodiments the panel contains at least 4 proteins selected ALDOA, FRIL, LG3BP, IBP3, LRP1, ISLR, TSP, COIA1, GRP78, TETN, PRXD1 and CD14
Optionally, the panel further comprises one or more proteins selected from ERO1A, 6PGD, GSTP1, GGH, PRDX1, CD14, PTPA, ICAM1, FOLH1, SODM, FIBA, GSLG1, RAP2B, or C163A or one or more proteins selected from LRP1, COIA1, TSP1, ALDOA, GRP78, FRIL, LG3BP, BGH3, ISLR, PRDX1, FIBA, or GSLG. In preferred embodiments, the panel contains at least TSP1, LG3BP, LRP1, ALDOA, and COIA1. In more a preferred embodiment, the panel contains at least TSP1, LRP1, ALDOA and COIA1.
In specific embodiments, the diagnostic methods disclosed herein are used to rule in a treatment protocol for a subject by measuring the abundance of a panel of proteins in a sample obtained from the subject, calculating a probability of cancer score based on the protein measurements and ruling in the treatment protocol for the subject if the score determined in the sample is greater than a pre-determined score. In some embodiments the panel contains at least 4 proteins selected ALDOA, FRIL, LG3BP, IBP3, LRP1, ISLR or TSP1 or ALDOA, FRIL, LG3BP, IBP3, LRP1, ISLR, TSP COIA1, GRP78, TETN, PRXD1 and CD14. Optionally, the panel further comprises one or more proteins selected from ERO1A, 6PGD, GSTP1, COIA1, GGH, PRDX1, SEM3G, GRP78, TETN, AIFM1, MPRI, TNF12, MMP9 or OSTP or COIALTETN, GRP78, APOE or TBB3.
In certain embodiments, the diagnostic methods disclosed herein can be used in combination with other clinical assessment methods, including for example various radiographic and/or invasive methods. Similarly, in certain embodiments, the diagnostic methods disclosed herein can be used to identify candidates for other clinical assessment methods, or to assess the likelihood that a subject will benefit from other clinical assessment methods.
The high abundance of certain proteins in a biological sample such as plasma or serum can hinder the ability to assay a protein of interest, particularly where the protein of interest is expressed at relatively low concentrations. Several methods are available to circumvent this issue, including enrichment, separation, and depletion. Enrichment uses an affinity agent to extract proteins from the sample by class, e.g., removal of glycosylated proteins by glycocapture. Separation uses methods such as gel electrophoresis or isoelectric focusing to divide the sample into multiple fractions that largely do not overlap in protein content. Depletion typically uses affinity columns to remove the most abundant proteins in blood, such as albumin, by utilizing advanced technologies such as IgY14/Supermix (SigmaSt. Louis, Mo.) that enable the removal of the majority of the most abundant proteins.
In certain embodiments of the methods provided herein, a biological sample may be subjected to enrichment, separation, and/or depletion prior to assaying biomarker or putative biomarker protein expression levels. In certain of these embodiments, blood proteins may be initially processed by a glycocapture method, which enriches for glycosylated proteins, allowing quantification assays to detect proteins in the high pg/ml to low ng/ml concentration range. Exemplary methods of glycocapture are well known in the art (see, e.g., U.S. Pat. No. 7,183,188; U.S. Patent Appl. Publ. No. 2007/0099251; U.S. Patent Appl. Publ. No. 2007/0202539; U.S. Patent Appl. Publ. No. 2007/0269895; and U.S. Patent Appl. Publ. No. 2010/0279382). In other embodiments, blood proteins may be initially processed by a protein depletion method, which allows for detection of commonly obscured biomarkers in samples by removing abundant proteins. In one such embodiment, the protein depletion method is a GenWay depletion method.
In certain embodiments, a biomarker protein panel comprises two to 100 biomarker proteins. In certain of these embodiments, the panel comprises 2 to 5, 6 to 10, 11 to 15, 16 to 20, 21-25, 5 to 25, 26 to 30, 31 to 40, 41 to 50, 25 to 50, 51 to 75, 76 to 100, biomarker proteins. In certain embodiments, a biomarker protein panel comprises one or more subpanels of biomarker proteins that each comprise at least two biomarker proteins. For example, biomarker protein panel may comprise a first subpanel made up of biomarker proteins that are overexpressed in a particular lung condition and a second subpanel made up of biomarker proteins that are under-expressed in a particular lung condition.
In certain embodiments of the methods, compositions, and kits provided herein, a biomarker protein may be a protein that exhibits differential expression in conjunction with lung cancer. For example, in certain embodiments a biomarker protein may be one of the proteins associated with lung cancer set forth in Table 6.
In other embodiments, the diagnosis methods disclosed herein may be used to distinguish between two different lung conditions. For example, the methods may be used to classify a lung condition as malignant lung cancer versus benign lung cancer, NSCLC versus SCLC, or lung cancer versus non-cancer condition (e.g., inflammatory condition).
In certain embodiments, kits are provided for diagnosing a lung condition in a subject. These kits are used to detect expression levels of one or more biomarker proteins. Optionally, a kit may comprise instructions for use in the form of a label or a separate insert. The kits can contain reagents that specifically bind to proteins in the panels described, herein. These reagents can include antibodies. The kits can also contain reagents that specifically bind to mRNA expressing proteins in the panels described, herein. These reagents can include nucleotide probes. The kits can also include reagents for the detection of reagents that specifically bind to the proteins in the panels described herein. These reagents can include fluorophores.
The following examples are provided to better illustrate the claimed invention and are not to be interpreted as limiting the scope of the invention. To the extent that specific materials are mentioned, it is merely for purposes of illustration and is not intended to limit the invention. One skilled in the art may develop equivalent means or reactants without the exercise of inventive capacity and without departing from the scope of the invention
A retrospective, case-control study design was used to identify biomarker proteins and panels thereof for diagnosing various lung diseases in pre-defined control and experimental groups. The first goal of these studies was to demonstrate statistically significant differential expression for individual proteins between control and experimental groups. The second goal is to identify a panel of proteins which all individually demonstrate statistically significant differential expression between control and experimental groups. This panel of proteins can then be used collectively to distinguish between dichotomous disease states.
Specific study comparisons may include 1) cancer vs. non-cancer, 2) small cell lung cancer versus non-small cell lung cancer (NSCLC), 3) cancer vs. inflammatory disease state (e.g., infectious granuloma), or 4) different nodule size, e.g., <10 mm versus ≧10 mm (alternatively using 10, 15 or 20 mm cut-offs depending upon sample distributions).
Data for each subject consisted of the following:
Archived plasma samples from subjects previously enrolled in Institute Review Board (IRB)-approved studies was used to identify biomarker proteins and biomarker panels for distinguishing lung malignancies from non-malignancies. Plasma samples were originally obtained by routine phlebotomy, aliquotted, and stored at −80° C. or lower. Sample preparation, assignment of subject identification codes, initial subject record entry, and specimen storage were performed as per IRB study protocols. Sample eligibility is based on clinical parameters, including the subject, PN, and clinical staging parameters. Parameters for inclusion and exclusion are set forth in Table 7.
The assignment of a sample to a control or experimental group, and its further stratification or matching to other samples within and between these groups, is dependent on various clinical data about the subject. This data includes, for example, demographic information such as age, gender, and clinical history (e.g., smoking status), co-morbid conditions, PN characterization, and pathologic interpretation of resected lesions and tissues (Table 8).
The study design and analytical plan prioritizes the control: experimental group pairings set forth in Table 9. Additional clinical and molecular insights may be gained by selective inclusion of phenotypes, e.g. effect of smoking, in the assignment of experimental and control groups. Demographic information available in the clinical database will enable further refinements in sample selection via the stratification or matching of samples in the case-control analyses with respect to clinical parameters, e.g., age and nodule size.
LC-SRM-MS is performed to identify and quantify various plasma proteins in the plasma samples. Prior to LC-SRM-MS analysis, each sample is depleted using IgY14/Supermix (Sigma) and then trypsin-digested. Samples from each control or experimental group are batched randomly and processed together on a QTrap 5500 instrument (AB SCIEX, Foster City, Calif.) for unbiased comparisons. Each sample analysis takes approximately 30 minutes. Peak areas for two transitions (native and heavy label) are collected and reported for all peptides and proteins. The data output for each protein analyzed by LC-SRM-MS typically yields four measurements consisting of two transition measurements from each of two peptides from the same protein. These measurements enable an inference of the relative abundance of the target protein, which will be used as its expression level in the bioinformatics and statistical analyses.
Identification of biomarker proteins having differential expression levels between the control and experimental groups yields one or more novel proteomic profiles. For example, biomarker proteins are identified with expression levels that differ in subjects with PNs who are diagnosed with NSCLC versus those without an NSCLC diagnosis, or in subjects with PNs who are diagnosed with NSCLC versus an inflammatory disorder. Panels of biomarker proteins are also identified which can collectively discriminate between dichotomous disease states.
Analyses may be (a priori) powered appropriately to control type 1 and type 2 errors at 0.05 and to detect inter-cohort differences of 25% per analyte. The diagnostic power of individual proteins is generally assessed to distinguish between two cohorts, assuming a one-sided paired non-parametric test is used. This provides a lower bound on the sample size required to demonstrate differential expression between experimental and control groups. Multiple testing effects apply for the identification of panels of proteins for assessing diagnostic efficacy, which requires larger sample sizes.
The sequence of steps for determining statistical significance for differential expression of an individual protein includes the following: 1) assessing and correlating the calibrated values of transitions of a single protein (a quality control measure); 2) comparing paired analysis of groups to control for other influences using the Mann-Whitney U-test (rank sum) to determine statistical significance; and 3) determining its significance based on a pre-defined significance threshold. Transitions within a protein that are not correlated across samples (e.g., Pearson correlation<0.5) will be deemed unreliable and excluded from the analysis.
Comparison of calibrated samples between two cohorts, e.g., cancer and non-cancer, requires pairing or matching using a variety of clinical parameters such as nodule size, age and gender. Such pairing controls for the potential influence of these other parameters on the actual comparison goal, e.g. cancer and non-cancer. A non-parametric test such as the Mann-Whitney U-test (rank sum) will then be applied to measure the statistical difference between the groups. The resulting p value can be adjusted using multiple testing corrections such as the false discovery rate. Permutation tests can be used for further significance assessments.
Significance will be determined by the satisfaction of a pre-defined threshold, such as 0.05, to filter out assays, with the potential use of higher threshold values for additional filtering. An additional significance criterion is that two of three replicate assays must individually be significant in order for the assay, e.g., single protein, to be significant.
Panels of proteins that individually demonstrate statistically significant differential expression as defined above and which can collectively be used to distinguish dichotomous disease states are identified using statistical methods described herein. This requires developing multivariate classifiers and assessing sensitivity, specificity, and ROC AUC for panels. In addition, protein panels with optimal discriminatory performance, e.g., ROC AUC, are identified and may be sufficient for clinical use in discriminating disease states.
The sequence of steps for determining the statistical significance of the discriminatory ability of a panel of proteins includes 1) developing multivariate classifiers for protein panels, and 2) identifying a protein panel with optimal discriminatory performance, e.g. ROC AUC, for a set of disease states.
A multivariate classifier (e.g., majority rule) will be developed for protein panels, including single protein assays deemed to be significant. The sensitivity and specificity of each classifier will be determined and used to generate a receiver operating characteristics (ROC) curve and its AUC to assess a given panel's discriminatory performance for a specific comparison, e.g. cancer versus non-cancer.
1. Review clinical data from a set of subjects presenting with lung disease.
2. Provide plasma samples from the subjects wherein the samples are either benign, cancerous, COPD or another lung disease.
3. Group the plasma samples that are benign or cancerous by PNs that are separated by size of the nodule.
4. Target a pool of 371 putative lung cancer biomarker proteins consisting of at least two peptides per protein and at least two LC-SRM-MS transitions per peptide. Measuring the LC-SRM-MS transitions in each specimen along with 5 synthetic internal standards consisting of 10 transitions to compare peptide transitions from the plasma to the synthetic internal standards by LC-SRM-MS mass spectroscopy.
5. Quantitate the intensity of each transition.
6. Normalize the quantitated transitions to internal standards to obtain a normalized intensity.
7. Review the measured peptide transitions for correlations from the same peptide, rejecting discordant transitions.
8. Generate an ROC for each transition by comparing cancerous with benign samples. (ROC compare specificity (true positive) to (1-sensitivity) false positive).
9. Define the AUC for each transition. (An AUC of 0.5 is a random classifier; 1.0 is a perfect classifier).
10. Determine an AUC cut-off point to determine transitions that are statistically significant.
11. Define the transitions that exceed the AUC cutoff point.
12. Combine all pairings of significant transitions.
13. Define a new AUC for each transition pair by means of logistical regression.
14. Repeat pairing combinations into triples, quad, etc.; defining a new AUC based upon the logistical regression of combined transitions until a panel of biomarker transitions with combined desired performance (sensitivity & specificity) have been achieved.
15. The panel of biomarker transitions is verified against previously unused set of plasma panels.
Plasma samples will be obtained from one or more subjects presenting with PNs to evaluate whether the subjects have a lung condition. The plasma samples will be depleted using IgY14/Supermix (Sigma) and optionally subjected to one or more rounds of enrichment and/or separation, and then trypsinized. The expression level of one or more biomarker proteins previously identified as differentially expressed in subjects with the lung condition will be measured using an LC-SRM-MS assay. The LC-SRM-MS assay will utilize two to five peptide transitions for each biomarker protein. For example, the assay may utilize one or more of the peptide transitions generated from any of the proteins listed in Table 6. Subjects will be classified as having the lung condition if one or more of the biomarker proteins exhibit expression levels that differ significantly from the pre-determined control expression level for that protein.
A panel of 15 proteins was created where the concentration of these 15 proteins relative to the concentration of 6 protein standards is indicative of likelihood of cancer. The relative concentration of these 15 proteins to the 6 protein standards was measured using a mass spectrometry methodology. A classification algorithm is used to combine these relative concentrations into a relative likelihood of the PN being benign or malignant. Further it has been demonstrated that there are many variations on these panels that are also diagnostic tests for the likelihood that a PN is benign or malignant. Variations on the panel of proteins, protein standards, measurement methodology and/or classification algorithm are described herein.
A Single Reaction Monitoring (SRM) mass spectrometry (MS) assay was developed consisting of 1550 transitions from 345 lung cancer associated proteins. The SRM-MS assay and methodology is described above. The goal of this study was to develop a blood-based diagnostic for classifying PNs under 2 cm in size as benign or malignant. The study design appears in Table 10.
The study consisted of 242 plasma samples from three sites (Laval, UPenn and NYU). The number of benign and malignant samples from each site are indicated in Table 10. The study consisted of 144 plasma samples from patients with PNs of size 2 cm or less and of 98 samples from patients with PNs of size larger than 2 cm. This resulted in an estimated power of 94% for discovering proteins with blood concentrations of 1.5 fold or more between benign and malignant cancer samples of size 2 cm or less. Power is 74% for PNs of size larger than 2 cm.
This study was a retrospective multisite study that was intended to derive protein biomarkers of lung cancer that are robust to site-to-site variation. The study included samples larger than 2 cm to ensure that proteins not detectable due to the limit of detection of the measurement technology (LC-SRM-MS) for tumors of size 2 cm or less could still be detected in tumors of size 2 cm or larger.
Samples from each site and in each size class (above and below 2 cm) were matched on nodule size, age and gender.
Each sample was analyzed using the LC-SRM-MS measurement methodology as follows:
1. Samples were depleted of high abundance proteins using the IGy14 and Supermix depletion columns from Sigma-Aldrich.
2. Samples were digested using trypsin into tryptic peptides.
3. Samples were analyzed by LC-SRM-MS using a 30 minute gradient on a Waters nanoacuity LC system followed by SRM-MS analysis of the 1550 transitions on a AB-Sciex 5500 triple quad device.
4. Raw transition ion counts were obtained and recorded for each of the 1550 transitions.
It is important to note that matched samples were processed at each step either in parallel (steps 2 and 4) or back-to-back serially (steps 1 and 3). This minimizes analytical variation. Finally, steps 1 and 2 of the sample analysis are performed in batches of samples according to day of processing. There were five batches of ‘small’ samples and four batches of ‘large’ samples as denoted in Table 10.
A shortlist of 68 proteins reproducibly diagnostic across sites was derived as follows. Note that each protein can be measured by multiple transitions.
Step 1: Normalization
Six proteins were identified that had a transition detected in all samples of the study and with low coefficient of variation. For each protein the transition with highest median intensity across samples was selected as the representative transition for the protein. These proteins and transitions are found in Table 11.
We refer to the transitions in Table 11 as normalizing factors (NFs). Each of the 1550 transitions were normalized by each of the six normalizing factors where the new intensity of a transition t in a sample s by NF f, denoted New(s,t,f), is calculated as follows:
New(s,t,f)=Raw(s,t)*Median(f)/Raw(s,f)
where Raw(s,t) is the original intensity of transition t in sample s; Median(f) is the median intensity of the NF f across all samples; and Raw(s,f) is the original intensity of the NF f in sample s.
For each protein and normalized transition, the AUC of each batch was calculated. The NF that minimized the coefficient of variation across the 9 batches was selected as the NF for that protein and for all transitions of that protein. Consequently, every protein (and all of its transitions) are now normalized by a single NF.
Step 2: Reproducible Diagnostic Proteins
For each normalized transition its AUC for each of the nine batches in the study is calculated as follows. If the transition is detected in fewer than half of the cancer samples and in fewer than half of the benign samples then the batch AUC is ‘ND’. Otherwise, the batch AUC is calculated comparing the benign and cancer samples in the batch.
The batch AUC values are transformed into percentile AUC scores for each transition. That is, if a normalized transition is in the 82nd percentile of AUC scores for all transitions then it is assigned percentile AUC 0.82 for that batch.
Reproducible transitions are those satisfying at least one of the following criteria:
1. In at least four of the five small batches the percentile AUC is 75% or more (or 25% and less).
2. In at least three of the five small batches the percentile AUC is 80% or more (or 20% and less) AND the remaining percentile AUCs in the small batches are above 50% (below 50%).
3. In all five small batches the percentile AUC is above 50% (below 50%).
4. In at least three of the four large batches the percentile AUC is 85% or more (or 15% and less).
5. In at least three of the four large batches the percentile AUC is 80% or more (or 20% and less) AND the remaining percentile AUCs in the large batches are above 50% (below 50%).
6. In all four large batches the percentile AUC is above 50% (below 50%).
These criteria result in a list of 67 proteins with at least one transition satisfying one or more of the criteria. These proteins appear in Table 12.
Step 3: Significance and Occurrence
To find high performing panels, 10,000 trials were performed where on each trial the combined AUC of a random panel of 15 proteins selected from Table 12 was estimated. To calculate the combined AUC of each panel of 15 proteins, the highest intensity normalized transition was utilized. Logistic regression was used to calculate the AUC of the panel of 15 across all small samples. 131 panels of 15 proteins had combined AUC above 0.80, as shown in
To calculate the combined AUC of each panel of 15 proteins, the highest intensity normalized transition was utilized. Logistic regression was used to calculate the AUC of the panel of 15 across all small samples. 5 panels of 15 proteins had combined AUC above 0.80.
Finally, the frequency of each of the 67 proteins on the 131 panels listed in Table 13 is presented in Table 12 both as raw counts (column 2) and percentage (column 3). It is an important observation that the panel size of 15 was pre-selected to prove that there are diagnostic proteins and panels. Furthermore, there are numerous such panels. Smaller panels selected from the list of 67 proteins can also be formed and can be generated using the same methods here.
In Table 14 a logistic regression classifier trained on all small samples is presented.
The classifier has the structure
Where C0 and Ci are logistic regression coefficients, Pi are logarithmically transformed normalized transition intensities. Samples are predicted as cancer if Probability≧0.5 or as benign otherwise. In Table 14 the coefficients Ci appear in the sixth column, C0 in the last row, and the normalized transitions for each protein are defined by column 2 (protein transition) and column 4 (the normalizing factor).
The performance of this classifier, presented as a ROC plot, appears in
Using a subset of 35 proteins (Table 15) from the 67 proteins identified as a diagnostic panel (Table 13), a backward systems analysis was performed. Two networks were queried that are identified as cancer networks with the identified 35 proteins. The results show that the networks that have the highest percentage of “hits” when the proteins are queried that are found in the blood of patients down to the level of the nucleus are initiated by transcription factors that are regulated by either cigarette smoke or lung cancer among others. See also Table 16 and
These results are further evidence that the proteins that were identified using the methods of the invention as diagnostic for lung cancer are prognostic and relevant.
To achieve unbiased discovery of cooperative proteins, selected reaction monitoring (SRM) mass spectrometry (Addona, Abbatiello et al. 2009) was utilized. SRM is a form of mass spectrometry that monitors predetermined and highly specific mass products of particularly informative (proteotypic) peptides of selected proteins. These peptides are recognized as specific transitions in mass spectra. SRM possesses the following required features that other technologies, notably antibody-based technologies, do not possess:
Labeled and unlabeled SRM peptides are commercially available, together with an open-source library and data repository of mass spectra for design and conduct of SRM analyses. Exceptional public resources exist to accelerate assay development including the PeptideAtlas [11] and the Plasma Proteome Project [12, 13], the SRM Atlas and PASSEL, the PeptideAtlas SRM Experimental Library (www.systemsbiology.org/passel).
Two SRM strategies that enhance technical performance were introduced. First, large scale SRM assay development introduces the possibility of monitoring false signals. Using an extension of expression correlation techniques [14], the rate of false signal monitoring was reduced to below 3%. This is comparable and complementary to the approach used by mProphet (Reiter, Rinner et al. 2011).
Second, a panel of endogenous proteins was used for normalization. However, whereas these proteins are typically selected as “housekeeping” proteins (Lange, Picotti et al. 2008), proteins that were strong normalizers for the technology platform were identified. That is, proteins that monitored the effects of technical variation so that it could be controlled effectively. This resulted, for example, in the reduction of technical variation due to sample depletion of high abundance proteins from 23.8% to 9.0%. The benefits of endogenous signal normalization has been previously discussed (Price, Trent et al. 2007).
The final component of the strategy was to carefully design the discovery and validation studies using emerging best practices. Specifically, the cases (malignant nodules) and controls (benign nodules) were pairwise matched on age, nodule size, gender and participating clinical site. This ensures that the candidate markers discovered are not markers of age or variations in sample collection from site to site. The studies were well-powered, included multiple sites, a new site participated in the validation study, and importantly, were designed to address the intended use of the test. The careful selection and matching of samples resulted in an exceptionally valuable feature of the classifier. The classifier generates a score that is independent of nodule size and smoking status. As these are currently used risk factors for clinical management of IPNs, the classifier is a complementary molecular tool for use in the diagnosis of IPNs.
Selection of Biomarker Candidates for Assay Development
To identify lung cancer biomarkers in blood that originate from lung tumor cells, resected lung tumors and distal normal tissue of the same lobe were obtained. Plasma membranes were isolated from both endothelial and epithelial cells and analyzed by tandem mass spectrometry to identify cell surface proteins over expressed on tumor cells. Similarly, Golgi apparatus were isolated to identify over-secreted proteins from tumor cells. Proteins with evidence of being present in blood or secreted were prioritized resulting in a set of 217 proteins. See Example 7: Materials and Methods for details.
To ensure other viable lung cancer biomarkers were not overlooked, a literature search was performed and manually curated for lung cancer markers. As above, proteins with evidence of being present in blood or secreted were prioritized. This resulted in a set of 319 proteins. See Example 7: Materials and Methods for details.
The tissue (217) and literature (319) candidates overlapped by 148 proteins resulting in a final candidate list of 388 protein candidates. See Example 7: Materials and Methods.
Development of SRM Assays
SRM assays for the 388 proteins were developed using standard synthetic peptide techniques (See Example 7: Materials and Methods). Of the 388 candidates, SRM assays were successfully developed for 371 candidates. The 371 SRM assays were applied to benign and lung cancer plasma samples to evaluate detection rate in blood. 190 (51% success rate) of the SRM assays were detected. This success rate compares favorably to similar attempts to develop large scale SRM assays for detection of cancer markers in plasma. Recently 182 SRM assays for general cancer markers were developed from 1172 candidates (16% success rate) [15]. Despite focusing only on lung cancer markers, the 3-fold increase in efficiency is likely due to sourcing candidates from cancer tissues with prior evidence of presence in blood. Those proteins of the 371 that were previously detected by mass spectrometry in blood had a 64% success rate of detection in blood whereas those without had a 35% success rate. Of the 190 proteins detected in blood, 114 were derived from the tissue-sourced candidates and 167 derived from the literature-sourced candidates (91 protein overlap). See Example 7: Materials and Methods and Table 6.
Typically, SRM assays are manually curated to ensure assays are monitoring the intended peptide. However, this becomes unfeasible for large scale SRM assays such as this 371 protein assay. More recently, computational tools such as mProphet (Reiter, Rinner et al. 2011) enable automated qualification of SRM assays. A complementary strategy to mProphet was introduced that does not require customization for each dataset set. It utilizes correlation techniques (Kearney, Butler et al. 2008) to confirm the identity of protein transitions with high confidence. In
Classifier Discovery
A summary of the 143 samples used for classifier discovery appears in Table 17. Samples were obtained from three sites to avoid overfitting to a single site. Participating sites were Laval (Institut Universitaire de Cardiologie et de Pneumologie de Quebec), NYU (New York University) and UPenn (University of Pennsylvania). Samples were also selected to be representative of the intended use population in terms of nodule size (diameter), age and smoking status.
Benign and cancer samples were paired by matching on age, gender, site and nodule size (benign and cancer samples were required to have a nodule identified radiologically). The benign and cancer samples display a bias in smoking (pack years), however, the majority of benign and cancer samples were current or past smokers. In comparing malignant and benign samples, the intent was to find proteins that were markers of lung cancer; not markers of age, nodule size or differences in site sample collection. Note that cancer samples were pathologically confirmed and benign samples were either pathologically confirmed or radiologically confirmed (no tumor growth demonstrated over two years of CT scan surveillance).
The processing of samples was conducted in batches. Each batch contained a set of randomly selected cancer-benign pairs and three plasma standards, included for calibration and quality control purposes.
All plasma samples were immunodepleted, trypsin digested and analyzed by reverse phase HPLC-SRM-MS. Protein transitions were normalized using an endogenous protein panel. The normalization procedure was designed to reduce overall variability, but in particular, the variability introduced by the depletion step. Overall technical variability was reduced from 32.3% to 25.1% and technical variability due to depletion was reduced from 23.8% to 9.0%. Details of the sample analysis and normalization procedure are available in Example 7: Materials and Methods.
To assess panels of proteins, they were fit to a logistic regression model. Logistic regression was chosen to avoid the overfitting that can occur with non-linear models, especially when the number of variables measured (transitions) is similar or larger than the number of samples in the study. The performance of a panel was measured by partial area under the curve (AUC) with sensitivity fixed at 90% (McClish 1989). Partial AUC correlates to high NPV performance while maximizing ROR.
To derive the 13 protein classifier, four criteria were used:
Finally, the 13 protein classifier was trained to a logistic regression model by Monte Carlo cross validation (MCCV) with a hold out rate of 20% and 20,000 iterations. The thirteen proteins for the rule-out classifier are listed in Table 18 along with their highest intensity transition and model coefficient.
Validation of the Rule-Out Classifier
52 cancer and 52 benign samples (see Table 17) were used to validate the performance of the 13 protein classifier. All samples were independent of the discovery samples, in addition, over 36% of the validation samples were sourced from a new fourth site (Vanderbilt University). Samples were selected to be consistent with intended use and matched in terms of gender, clinical site and nodule size. We note a slight age bias, which is due to 5 benign samples from young patients. Anticipating a NPV of 90%, the 95% confidence interval is +/−5%.
At this point we refer to the 13 protein classifier trained on 143 samples the Discovery classifier. However, once validation is completed, to find the optimal coefficients for the classifier, it was retrained on all 247 samples (discovery and validation sets) as this is most predictive of future performance. We refer to this classifier as the Final classifier. The coefficients of the Final classifier appear in Table 21.
The performance of the Discovery and Final classifiers is summarized in
NPV and ROR are each reported as a fraction from 0 to 1. Similarly, the classifier produces a score between 0 and 1, which is the probability of cancer predicted by the classifier.
The discovery and validation curves for NPV and ROR are similar with the discovery curves superior as expected. This demonstrates the reproducibility of performance on an independent set of samples. A Discovery classifier rule out threshold of 0.40 achieves NPV of 96% and 90%, whereas ROR is 33% and 23%, for the discovery samples and the validation samples, respectively. Final classifier rule threshold of 0.60 achieves NPV of 91% and 90%, whereas ROR is 45% and 43%, for all samples and all samples restricted to be 8 mm-20 mm, respectively.
First, note the density of cancer samples with high classifier scores. The classifier has been designed to detect a cancer signature in blood with high sensitivity. As a consequence, to the left of the rule out threshold (0.60) there are very few (<10%) cancer samples, assuming cancer prevalence of 25% [16, 17].
Third is the observation that nodule size does not appear to increase with the classifier score. Both large and small nodules are spread across the classifier score spectrum. Similarly, although there are a few very heavy smokers with very high classifier scores, increased smoking does not seem to increase with classifier score. To quantify this observation the correlation between the classifier score and nodule size, smoking and age were calculated and appear in Table 19. In all cases there is no significant relationship between the classifier score and the risk factors. The one exception is a weak correlation between benign classifier scores and benign ages. However, this correlation is so weak that the classifier score increases by only 0.04 every 10 years.
This lack of correlation has clinical utility. It implies that the classifier provides molecular information about the disease status of an IPN that is incremental upon risk factors such as nodule size and smoking status. Consequently, it is a clinical tool for physicians to make more informed decisions around the clinical management of an IPN.
To visual how this might be accomplished, we demonstrate how the cancer probability score generated by the classifier can be related to cancer risk (see
At a given classifier score, some percentage of all cancer nodules will have a smaller score. This is the sensitivity of the classifier. For example, at classifier score 0.8, 47% of cancer patients have a lower score, at classifier score 0.7, 28% of cancer patients have a lower score, at classifier score 0.5, only 9% are lower and finally at score 0.25, only 4% are lower. This enables a physician to interpret a patient's classifier score in terms of relative risk.
The Molecular Foundations of the Classifier
The goal was to identify the molecular signature of a malignant pulmonary nodule by selecting proteins that were the cooperative, robustly detected by SRM and stable within the classifier. How well associated with lung cancer is the derived classifier? Is there a molecular foundation for the perturbation of these 13 proteins in blood? And finally, how unique is the classifier among other possible protein combinations?
To answer these questions the 13 proteins of the classifier were submitted for pathway analysis using IPA (Ingenuity Systems, www.ingenuity.com). The first step was to work from outside the cell inwards to identify the transcription factors most likely to cause a modulation of these 13 proteins. The five most significant were FOS, NRF2, AHR, HD and MYC. FOS is common to many forms of cancer. However, NRF2 and AHR are associated with lung cancer, response to oxidative stress and lung inflammation. MYC is associated with lung cancer and response to oxidative stress while HD is associated with lung inflammation and response to oxidative stress.
The 13 classifier proteins are also highly specific to these three networks (lung cancer, response to oxidative stress and lung inflammation). This is summarized in
To address the question of classifier uniqueness, every classifier from the 21 robust and cooperative proteins was formed (Table 20). Due to the computational overhead, these classifiers could not be fully trained by Monte Carlo cross validation, consequently, only estimates of their performance could be obtained. Five high preforming alternative classifiers were identified and then fully trained. The classifier and the five high performing alternatives appear in Table 20. The frequency of each protein appears in the tally column, in particular, the first 11 proteins appear in 4 out of the 6 classifiers. These 11 proteins have significantly higher cooperative scores than the remaining proteins. By this analysis it appears that there is a core group of proteins that form the blood signature of a malignant nodule.
This result suggests that there is a core group of proteins that define a high performance classifier, but alternative panels exist. However, changes in panel membership affect the tradeoff between NPV and ROR.
Assay Development Candidates Sourced from Tissue
Patient samples obtained from fresh lung tumor resections were collected from Centre Hospitalier de l'Universite de Montreal and McGill University Health Centre under IRB approval and with informed patient consent. Samples were obtained from the tumor as well as from distal normal tissue in the same lung lobe. Plasma membranes of each pair of samples were then isolated from the epithelial cells of 30 patients (19 adenocarcinoma, 6 squamous, 5 large cell carcinoma) and endothelial cells of 38 patients (13 adenocarcinoma, 18 squamous, 7 large cell carcinoma) using immune-affinity protocols. Golgi apparatus were isolated from each pair of samples from 33 patients (18 adenocarcinoma, 14 squamous, 1 adenosquamous) using isopycnic centrifugation followed by ammonium carbonate extraction. Plasma membrane isolations and Golgi isolations were then analyzed by tandem mass spectrometry to identify proteins overexpressed in lung cancer tissue over normal tissue, for both plasma membranes and Golgi.
Assay Development Candidates Sourced from Literature
Candidate lung cancer biomarkers were identified from two public and one commercial database: Entrez (www.ncbi.nlm.nih.gov/books/NBK3836), UniProt (www.uniprot.org) and NextBio (www.nextbio.com). Terminologies were predefined for the database queries which were automated using PERL scripts. The mining was carried out on May 6, 2010 (UniProt), May 17, 2010 (Entrez) and Jul. 8, 2010 (NextBio), respectively. Biomarkers were then assembled and mapped to UniProt identifiers.
Evidence of Presence in Blood
The tissue-sourced and literature-source biomarker candidates were required to have evidence of presence in blood. For evidence by mass spectrometry detection, three datasets were used. HUP09504 contains 9504 human proteins identified by tandem mass spectrometry [13]. HUP0889, a higher confidence subset of HUP09504, contains 889 human proteins [18]. The PeptideAtlas (November 2009 build) was also used. A biomarker candidate was marked as previously detected if it contained at least one HUP0889, or at least two HUP09504 peptides, or at least two PeptideAtlas peptides.
In addition to direct evidence of detection in blood by mass spectrometry, annotation as secreted proteins or as single-pass membrane proteins [19] were also accepted as evidence of presence in blood. Furthermore, proteins in UniProt or designation as plasma proteins three programs for predicting whether or not a protein is secreted into the blood were used. These programs were TMHMM [20], SignalP [21] and SecretomeP [22]. A protein was predicted as secreted if TMHMM predicted the protein had one transmembrane domain and SignalP predicted the transmembrane domain was cleaved; or TMHMM predicted the protein had no transmembrane domain and either SignalP or SecretomeP predicted the protein was secreted.
SRM Assay Development
SRM assays for 388 targeted proteins were developed based on synthetic peptides, using a protocol similar to those described in the literature [15, 23, 24]. Up to five SRM suitable peptides per protein were identified from public sources such as the PeptideAtlas, Human Plasma Proteome Database or by proteotypic prediction tools [25] and synthesized. SRM triggered MS/MS spectra were collected on an ABSciex 5500 QTrap for both doubly and triply charged precursor ions. The obtained MS/MS spectra were assigned to individual peptides using MAS-COT (cutoff score≧15) [26]. Up to four transitions per precursor ion were selected for optimization. The resulting corresponding optimal retention time, declustering potential and collision energy were assembled for all transitions. Optimal transitions were measured on a mixture of all synthetic peptides, a pooled sample of benign patients and a pooled sample of cancer patients. Transitions were analyzed in batches, each containing up to 1750 transitions. Both biological samples were immuno-depleted and digested by trypsin and were analyzed on an ABSciex 5500 QTrap coupled with a reversed-phase (RP) high-performance liquid chromatography (HPLC) system. The obtained SRM data were manually reviewed to select the two best peptides per protein and the two best transitions per peptide. Transitions having interference with other transitions were not selected. Ratios between intensities of the two best transitions of peptides in the synthetic peptide mixture were also used to assess the specificity of the transitions in the biological samples. The intensity ratio was considered as an important metric defining the SRM assays.
Processing of Plasma Samples
Plasma samples were sequentially depleted of high- and medium-abundance proteins using immuno-depletion columns packed with the IgY14-Supermix resin from Sigma. The depleted plasma samples were then denatured, digested by trypsin and desalted. Peptide samples were separated using a capillary reversed-phase LC column (Thermo BioBasic 18 KAPPA; column dimensions: 320 μm×150 mm; particle size: 5 μm; pore size: 300 Å) and a nano-HPLC system (nanoACQUITY, Waters Inc.). The mobile phases were (A) 0.2% formic acid in water and (B) 0.2% formic acid in acetonitrile. The samples were injected (8 μl) and separated using a linear gradient (98% A to 70% A over 19 minutes, 5 μl/minute). Peptides were eluted directly into the electrospray source of the mass spectrometer (5500 QTrap LC/MS/MS, AB Sciex) operating in scheduled SRM positive-ion mode (Q1 resolution: unit; Q3 resolution: unit; detection window: 180 seconds; cycle time: 1.5 seconds). Transition intensities were then integrated by software MultiQuant (AB Sciex). An intensity threshold of 10,000 was used to filter out noisy data and undetected transitions.
Plasma Samples Used for Discovery and Validation Studies
Aliquots of plasma samples were provided by the Institut Universitaire de Cardiologie et de Pneumologie de Quebec (IUCPQ, Hospital Laval), New York University, the University of Pennsylvania, and Vanderbilt University (see Table 17). Subjects were enrolled in clinical studies previously approved by their Ethics Review Board (ERB) or Institutional Review Boards (IRB), respectively. In addition, plasma samples were provided by study investigators after review and approval of the sponsor's study protocol by the respective institution's IRB as required. Sample eligibility for the proteomic analysis was based on the satisfaction of the study inclusion and exclusion criteria, including the subject's demographic information, the subject's corresponding lung nodule radiographic characterization by chest computed tomography (CT), and the histopathology of the lung nodule obtained at the time of diagnostic surgical resection. Cancer samples had a histopathologic diagnosis of either non-small cell lung cancer (NSCLC), including adenocarcinoma, squamous cell, large cell, or bronchoalveolar cell carcinoma and a radiographic nodule of 30 mm or smaller. Benign samples, including granulomas, hamartomas and scar tissue, were also required to have a radiographic nodule of 30 mm or smaller and either histopathologic confirmation of being non-malignant or radiological confirmation in alignment with clinical guidelines. To ensure the accuracy of the clinical data, independent monitoring and verification of the clinical data associated with both the subject and lung nodule were performed in accordance with the guidance established by the Health Insurance Portability and Accountability Act (HIPAA) of 1996 to ensure subject privacy.
Study Design
The objective of the study design was to eliminate clinical and technical bias. Clinically, cancer and benign samples were paired so that they were from the same site, same gender, nodule sizes within 10 mm, age within 10 years, and smoking history within 20 pack years. Up to 15 pairs of matched cancer and benign samples per batch were assigned iteratively to processing batches until no statistical bias was demonstrable based on age, gender or nodule size.
Paired samples within each processing batch were further randomly and repeatedly assigned to positions within the processing batch, until the absolute values of the corresponding Pearson correlation coefficients between position and gender, nodule size, and age were less than 0.1. Afterwards, each pair of cancer and benign samples was randomized to their relative positions. To provide a control for sample batching, three 200 μl aliquots of a pooled human plasma standard (HPS) (Bioreclamation, Hicksville, N.Y.) were positioned at the beginning, middle and end of each processing batch, respectively. Samples within a batch were analyzed together.
Logistic Regression Model
The logistic regression classification method [27] was used to combine a panel of transitions into a classifier and to calculate a classification probability score between 0 and 1 for each sample. The probability score (Ps) of a sample was determined as Ps=1/[1+exp(−α−Σi=1Nβi*{hacek over (I)}i,s)], where {hacek over (I)}i,s was the logarithmically transformed (base 2), normalized intensity of transition i in sample s, βi was the corresponding logistic regression coefficient, a was a classifier-specific constant, and N was the total number of transitions in the classifier. A sample was classified as benign if Ps was less than a decision threshold. The decision threshold can be increased or decreased depending on the desired NPV. To define the classifier, the panel of transitions (i.e. proteins), their coefficients, the normalization transitions, classifier coefficient α and the decision threshold must be learned (i.e. trained) from the discovery study and then confirmed using the validation study.
Discovery of the Rule-Out Classifier
A summary of the 143 samples used for classifier discovery appears in Table 17 and processed as described above.
Protein transitions were normalized as described above. Transitions that were not detected in at least 50% of the cancer samples or 50% of the benign samples were eliminated leaving 117 transitions for further consideration. Missing values for these transitions were replaced by half the minimum detected value over all samples for that transition.
The next step was finding the set of most cooperative proteins. The cooperative score of a protein is the number of high performing panels it participates in divided by the number of such panels it could appear on by chance alone. Hence, a cooperative score above 1 is good, and a score below 1 is not. The cooperative score for each protein is estimated by the following procedure:
One million random panels of 10 proteins each, selected from the 117 candidates, were generated. Each panel of 10 proteins was trained using the Monte Carlo cross validation (MCCV) method with a 20% hold-off rate and one hundred sample permutations per panel) to fit a logistic regression model and its performance assessed by partial AUC [28].
By generating such a large number of panels, we sample the space of classifiers sufficiently well to find some high performers by chance. The one hundred best random panels (see Table 2) out of the million generated were kept and for each of the 117 proteins we determined how frequently each occurred on these top panels. Of the 117 proteins, 36 had frequency more than expected by chance, after endogenous normalizers were removed. (Table 22) The expected number of panels on which a protein would appear by chance is 100*10/117=8.33. The cooperative score for a protein is the number of panels it appears on divided by 8.33.
The 36 most cooperative proteins are listed in Table 22.
The set of 36 cooperative proteins was further reduced to a set of 21 proteins by manually reviewing raw SRM data and eliminating proteins that did not have robust SRM transitions due to low signal to noise or interference.
Proteins were iteratively eliminated from the set of 21 proteins until a classifier with the optimal partial AUC was obtained. The criteria for elimination was coefficient stability. In a logistic regression model each protein has a coefficient. In the process of training the model the coefficient for each protein is determined. When this is performed using cross validation (MCCV), hundreds of coefficient estimates for each protein are derived. The variability of these coefficients is an estimate of the stability of the protein. At each step the proteins were trained using MCCV (hold out rate 20%, ten thousand sample permutations per panel) to a logistic regression model and their stability measured. The least stable protein was eliminated. This process continued until a 13 protein classifier with optimal partial AUC was reached.
Finally, the 13 protein classifier was trained to a logistic regression model by MCCV (hold out rate 20%, twenty thousand sample permutations). The thirteen proteins for the rule-out classifier are listed in Table 18 along with their highest intensity transition and model coefficient.
Selection of a Decision Threshold
Assuming the cancer prevalence of lung nodules is prev, the performance of a classifier (NPV and ROR) on the patient population with lung nodules was calculated from sensitivity (sens) and specificity (spec) as follows:
The threshold separating calls for cancer or benign samples was then selected as the probability score with NPV≧90% and ROR≧20%. As we expect the classifier's performance measured on the discovery set to be an overestimate, the threshold is selected to be a range, as performance will usually degrade on an independent validation set.
Validation of the Rule-Out Classifier
52 cancer and 52 benign samples (see Table 17) were used to validate the performance of the 13 protein classifier. Half of the samples were placed in pre-determined processing batches analyzed immediately after the discovery samples and the other half of samples were analyzed at a later date. This introduced variability one would expect in practice. More specifically, the three HPS samples run in each processing batch were utilized as external calibrators. Details on HPS calibration are described below.
Calibration by HPS Samples
For label-free MS approach, variation on signal intensity between different experiments is expected. To reduce this variation, we utilized HPS samples as an external standard and calibrated the intensity between the discovery and validation studies. Assume that {hacek over (I)}i,s is the logarithmically transformed (base 2), normalized intensity of transition i in sample s, {hacek over (I)}i,dis and {hacek over (I)}i,val are the corresponding median values of HPS samples in the discovery and the validation studies, respectively. Then the HPS corrected intensity is
Ĩ
i,s
={hacek over (I)}
i,s
−{hacek over (I)}
i,val
+{hacek over (I)}
i,dis
Consequently, assume that the probability for cancer of a clinical sample in the validation study is predicted as prob by the classifier. Then the HPS corrected probability of cancer of the clinical sample is calculated as follows:
Here SHPS,dis and SHPS,val were the median value of S of all HPS samples in the discovery and validation studies, respectively.
Statistical Analysis
All statistical analyses were performed with Stata, R and/or MatLab.
Depletion Column Drift
We observed an increase of signal intensity as more and more samples were depleted by the same column. We used transition intensity in HPS samples to quantify this technical variability. Assuming Ii,s was the intensity of transition i in a HPS sample s, the drift of the sample was defined as
where Îi was the mean value of Ii,s among all HPS samples that were depleted by the same column and the median was taken over all detected transitions in the sample. Then the drift of the column was defined as
driftcol=median(drifts>0)−median(drifts<0).
Here the median was taken over all HPS samples depleted by the column. If no sample drift was greater or less than zero, the corresponding median was taken as 0. The median column drift was the median of drifts of all depletion columns used in the study.
Identification of Endogenous Normalizing Proteins
The following criteria were used to identify a transition as a normalizer:
Data Normalization
A panel of six normalization transitions (see Table 23) were used to normalize raw SRM data for two purposes: (A) to reduce sample-to-sample intensity variations within same study and (B) to reduce intensity variations between different studies. For the first purpose, a scaling factor was calculated for each sample so that the intensities of the six normalization transitions of the sample were aligned with the corresponding median intensities of all HGS samples. Assuming that Ni,s is the intensity of a normalization transition i in sample s and {circumflex over (N)}i the corresponding median intensity of all HGS samples, then the scaling factor for sample s is given by Ŝ/Ss, where
is the median of the intensity ratios and Ŝ is the median of Ss over all samples in the study. For the second purpose, a scaling factor was calculated between the discovery and the validation studies so that the median intensities of the six normalization transitions of all HGS samples in the validation study were comparable with the corresponding values in the discovery study. Assuming that the median intensities of all HGS samples in the two studies are {circumflex over (N)}i,dis and {circumflex over (N)}i,val, respectively, the scaling factor for the validation study is given by
Finally, for each transition of each sample, its normalized intensity was calculated as
Ĩ
i,s
=I
i,s
*R*Ŝ/S
s
where Ii,s was the raw intensity.
Isolation of Membrane Proteins from Tissues
Endothelial plasma membrane proteins were isolated from normal and tumor lung tissue samples that were obtained from fresh lung resections. Briefly, tissues were washed in buffer and homogenates were prepared by disrupting the tissues with a Polytron. Homogenates were filtered through a 180-μm mesh and filtrates were centrifuged at 900×g for 10 min, at 4° C. Supernatants were centrifuged on top of a 50% (w:v) sucrose cushion at 218,000×g for 60 min at 4° C. to pellet the membranes. Pellets were resuspended and treated with micrococcal nuclease. Membranes from endothelial cells were incubated with a combination of anti-thrombomodulin, anti-ACE, anti-CD34 and anti-CD144 antibodies, and then centrifuged on top of a 50% (w:v) sucrose cushion at 280,000×g for 60 min at 4° C. After pellets were resuspended, endothelial cell plasma membranes were isolated using MACS microbeads, treated with potassium iodide to remove cytoplasmic peripheral proteins.
Epithelial plasma membrane proteins from normal and tumor lung tissue samples were isolated from fresh lung resections. Tissues were washed and homogenates as described above for endothelial plasma membrane proteins preparation. Membranes from epithelial cells were labeled with a combination of anti-ESA, anti-CEA, anti-CD66c and anti-EMA antibodies, and then centrifuged on top of a 50% (w:v) sucrose cushion at 218,000×g for 60 min at 4° C. Epithelial cell plasma membranes were isolated using MACS microbeads and the eluate was centrifuged at 337,000×g for 30 minutes at 4° C. over a 33% (w:v) sucrose cushion. After removing the supernatant and sucrose cushion, the pellet was resuspended in Laemmli/Urea/DTT.
Isolation of Secreted Proteins from Tissues
Secreted proteins were isolated from normal and tumor lung tissue samples that were isolated from fresh lung resections. Tissues were washed and homogenized using a Polytron homogenization. The density of the homogenates was adjusted to 1.4 M with concentrated sucrose prior to isolating the secretory vesicles by isopycnic centrifugation at 100,000×g for 2 hr at 4° C. on a 0.8 and 1.2 M discontinuous sucrose gradient. Vesicles concentrating at the 0.8/1.2 M interface were collected and further incubated for 25 minutes with 0.5 M KCl (final concentration) to remove loosely bound peripheral proteins. Vesicles were recuperated by ultracentrifugation at 150,000×g for one hour at 4° C. and then opened with 100 mM ammonium carbonate pH 11.0 for 30 minutes at 4° C. Secreted proteins were recovered in the supernatant following a 1-hour ultracentrifugation at 150,000×g at 4° C.
Preparation of IgY14-SuperMix Immunoaffinity Columns
Immunoaffinity columns were prepared in-house using a slurry containing a 2:1 ratio of IgY14 and SuperMix immunoaffinity resins, respectively (Sigma Aldrich). Briefly, a slurry (10 ml, 50%) of mixed immunoaffinity resins was added to a glass chromatography column (Tricorn, GE Healthcare) and the resin was allowed to settle under gravity flow, resulting in a 5 ml resin volume in the column. The column was capped and placed on an Agilent 1100 series HPLC system for further packing (20 minutes, 0.15M ammonium bicarbonate, 2 ml/min). The performance of each column used in the study was then assessed by replicate injections of aliquots of HPS sample. Column performance was assessed prior to beginning immunoaffinity separation of each batch of clinical samples.
IgY14-Sumermix Immunoaffinity Chromatography
Plasma samples (60 μl) were diluted (0.15M ammonium bicarbonate, 1:2 v/v, respectively) and filtered (0.2 μm AcroPrep 96-well filter plate, Pall Life Sciences) prior to immunoaffinity separation. Dilute plasma (90 μl) was separated on the IgY14-SuperMix column connected to an Agilent 1100 series HPLC system using a three buffers (loading/washing: 0.15M ammonium bicarbonate; stripping/elution: 0.1M glycine, pH 2.5; neutralization: 0.01M Tris-HCl, 0.15M NaCl, pH 7.4) with a load-wash-elute-neutralization-re-equilibration cycle (36 minutes total time). The unbound and bound fractions were monitored using a UV absorbance (280 nm) and were baseline resolved after separation. Only the unbound fraction containing the low abundance proteins was collected for downstream processing and analysis. Unbound fractions were lyophilized prior to enzymatic digestion.
Enzymatic Digestion of Low Abundance Proteins
Low abundance proteins were reconstituted under mild denaturing conditions (200 μl of 1:1 0.1M ammonium bicarbonate/trifluoroethanol v/v) and allowed to incubate (30 minutes, room temperature, orbital shaker). Samples were then diluted (800 μl of 0.1M ammonium bicarbonate) and digested with trypsin (Princeton Separations; 0.4 μg trypsin per sample, 37° C., 16 hours). Digested samples were lyophilized prior to solid-phase extraction.
Solid-Phase Extraction
Solid phase extraction was used to reduce salt and buffer contents in the samples prior to mass spectrometry. The lyophilized samples containing tryptic peptides were reconstituted (350 μl 0.01M ammonium bicarbonate) and allowed to incubate (15 minutes, room temperature, orbital shaker). A reducing agent was then added to the samples (30 μl 0.05M TCEP) and the samples were incubated (60 minutes, room temperature). Dilute acid and a low percentage of organic solvent (375 μl 90% water/10% acetonitrile/0.2% trifluoroacetic acid) were added to optimize the solid phase extraction of peptides. The extraction plate (Empore C18, 3M Bioanalytical Technologies) was conditioned according to manufacturer protocol. Samples were loaded onto the solid phase extraction plate, washed (500 μl 95% water/5% acetonitrile/0.1% trifluoroacetic acid) and eluted (200 μl 52% water/48% acetonitrile/0.1% trifluoroacetic acid) into a collection plate. The eluate was split into two equal aliquots and each aliquot was taken to dryness in a vacuum concentrator. One aliquot was used immediately for mass spectrometry, while the other was stored (−80° C.) and used as needed. Samples were reconstituted (12 μl 90% water/10% acetonitrile/0.2% formic acid) just prior to LC-SRM MS analysis.
Inclusion and Exclusion Criteria
Plasma samples were eligible for the studies if they were (A) obtained in EDTA tubes, (B) obtained from subjects previously enrolled in IRB-approved studies at the participating institutions, and (C) archived, e.g. labeled, aliquotted and frozen, as stipulated by the study protocols. The samples must also satisfy the following inclusion and exclusion criteria:
Power Analysis for the Discovery Study
The power analysis for the discovery study was based on the following assumptions: 1) The overall false positive rate (α) was set to 0.05. 2) Ŝidák correction for multiple testing was used to calculate the effective αeff for testing 200 proteins, i.e.,
3) The effective sample size was reduced by a factor of 0.864 to account for the larger sample requirement for the Mann-Whitney test than for the t-test. 4) The overall coefficient of variation was set to 0.43 based on a previous experience. 5) The power (1−β) of the study was calculated based on the formula for the two-sample, two-sided t-test, using effective αeff and effective sample size. The power for the discovery study was tabulated in Table 24 by the sample size per cohort and the detectable fold difference between control and disease samples.
Power Analysis for the Validation Study
Sufficient cancer and benign samples are needed in the validation study to confirm the performance of the rule-out classifier obtained from the discovery study. We are interested in obtaining the 95% confidence intervals (CIs) on NPV and ROR for the rule-out classifier. Using the Equations in the Selection of a Decision Threshold section herein, one can derive sensitivity (sens) and specificity (spec) as functions of NPV and ROR, i.e.,
sens=1−ROR*(1−NPV)/prev,
spec=ROR*NPV/(1−prev),
where prev is the cancer prevalence in the intended use population. Assume that the validation study contains NC cancer samples and NB benign samples. Based on binomial distribution, variances of sensitivity and specificity are given by
var(sens)=sens*(1−sens)/NC
var(sens)=sens*(1−sens)/NB
Using the Equations in the Selection of a Decision Threshold section herein, the corresponding variances of NPV and ROR can be derived under the large-sample, normal-distribution approximation as
The two-sided 95% CIs of NPV and ROR are then given by ±zα/2√{square root over (var(NPV))} and ±zα/2√{square root over (var(ROR))}, respectively, where zα/2=1.959964 is the 97.5% quantile of the normal distribution. The anticipated 95% CIs for the validation study were tabulated in Table 24 by the sample size (NC=NB=N) per cohort.
Calculation of Q-Values of Peptide and Protein Assays
To determine the false positive assay rate the q-values of peptide SRM assays were calculated as follows. Using the distribution of Pearson correlations between transitions from different proteins as the null distribution (
The q-values of protein SRM assays were calculated in the same way except Pearson correlations of individual proteins were calculated as those between two transitions from different peptides of the protein. For proteins not having two peptides detected in five or more common samples, their q-values could not be properly evaluated and were assigned ‘NA’.
Impact of Categorical Confounding Factors
Impact of Continuous Confounding Factors
This application claims priority and benefit of U.S. Provisional Application No. 61/578,712 filed Dec. 21, 2011, U.S. Provisional Application No. 61/589,920 filed Jan. 24, 2012, U.S. Provisional Application No. 61/676,859 filed Jul. 27, 2012 and U.S. Provisional Application No. 61/725,153 filed Nov. 12, 2012, the contents of each of which are incorporated herein by reference in their entireties.
| Number | Date | Country | |
|---|---|---|---|
| 61578712 | Dec 2011 | US | |
| 61589920 | Jan 2012 | US | |
| 61676859 | Jul 2012 | US | |
| 61725153 | Nov 2012 | US |
