Predictive testing and classifier development approach for lung cancer patients who would benefit from PD-1 antibody drugs
Through mass spectrometry analysis and computer classifier methods, we predict the response of melanoma patients to block the PD-1 ligand activation antibody drugs, solving the problem of difficult prediction of treatment response in the prior art, and improving the targeted and economical treatment of treatment sex.
Patent Information
- Application Number
- CN202011108535.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Priority Date
- 2016-05-24
- Filing Date
- 2016-07-12
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2036-07-12
AI Technical Summary
The existing anti-PD-1 treatments are difficult to predict responses to melanoma patients, resulting in high-cost and unnecessary treatments, and the lack of effective prediction tools to select patients who may benefit from these treatments.
Through the combination of mass spectrometry analysis and computer classifiers, a method is developed to predict the response of melanoma patients to antibody drugs that block PD-1 ligand activation, using mass spectrometry characteristics in blood samples to generate category markers to guide Treatment plan.
Predictions that patients with melanoma may benefit from anti-PD-1 treatment are achieved, reducing the occurrence of unnecessary treatments and improving the targeted and economical treatment.
Smart Images

Figure CN112710723B_ABST
Abstract
Description
[0001] This application is a divisional application of a patent application with the original application date of July 12, 2016, application number 201680052342.5 (international application number PCT / US2016 / 041860), and invention name “Predictive testing and classifier development methods for melanoma patients who benefit from antibody drugs that block ligand activation of T cell programmed cell death 1 (PD-1) checkpoint protein”.
[0002] priority
[0003] This application claims priority to four U.S. provisional applications, namely, Serial Nos. 62 / 191,895, filed on July 13, 2015; 62 / 289,587, filed on February 1, 2016; 62 / 340,727, filed on May 24, 2016; and 62 / 319,958, filed on April 8, 2016. The contents of each of the four provisional applications (including their appendices) are incorporated herein by reference in their entirety.
[0004] field
[0005] The present invention relates to a method for predicting whether a cancer patient is likely to benefit from the administration of immune checkpoint inhibitors (including, for example, anti-PD-1 and / or anti-CTLA4 agents) before treatment, which allow the immune system to attack the tumor. The application also relates to a method for developing, i.e., training, a classifier implemented by a computer from a development sample set. background
[0006] Melanoma is a type of cancer that primarily affects the skin and develops from pigment-containing cells called melanocytes. The primary cause of melanoma is ultraviolet (UV) radiation exposure in people with low levels of skin pigment, which damages the DNA in skin cells. UV radiation can come from the sun or tanning devices. Approximately 25% of melanomas arise from moles. Individuals with many moles, a family history of the condition, or poor immune function are at risk of melanoma. A number of rare genetic defects also increase the risk of melanoma. Diagnosis of melanoma is usually made through a visual inspection of any concerning lesions, followed by a biopsy.
[0007] Melanoma is usually treated with surgical removal. Patients with larger cancers may be tested for spread to nearby lymph nodes. If no spread has occurred, most patients are cured. For patients whose melanoma has spread, immunotherapy, biological therapy, radiation therapy, or chemotherapy can improve survival. In the United States, with treatment, the five-year survival rate for patients with localized disease is 98%, while for patients who have spread, the five-year survival rate is only 17%. Melanoma is considered the most dangerous skin cancer. Globally, in 2012, it affected 232,000 people and resulted in 55,000 deaths.
[0008] Tumor mutations, including those associated with melanoma, generate specific neoantigens that can be recognized by the immune system. Approximately 50% of melanomas are associated with endogenous T cell responses. Cytotoxic T cells (CT cells or cytotoxic T lymphocytes (CTLs)) are white blood cells that destroy virus-infected cells and tumor cells and are also involved in transplant rejection. CTLs that express the CD8 glycoprotein on their surface are also called CD8+ T cells or CD8+ CTLs. Tumors have developed a variety of immune evasion mechanisms, including local immunosuppression in the tumor microenvironment, induction of T cell tolerance, and immunoediting. As a result, even if T cells infiltrate the tumor, they are unable to kill cancer cells. One example of this immunosuppression in cancer is mediated by a protein called programmed cell death 1 (PD-1), which is expressed on the surface of activated T cells. If another molecule called programmed cell death 1 ligand 1 or programmed cell death 1 ligand 2 (PD-L1 or PD-L2) binds to PD-1, the T cell becomes inactive. The production of PD-L1 and PD-L2 is one of the body's natural ways of regulating the immune system. Many cancer cells produce PD-L1, hijacking this natural system so that the cancer cells suppress T cells from attacking the tumor.
[0009] One approach to treating cancer is to interfere with inhibitory signals produced by cancer cells, such as PD-L1 and PD-L2, to effectively prevent tumor cells from restricting the immune system. Recently, the U.S. Food and Drug Administration approved a drug called nivolumab as Opdivo. Nivolumab is an anti-PD-1 monoclonal antibody marketed as a treatment for patients with unresectable or metastatic melanoma that has not responded to other drugs. In addition, nivolumab is approved for the treatment of squamous and non-squamous non-small cell lung cancer and renal cell carcinoma. Nivolumab is also approved for the treatment of melanoma in combination with the anti-cytotoxic T lymphocyte-associated protein 4 (CTLA4) antibody ipilimumab. Nivolumab acts as an immunomodulator by blocking ligand activation of the PD-1 receptor on activated T cells. In contrast to traditional chemotherapy and targeted anticancer therapies, which exert their effects through direct cytotoxicity or tumor growth inhibition, nivolumab works by blocking negative regulators of T cell activation and response, thereby allowing the immune system to attack the tumor. PD-1 blockade appears to unleash the immune system only around the tumor, rather than more broadly, which may reduce the side effects of these drugs.
[0010] Current clinical results with anti-PD-1 therapy in melanoma patients are encouraging, with overall results yielding superior progression-free and overall survival outcomes compared to alternative therapies. However, the real promise of these therapies lies in durable responses and long-term clinical benefit in a subgroup of approximately 40% of melanoma patients. A subset of the remaining approximately 60% of patients may do better with alternative therapies. Being able to select patients who derive minimal benefit from anti-PD-1 therapy from pre-treatment samples would enable better clinical understanding and enhance the development of alternative therapies for these patients. The costs associated with these therapies are also considerable. For example, the recently approved combination of ipilimumab and nivolumab for melanoma, while showing impressive results, is effective in only approximately 55% of patients and costs approximately $295,000 per course of treatment. (Leonard Saltz, MD, at the ASCO 2015 plenary session, stated: "The price of the Opdivo + Yervoy combination is approximately 4,000 times the price of gold ($158 / mg).") This results in an out-of-pocket cost of $60,000 for patients under standard health insurance plans. By limiting these treatments to patients who are likely to benefit from them, such costs could be avoided, resulting in substantial savings for the healthcare system and patients. It is currently unclear whether the benefit of the nivolumab and ipilimumab combination results from a synergistic effect or is simply the sum of different patient populations responding to nivolumab or ipilimumab. In any case, testing the benefit of nivolumab could clarify this issue.
[0011] Many studies have been conducted using PD-L1 expression measured by immunohistochemistry (IHC) as a biomarker for selecting anti-PD-1 therapy. Correlations between anti-PD-1 efficacy and outcome have been observed in some studies, but not in others. Particularly problematic is the current lack of standardized and universally accepted cutoff values for IHC staining, which makes comparison of these data difficult. A more significant issue is the observation that PD-L1 expression appears to be a dynamic marker, with IHC expression changing during tumor evolution and treatment. If PD-L1 expression obtained by IHC is used in a rigid manner, multiple repeat biopsies would be required, with correspondingly high risks and costs for the patient. In contrast, serum-based tests are not subject to these effects.
[0012] Assignee Biodesix has developed a classifier for predicting which patients may or may not benefit from certain anti-cancer drugs using mass spectrometry of blood samples. Representative patents include U.S. Patents 7,736,905, 8,914,238, 8,718,996, 7,858,389, 7,858,390, and U.S. Patent Application Publication Nos. 2013 / 0344111 and 2011 / 0208433.
[0013] Overview
[0014] In one aspect, and as will be described in more detail in Examples 9 and 10, a practical method for guiding the treatment of melanoma patients with immunotherapy drugs is disclosed. The method comprises the following steps: a) performing mass spectrometry analysis on a patient's blood sample and obtaining mass spectrometry data; (b) obtaining integrated intensity values of multiple mass spectrometry features in the mass spectrometry data; and (c) operating the mass spectrometry data using a program-controlled computer implementing a classifier. In the operating step, the classifier uses a classification algorithm to compare the integrated intensity value with the feature value of a reference set of class-labeled mass spectrometry data obtained from blood samples obtained from multiple other melanoma patients treated with antibody drugs that block ligand activation of programmed cell death 1 (PD-1), and generates a class label for the sample. The category label "good" or equivalent (e.g., late in the description of Example 10) predicts that the patient is likely to obtain similar benefits from the combination therapy comprising an antibody drug that blocks ligand activation of PD-1 and an antibody drug targeting CTLA4, and is therefore directed to monotherapy with the antibody drug that blocks ligand activation of PD-1 (e.g., nivolumab), while the category label "bad" or equivalent (e.g., early in the description of Example 10) indicates that the patient is likely to obtain greater benefits from the combination therapy than from the monotherapy with the antibody drug that blocks ligand activation of programmed cell death 1 (PD-1), and is therefore directed to the combination therapy.
[0015] In another embodiment, a method of treating a melanoma patient is disclosed. The method comprises performing the above method and, if the class marker is good or equivalent, administering to the patient a monotherapy of an antibody drug that blocks ligand activation of PD-1 (e.g., nivolumab), and if the class marker is reported as "bad" or equivalent, administering to the patient a combination therapy of an antibody drug that blocks ligand activation of PD-1 and an antibody drug that targets CTLA4 (e.g., a combination of nivolumab and ipilimumab).
[0016] In one embodiment, the mass spectrometric signature includes many of the features listed in Appendix A, Appendix B, or Appendix C or features associated with acute response to biological functions and wound healing (see Examples 1, 6, and 10). In a preferred embodiment, the classifier is obtained from filtering small classifiers combined using a regularized combination method, such as using Figure 8 or Figure 54The regularized combination method can take the form of repeatedly performing logistic regression with extreme value dropout on the filtered small classifiers. In one example, the small classifiers are filtered according to the criteria listed in Table 10. As disclosed in Example 9, the classifiers can take the form of clusters of tumor classifiers (each with a different proportion of patients with large tumors and small tumors) combined in a hierarchical manner. In the illustrated embodiment of Example 9, if any one of the tumor classifiers returns early stage or equivalent, then a bad or equivalent class label is reported, while if all tumor classifiers return late stage class labels, then a good or equivalent class label is reported.
[0017] In this approach, a relatively greater benefit from combination therapy signifies significantly greater (longer) overall survival compared to monotherapy.
[0018] In another aspect, the reference set takes the form of a class-labeled mass spectrometry dataset of a development sample set having an early or equivalent or a late or equivalent class label, wherein samples having the early class label include samples having relatively shorter overall survival when treated with nivolumab compared to samples having the late class label.
[0019] In a preferred embodiment, mass spectral data is acquired from at least 100,000 laser shots performed on a sample using MALDI-TOF mass spectrometry. The method is described in Example 1 and in the prior art patents cited in Example 1.
[0020] In one embodiment, as shown in Examples 6 and 10, mass spectrometric features are selected based on their association with at least one biological function, such as a set of features associated with the biological functions acute response and wound healing.
[0021] On the other hand, a practical test method is disclosed for predicting the response of melanoma patients to antibody drugs that block PD-1 ligand activation. The method comprises the following steps: a) performing mass spectrometry analysis on a blood sample of a melanoma patient and obtaining mass spectral data; (b) obtaining the integral intensity values of a plurality of predetermined mass spectral features in the mass spectral data; and (c) operating the mass spectral data using a program-controlled computer implementing a classifier. In the operation step, the classifier uses a classification algorithm to compare the integral intensity value obtained in step (b) with the feature values of a reference set of class-labeled mass spectral data obtained from a plurality of other melanoma patients treated with the drug, and generates a class label for the sample. The class label "early stage" or its equivalent predicts that the patient may obtain relatively less benefit from the antibody drug, while the class label "late stage" or its equivalent indicates that the patient may obtain relatively greater benefit from the antibody drug. The method for generating a classifier used in this test from a development sample data set is described in detail in the present disclosure.
[0022] On the other hand, a machine for predicting that melanoma patients benefit from ligand-activated antibody drugs that block programmed cell death 1 (PD-1) is described. The machine includes a memory that stores a reference set in the form of characteristic values of multiple mass spectral features obtained from mass spectrometry analysis of blood samples of multiple melanoma patients treated with antibody drugs. The memory further stores a set of codes defining a set of main classifiers, each of which is generated from multiple filtered small classifiers combined using a regularized combination method. The machine also includes a central processing unit, which operates on the reference set and mass spectral data obtained from the code set and the blood sample of a melanoma patient, and generates a class mark of the blood sample in response, wherein the class mark "early stage" or equivalent predicts that the patient may obtain relatively less benefit from the antibody drug, while the class mark "late stage" or equivalent indicates that the patient may obtain relatively greater benefit from the antibody drug.
[0023] In another aspect, a system for predicting that a patient will benefit from an antibody drug that blocks ligand activation of PD-1 is disclosed, the system being in the form of a mass spectrometer for performing mass spectrometric analysis on a patient's blood sample and a machine as described in the previous paragraph.
[0024] In another aspect, a method for generating a classifier for predicting whether a patient will benefit from an antibody drug that blocks ligand activation of programmed cell death 1 (PD-1) is disclosed. The method comprises the following steps:
[0025] 1) obtaining mass spectrometry data from a development set of blood samples obtained from melanoma patients treated with an antibody drug, wherein mass spectra of at least 100,000 laser shots are obtained from each member of the set;
[0026] 2) performing spectral preprocessing operations on mass spectrometric data from the development sample set, including background estimation and subtraction, alignment, batch correction, and regularization;
[0027] 3) Execution Figure 8 The process of steps 102-150 includes generating a main classifier based on a regularized combination of the filtered small classifier set;
[0028] 4) evaluating the performance of the main classifier generated according to step 3); and
[0029] 5) Define the final classifier based on the main classifier generated in step 3).
[0030] In a preferred embodiment, the final classifier includes a training set that includes feature values of the feature set listed in Appendix A, Appendix B, or Appendix C. In one possible embodiment, the method may include deselecting features that do not contribute to classifier performance from the feature list in Appendix A and performing steps 3), 4), and 5) using a reduced feature list. Such a reduced feature list may take the form of a feature list of the set in Appendix B or a feature list in Appendix C.
[0031] In another embodiment, a method for treating a melanoma patient is disclosed, comprising performing the method for predicting whether a patient will benefit from an antibody drug that blocks PD-1 ligand activation as described above, and administering the antibody drug to the patient if the patient has a late-stage or equivalent class marker in their blood sample.
[0032] On the other hand, an improved general-purpose computer is disclosed, which is configured as a classifier for classifying blood samples from human cancer patients so as to predict the relative likelihood of survival or benefit from a drug for the patient. The improvement is in the form of a memory, which stores a reference set in the form of characteristic values of multiple mass spectral features obtained from mass spectrometry analysis of blood samples of multiple melanoma patients treated with immune checkpoint inhibitors, and the associated category label of each blood sample in the reference set. The data of the blood sample forms a set for developing a classifier. The memory further stores a set of computer executable codes that define a final classifier based on multiple main classifiers, each main classifier being generated from a set of filtered small classifiers that execute a classification algorithm and are combined using regularization combination methods such as extreme value dropout and logistic regression. Multiple main classifiers are obtained by separating the development set into many different implementations of classifier training and test sets. The improvement also includes a central processing unit that operates on the code set, the reference set, and the mass spectral data obtained from the blood sample of the cancer patient to be tested, and generates a category label for the blood sample.
[0033] In one embodiment, the memory stores feature values for at least 50 features listed in Appendix A. In another embodiment, the memory stores feature values for a reduced set of features, such as features of one of the methods listed in Appendix B or the feature list of Appendix C.
[0034] In one embodiment, the immune checkpoint inhibitor comprises an antibody that blocks activation of the PD-1 ligand. In another embodiment, the immune checkpoint inhibitor comprises an antibody that blocks activation of the CTLA4 ligand.
[0035] In a still further aspect, a laboratory testing center is described that includes a mass spectrometer for performing mass spectrometric analysis of a blood sample from a cancer patient and a machine configured as a classifier and storing a reference set of mass spectrometric data as described herein.
[0036] A general extension of classifier development is further described in Example 9 below, to design a development set of classifier clusters for exploring different clinical groups, such as patients with different proportions of large and small tumors. In one embodiment, a method for generating a classifier cluster from a patient sample set is disclosed, comprising the following steps:
[0037] a. defining a plurality of classifier development sample sets from the patient sample set, wherein each sample set has different clinical characteristics (eg, the proportion of patients with large or small tumors or other relevant clinical groupings);
[0038] b. performing mass spectrometry analysis on the patient sample set and storing mass spectrometry data;
[0039] c. using a programmed computer, performing a classifier development exercise using the mass spectrometry data of each of the development sets defined in step a., and storing the parameters of the classifier thus generated in a memory associated with the computer, thereby generating a cluster of classifiers; and
[0040] d. Define a rule or set of rules for generating a class label for a test sample undergoing classification using the cluster of classifiers generated in step c. Examples 8 and 9 also disclose a method for testing a sample using the cluster of classifiers generated according to this method. In this method, step b. can be performed before or after step a.
[0041] Although the findings in melanoma and anti-PD-1 and anti-CTLA4 antibody drugs are described in detail in the Examples, the protein studies associated with the class markers described in detail below enable the generalization of these findings. Specifically, it can be expected that the classifiers of Examples 1, 2, and 3 may be relevant / applicable to a variety of drugs that affect the patient's immunological state, such as various immune checkpoint inhibitors, high-dose IL2, vaccines, and / or combination therapies, such as anti-PD-1 and anti-CTLA4 combination therapies. In addition, since the effects measured in serum reflect the state of the entire organism, and the complement system associated with these findings was found to affect innate and adaptive immunity not only in the tumor site, but also at a global level, the classifier is expected to have similar performance in different indications (e.g., lung cancer, kidney cancer), and is not limited to melanoma.
[0042] In another aspect, a method for generating a classifier is described, comprising the steps of:
[0043] a) obtaining physical measurement data from a development sample set and providing the measurement data to a general purpose computer, where each sample is further associated with clinical data;
[0044] b) identifying multiple different clinical subgroups 1...N in the development set based on clinical data;
[0045] c) for each different clinical subgroup, performing a classifier generation process based on the measurement data of each member of the development set associated with these clinical subgroups, thereby generating clinical subgroup classifiers C1...CN; and
[0046] d) storing in a memory of the computer a classification program involving all classifiers C1 . . . CN developed in step c), each classifier being associated with a reference set comprising the samples in the development set used to generate the classifier and the associated measurement data.
[0047] In another aspect, a multi-stage classifier is disclosed that includes a programmable computer that implements a hierarchical classifier structure that operates on mass spectral data of a test sample stored in a memory and utilizes a reference set of class-labeled mass spectral data stored in the memory. The classifier includes (a) a first-stage classifier that stratifies the test mass spectral data into an early or late group (or equivalent, the name is not important); (b) a second-stage classifier that further stratifies the early group of the first-stage classifier into an early and late group (or earlier and later groups, or equivalent), executing the second stage if the first-stage classifier classifies the test mass spectral data into the early group, and the early class label generated by the second-stage classifier is associated with an abnormally poor prognosis; and (c) a third-stage classifier that further stratifies the late group of the first-stage classifier into an early and late group (or earlier and later groups, or equivalent). If the first stage classifier classifies the test mass spectrometry data into the late stage group, a third stage classifier is implemented, wherein the late stage category signature (or more late stage or equivalent) generated by the third stage classifier is associated with a particularly good prognosis.
[0048] In one embodiment, the third-level classifier includes one or more classifiers developed from one or more different clinical subgroups of the classifier development set used to generate the first-level classifier. In one example, the third-level classifier includes at least four different classifiers C1, C2, C3, and C4, each classifier developed from a different clinical subgroup. In a specific embodiment, the multi-level classifier is configured to predict whether an ovarian cancer patient is likely or unlikely to benefit from platinum chemotherapy, and wherein the classifiers C1, C2, C3, and C4 are developed from the following clinical subgroups:
[0049] C1: developed from a subset of patients with non-serous histology or serous histology with unknown FIGO score;
[0050] C2: developed from a subset of patients with serous histology that was not used to develop classifier C1;
[0051] C3: developed from a subset of patients with residual tumor after surgery;
[0052] C4: Developed from a subset of patients who had no residual tumor after surgery.
[0053] In yet another aspect, a method has been discovered for generating a classifier for classifying test samples from a development sample set, each sample being associated with clinical data. The method comprises the following steps:
[0054] (a) dividing the development sample set into different clinical subgroups 1...N based on clinical data, where N is an integer of at least 2;
[0055] (b) Perform the classifier development process for each different clinical subgroup 1...N (e.g. Figure 8 process), thereby generating different classifiers C1...CN; and
[0056] (c) Define the final classification process, where the patient samples are classified by classifiers C1...CN.
[0057] In another aspect, a method for generating a classifier for classifying a test sample is found, comprising the steps of:
[0058] (a) generating a first classifier from the measurement data of the development sample set using a classifier development process;
[0059] (b) classifying the measurement data of the development sample set using the first classifier to assign each member of the development sample set a class label (early / late or equivalent) in a binary classification scheme; and
[0060] (c) generating a second classifier using a classifier development process, wherein the input classifier development set is members of the development set that are assigned one of the two class labels (e.g., the early group) in the binary classification scheme by the first classifier, so that the second classifier stratifies the members of the set having the first class label into two additional subgroups. The method optionally includes the following steps: (d) dividing the development sample set into different clinical subgroups 1...N, where N is an integer of at least 2; and (e) repeating the classifier development process for each different clinical group 1...N, thereby generating a different third classifier C1...CN; and (f) defining a hierarchical classification process, wherein:
[0061] i. The patient sample is first classified by the first classifier generated in step a);
[0062] ii. If the class label assigned by the first classifier is the class label used to generate the second classifier, the patient sample is classified using the second classifier; and
[0063] iii. If the class label assigned by the first classifier is not the class label used to generate the second classifier, the patient sample is classified using the third classifier C1 ... CN; and
[0064] iv. As a result of classification step ii or step iii, a final label is assigned.
[0065] Example 6 below describes the ability to correlate specific mass spectrometric features with functional groups of proteins circulating in serum and use this correlation to train a classifier or monitor changes in a biological process. In one embodiment, a method for training a classifier is disclosed, comprising the following steps:
[0066] a) obtaining a development sample set from a subject population, and optionally obtaining a second independent sample set from a similar, but not necessarily identical, subject population;
[0067] b) performing mass spectrometry analysis on the development sample set and optionally on the second sample set, and identifying mass spectrometry features present in the mass spectra of the sample sets;
[0068] c) for each sample in the development sample set or optionally each sample in the second sample set, obtaining protein expression data from a larger set of proteins covering the biological function of interest;
[0069] d) using a gene set enrichment analysis method, identifying statistically significant associations between one or more of the mass spectrometric features and combinations of proteins grouped according to their biological function; and
[0070] e) training a classifier based on the development sample set using one or more mass spectral features identified in step d) with computer assistance, wherein the classifier is in the form of a set of parameters that assign class labels to samples of the same type as the development sample set according to program instructions.
[0071] In one embodiment, the classifier is in the form of a combination of filtered small classifiers that are subjected to a regularization procedure.The samples in the development set and the optional second sample set are blood samples, such as serum or plasma samples from human patients.
[0072] On the other hand, a classifier development system is disclosed, comprising: a mass spectrometer for performing mass spectrometry analysis on a development sample set and, optionally, a second independent sample set to generate mass spectrometry data, the data comprising a plurality of mass spectrometry features; a platform for performing gene set enrichment analysis on the development sample set or an optional second independent sample set and identifying statistically significant associations of one or more mass spectrometry features with a group of proteins grouped according to their biological function; and a computer programmed to train a classifier for the development sample set using the one or more mass spectrometry features identified by the platform, the classifier being in the form of a set of parameters that assign class labels to samples of the same type as the development sample set according to program instructions. In a preferred embodiment, the development sample set and the optional second independent sample set are blood samples from humans. For example, the blood samples used in the development sample set are obtained from melanoma patients obtained prior to treatment with an immunotherapy drug, such as nivolumab. BRIEF DESCRIPTION OF THE DRAWINGS
[0073] Figure 1 shows the time to progression (TTP) in a cohort of 119 melanoma patients treated with nivolumab ( Figure 1A ) and overall survival (OS) ( Figure 1B ) where available clinical data and spectra were from pre-treatment samples.
[0074] Figure 2A and 2B Figure 2 shows Kaplan-Meier plots of time-to-event data (TTP and OS) for all 119 patients with available clinical data and spectra from the pre-treatment sample categorized by previous treatment (no previous ipilimumab use, previous ipilimumab use). The results were not statistically significant.
[0075] Figure 3A and 3B Kaplan-Meier plots of time-to-event data (TTP and OS) for all 119 patients with available clinical data and spectra from pre-treatment samples of patients in cohort 5 who showed relatively good outcomes, respectively.
[0076] Figure 4A and Figure 4B is a Kaplan-Meier plot showing time-to-event data for all 119 patients for whom available clinical data and spectra were from a pre-treatment sample ("DEV1") divided into development (N=60) and validation (N=59) sets; Figure 4C and 4DFigure 5 is a Kaplan-Meier plot of time-to-event data for all 119 patients for whom clinical data and spectra were available from a pretreatment sample that was split into development (N=60) and validation (N=59) sets, and a second split of the sample into development and validation sets ("DEV2").
[0077] Figure 5 is a plot of the interval regularization scalar as a function of disease control (DC); the interval method was used to compare the regularization scalars between clinical groups of interest to ensure that the window used for classification was not used for partial ion current regularization.
[0078] Figure 6 are mass spectra from multiple samples showing several feature definitions defined within the m / z range of interest; different clinical performance groups are shown with contrasting linear conventions.
[0079] Figure 7 is a plot of the interval regularization scalar as a function of DC for the partial ion current regularization performed on the features in the final feature table used for classifier generation.
[0080] Figures 8A-8B is a flow chart of the classifier development process used to develop the melanoma / nivolumab classifier of the present disclosure from a sample set of 119 serum samples from melanoma patients on a nivolumab trial.
[0081] Figure 9-12 shows the use of Figure 8 Kaplan-Meier plot of the classifier performance of the classifier developed in Example 1.
[0082] Figure 9A and 9B Kaplan-Meier plots of OS and TTP for the early and late category groups, respectively, are shown for the development set "Method 1" (see Table 10). Figure 9C and 9D Kaplan-Meier plots of OS and TTP are shown for the early and late category groups, respectively, for method 1 (see Table 10) of the validation set.
[0083] Figure 10A and 10B Kaplan-Meier plots of OS and TTP for the early and late category groups for the development set "Method 2" (see Table 10) are shown, respectively. Figure 10C and 10D Kaplan-Meier plots of OS and TTP by early and late category groups, respectively, are shown for the validation set of Approach 2 (see Table 10).
[0084] Figure 11A and11B Kaplan-Meier plots of OS and TTP for the early and late category groups, respectively, are shown for the development set of "Method 3" (see Table 10). Figure 11C and 11D Kaplan-Meier plots of OS and TTP are shown for the early and late category groups, respectively, for method 3 (see Table 10) of the validation set.
[0085] Figure 12A and 12B Kaplan-Meier plots showing OS and TTP for the early and late category groups for "Method 4" (see Table 10) for the development set, Figure 12C and 12D Kaplan-Meier plots showing OS and TTP for the early and late category groups for Method 4 (see Table 10) for the validation set.
[0086] Figure 13A and 13B The Kaplan-Meier plots of OS and TTP for the early and late category groups of the first approach shown in Table 13 are shown, respectively, which were applied as the first approach according to Figure 8 The classifier generates a set of 119 samples from the development set (100). Figure 13C and 13B The Kaplan-Meier plots of OS and TTP for the early and late category groups of the second method shown in Table 13 are shown, respectively. The second method was applied to the early and late category groups according to the second method. Figure 8 The classifier generates a set of 119 samples from the development set (100).
[0087] Figure 14 This is a Kaplan-Meier plot for the Yale cohort of patients treated with anti-PD-1 antibodies, which was used to validate the Figure 8 An independent sample set of the classifier of Example 1 was developed.
[0088] Figure 15 is an illustration of a laboratory testing center that includes machinery in the form of a mass spectrometer and a general-purpose computer used to predict which melanoma patients will benefit from an antibody drug that blocks ligand activation of PD-1.
[0089] Figure 16A and 16BKaplan-Meier plots of progression-free survival (PFS) and overall survival (OS) for all 173 non-small cell lung cancer (NSCLC) patients from the ACORN NSCLC cohort, respectively, for whom clinical data and spectra were available from pre-treatment samples. The ACORN NSCLC cohort was used to develop and tune the classifier of Example 2 to predict whether nivolumab would be beneficial for melanoma patients.
[0090] Figure 17A and 17B It is a figure of progression-free survival PFS and overall survival OS of three different subsets of ACORN NSCLC group.In the generation of the classifier of embodiment 2, a subset ("additional filtering" in the figure) is used to filter the small classifier.A subset ("test" in the figure) is used to test the classifier performance together with the melanoma development sample set.Another subset ("validation" in the figure) is used as an internal validation set, in addition to the melanoma subset that has been maintained for this purpose. Figure 17A and 17B The similarity in the plots indicates that the survival data for the “additional filtering” subset is representative of the other subsets.
[0091] Figures 18A-18H is the Kaplan-Meier plot of OS and TTP or PFS for the sample development and internal validation sets. Specifically, Figures 18A-18D Figure 2 is a plot of OS and TTP for the development and validation sets of the melanoma / nivolumab cohort. Note that Figures 18A-18D The separation of samples labeled as early and late by the classifier of Example 2 in terms of survival plots is shown. Figures 18E-18G ) are the OS and PFS plots for the development and validation sets of the ACORN NSCLC chemotherapy cohort. Note that for samples labeled as early and late stages by the classifier of Example 2, the plots in the ACORN NSCLC cohort ( Figures 18E-18G ) showed similar OS and PFS.
[0092] Figure 19 is a Kaplan-Meier plot of an independent validation cohort of patients treated with anti-PD-1 antibody (Yale anti-PD-1 cohort), showing separation of OS profiles of patient samples labeled as early-stage and late-stage by the classifier of Example 2.
[0093] Figure 20 is a Kaplan-Meier plot of an independent Yale validation cohort of melanoma patients treated with anti-CTLA4 antibodies, showing separation of OS profiles for patient samples labeled as early-stage and advanced-stage by the classifier of Example 2.
[0094] Figure 21A and 21BThe OS of an independent validation cohort of ovarian cancer patients treated with platinum-doublet chemotherapy after surgery was ( Figure 21A ) and disease-free survival (DFS, Figure 21B ). Figures 18E-18G Similar to the graph, Figure 21A and 21B The graph shows that there is no separation in OS and DFS of ovarian cancer patients whose samples are classified as early or late stage by the classifier of Example 2.
[0095] Figure 22 is a Kaplan-Meier plot of overall survival for the Yale anti-CTLA4 cohort.
[0096] Figure 23 is a Kaplan-Meier plot for the anti-CTLA4 antibody-treated Yale melanoma patient cohort for the "Encyclopedia" classifier of Example 1, see Table 13 below. Figure 23 and Figure 20 1 and 2 are similar in that both figures show that the classifiers of Example 1 and Example 2 of the present disclosure are able to predict whether melanoma patients will have relatively better or worse outcomes when treated with anti-CTLA4 antibodies. Figure 23A and 23B is a Kaplan-Meier plot of classifier performance of a second anti-CTLA4 antibody classifier developed from the Yale anti-CTLA4 cohort.
[0097] Figure 24A and 24B Kaplan-Meier plots of OS and PFS, respectively, for the ACORN NSCLC cohort used to develop the classifier of Example 2.
[0098] Figure 25A and 25B Kaplan-Meier plots of OS and PFS, respectively, for the ACORN NSCLC cohort, classified by the classification generated by the "full set" classifier of Example 1. Note that Figure 25A In , there is a clear separation in the overall survival plot between samples classified as early and late stages by the ensemble classifier of Example 1. Figure 25B The PFS separation between early and late category samples is also shown in .
[0099] Figure 26A and 26BFigure 1 is a Kaplan-Meier plot of the ACORN NSCLC cohort by classification and VeriStrat signature (assigned according to the classifier and training set of U.S. Patent 7,736,905). Note that there is essentially no separation between VeriStrat good (VS-G) and poor (VS-P) samples classified as early stage by the ensemble classifier of Example 1, and there is a clear separation in the survival plot between samples classified as late stage and samples classified as early stage by the ensemble classifier of Example 1.
[0100] Figure 27A and 27B Figure 5 is a Kaplan-Meier plot of OS and DFS for the ovarian cancer chemotherapy cohort of 138 patients used for internal validation of the classifier of Example 2.
[0101] Figure 28A and 28B The Kaplan-Meier plots of OS and DFS for the ovarian cancer chemotherapy cohort classified by the "full set" classifier of Example 1 are shown in Figure 1. Figure 28A In , there is a clear separation in the overall survival plot between samples classified as early and late by the ensemble classifier of Example 1. Figure 28B It also shows a clear separation in DFS between samples in the early and late classes.Thus, these figures show the ability of the ensemble classifier of Example 1 to predict survival of ovarian cancer on chemotherapy.
[0102] Figure 29A and 29B Figure 2 is a Kaplan-Meier plot of the ovarian cancer chemotherapy cohort by classification and VeriStrat labeling. Note that there is essentially no separation between samples classified as early stage by the full set classifier of Example 1 and tested as VeriStrat good (VS-G) and poor (VS-P), and there is a clear separation in the survival plot between samples classified as late stage by the full set classifier of Example 1 and samples classified as early stage. As with the NSCLC patient cohort, it is clear that within the group of patients classified as early stage, the results between the VeriStrat subgroups are similar.
[0103] Figures 30A-30F is the initial (before treatment, herein referred to as "baseline") Figure 30A and 30B As shown in the figure, after 7 weeks of treatment (WK7), Figure 30C and 30D As shown, and 13 weeks later (WK13) as Figure 30E and 30F Shown is a Kaplan-Meier plot generated by the full set classifier of Example 1, classified into early and late groups.
[0104] Figure 31A and 31B Figures are Kaplan-Meier plots of overall survival and time to progression (TTP) grouped by baseline, WK7, and WK13 ternary classifications, respectively, generated by the full set classifier of Example 1. Other marker combinations had too few patients for meaningful analysis.
[0105] Figure 32A and 32B Kaplan-Meier plots of overall survival and time to progression (TTP), respectively, showing the results when patients were grouped according to the baseline, WK7, and WK13 ternary classifications generated by the classifier of Example 2. (E = early, L = late)
[0106] Figure 33A and 33B Kaplan-Meier plot for a subset of 104 patients from the development sample set of Example 1, showing patients with early and late marking from the initial classification produced by the full set classifier of Example 1 ("Initial") and the classification produced by the new "large" and "small" tumor classifiers ("TS-adjusted").
[0107] Figure 34A and 34B Figure 1 is a Kaplan-Meier diagram of overall survival and time to progression (TTP) of 47 early stage patients (classified according to the complete set of method 1 classifiers of Example 1) that were classified as earlier stage and later stage groups using a classifier that used these 47 samples as its development set. Patients in the earlier stage category group were removed from the development sample set that generated the classifier, which took tumor size into account.
[0108] Figure 35A and 35B 34 ) are classified as early stage (their initial embodiment 1 full set method 1 classification). These groups are labeled "final" and compared with the groups generated by the embodiment 1 full set classifier ("initial").
[0109] Figure 36A and 36B is a waterfall plot showing class identification of melanoma patients who had increasing or decreasing tumor size during treatment with nivolumab. Figure 36AData are shown for a large and small tumor classifier that classifies members of the development sample set (after removing rapidly progressing patients) and patients with rapidly progressing disease, using available tumor size change data classified as early ("final"), whereas Figure 36B The data of the full set classifier of Example 1 is shown.
[0110] Figure 37 is a flowchart showing the process of developing small tumor and large tumor classifiers from a development sample set.
[0111] Figure 38 Is displayed based on Figure 37 Flowchart of how the resulting large and small tumor classifiers are used to test samples from cancer patients, depending on whether the patient has a large or small tumor.
[0112] Figure 39 is a flowchart illustrating a method for removing fast-progressing samples from a development set, Figure 37 Step 3702.
[0113] Figure 40A and 40B Kaplan-Meier plots of overall survival and time to progression, respectively, for the classifications generated by a cluster of seven classifiers, each based on a different classifier development set with different clinical groupings (in this example based on tumor size) and according to Figure 8 A cluster of classifiers was generated from the 119 patient samples described in Example 1. A set of rules defines the class labels of the labels generated by the classifier cluster, such as "poor," "good," and "other," which can be used to guide melanoma patient treatment, as explained in Example 9 below.
[0114] Figure 41 is a Kaplan-Meier plot of overall survival of the classification obtained by testing the cluster consisting of 7 classifiers of Example 9 for 30 samples in the anti-PD-1 treatment validation cohort.
[0115] Figure 42 Figure 8 is a Kaplan-Meier plot of overall survival of good and bad class signatures for melanoma patients treated with anti-PD1 monotherapy (nivolumab) and melanoma patients treated with ipilimumab and nivolumab combination therapy, generated by the cluster of 7 classifiers of Example 8. The plot shows that nivolumab patients with a good class signature had very similar survival compared to patients using nivolumab + ipilimumab combination therapy.
[0116] Figure 43 is targeted at a protein group S lThe running sum (RS) score (RS(S)) calculated in gene set enrichment analysis (GSEA) l,p ))'s graph example.
[0117] Figure 44A and 44B are examples of two defined null-point distributions of the enrichment score (ES) used in GSEA, showing the calculated ES and the regions evaluated to determine the p-value.
[0118] Figure 45A and 45B is a “heat map,” a plot of p-values generated by GSEA, that relates all 351 defined mass spectral features from the nivolumab study of Examples 1 and 6 (Appendix A) to protein functional groups. Figure 45A It is the heat map of ES definition 1, Figure 45B is a heatmap of ES definition 2. Only every 5th spectral feature is labeled on the x-axis.
[0119] Figure 46 This is an overview of the final classification procedure of Example 6.
[0120] Figure 47A and 47B The overall survival (OS) of the Melanoma / Immunotherapy New Classifier Development (NCD) cohort from the classification panel of Classifier 1 of Example 6 is shown ( Figure 47A ) and time to progression (TTP) ( Figure 47B ) Kaplan-Meier plot.
[0121] Figure 48A and 48B Kaplan-Meier plots of OS and TTP for the classification group of classifier 2 from Example 6 are shown for 68 samples not classified as “early stage” by classifier 1, respectively.
[0122] Figure 49A and 49B Shown respectively by Figure 46 Figure 4. Kaplan-Meier plots of OS and TTP for all 119 samples of the overall classification generated by the classifier.
[0123] Figures 50A-50D Shows the classifier developed in Example 6 ("Current") compared to Example 1 ( Figure 50A and 50B OS and TTP, respectively) ("IS2") and Example 8 ( Figure 50C and 50D Kaplan-Meier plots comparing the performance of the classifiers developed in
[15] for OS and TTP, respectively) ("IS6").
[0124] Figure 51 It is a Kaplan-Meier plot of OS according to classification group in the validation cohort of Example 6.
[0125] Figure 52 is a block diagram of a classifier training system comprising: a mass spectrometer for performing mass spectrometric analysis on a development sample set (not shown, such as serum or blood samples); a GSEA platform comprising a protein measurement system and a computer having a GSEA analysis module; and a computer programmed to perform a classifier training procedure using multiple sets of mass spectrometric peaks in the development sample set that are associated with specific protein functional groups of a biological process of interest.
[0126] Figure 53A and 53B The disease-free survival (DFS) of a cohort of 138 ovarian cancer patients with available clinical and mass spectrometry data was Figure 53A ) and overall survival (OS, Figure 53B ), which was used to develop the ovary classifier of Example 9.
[0127] Figure 54A and 54B is a flow chart illustrating a computer-implemented procedure for developing a classifier from a development sample set. In Example 9, the classifier is executed for different configurations or subsets of the development sample set. Figure 54A and 54B The procedure (up to and including step 350) is repeated different times to result in the creation of a hierarchical or graded series of classifiers (referred to as classifiers A, B, and C), which will be explained in more detail in the description of Example 9.
[0128] Figure 55A and 55B is a Kaplan-Meier plot of event time data for 129 patients in the ovarian development sample set of Example 9, who had available clinical data, DFS>1 month, and mass spectrometry data from pre-treatment samples, showing the division of the sample set into development (N=65) and validation (N=64) sets. Figure 55A The DFS plot is shown; Figure 55B OS plots are shown. Note that the plots for the development and validation sample sets are essentially identical.
[0129] Figures 56A-56D is a Kaplan-Meier plot of OS and DFS for early and late category groups generated by the first tier or "Classifier A" classifier of Example 9 for 129 patients divided into development and validation sets. Figure 56A is the OS graph of the development set; Figure 56B is the DFS graph of the development set; Figure 56Cis the OS graph of the validation set; Figure 56D is the DFS graph of the validation set.
[0130] Figure 57A and 57B is a Kaplan-Meier plot of OS and DFS for the early and late category groups for "Classifier A" of Example 9 run on all 138 samples.
[0131] Figure 58A and 58B Kaplan-Meier plots of OS and DFS for the classification group generated by Classifier B of Example 9, respectively, for the subset of the development sample set for developing Classifier B.
[0132] Figure 59 1 is a flowchart showing the process of generating a second-layer classifier ("Classifier B") from those development set samples classified as "early" by the first-layer "Classifier A" classifier of Example 9.
[0133] Figure 60 is a flow chart illustrating the process of generating the third layer classifier C of Example 9; in this particular instance, the third layer consists of several different classifiers, each based on a different and clinically unique subset of the development sample set.
[0134] Figure 61 is a diagram showing the structure of the third layer classifier C and how a “good” class label can be generated based on the classification results of each member of the third layer.
[0135] Figure 62 is a diagram or outline showing the structure of a final classifier composed of three levels of hierarchical classifiers.
[0136] Figure 63 is a diagram or summary illustrating the structure of an alternative final classifier, in which the third level of a three-level hierarchical classifier consists of four individual classifiers developed from clinically distinct subgroups.
[0137] Figure 64A and 64B Use Figure 63 Kaplan-Meier plots of OS and DFS for the classification groups generated by the final classifier structure of Example 9 for the development sample set.
[0138] Figures 65A-65C It is an acute response to biological function ( Figure 65A ), wound healing ( Figure 65B ) and the complement system ( Figure 65C), mean percentage of variance of the “PROSE-erlotinib” data set explained by each principal component (PC) as a function of the PC index (variance descending) from Example 10.
[0139] Figures 66A-66D is the distribution of acute reaction scores in the four sample sets described in Example 10. Figures 66A-66D In the results shown, twenty-nine mass spectrometric features were identified as being associated with AR biological functions and used to calculate the corresponding scores.
[0140] Figures 67A-67D is the distribution of acute response scores in four sample sets: “PROSE-erlotinib” (85 samples with available IS2 markers), Figure 67A ; "PROSE-chemotherapy" (contains 122 samples with available IS2 markers), Figure 67B ; "Moffitt", Figure 67C ; and "Moffitt-7 Weeks," Figure 67D . Figures 67A-67D Also shown are the scores for samples from the Example 1 classifier divided by IS2 class labeling.
[0141] Figures 68A-68D This is the Kaplan-Meier graph of OS and PFS for the PROSE-erlotinib sample set. Figures 68A-68B , Kaplan-Meier plots of OS and PFS in the “PROSE-chemotherapy” set, Figures 68C-68D is the Kaplan-Meier plot of OS and TTP in the "Moffitt" set, Figures 68E-68F The groups were defined according to the AR score thresholds defined as shown in the figures. The corresponding number of samples in each group, hazard ratio (HR), log-rank p-value and median are shown below each figure.
[0142] Figures 69A-69D is a graphic representation of the evolution over time of acute response scores for 107 patients in the Moffitt and Moffitt-7 week collections grouped by combined IS2 markers at baseline (before treatment) and week 7 (after treatment). Each line in the graph represents the score for a single patient. Figure 69A is a plot of patients with early category markers at baseline and week 7; Figure 69B is a plot of patients with late category markers at baseline and week 7; Figure 69C is a plot of patients with early category markers at baseline and late category markers at week 7, and Figure 69D is a graph of patients with late category markers at baseline and early category markers at week 7.
[0143] Figures 70A-70Cis a graph showing the evolution of acute response scores for 107 patients with samples from the "Moffitt" and "Moffitt-7 weeks" collections, grouped according to treatment response. Each line represents one patient. Figure 70A It is a graph of PD: progressive disease treatment response; Figure 70B is a graph of PR: partial response to treatment, Figure 70C is a graph of SD: stable disease treatment response.
[0144] Figure 71A and 71B Kaplan-Meier plots of OS and TTP for samples in the "Moffitt" and "Moffitt-7 weeks" collections, grouped according to whether the AR score increased or had a small change or decrease, respectively. The figure shows that the change in AR score has prognostic significance.
[0145] Figures 72A-72D is a distribution graph of the wound healing scores of the four sample sets described in Example 10.
[0146] Figures 73A-73D is a graphic representation of the evolution over time of wound healing scores for 107 patients in the Moffitt and Moffitt-7 week collections grouped by combined IS2 markers at baseline (before treatment) and week 7 (after treatment). Each line in the graph represents the score for a single patient. Figure 73A is a plot of patients with early category markers at baseline and week 7; Figure 73B is a plot of patients with late category markers at baseline and week 7; Figure 73C is a plot of patients with early category markers at baseline and late category markers at week 7, and Figure 73D is a graph of patients with late category markers at baseline and early category markers at week 7.
[0147] Figures 74A-74C Figure 2 is a graph showing the evolution of wound healing scores for 107 patients with samples from the "Moffitt" and "Moffitt-7 weeks" collections, grouped according to treatment response. Each line represents one patient. Figure 74A It is a graph of PD: progressive disease treatment response; Figure 74B is a graph of PR: partial response to treatment, Figure 74C is a graph of SD: stable disease treatment response.
[0148] Figures 75A-75D is a distribution graph of the complement system scores for the four sample sets described in Example 10.
[0149] Figures 76A-76Dis a graphical representation of the change in complement system scores over time for 107 patients in the Moffitt and Moffitt-7 week collections grouped by combined IS2 markers at baseline (before treatment) and week 7 (after treatment). Each line in the graph represents the score for an individual patient. Figure 76A is a plot of patients with early category markers at baseline and week 7; Figure 76B is a plot of patients with late category markers at baseline and week 7; Figure 76C is a plot of patients with early category markers at baseline and late category markers at week 7, and Figure 76D is a graph of patients with late category markers at baseline and early category markers at week 7.
[0150] Figures 77A-77C Figure 2 is a graph showing the evolution of wound healing scores for 107 patients with samples from the "Moffitt" and "Moffitt-7 weeks" collections, grouped according to treatment response. Each line represents one patient. Figure 77A It is a graph of PD: progressive disease treatment response; Figure 77B is a graph of PR: partial response to treatment, Figure 77C is a graph of SD: stable disease treatment response.
[0151] Figures 78A-78D is OS, PFS ("PROSE-chemotherapy"), Figures 78A-78B is the Kaplan-Meier plot of OS and TTP (“Moffitt” set), Figure 78C and 78D The groups are defined by the AR score thresholds defined by tertiles in the PROSE set. The corresponding number of samples in each group, hazard ratio (HR), log-rank p-value and median are shown below each figure.
[0152] Figure 79A and 79B It is based on Figure 8 Kaplan-Meier plots of the classification groups early and late obtained with the classifier developed by the program of Example 10, but instead of using mass spectrometry features, the features used for classification are the biological function scores of Example 10, in this case the acute response score, the wound healing score and the complement score.
[0153] Figure 80 is a schematic diagram of the system for generating the biological function score of Example 10 for a sample set.
[0154] Figure 81 is a flow chart showing the steps for calculating the biological function score of Example 10 from mass spectrometry data.
[0155] Figure 82Table F is a partial feature table of the sample set ss used to calculate the biological function score of Example 10. ss icon.
[0156] Figure 83 is the average first principal component vector used to calculate the biological function score of Example 10 icon.
[0157] Details
[0158] Disclosed in this document is a practical test (method) for predicting whether a cancer patient is likely to benefit from the administration of an immune checkpoint inhibitor, such as a monoclonal antibody drug that blocks the activation of a ligand of PD-1 on activated T cells, such as nivolumab. The method utilizes a mass spectrometer of a serum or plasma sample of a patient obtained before treatment, and a general-purpose computer configured as a classifier that assigns a class label to the mass spectrometer. The class label can take the form of "early stage" or an equivalent, or "late stage" or an equivalent, with the class label "late stage" indicating that the patient is a member of a class of patients who are likely to receive a relatively greater benefit than members of a class of patients having an "early stage" class label. The specific names used for the class labels are not particularly important.
[0159] Overall survival is the primary indicator for evaluating the benefit of antibody drugs that block ligand activation of PD-1. Therefore, when considering the meaning of early and late markers, in a preferred embodiment, a "relatively greater benefit" associated with a late marker means that patients whose samples are assigned to the late marker are likely to have significantly greater (longer) overall survival compared to patients with early category markers.
[0160] The term "antibody drug that blocks ligand activation of PD-1" is intended to include not only antibodies that bind to this PD-1, but also those that bind to the ligands (PD-L1 and PD-L2). Anti-PD-L1 or anti-PD-L2 monoclonal antibodies (mAbs) also block ligand activation of PD-1. The term "immune checkpoint inhibitor" is intended to include inhibitors that block PD-1 as well as inhibitors that block CLTA-4. The term immune checkpoint inhibitor or "checkpoint blocker" is defined as an immunomodulatory mAb that targets CTLA4-like receptors and their ligands. See Galluzzi L, Kroemer G, Eggemont A. Novel immune checkpoint blocker approved for the treatment of advanced melanoma. Oncoimmunology 2014; 3: e967147. FDA-approved drugs of this type include ipilimumab (anti-CTLA4), nivolumab (anti-PD-1), and pembrolizumab (anti-PD-1). Several publications / abstracts published in the past 13 months report on the involvement of additional checkpoint blockers (e.g., CTLA4-targeting mAb tremelimumab, PD-1-targeting mAb pidilizumab, and PD-L1-targeting mAbs MEDI4736 (durvalumab), MPDL3280A (atezolizumab), and MSB0010718C (avelumab). See Buqué A, Bloy N, Aranda F, Castoldi F, Eggermont A, Cremer I, Fridman WH, Fucikova J, Galon J, Marabelle A, Spisek R, Tartour E, Zitvogel L, Kroemer G, Galluzzi L. Trial Watch: Immunomodulatory monoclonal antibodies for oncological indications Oncoimmunology. 2015 Mar 2; 4(4). The contents of the above-cited scientific publications are incorporated herein by reference.
[0161] Example 1 explains the development of a classifier from the melanoma / nivolumab sample set and provides details of the classifier performance.
[0162] Example 2 explains the redevelopment of Example 1, where the classifier was adjusted to be more predictive and less prognostic for patients who will benefit from nivolumab.
[0163] Example 3 explains the development of a classifier that predicts melanoma patients who will benefit from another immune checkpoint inhibitor, anti-CTLA4 antibody.
[0164] Example 4 explains how the classifier developed according to Example 1 can also predict whether non-small cell lung cancer (NSCLC) and ovarian cancer patients are likely to have relatively higher or lower overall survival from chemotherapy.
[0165] Example 5 and Figure 15 A practical testing environment is described for performing the disclosed test on a blood sample from a patient prior to treatment.
[0166] Example 6 describes the study of proteins associated with early and late class markers in the classifiers of Example 1 and Example 2 using gene set enrichment analysis, which allows the generalization of these findings to other immune checkpoint inhibitors and other types of cancer, and the generation of classifiers based on mass spectrometric features associated with specific functional groups of proteins.
[0167] Example 7 describes a longitudinal study of patient samples from Example 1 and, in particular, how changes in the class labels produced by the classifier can be used to monitor treatment efficacy and guide therapy.
[0168] Example 8 describes clusters of classifiers generated from different clinical subsets of the development sample population of Example 1, in this example, melanoma patients with large and small tumors.
[0169] Example 9 further provides further examples of developing classifier clusters from clinically diverse development sets, including a first example of a classifier cluster from the nivolumab / melanoma set of the Examples and a second example from a study of ovarian cancer patients treated with chemotherapy.
[0170] Example 10 describes a method for measuring a biological function score using mass spectrometry data. Example 10 further describes measuring a biological function score in four different sample sets, each blood sample from a person with cancer. Example 10 describes how to use the score to guide treatment and use the biological function score as a feature for classifier training to build a classifier, such as using Figure 8 Example 10 also builds upon the findings described in Example 6, including the correlation of biological functions with mass spectrometry peaks.
[0171] Example 1
[0172] Classifiers and methods for predicting melanoma patients who will benefit from antibody drugs that block ligand activation of PD-1
[0173] The classifier of Example 1 was developed using samples obtained from a clinical trial of nivolumab for the treatment of melanoma. This trial is briefly described below. The patient samples obtained are then described, along with the mass spectrometry method for obtaining spectra from the samples, including a spectral processing step. These mass spectrometry procedures are preferably based on the so-called "deep MALDI" method described in U.S. Patent No. 9,279,798, the contents of which are incorporated herein by reference. A classifier generation process is then described in detail that is used to define the final classifier, which can assign class labels to spectra based on the test. The results of the classifier generation process are also described below, and it has been demonstrated that it is capable of assigning class labels to mass spectra from blood samples, thereby predicting whether the patient providing the sample is likely to obtain a relatively large or small benefit from the antibody drug.
[0174] The classifier generation method uses what is called "Combination of Small Classifiers with Dropout Regularization" or CMC / D, which is described in co-pending U.S. patent application Ser. No. 14 / 486,442, filed Sep. 15, 2014, published as U.S. Patent Application Publication No. 2015 / 0102216, the contents of which are incorporated herein by reference. This program for developing classifiers is also referred to herein as DIAGNOSTIC CORTEX, a trademark of Biodesix, Inc. See below. Figure 8 By applying this procedure to pre-treatment spectra of serum from melanoma patients obtained using deep MATLDI spectral acquisition, "early" and "late" clinical groups were identified. Following treatment with the anti-PD-1 therapy nivolumab, these groups showed significant differences in outcomes (time to progression (TTP) and overall survival (OS)). See Figures 9-14 and the discussion below. A testing procedure has been proposed to identify these groups from pre-treatment samples and to validate the results in internal and external validation sets (see Figure 14 ).
[0175] Patients whose serum classification was "early" showed significantly faster progression and shorter survival compared with patients whose serum classification was "advanced," making this test suitable as a biomarker for nivolumab therapy. The "early" and "advanced" clinical groups were not associated with PD-L1 expression, and this classification remained a significant predictor of outcome (TTP and OS) even when other clinical characteristics were included in multivariate analysis.
[0176] While the methods associated with test development do not provide insight into the biology that could explain the differences between the identified groups, we have conducted some preliminary studies that link the two distinct groups to differences in acute phase reactants and the complement system. These studies are described in Example 1 and subsequently in Example 6 of this document. The success of this project exemplifies how deep MALDI spectral acquisition combined with our innovative classifier development approach can lead to the construction of a clinically useful, practical test.
[0177] clinical trials
[0178] The trial from which samples were used for the development of the new classifier was a study of nivolumab with or without a peptide vaccine in patients with unresectable stage III or IV melanoma. Melanoma, described in a paper in J. Clin. Oncol. vol. 31 pp. 4311-4318 (2013), the contents of which are incorporated herein by reference. The patients participating in the trial experienced progression after previously receiving at least one treatment, but had not previously received PD-1 or PD-L1 treatment. The trial consisted of 6 patient groups. Groups 1-3 enrolled patients who had not received ipilimumab treatment, while patients in groups 4-6 had previously progressed after receiving ipilimumab therapy. In addition to nivolumab, groups 1-5 received a peptide vaccine, and group 6 received nivolumab alone. Group 5 only enrolled patients who experienced grade 3 dose-limiting toxicity when receiving ipilimumab treatment, while patients in groups 4 and 6 only experienced the highest grade 2 dose-limiting ipilimumab toxicity. The doses of nivolumab in groups 1-3 were different (1 mg / kg, 3 mg / kg, or 10 mg / kg). There was no limit on the number of previous treatment regimens. All patients had an ECOG performance status (PS) of 0-1.
[0179] For the purposes of this new classifier development project, the dose of nivolumab and whether the peptide was administered with nivolumab were not considered to be very important.
[0180] sample
[0181] The samples available for this study were pre-treatment serum samples. Clinical data and spectra were available for 119 patients. Available outcome data included time to progression (TTP), overall survival (OS), and response. Table 1 lists the baseline clinical characteristics of patients for whom spectra from pre-treatment samples were available. (Lactate dehydrogenase (LDH) is known to have prognostic significance in many cancer types and is often used as an important factor in assessing prognosis in melanoma patients.)
[0182] Table 1: Baseline characteristics of patients with available spectra
[0183]
[0184]
[0185] *Further data review required; See Appendix D of the previous provisional application for details.
[0186] x Not suitable for one patient,
[0187] xx ULN = upper limit of normal range
[0188] Figure 1A and 1B Kaplan-Meier plots of time to progression (TTP) and overall survival (OS) are shown for the cohort of 119 patients with spectra obtained for baseline samples and pre-treatment samples, respectively. Note: Of the 14 patients on the balanced segment of the TTP Kaplan-Meier plot, 3 (21%) had an objective response of SD rather than PR.
[0189] Figure 2A and 2B Kaplan-Meier plots of time-to-event data are shown for all 119 patients with available clinical data and spectra from pretreatment samples, grouped by previous treatment (not previously ipilimumab-naive; "Ipi-naive" vs. previously ipilimumab-naive). The results were not statistically significant.
[0190] Figure 3A and 3BKaplan-Meier plots of time-to-event data (TTP and OS) for all 119 patients with available clinical data and spectra from pre-treatment samples, respectively, show particularly good outcomes for patients in cohort 5, both in absolute terms and compared with patients who did not receive ipilimumab and other patients previously treated with ipilimumab. It should be recalled that cohort 5 in the nivolumab study involved patients who had progressed on or after ipilimumab treatment and enrolled only patients who had experienced grade 3 dose-limiting toxicity while they were previously receiving ipilimumab.
[0191] The relatively large number of samples obtained from the study (119) allowed the samples to be divided into a development set and an internal validation set for classifier development. Two different partitioning methods were studied. The first partitioning method (referred to as "DEV1" in the following discussion) was to stratify by VeriStrat-like classification, response, review for TTP, and TTP. The second partitioning method (referred to as "DEV2" in the following discussion) was to stratify by group, VeriStrat-like classification, response, review for TTP, and TTP. ("VeriStrat-like classification" refers to the use of the classification algorithm and training set of the VeriStrat test described in U.S. Patent 7,736,905 to assign "good" or "poor" category labels to mass spectra. The assignment of individual samples to the validation set or development set is listed in Appendix E of the previous provisional application serial number 62 / 289,587. The clinical characteristics used for the development and validation partitioning are listed in Tables 2A and 2B and are shown in Tables 2A and 2B. Figures 4A-4D Comparison of event time data between the development and validation sets is shown in .
[0192] Table 2A: Baseline Characteristics of Patients with Available Spectra Divided into Development and Internal Validation Sets ("DEV1")
[0193]
[0194]
[0195] Table 2B: Baseline Characteristics of Patients with Available Spectra Divided into Development and Internal Validation Sets ("DEV2")
[0196]
[0197] Figure 4A 、 4B Kaplan-Meier plots of time to progression (TTP) and overall survival (OS) for the development and validation sets are shown in 4C and 4D. Specifically, Figure 4A and 4BTime-to-event data for all 119 patients with available clinical data and spectra from pretreatment samples, divided into development (N=60) and validation (N=59) sets according to the first partitioning method (“DEV1”) are shown; Figure 4C and 4D Shown are event time data for all 119 patients with available clinical data and spectra from pre-treatment samples, divided into development (N=60) and validation (N=59) sets according to the second partitioning method ("DEV2").
[0198] Sample preparation
[0199] Serum samples were thawed and 3 μl aliquots of each test sample (from patients treated with nivolumab) and quality control serum (a pooled sample obtained from sera of 5 healthy patients, purchased from ProMedDx, "SerumP3") were dropped onto VeriStrat™ cellulose serum cards (Therapak). The cards were allowed to dry for 1 hour at ambient temperature, and then all serum spots were punched out using a 6 mm skin biopsy punch (Acuderm). Each punch was placed into a centrifugal filter with a 0.45 μm nylon membrane (VWR). 100 μl of HPLC grade water (JT Baker) was added to the centrifugal filter containing the punch. The punch was gently vortexed for 10 minutes and then centrifuged at 14,000 rcf for 2 minutes. The effluent was removed and transferred back to the punch for a second round of extraction. For the second round of extraction, the punches were gently vortexed for three minutes and then centrifuged at 14,000 rcf for two minutes. Twenty microliters of filtrate from each sample was then transferred to a 0.5 ml Eppendorf tube for MALDI analysis.
[0200] All subsequent sample preparation steps were performed in a custom-designed humidity and temperature controlled chamber (Coy Laboratory), with the temperature set at 30°C and relative humidity at 10%.
[0201] An equal volume of freshly prepared matrix (25 mg of sinapinic acid per 1 ml of 50% acetonitrile:50% water plus 0.1% TFA) was added to each 20 μl serum extract, and the mixture was vortexed for 30 seconds. The first three aliquots (2 x 2 μl) of the sample:matrix mixture were discarded into the cap of the tube. Eight 2 μl aliquots of the sample:matrix mixture were then dropped onto a stainless steel MALDI target plate (SimulTOF). The MALDI target was allowed to dry in the chamber before placement in the MALDI mass spectrometer.
[0202] This sample set was processed in three batches for MALDI analysis. QC samples were added at the beginning (two preparations) and end (two preparations) of each batch run.
[0203] Spectral acquisition
[0204] MALDI spectra were obtained using a MALDI-TOF mass spectrometer (SimulTOF 100s / n:LinearBipolar 11.1024.01 from Virgin Instruments, Sudbury, MA, USA). The instrument was set to operate in positive ion mode using ions produced by a 349nm, diode-pumped, frequency-tripled Nd:YLF laser operating at a laser repetition rate of 0.5kHz. External calibration was performed using a standard protein mixture (Bruker Daltonics, Germany) consisting of insulin (m / z 5734.51Da), ubiquitin (m / z, 8565.76Da), cytochrome C (m / z 12360.97Da) and myoglobin (m / z 16952.30Da).
[0205] As the stage moved at a speed of 0.25 mm / s and the laser illuminated the spot continuously, the spectrum from each MALDI spot (8 spots per sample) was collected as 800 emission spectra, which were "hardware averaged." A minimum intensity threshold of 0.01 V was used to discard any "flat line" spectra. All 800 emission spectra with intensities above this threshold were acquired without any further processing.
[0206] MALDI-TOF mass spectrometry data acquisition and processing (for the purpose of obtaining a set of data for classifier development and performing tests on samples for the purpose of determining patient benefit) is optionally based on The so-called "deep MALDI" method described in U.S. Patent No. 9,279,798 to et al., the contents of which are incorporated herein by reference, is performed. The '798 patent describes the surprising discovery that collecting many laser shots (typically 100,000 to 500,000 or more) from the same MALDI spot and averaging these shots or combining the accumulated spectra from multiple spots of the same sample results in a reduction in the relative levels of noise to signal and the revealing of a wealth of additional spectral information from the mass spectra of complex biological samples. This document also shows that hundreds of thousands of shots can be run for a single spot before the protein content of the spot is completely depleted. Second, the reduction of noise by averaging many shots results in the appearance of previously invisible peaks (i.e., peaks that were not apparent in the spectrum produced by a typical 1,000 laser shots). Even previously visible peaks become more sharply defined, and this allows for more reliable measurements of peak intensities and comparisons between samples when the sample is subjected to very many shots. The classifier of the present disclosure utilizes the deep MALDI method to deeply interrogate the proteome of a serum sample and uses a relatively large number of peaks (hundreds) for classification that would otherwise be unnoticeable in a spectrum obtained from a typical 1,000 laser shots. These peaks are not observable in conventional "dilution and illumination" spectra obtained by illumination mass spectrometry. See, for example, Appendix A for definitions of classification feature values.
[0207] The following sections of this document describe the spectral processing used on the initial spectrum from the mass spectrometer to construct the feature table used for classifier generation. The following procedures are executed in software on a general-purpose computer that receives the spectrum from the mass spectrometer. Some steps, such as defining the features used for classification, can be performed partially or entirely by a human operator by examining a plot of the mass spectral data.
[0208] Spectral processing
[0209] Grid spectral preprocessing
[0210] Recalibration
[0211] Instrument calibration can cause dramatic shifts in peak positions (mass (m) / charge (z) = m / z) that are most pronounced between batches in the high mass region. This makes it impossible to consistently use predefined workflows to process data that rely on peak positions and set alignment tolerances. To overcome this problem, a recalibration of the m / z data that requires a standard reference spectrum can be performed. This standard is compared to the spectrum of the current batch to determine if there has been a change in the position of the common serum peaks. The m / z positions are borrowed from the reference, but any "shifts" are applied to recalibrate the spectra. The result is a recalibrated spectrum with comparable m / z between batches. In a sense, this is a batch correction procedure for coarse alignment problems.
[0212] Alignment and filtering
[0213] This workflow performs a ripple filter because the resulting average was observed to improve noise. The spectra are then background-subtracted and peak-finded to perform alignment. The spectra used for averaging are aligned ripple-filtered spectra without any other preprocessing. The alignment step uses a set of 43 alignment points listed in Table 3 below. Additional filtering parameters require that the spectra have at least 20 peaks and that at least 5 alignment points be used.
[0214] Table 3: Alignment points used to align the grid spectra
[0215]
[0216]
[0217] Raster Averaging
[0218] An average was created from a pool of rescaled, aligned, and filtered gridded spectra. 500 spectra were randomly selected and averaged to produce a final sample spectrum of 400,000 exposures. Multiple 800 exposure spectra were collected per point, resulting in a pool of over 500 spectra, totaling 800 exposure grid spectra from 8 points per sample. 500 spectra were randomly sampled from this pool and averaged to obtain a final MALDI spectrum of 400,000 exposures averaged depth.
[0219] Deep MALDI average spectrum preprocessing
[0220] Background estimation and subtraction
[0221] Background estimation was performed with additional consideration of high-quality regions using a two-window approach for background estimation and subtraction (Table 4).
[0222] Table 4: Background estimation window
[0223] Wide window m / Z width 3000 80000 30000 80000 31000 160000 Medium window 3000 5000 30000 5000 31000 10000
[0224] Details of mass spectrometry background subtraction are known in the art and are described in prior US Patent No. 7,736,905, such description being incorporated herein by reference.
[0225] Regularization via interval methods
[0226] The interval method is used to compare the clinical groups of interest to ensure that no regularization window useful for classification is selected. The feature definitions used in this analysis are included in Appendix C of the previous provisional application serial number 62 / 289,587. The method compares feature values according to clinical groups and calculates the coefficient of variance (CV) of the features of all samples. Thresholds are set for p-values and CV to remove any regions that significantly distinguish the groups of interest or have intrinsic instability (high CV). Progressive disease (PD) is used compared to stable disease (SD) and partial response (PR), that is, whether there is clinical group comparison of disease control (DC) to calculate univariate p-values. A second comparison using PR compared to PD and SD is added. For both comparison methods, the p-value cutoff is set to 0.22. In order to calculate the regularization scalar, feature intervals with p-values less than 0.22 are not included. The CV cutoff used is 1.0. Features with CVs higher than 1.0 are also excluded. Features above 25kDa are manually removed, and features known to be intrinsically unstable (17kDa regions) are also removed. A total of 16 intervals were determined to be included as the regularization window (see Table 5).
[0227] Table 5: Iteration #1 regularization interval
[0228] Left side central right side 3530.679 3657.668 3784.658 3785.029 3931.884 4078.739 4220.21 4271.637 4323.065 4875.581 4909.742 4943.903 5260.635 5348.079 5435.524 5436.47 5559.451 5682.433 6050.421 6213.614 6376.807 6510.852 6555.966 6601.081 7751.414 7825.12 7898.826 10606.12 10751.66 10897.2 10908.61 11132.56 11356.51 12425.27 12476.27 12527.26 17710.35 18107.52 18504.69 19212.92 19978.37 20743.82 22108.95 22534.05 22959.15 23738.5 24238.77 24739.04
[0229] A second iteration of regularization according to intervals was performed on the spectra regularized using the Iteration #1 interval. The CV after the first iteration was generally lower. This allowed a new threshold of CV of 0.68 to be set. The cutoff value of the p-value was increased to increase the stringency of the requirement. The same clinical group was used in the evaluation. The second iteration resulted in the inclusion of 9 windows as regularization intervals (see Table 6).
[0230] Table 6: Iteration #2 regularization partition interval
[0231] Left side central right side 4168.226 4194.033 4219.839 4875.581 4909.742 4943.903 4946.131 5011.854 5077.576 5080.918 5170.405 5259.892 5260.635 5348.079 5435.524 6510.852 6555.966 6601.081 7751.414 7825.12 7898.826 10606.12 10751.66 10897.2 10908.61 11132.56 11356.51
[0232] For each spectrum, the scalars obtained using these windows were found and compared against the disease control group, i.e., the scalars from the spectra of patients with disease control were compared to the spectra of patients without disease control to ensure that there were no significant differences in the scalars depending on the clinical group. Figure 5 The plots of the regularized scalars shown in show that the distribution of the resulting scalars does not differ significantly between the clinical groups, and therefore the regularization intervals are not useful for classification. Over these windows, the spectra were regularized using the partial ion current (PIC).
[0233] Average spectral alignment
[0234] Peak alignment of the averaged spectra was generally very good; however, a fine-tuning alignment step was performed to resolve small differences in peak positions in the spectra. A set of alignment points was determined and applied to the analytical spectra (Table 7).
[0235] Table 7: Alignment points used to calibrate the spectral average
[0236]
[0237]
[0238] Feature Definition
[0239] After all the above preprocessing steps are completed, the process of classifier development continues by identifying and defining features (m / z regions) that are useful for classification. All 119 averaged spectra are viewed simultaneously according to the clinical groups PD, SD and PR to select features for classification. This approach ensures that features found in only a portion of the spectra are not missed and that the feature definitions are broad enough to cover variations in peak width or position. A total of 351 features were defined for the dataset. See Appendix A. These feature definitions were applied to all spectra to create a feature table of feature values (integrated intensity values over each feature) for each of the 119 spectra. An example of the selected features is Figure 6 As shown, the shaded area represents the m / z region defined for each feature.
[0240] Batch calibration of analytical spectra
[0241] Analysis of the reference sample "SerumP3"
[0242] At the beginning (1, 2) and end (3, 4) of each sample run through the MALDI-TOF mass spectrometer, two preparations of the reference sample SerumP3 were plated. The purpose of these samples is to ensure that batch variations due to slight variations in instrument performance (e.g., detector aging) can be corrected. To perform batch calibration, a spectrum must be used as a reference for the batch, which is the average of one preparation from the beginning and one preparation from the end of the batch. The procedure for selecting the pair is first described.
[0243] The reference sample was preprocessed as described above. All 351 features (Appendix A) were used to evaluate possible combinations (1-3, 1-4, 2-3, 2-4). The function was used to compare each possible repeat combination:
[0244] A=min(abs(1-ftrval1 / ftrval2),abs(1-ftrval2 / ftrval1))
[0245] Where ftrval 1 (ftrval2) is the feature value of the first (second) replicate of the replicate pair. This quantity A gives a measure of how similar the replicates of this pair are. For each feature, A is reported. If this value is greater than 0.5, then the feature is determined to be inconsistent, or "bad". For each possible combination, the count of bad features is reported. If the value of A is <0.1, then the feature is determined to be consistent and reported as "good". For each possible combination, the count of good features is reported. Using the counts of bad and good features from each possible combination, the bad / good ratio is calculated. The combination with the lowest ratio is reported as the most similar combination, and it is unlikely to contain any systematic or local outlier behavior in either reference spectrum. If no ratio less than 0.12 is found, then the batch is considered a failure. Table 8 reports the most similar combinations found for each batch.
[0246] Table 8: Serum P3 preparations found to be most similar by batch
[0247] batch combination 1 2_4 2 1_4 3 2_4
[0248] Batch calibration
[0249] Use batch 1 as the reference batch to calibrate all other batches. Use the reference sample to find the calibration factor for each of batches 2 and 3 using the following procedure.
[0250] In each batch j (2≤j≤3), for (m / z) i For each i-th feature centered on and average amplitude in is the mean reference spectral amplitude of feature i in the batch being calibrated and is the reference spectral amplitude of feature i in batch 1 (reference standard). Assume that the amplitude ratio between the two batches depends on:
[0251]
[0252] Based on each batch, by minimizing the sum of squared residuals The experimental data of the reference samples were used to construct the continuous fits. The SerumP3 reference samples were used to calculate the correction function. Care was taken to exclude outliers to avoid bias in the parameter estimates. The values of the coefficients a0, a1, b0, b1, and c0 obtained for the different batches are listed in Appendix B (Table B.1) of the previous provisional application serial number 62 / 289,587. Relative to (m / z) iThe mapping in the plane and the surface defined by the fit itself are shown in Figure B.1 of Appendix B of prior provisional application serial number 62 / 289,587.
[0253] Once the final fit for each batch is determined The next step is to All features of all samples are corrected (amplitude A in (m / z)). After this correction, the correction calculated for the reference spectrum The characteristic values are located around the horizontal line defined by r = 1, as shown in Figure B.1 of Appendix B of prior provisional application serial number 62 / 289,587. Corrected coefficients are calculated for comparison with quality control thresholds. These coefficients can be found in Table B.2 of Appendix B and in the corresponding graph in Figure B.2 of prior provisional application serial number 62 / 289,587.
[0254] Using 351 features from all batches and all SerumP3 samples, the reproducibility of feature values was assessed before and after batch correction. Overall, the median and mean CVs before batch correction were 14.8% and 18.3%, respectively. After batch correction, the median and mean CVs were 15.0% and 18.2%, respectively. As shown in the figure found in Appendix B of prior provisional application serial number 62 / 289,587, the batches were very similar, requiring minimal correction. This is reflected in the lack of improvement in the CV features across all SerumP3 samples.
[0255] Partial ion current (PIC) regularization
[0256] It was found that performing spectral regularization before batch correction was beneficial (see above). However, it was found that after batch correction, if another regularization was performed, the coefficient of variance (CV) of the features could be improved and better results were obtained. This second PIC regularization was based on a smaller window around the individual peaks determined above.
[0257] The spectrum is regularized using the partial ion current (PIC) regularization method. Background information about partial ion current regularization is described in previous U.S. Patent 7,736,905, which is incorporated herein by reference. A complete feature table is checked to find the intrinsic stability region used as the final regularization window. First, the DC group is compared according to the feature, thereby finding the univariate p value. Features with a p value less than 0.15 are excluded from the PIC analysis because these features may provide meaningful information for the test to be developed. A set of 221 features is used in the PIC analysis, of which 30 features are used for the final regularization (Table 9).
[0258] Table 9: Features used for PIC regularization
[0259]
[0260]
[0261] For regularization, the listed features are summed to find the regularization factor for each sample. All feature values are then divided by the regularization factor to obtain Figure 8 The final feature table used in the subsequent classifier generation method of . The regularization factors were checked according to the DC group to test that the calculated factors were not significantly correlated. Figure 7 The plots of show the distribution of these factors. The plots for both groups are very similar, indicating that the regularization scalar is suitable for use.
[0262] Classifier Development Using Diagnostic Cortex™
[0263] After creating the signature table of mass spectra for the 119 samples (as described above), we started using Figure 8 The classifier generation method shown in the flowchart in the figure is used to develop a classifier. This method, called "Combination of Small Classifiers with Dropout Regularization" or "CMC / D" or DIAGNOSTIC CORTEX TM, is described in et al., which is described in detail in their pending U.S. Patent Application Publication No. 2015 / 0102216, the entire contents of which are incorporated herein by reference. An overview of the method will be provided first, followed by Figure 8 The method to generate the melanoma / nivolumab classifier is described in detail.
[0264] Compared to standard applications of machine learning, which focus on developing classifiers when large training datasets are available, the big data challenges in the context of biological life science problems are different. Here, there is the problem that the number of available samples (n), typically from clinical studies, is often limited, and the number of attributes (measurements) (p) per sample often exceeds the number of samples. In these deep data problems, rather than deriving information from many instances, one seeks to derive information from deep descriptions of individual instances. The current approach exploits this insight and is particularly useful in problems where p>>n.
[0265] The method comprises a first step a): obtaining measurement data for classification from a plurality of samples, i.e., measurement data reflecting some physical properties or characteristics of the samples. The data for each sample consists of a plurality of eigenvalues and a category label. In this example, the data is in the form of mass spectral data, which is in the form of eigenvalues (integrated peak intensity values at a plurality of m / z ranges or peaks, see Appendix A) and a label indicating a certain attribute of the sample (e.g., early or late death or disease progression of the patient). In this example, after investigating the clinical data associated with the sample, the category label is assigned to each sample by the operator. The development sample set is then divided into a training set and a test set, and the training set is used in the following steps b), c) and d).
[0266] The method continues with step b): constructing multiple individual small classifiers using a set of feature values from the sample up to a feature set size s (s=integer 1...n). For example, multiple individual small or atomic classifiers can be constructed using a single feature (s=1) or a feature pair (s=2) or a triplet of features (s=3) or even higher-order combinations of more than 3 features. The value of s is typically chosen to be small enough to allow the code implementing the method to run in a reasonable time, but may be larger in some cases or where longer code run times are acceptable. The choice of the value of s may also be determined by the number of measurement data values (p) in the dataset, and when p is in the hundreds, thousands, or even tens of thousands, s will typically be 1 or 2 or possibly 3, depending on the available computing resources. The small classifiers perform a supervised learning classification algorithm, such as k-nearest neighbor (kNN), in which the value of a feature, feature pair, or triplet of features of a sample instance is compared to the value of the same feature in the training set or the nearest neighbor in an s-dimensional feature space is identified (e.g., k=9), and a class label is assigned to the sample instance of each small classifier by majority voting. In practice, there may be thousands of such small classifiers, depending on the number of features used for classification.
[0267] The method continues with a filtering step c), i.e., testing the performance (e.g., accuracy) of each individual small classifier to correctly classify the sample, or measuring the performance of the individual small classifiers by some other metric (e.g., the difference between the risk ratios (HR) obtained between the groups defined by the classification of the training set samples by the individual small classifiers), and retaining only those small classifiers whose classification accuracy, predictive power, or other performance metrics exceed a predefined threshold, thereby obtaining a filtered (pruned) set of small classifiers. If the performance metric selected for small classifier filtering is classification accuracy, the class labels obtained from the classification operation can be compared with the class labels of the samples known in advance. However, other performance metrics can be used and evaluated by the class labels generated by the classification operation. Only those small classifiers that perform well under the selected performance metric can be maintained. Alternative supervised classification algorithms can be used, such as linear discriminants, decision trees, probabilistic classification methods, edge-based classifiers such as support vector machines, and any other classification methods for training classifiers from a set of labeled training data.
[0268] To overcome the bias of some univariate feature selection methods due to subset bias, a large fraction of all possible features is considered as candidate small classifiers. Then, all possible kNN classifiers are constructed using a feature set of a preselected size (parameter s). This yields many "small classifiers": for example, if one starts with 100 features per sample (p = 100), 4,950 "small classifiers" are obtained from all possible combinations of these feature pairs (s = 2), 161,700 small classifiers are obtained from all possible combinations of three features (s = 3), and so on. Other methods of exploring the space of possible small classifiers and defining their characteristics are certainly possible and could be used instead of this hierarchical approach. Of course, many of these "small classifiers" will have poor performance, so only those that pass predetermined criteria are used in the filtering step c). These filtering criteria are chosen based on the specific problem: if one has a two-class classification problem, only those small classifiers whose classification accuracy exceeds a predefined threshold, i.e., those that are reasonably predictive, are selected. Even filtering for “small classifiers” results in tens of thousands of “small classifier” candidates, with performance ranging from ambiguous to fairly good to excellent.
[0269] The method continues with step d): generating a main classifier (MC) by combining the filtered small classifiers using a regularized combination method. In one embodiment, this regularized combination method uses a set of filtered small classifiers to repeat logical training relative to the class label of the sample. This method is carried out by the following steps: owing to performing extreme value discarding (herein referred to as the technology of "discarding regularization") from the filtered small classifier set, a small part of the filtered small classifiers is randomly selected, and logical training is carried out for these selected small classifiers. Although similar in spirit to the standard classifier combination method (see, for example, S.Tulyakov et al., Review of ClassifierCombination Methods, Studies in Computational Intelligence, Vol. 90, 2008, pp. 361-386), a specific problem occurs, that is, some "small classifiers" may be artificially perfect due to random factors, and therefore will dominate in the combination. In order to avoid this overfitting to a particularly dominant "small classifier", many logical training steps are generated by randomly selecting a small part of "small classifiers" for each of these logical training steps. This is a regularization of the problem in the spirit of dropout from deep learning theory. In this case, with many small classifiers and a small training set, extreme value dropout is used where more than 99% of the filtered small classifiers are dropped at each iteration.
[0270] In more detail, the result of each small classifier is one of two values, in this example "early" or "late". Logistic regression can then be used to combine the results of the small classifiers in the spirit of logistic regression by defining the probability of getting the "early" label using standard logistic regression (see, for example, http: / / en.wikipedia.org / wiki / Logistic_regression).
[0271] Equation (1)
[0272]
[0273] where I(mc(eigenvalue)) = 1 if the small classifier mc applied to the eigenvalue of the sample returns "early", and returns 0 if the small classifier returns "late". The weight w of the small classifier mcis unknown and needs to be determined by fitting the above formula to all examples in the training set. Accordingly, for late-labeled examples in the training set, the left-hand side of the formula is assigned a value of +1, and for early-labeled examples, a value of 0. Because there are more small classifiers, and therefore more weights, than examples—typically thousands of small classifiers with only a few dozen examples—such a fit always results in near-perfect classification and can easily be dominated by small classifiers that, due to random chance, are well-suited to a particular problem. It is undesirable for the final test to be dominated by a single small classifier that performs well for this particular set but does not generalize well. Therefore, a method was devised to regularize this behavior: instead of fitting all weights of all small classifiers to the training data simultaneously in an ensemble regression, only some of the small classifiers are used for regression, but this process is repeated multiple times to generate a master classifier. For example, three small classifiers are randomly selected, regression is performed on their three weights, another set of three small classifiers are selected and their weights determined, and this process is repeated multiple times to generate many randomly selected, i.e., three realizations of the small classifier. The final weight of the master classifier is defined as the average of the weights of all these realizations. The number of realizations should be large enough so that each small classifier is likely to be picked at least once throughout the process. This approach is similar in spirit to "dropout" regularization, a method used in the deep learning community to add noise to neural network training to avoid getting stuck in local minima of the objective function.
[0274] Other methods for performing the regularized combination method in step (d) that may be used include:
[0275] Logistic regression, with a penalty function like ridge regression (based on Tikhonov regularization, Tikhonov, Andrey Nikolayevich (1943). [On thestability of inverse problems].Doklady Akademii Nauk SSSR 39(5):195-198. )
[0276] Lasso method (Tibshirani, R. (1996). Regression shrinkage and selection via the lasso. J. Royal. Statist. Soc B., Vol. 58, No. 1, pages 267-288).
[0277] Neural Networks Regularized with Dropout (Nitish Shrivastava, "Improving Neural Networks with Dropout", Master's Thesis, Graduate Department of Computer Science, University of Toronto), available from the University of Toronto Department of Computer Science website.
[0278] General regularized neural networks (Girosi F. et al, Neural Computation, (7), 219 (1995)).
[0279] The aforementioned publication is incorporated herein by reference. Methods using dropout regularization have shown promise in avoiding overfitting and increasing the likelihood of producing generalizable tests, i.e., tests that can be validated on an independent set of samples. The performance of the main classifier is then evaluated by how well it classifies a subset of samples that form the test set.
[0280] In step e), steps b)-d) are repeated on a program-controlled computer to separate the sample set into a test set and a training set at different implementations, thereby generating multiple master classifiers, each master classifier implementing a separation of the sample set into a training set and a test set. The performance of the classifier is evaluated by all implementations of the separation of the development sample set into a training set and a test set. If there are some persistently misclassified samples in the test set, the process optionally loops back and repeats steps b), c), d), and e) for these misclassified samples using flipped class labels.
[0281] The method then proceeds to step f): defining a final classifier from a combination of one or more of the plurality of primary classifiers. In this example, the final classifier is defined as a majority vote or average probability cutoff of all primary classifiers resulting from each separation of the sample set into a training set and a test set.
[0282] Now turn Figure 8 , the classifier development process will be described in further detail in the context of the melanoma / nivolumab classifier.
[0283] The available set of 119 samples was initially randomly divided into two subsets: a set of 59 samples for validation of the generated classifier and a development set of 60 samples (100). As mentioned above, this division was performed twice (see the discussion of DEV1 and DEV2 above), with stratification.
[0284] At step 102, the definition of two class labels (or groups) for the samples in the development set 100 is performed. Although some initial approaches for classifier development have employed well-defined class labels, such as reaction categories, these approaches have proven unsuccessful. All of the methods discussed in this application utilize event time data for classifier training. In this case, the initial class label definition (early, late) is not obvious, such as Figure 8 As shown, the process uses an iterative approach to optimize the class labels while creating the classifier (see loop 146 discussed below). At the beginning, an initial guess is made for the class labels. Typically, the samples are sorted based on TTP or OS, and half of the samples with the lowest event time results are assigned an "early" class label (early death or progression, i.e., poor outcome), while the other half are assigned a "late" class label (late death or progression, i.e., good outcome). The classifier is then built using the outcome data and these class labels. This classifier can then be used to generate classifications for all development set samples, which are then used as new class labels for the second iteration of the classifier building step. This process iterates until convergence.
[0285] Although “early” and “late” can be defined by setting cutoff values for clinical data, we initially used TTP (time to progression) data to assign training labels. The 30 patients with the lowest TTP were assigned the “early” category label, and the 30 patients with the highest TTP were assigned the “late” category label. It should be noted that in the following procedures, there is a procedure to flip the category label for samples that are consistently misclassified, so these initial category label assignments are not necessarily fixed. After reaching the initial category label definition, based on Figure 8 The resulting data indicated by the two groups 104 and 106 in FIG. 5 assigns early and late categories to the samples.
[0286] At step 108, the early and late samples from the development set (100) are then randomly divided into training (112) and test sets (110), each with 30 patients. This division is performed in a stratified manner. Typically, twenty samples from each class are assigned to the training set, and the remainder are assigned to the test set. Occasionally, during the class label flipping optimization process, the number of samples in the early group drops too low to assign 20 samples to the training set and still have a reasonable number of samples in the test set (e.g., more than 7 or 8). In these cases, a small number of samples from each class are assigned to the training set, such as 17 or 18 samples. In all cases, the number of samples assigned to the training set from each class is the same. The training set (112) then undergoes steps 120, 126, and 130. In step 120, a subset of features from the identified 351 mass spectral features (see Appendix A) is used to construct (define) a number of k-nearest neighbor (kNN) small classifiers (mC) using the training set as its reference set. For many of the investigations conducted, all possible single features and feature pairs (s=2) were examined; however, when fewer features were used, triplets were sometimes also considered (s=3). Although different values of k in the kNN algorithm were tried in preliminary studies, the methods described in Example 1 all used k=9. In order to be able to consider subsets of single, two, or three features and improve classifier performance, it was necessary to deselect features that were not useful for classification from the 351 feature set. This was done using the bagged feature selection method outlined in Appendix F of the previous provisional application serial number 62 / 289,587. Further details about this method, its basic principles and benefits are described in U.S. patent application serial number 15 / 091,417 filed on April 5, 2016 by J. Roder et al. and U.S. provisional application serial number 62 / 319,958 filed on April 8, 2016, the contents of which are incorporated herein by reference. Appendix B lists a list of selected features used for different classification methods.
[0287] In step 126, a filtering process is used to select only small classifiers (mC) that have useful or good performance characteristics. This can be done in Figure 8 This is understood by the spectrum 124 containing many individual features (shown by the hatched area), and the individual and paired features are indicated in the feature space 122. For some kNN small classifiers, these features (individual or paired) perform well for classifying the sample, and these small classifiers are retained (shown by Figure 8 The small classifiers are indicated by the “+” sign at 128 in
[15] , while the other small classifiers indicated by the “-” sign are not retained.
[0288] To target a final classifier with certain performance characteristics, these mCs are filtered as follows. Each mC is applied to its training set, and a performance metric is calculated based on the resulting classification of the training set. Only mCs that meet these performance metric thresholds pass the filtering process and are subsequently used. mCs that fail the filtering are discarded. For this project, both accuracy and hazard ratio filtering were used. Sometimes, simple individual filters are used, while other times, composite filters are constructed by combining multiple individual filters using a logical AND operation. For accuracy filtering, the classifier is applied to the training set or a subset of the training set, and the resulting classification accuracy must fall within a preset range for the mC to pass the filtering. For hazard ratio filtering, the classifier is applied to the training set or a subset of the training set. The hazard ratio for a specified outcome (TTP or OS) is then calculated between the group classified as early-stage and the remaining group classified as late-stage. The hazard ratio must fall within a specific range for the mC to pass the filtering. In this specific classifier exercise, small classifiers were filtered based on the hazard ratio (HR) between early-stage and late-stage and the accuracy of the early-stage category (defined as TTP <100 days) and late-stage (defined as TTP >365 days).
[0289] Table 10: Parameters used in small classifiers and filtering
[0290]
[0291]
[0292] Here, “methods” means different classifier development exercises. Essentially, Figure 8 The process is repeated for different times, using different values of the depth parameter s in each iteration, and as shown in Figure 10, Figure 8 Different small classifier filtering criteria are used in step 126. Note that this exercise is performed twice for each split of the entire set of 119 samples into development and validation sets (DEV1 and DEV2).
[0293] For each realization of separating the development set into training and test sets in step 108, a main classifier (MC) is generated in step 130. Once the filtering of the mCs is completed, the mCs are merged into one main classifier (MC) in step 132 using the logistic regression trained on the training set class labels, step 132. See the previous discussion on dropout regularization. To avoid overfitting, the regression is regularized using extreme value dropout, and only a few randomly selected mCs are included in each logistic regression iteration. The number of dropout iterations is selected based on the typical number of mCs that pass the filtering to ensure that each mC is likely to be included in the dropout process multiple times. All methods outlined in Example 1 leave 10 randomly selected mCs in each iteration. The methods using only single and pairwise features used 10,000 dropout iterations; the methods using single, pairwise and triplet features used 100,000 dropout iterations.
[0294] At step 134, the performance of the MC achieved in step 132 is evaluated for its ability to classify the test set (110) of samples. With each iteration of steps 120, 126, 130, 134, the performance of the resulting MC is evaluated for its ability to classify the members of the test set 110.
[0295] After the evaluation step 134, the process loops back to step 108 via loop 135 and generates different implementations of splitting the development set into training and test sets. The process of steps 108, 120, 126, 130, 132, 134 and looping back at 135 to split the development set into training and test sets again (step 108) is performed multiple times and 625 different implementations (loops) are used in this project. When the training set classes have the same number of samples, Figure 8 The method works best. Therefore, if the number of members of the classes is different, they are divided into test and training in different proportions. Using multiple train / test splits (loop 135 and subsequent execution of steps 120, 126, 130, 132 and 134) avoids selecting a single particularly favorable or difficult training set split for classifier creation and avoids bias in performance evaluation caused by testing on a test set that may be particularly easy or difficult to classify.
[0296] At step 136, there is an optional procedure to analyze the data from the training and test splits and obtain the performance characteristics of the MC and its classification results from each training / test set split as shown in box 138. Optional steps 136 and 138 are not performed in this project.
[0297] At step 144, it is determined whether there are samples in the test set 110 that have been consistently misclassified during many iterations of loop 135. If so, the class labels of these misclassified samples are flipped, and in loop 146 the process returns to the beginning of step 102 and repeats. Figure 8 The method shown in .
[0298] If there are no persistently misclassified samples at step 144, then proceed to step 150 and define the final classifier in one of several ways, including (i) a majority vote of each main classifier (MC) that implements the split of the development set into training and test sets, or (ii) an average probability cutoff. The output of the logistic regression that defines each MC (step 132) may be one of the two training classes (early or late). These MC probabilities can be averaged to produce an average probability for the sample. When using the development set (100), this approach is adjusted so that the MCs for a given sample are not included in the training set are averaged ("out-of-bag" estimates). These average probabilities can be converted to binary classifications by applying a threshold (cutoff). During the iterative classifier construction and label optimization process, the classification is assigned by a majority vote of the individual MC labels obtained with a cutoff of 0.5. This process is modified so that only mCs are selected where, for samples in the development set, the sample is not in the training set (modified or "out-of-bag" majority vote). This procedure gives a classification that is very similar to using a cutoff of 0.5 for the average probability on the MC.
[0299] After the final classifier is defined in step 150, the process optionally continues with a validation step 152, where the final classifier defined in step 150 is tested on an internal validation set of samples (if available). In this example, the initial set of 119 samples is divided into a development set (100) and a separate internal validation set, so this validation set exists and undergoes the validation step 152. Ideally, in step 154, this final classifier defined in step 150 is also validated on an independent sample set. Validation on an independent sample set is described later in this Example 1.
[0300] Example 1 Classifier Results
[0301] The goal of this exercise is to determine the Figure 8Can the final classifier defined demonstrate separation between the early and late class markers, in other words, predict from the mass spectra of pre-treatment serum samples whether the patient providing the sample is likely to benefit from treatment with nivolumab for cancer. This goal has been achieved, as shown by the classifier performance data in this section of the document. The performance of the classifier was evaluated using Kaplan-Meier plots of TTP and OS between samples classified as early and late, along with the corresponding hazard ratios (HRs) and log-rank p-values. Table 11 summarizes the results of the four methods in Table 10. This table lists the development set ( Figure 8 The classifier performance of the other half of the set of 119 samples (i.e., the validation set) is shown in Table 11. Figure 8 The significant differences in median OS and TTP between the early and late groups identified by the developed classifier in Figure 5 demonstrate the ability of the classifier to identify clinically useful groups.
[0302] Table 11: Summary of final classifier performance of the four methods in Table 10
[0303]
[0304]
[0305] Kaplan-Meier plots corresponding to the data in Table 11 are shown in Figures 9-12 for the four methods. The classification of each sample for each method is listed in Appendix G of the previous provisional application serial number 62 / 289,587. For each of the four methods, the baseline clinical characteristics are summarized by classification group in Table 12.
[0306] Table 12: Clinical characteristics of the classification groups
[0307]
[0308]
[0309] The data in Tables 11-12 and Figures 9-12 demonstrate that it is possible to construct a classifier that can identify patients with better and worse outcomes with nivolumab treatment using mass spectra generated from pre-treatment serum samples. Patients classified as advanced had better OS and TTP than those classified as early. Additionally, patients classified as advanced were more likely to have a partial response or stable disease on nivolumab treatment than those classified as early. Table 12 illustrates that the proportion of patients who were PD-L1 positive was similar in the early and late class groups, so these two quantities are not significantly correlated, and the serum classifier of Example 1 will provide additional information to any information provided by PD-L1 expression levels.
[0310] These results show consistency in identifying patients with better and worse outcomes with nivolumab therapy across the development / validation set split, with good generalization between validation and development set performance. Having demonstrated impressive validation performance, the two small classifier filtering methods used were applied as Figure 8 The results are summarized in Table 13 and the Kaplan-Meier plot is shown in Figure 13. (According to Figure 8 The constructed classifier uses all 119 samples in the development set explained here and in Table 13, and is referred to as the "full set classifier of Example 1" hereinafter.
[0311] Table 13: Performance of two classifiers built on the full set of 119 samples as the development set
[0312]
[0313]
[0314] The performance of the two methods is very similar, and both are in good agreement with the classifier built on the development / validation split of 119 samples. This further indicates that Figure 8 The procedure produces reliable classifiers with good generalization properties.
[0315] In Table 14, clinical characteristics are summarized according to the classification groups given by method 1 (simple OS filtering). The early and late categories were not associated with PD-L1 expression or with any cutoff or with sex or age. More than 70% of patients in cohort 5 were classified as late, while patients in cohorts 4 and 6 were roughly evenly divided between the two classification groups. LDH was significantly higher in the early group than in the late group (Mann-Whitney p value = 0.003).
[0316] Table 14: Baseline characteristics of the classification group for the full set classifier method 1
[0317]
[0318]
[0319] x One patient's data is missing xx ULN = upper limit of normal range
[0320] Multivariate analysis showed that the early / late stage category remained independently significant when adjusting for other clinical factors.
[0321] Table 15: Multivariate analysis of OS and TTP
[0322]
[0323] The table shows that serum testing adds information to other available clinical features, and even when adjusting for other available features (including PD-L1 expression level and LDH level), the serum class marker remains a significant predictor of both TTP and OS. Of particular note, although LDH level and test class are correlated with each other, they are both independently significant predictors of both OS and TTP. In addition, the dependence of early / late class on tumor size was examined. Classification was significantly associated with tumor size (Mann-Whitney p < 0.001), with the median tumor size of 53 cm in the early class group and 16 cm in the late class group. However, although tumor size only showed a trend towards significance for OS and TTP prediction in univariate analysis (p = 0.065 and p = 0.07, respectively), the early and late classes still had highly significant predictive power for OS and TTP after adjusting for tumor size (p < 0.001 for OS and p = 0.004 for TTP).
[0324] Given the differences in outcomes between the early and late stage category groups and the independence from other clinical information, the developed classifier may provide additional information to physicians and patients to inform the decision whether nivolumab (or another anti-PD-1 antibody) or alternative therapy is appropriate treatment for the patient. The preferred final classifier was the one generated from all 119 patient samples in the development set with OS filtering (Method 1 in Table 13, Figure 13C and 13D Kaplan-Meier plot of ). This classifier is referred to as the ensemble classifier or Example 1 or "IS2" later in this document.
[0325] A second sample set was used to independently verify the Figure 8 Generated classifier
[0326] As an independent validation cohort, a set of 30 pre-treatment samples from patients treated with anti-PD-1 antibodies from Yale University were available.
[0327] Deep MALDI spectra were generated from these samples and analyzed using the Figure 8 The same procedures used in the development of the classifier for the were used as described above. The classifier for "Method 1" (Table 13 Method 1) for the entire set was applied to the resulting feature table, generating a classification of "early" or "late" for each sample. Ten samples were classified as "early" and the remaining 20 as "late". The Kaplan-Meier curve for overall survival is shown in Figure 14 The OS analysis is summarized in Table 16. As in Figure 13, please note that Figure 14The clear separation between the early-intermediate and late-stage groups demonstrates the ability of the final classifiers in Table 13 and Figure 13 to correctly classify samples in the independent validation cohort.
[0328] Table 16: Summary of classifier performance in the Yale anti-PD-1 antibody treatment cohort
[0329]
[0330] The classifier of Example 1 was also well validated in an independent cohort of 48 melanoma patients treated with the anti-CTLA4 antibody ipilimumab. See Example 3 below.
[0331] Protein identification
[0332] The method of generating a classifier is based on the above combination Figure 8 The CMC / D process described here is a correlation analysis that relates peak intensities to clinical outcomes. Therefore, the proteins underlying the feature definition used in the classification may not be causally related to treatment outcomes. It is also not trivial to correlate the peaks measured in the MALDI-TOF experiment with previously identified proteins and their functions. However, by reviewing the literature on serum proteins observed in MALDI-TOF studies, it is still possible to give the names of some peaks used in the classification. Table 17 provides a tentative list:
[0333] Table 17: Tentative assignment of feature subsets for classification to proteins / protein fragments
[0334] Mass [Da] Experimental protein identification
[0335] 4133 c1 inhibitor
[0336] 4264 ITIH4, C1 fragment
[0337] 4381 β-defensin 4A
[0338] 5867 Leukocyte-specific transcript 1 protein
[0339] 5889 Leukocyte-specific transcript 1 protein
[0340] 5911 Gamma secretase C-terminal fragment of amyloid βA4 protein 50
[0341] 5997 Granulin-3 (Granulin-B)
[0342] 7318 CC motif chemokine 20
[0343] 8413 Complement C3a
[0344] 9109 Apolipoprotein C3
[0345] 9226 CDC42 small effector protein 2
[0346] 10285 B melanoma antigen 3
[0347] 11686 Serum amyloid A (SAA)
[0348] 23469 C-reactive protein (CRP)
[0349] While the presence of acute phase reactants such as SAA and CRP was expected, we were surprised to find evidence for proteins associated with the complement system (complement C3a, C1 inhibitor, C1 fragment). The complement system has been implicated in immune responses and cancer immunotherapy (Markiewski et al., Is complement good or bad for cancer patients? A new perspective on an old dilemma. Trends Immunol. 30: 286-292 (2009)) and has recently been implicated in PD-1 inhibition (Seng-Ryong Woo et al., Innate Immune Recognition of Cancer Annual Review of Immunology, Vol. 33. 445-474 (2015)).
[0350] It appears that at least part of the classification of patients into the early and late groups is related to the interaction between acute phase reactants and activation of the complement system. A more detailed explanation of the relationship between the classification features and the biological functions of the early and late category markers is set forth in Example 6 later in this document.
[0351] Conclusions about the classifier developed in Example 1
[0352] The DIAGNOSTIC CORTEX™ program ( Figure 8 ) was applied to pre-treatment serum spectra from melanoma patients obtained using deep MALDI spectral acquisition to identify clinical groups "early" and "late". After treatment with the anti-PD-1 therapy nivolumab, these groups showed significant differences in outcomes (TTP and OS). A testing procedure has been proposed to identify these groups from pre-treatment samples and validate the results in internal and external validation sets.
[0353] Patients whose serum classification was "early" showed significantly faster progression and shorter survival compared with patients whose serum classification was "advanced," making this test suitable as a biomarker for nivolumab therapy. The "early" and "advanced" clinical groups were not associated with PD-L1 expression, and this classification remained a significant predictor of outcome (TTP and OS) even when other clinical characteristics were included in multivariate analysis.
[0354] While the methods associated with test development do not provide insight into the biology that could explain the differences between the identified groups, we have conducted some preliminary studies that link the two distinct groups to differences in acute phase reactants and the complement system. The success of this project exemplifies how deep MALDI spectral acquisition combined with these classifier development methods can lead to the construction of clinically useful tests.
[0355] The following clauses are provided as further description of the invention disclosed in Example 1.
[0356] 1. A method for predicting that a melanoma patient will benefit from an antibody drug that blocks ligand activation of programmed cell death 1 (PD-1), comprising:
[0357] a) performing mass spectrometry analysis on a patient's blood sample and obtaining mass spectrometry data;
[0358] (b) obtaining integrated intensity values in the mass spectral data for a plurality of mass spectral features; and
[0359] (c) operating the mass spectrometry data using a programmable computer implementing the classifier;
[0360] Wherein, in the operation step, the classifier uses a classification algorithm to compare the integrated intensity value with feature values of a reference set of class-labeled mass spectrometry data obtained from blood samples obtained from multiple other melanoma patients treated with the drug, and generates a class label for the sample, wherein the class label "early stage" or an equivalent predicts that the patient is likely to obtain relatively less benefit from the antibody drug, and the class label "late stage" or an equivalent indicates that the patient is likely to obtain relatively greater benefit from the antibody drug.
[0361] 2. The method of clause 1, wherein the melanoma patient providing the blood sample has previously been treated with ipilimumab.
[0362] 3. The method of clause 2, wherein the patient is highly toxic to ipilimumab.
[0363] 4. A method as described in any of clauses 1-3, wherein the mass spectrometric characteristics include multiple characteristics listed in Appendix A, Appendix B or Appendix C.
[0364] 5. The method of any one of clauses 1-4, wherein the classifier is generated from an ensemble of filtered small classifiers using a regularized combination method.
[0365] 6. The method of any one of clauses 1 to 5, wherein the mass spectral data is acquired from at least 100,000 laser shots performed on the sample using MALDI-TOF mass spectrometry.
[0366] 7. The method of clause 6, wherein the small classifiers are filtered according to any one of the criteria listed in Table 10.
[0367] 8. The method of clause 5, wherein the classifier is defined based on a plurality of master classifiers generated from generating training and test sets from various splits of a development sample set.
[0368] 9. The method of clause 5, wherein the classifier is developed from a sample set comprising patients previously treated and not previously treated with ipilimumab.
[0369] 10. The method of clause 5, wherein the regularized combination method comprises repeatedly performing logistic regression with extreme value dropout.
[0370] 11. The method of any one of clauses 1 to 10, wherein the antibody drug comprises nivolumab.
[0371] 12. The method of any one of clauses 1 to 11, wherein the relatively greater benefit associated with the late-stage marker means significantly greater (longer) overall survival compared to early-stage markers.
[0372] 13. The method of any one of clauses 1 to 12, wherein the reference set is derived from class-labeled blood samples from melanoma patients treated with nivolumab, wherein the class-labeled blood samples in the reference set have the class label early or late, or equivalent, and wherein late samples have higher overall survival when treated with nivolumab compared to early samples.
[0373] 14. Machines for predicting whether melanoma patients will benefit from antibody drugs that block ligand activation of programmed cell death 1 (PD-1) (see Figure 15 ),include:
[0374] a memory storing a reference set in the form of feature values for a plurality of mass spectral features obtained from mass spectra of blood samples of a plurality of melanoma patients treated with the antibody drug;
[0375] The memory further stores a set of codes for defining a final classifier based on a plurality of main classifiers, each main classifier being generated from filtered mini-classifiers combined using a regularized combination method;
[0376] a central processing unit that operates on the code set and the reference set and mass spectrometry data obtained from a blood sample of a melanoma patient to be tested, and responsively generates a class label for the blood sample, wherein the class label "early stage" or an equivalent predicts that the patient is likely to derive relatively less benefit from the antibody drug, and the class label "late stage" or an equivalent indicates that the patient is likely to derive relatively greater benefit from the antibody drug.
[0377] 15. The machine of clause 14, wherein the mass spectral data for the development set and samples are obtained from performing at least 100,000 laser shots using MALDI-TOF mass spectrometry on the samples forming the development set and the samples to be tested.
[0378] 16. A machine as described in clause 14, wherein the small classifiers are filtered according to any one of the criteria listed in Table 10.
[0379] 17. The machine of clause 14, wherein the final classifier is defined from a plurality of master classifiers, each master classifier being generated from splitting a classifier development sample set into a training set and a test set.
[0380] 18. The machine of clause 14, wherein the reference set comprises mass spectrometry data from patients previously treated and not treated with ipilimumab.
[0381] 19. The machine of clause 14, wherein the regularized combination method comprises logistic regression with extreme value dropout.
[0382] 20. The machine of any one of clauses 14 to 19, wherein the drug comprises nivolumab.
[0383] 21. A machine as claimed in any one of clauses 14 to 20, wherein the mass spectrometry signature comprises a plurality of signatures listed in Appendix A, Appendix B or Appendix C.
[0384] 22. The machine of any one of clauses 14 to 21, wherein the relatively greater benefit associated with late-stage markers means significantly greater (longer) overall survival compared to early-stage markers.
[0385] 23. A system for predicting patients who will benefit from an antibody drug that blocks ligand activation of programmed cell death 1 (PD-1).
[0386] Mass spectrometer and
[0387] A machine as described in any of clauses 14-22.
[0388] 24. A method for generating a classifier for predicting whether a patient will benefit from an antibody drug that blocks ligand activation of programmed cell death 1 (PD-1), the method comprising:
[0389] 1) obtaining mass spectrometry data from a development set of blood samples obtained from melanoma patients treated with an antibody drug, wherein mass spectra of at least 100,000 laser shots are obtained from each member of the set;
[0390] 2) performing spectral preprocessing operations on mass spectrometric data from the development sample set, including background estimation and subtraction, alignment, batch correction, and regularization;
[0391] 3) Execution Figure 8 The process of steps 102-150 includes a regularized combination based on a filtered small set of classifiers to separate the development sample set into a training set and a test set each time, thereby generating a main classifier;
[0392] 4) evaluating the main classifier performance of the classifier generated according to step 3); and
[0393] 5) Define the final classifier based on the main classifier generated in step 3).
[0394] 25. The method of clause 24, wherein the final classifier comprises a reference set comprising feature values of a feature set listed in Appendix A, Appendix B, or Appendix C.
[0395] 26. A method as described in clause 24, wherein an integrated intensity value is obtained for each feature listed in Appendix A, and wherein the method further includes deselecting features that do not contribute to classifier performance from the feature list in Appendix A and using the reduced feature list to perform steps 3), 4) and 5).
[0396] 27. The method of clause 26, wherein the reduced feature list comprises a feature list in the set of Appendix B or a feature list in Appendix C.
[0397] 28. The method of clause 24, wherein the filtering criteria comprises the filtering criteria of Table 10.
[0398] 29. A method of treating a patient with melanoma, comprising
[0399] Patients are administered an antibody drug that blocks ligand activation of PD-1.
[0400] wherein the patient's blood sample has previously been assigned a class label of late stage or equivalent according to performing the method of any of clauses 1 to 13 on the blood sample.
[0401] 30. An improved general purpose computer configured as a classifier for classifying a blood sample from a human cancer patient to predict the patient's relative likelihood of survival or benefit from a drug, comprising:
[0402] a memory storing a reference set in the form of feature values for a plurality of mass spectral features derived from mass spectrometric analysis of blood samples from a plurality of melanoma patients treated with an immune checkpoint inhibitor, and an associated class label for each blood sample in a training set, the blood samples forming a classifier development set;
[0403] The memory further stores a set of computer executable codes for defining a final classifier based on a plurality of master classifiers, each master classifier being generated from a set of filtered small classifiers that perform a classification algorithm and combined using a regularized combination method; wherein the plurality of master classifiers are obtained from a number of different implementations of splitting the development set into classifier training and test sets; and
[0404] A central processing unit operates on the code set, the reference set, and mass spectrometry data obtained from a blood sample of a cancer patient to be tested, and generates a class label for the blood sample.
[0405] 31. The improved computer of clause 30, wherein the memory stores feature values for at least 50 features listed in Appendix A.
[0406] Example 2
[0407] Predictive classifier for melanoma patients who benefit from immune checkpoint inhibitors
[0408] The classifier of embodiment 1 including the classifier constructed from half of the sample set and the classifier constructed using the entire sample set showed similar performance and was well validated on two independent sample sets. It was found that the classifier of embodiment 1 also divided 173 first-line, advanced non-small cell lung cancer (NSCLC) patient groups treated with platinum doublet + cetuximab and another group of 138 ovarian cancer patients treated with platinum-doublet after surgery into groups with better and worse OS and progression-free survival (PFS). For more details about this discovery, see Example 4 below.
[0409] It is possible to argue that the classifiers of Example 1 above have a strong prognostic component because they appear to stratify patients based on their outcomes regardless of whether the treatment is immunotherapy or chemotherapy. While this may or may not be clinically important, Example 2 describes the development of a classifier and test that is able to stratify patients treated with immune checkpoint inhibitors based on their outcomes, while not stratifying outcomes for patients receiving chemotherapy—i.e., a predictive test between immune checkpoint inhibitor and chemotherapy regimens, which has a smaller prognostic component.
[0410] Patient samples
[0411] The samples used to develop the classifier in Example 2 were pre-treatment serum samples from two different cohorts. The first cohort consisted of 119 samples from melanoma patients treated with nivolumab and described in detail in Example 1. The baseline clinical characteristics are shown in Table 1. Figure 1A and 1B etc. This set was split into development and validation sets, as explained in Example 1 above.
[0412] The second cohort, referred to as the ACORN NSCLC cohort, is a collection of 173 pre-treatment serum samples, clinical data, and associated mass spectra from patients with non-small cell lung cancer (NSCLC) treated with platinum-doublet plus cetuximab. The purpose of the second cohort is to adjust the development of the classifier so that (a) patients are predicted to benefit from nivolumab, and (b) outcomes are not predicted in NSCLC patients treated with chemotherapy. Available outcome data include progression-free survival (PFS) and overall survival (OS). Selected clinical features of patients with available spectra from pre-treatment samples are listed in Table 18.
[0413] Table 18: Baseline characteristics of NSCLC patients with available spectroscopy
[0414]
[0415] Kaplan-Meier curves of progression-free survival (PFS) and overall survival (OS) for the second cohort of 173 NSCLC patients with baseline samples and acquired spectra are shown in Figure 2. Figure 16A and 16B shown.
[0416] The ACORN NSCLC cohort was divided into 3 subsets: 58 samples (referred to herein as the "AddFilt" subset) were designated for additional filtering by the small classifier, as explained in detail below; 58 samples (referred to herein as the "test" subset) were used to test classifier performance together with the melanoma development sample set; and 57 samples (referred to herein as the "validation" subset) were used as part of an internal validation set, in addition to the melanoma validation subset already maintained for this purpose. A method was implemented to select this partitioning while ensuring that the 3 subsets were balanced. The details of how these partitions were generated are not particularly important. Suffice it to say that mass spectral features associated with overall survival have been analyzed and subsets selected such that differences in these features across the three subsets are minimized. Clinical features are listed for each subset in Table 19 and are shown in Table 19. Figure 17A and 17B A comparison of event time data is shown in .
[0417] Table 19. Baseline characteristics of NSCLC patients with available spectra for each of the three created subsets.
[0418]
[0419]
[0420] Sample preparation
[0421] The sample preparation of serum samples from the ACORN NSCLC cohort was the same as that of the nivolumab cohort and is described in detail in Example 1.
[0422] Spectral acquisition
[0423] The mass spectrometry data for the ACORN NSCLC cohort were acquired identically to those for the nivolumab cohort and are described in detail in Example 1.
[0424] Spectral processing
[0425] Mass spectral data processing, including grid spectrum preprocessing, deep MALDI average spectrum preprocessing (background subtraction, regularization, alignment), and batch correction, was the same as described in Example 1. In the project of Example 2, feature 9109 was not used in the development of the classifier; however, all other 350 features listed in Appendix A were used.
[0426] Classifier development
[0427] use Figure 8 The classifier development process of Example 1 (described in detail in Example 1) was used to develop the classifier of Example 2.
[0428] Definition of category labels (step 102, Figure 8A )
[0429] The classifier development of Example 2 used OS data from the melanoma / nivolumab development set to initially assign classification group labels to samples in the development set. In this case, the class labels were not obvious, as Figure 8 As shown in , the classifier development process uses an iterative approach to refine the class labels while creating the classifier. Figure 8B , step 144 and loop 146. In the initial iteration of the method described in step 102, the class labels are initially guessed. The samples are sorted by OS, where half of the samples with the lowest event time results are assigned the "early" class label (early death, i.e., poor outcome), and the other half are assigned the "late" class label (late death, i.e., good outcome). Then, according to Figure 8A classifier is constructed using the resulting data and these class labels. This classifier is used to generate classifications for all development set samples. The labels of samples that are consistently misclassified (loop 146) in the clusters of the main classifier created by multiple partitioning into training and test sets are flipped. This optimized set of class labels is then used as the new class labels for the second iteration of the method in step 102. This process is iterated until convergence, exactly as in Example 1.
[0430] Select training and test sets (108, Figure 8A )
[0431] The development set samples are split into training and test sets in a number of different random realizations (step 108). 625 realizations are used (i.e., iterations through loop 135). Figure 8 The method works best when the early and late training classes have the same number of examples in each realization or iteration through loop 135. Therefore, if the number of members of the training classes is different, they are divided into testing and training in different proportions.
[0432] Creation and filtering of small classifiers (120, 126, Figure 8A )
[0433] In step 120, a subset of features is used to construct a number of k-nearest neighbor (kNN) mini-classifiers (mC) using the training set as their reference set. In this project, k = 9 was used. In order to be able to consider subsets of single and double features and improve classifier performance, features that are not useful for classification were eliminated from the set of 350. This was done using the bagged feature selection method described in Appendix F of the previous provisional application serial number 62 / 289,587 and in U.S. patent application serial number 15 / 091417 by J. Roder et al.
[0434] To target a final classifier with specific performance characteristics, Figure 8A These mCs are filtered in step 126 of . Each mC is applied to its training set and the "AddFilt" NSCLC subset, and performance metrics are calculated from the resulting classifications. Only mCs that meet these performance metric thresholds pass the filtering for further use in this process. mCs that fail the filtering are discarded. For this project, a hazard ratio based OS filtering is used. Table 20 shows the filtering criteria of step 126 used to deselect small classifiers in each marker flipping iteration (step 146) of classifier development. In each iteration, a composite filter is constructed by combining two single filters using an "and" operation, see Table 21. Such filtering criteria are designed so that patients treated with immunotherapy are divided according to their OS results, but patients treated with chemotherapy are not.
[0435] Table 20: Criteria used in the filtration step
[0436]
[0437] Table 21 Parameters used in small classifiers and filtering
[0438]
[0439] Combining small classifiers using logistic regression and dropout (steps 130, 132)
[0440] Once the mCs were filtered, they were combined into a main classifier (MC) using a logistic regression trained on the training set class labels. To avoid overfitting, the regression was regularized using extreme value dropout, and only a few randomly selected mCs were included in each logistic regression iteration. The number of dropout iterations was chosen based on the typical number of mCs that passed the filter, ensuring that each mC was likely to be included in the dropout process multiple times. In all label flipping iterations, 10 mCs were randomly selected in each of the 10,000 dropout iterations.
[0441] Although the ACORN NSCLC "AddFilt" sample subset was used for the filtering step as described above, such samples were not included in the training of the logistic regression in combination with mC (or the reference set for the kNN small classifier). Only the training subset extracted from the development set (melanoma samples) was used for this purpose.
[0442] Training / testing split ( Figure 8A Loop 135)
[0443] The use of multiple train / test splits in loop 135, and in this example 625 different train / test set implementations, avoids selecting a single particularly favorable or difficult training set for classifier creation and avoids bias in performance evaluation due to testing on a test set that may be particularly easy or difficult to classify.
[0444] In step 150, the final test is defined
[0445] The output of the logistic regression that defines each MC is the probability of being in one of the two training categories (early or late). These MC probabilities for all 625 training and test set partitions can be averaged to produce an average probability for the sample. When using the development set, this approach is adjusted so that the MCs that are not included in the training set for a given sample are averaged ("out-of-bag" estimation). These average probabilities can be converted to binary classifications by applying a threshold (cutoff). During the iterative classifier construction and labeling optimization process, the classification is assigned by majority voting of the individual MC labels obtained with a cutoff of 0.5. This process is modified so that only MCs are selected where, for samples in the development set, the sample is not in the training set (modified or "out-of-bag" majority voting). This procedure gives a classification that is very similar to using a cutoff of 0.5 for the average probability on the MC.
[0446] result
[0447] The performance of the classifier developed in Example 2 was evaluated using Kaplan-Meier plots of OS between samples classified as early and late stages and the corresponding hazard ratios (HRs) and log-rank p-values. This performance evaluation was performed separately for melanoma patients treated with nivolumab and NSCLC samples treated with chemotherapy. Table 22 summarizes the results for the available time endpoints for each sample set. Figures 18A-18H The Kaplan-Meier plot corresponding to the data in Table 22 is shown in FIG. In the case of the ACORN SLC set, Figure 18A The plot of -H and the corresponding results in Table 22 do not take into account any of the "AddFilt" subset samples, as they were used directly in the mC filtering step for all 625 training / testing implementations. The classifications for all 119 samples in the melanoma cohort are listed in Appendix I of prior provisional application Ser. No. 62 / 289,587, along with the classifications obtained from Example 1 for comparison. Appendix C gives the features used in the final label flipping iteration of this classifier development.
[0448] Table 22: Performance summary
[0449]
[0450] Note that in Figures 18A-18D In the NIVO / melanocortisone cohort, OS and TTP curves showed clear separation between the early and late classes in the development and validation sets, whereas in the NSCLC / chemotherapy cohort, there was little or no separation between the early and late classes in the development and validation sets. This suggests that using the NSCLC cohort to filter small classifiers in the development of the main classifier and in the definition of the final classifier can be used to create a classifier that predicts the benefit of nivolumab in melanoma, but does not stratify the outcomes of NSCLC patients treated with chemotherapy.
[0451] The baseline clinical characteristics of the melanoma / nivolumab sample set (development + validation) are listed in Table 23. Table 24 summarizes the clinical characteristics of the ACORN NSCLC sample set ("testing" + "validation") by classification group.
[0452] Table 23: Clinical characteristics of the classification groups (melanoma sample set)
[0453]
[0454]
[0455] x One patient's data was missing. xx ULN = upper limit of normal range
[0456] Table 24: Clinical Characteristics of Classification Groups (ACORN NSCLC Sample Set)
[0457]
[0458] The reader will note that the data in Table 14 differs slightly from the data in Table 23. Of particular note, the classifications obtained from the classifier of Example 2 are not exactly the same as those obtained from the classifier of Example 1; however, as expected, a significant proportion of the samples received the same label for both classifiers. See Appendix I of prior provisional application serial number 62 / 289,587 for the labels generated by the classifiers of Example 1 and Example 2 for all 119 samples. The reader will also note that Appendix C lists a subset of the 350 features used in the final classifier of Example 2.
[0459] The idea behind the reduced feature set of Appendix C of the classifier of Example 2 is that instead of taking only all mCs (and related features) that give the prognostic behavior of nivolumab treatment, in the classifier of Example 2, only a subset of the following mCs are used that do not show separation in the ACORN NSCLC cohort except for predicting benefit / no benefit from nivolumab treatment. Therefore, a smaller subset of all mCs is used because there are additional constraints on their behavior. Then, when these mCs are combined, a classifier with different sample classifications and different behavior is obtained (at least on the ACORN NSCLC set). In order to obtain different behavior on the ACORN group, some different markers must be obtained on the melanoma group - but not enough to destroy the good separation between early and late populations obtained with the classifier of Example 1. It is not clear a priori whether this can be done, but the results of Example 2 show that it is possible to produce such a classifier.
[0460] The results of the multivariate analysis of the melanoma sample set (development + validation) are shown in Table 25. Two samples for which sex was not available and one sample for which LDH level was not available were not considered in this analysis.
[0461] Table 25: Multivariate analysis of OS and TTP in the melanoma set (development + validation)
[0462]
[0463] Although the classification of Example 2 was significantly correlated with LDH levels (Table 23), both quantities were independently significant predictors of TTP and OS in multivariate analysis. In addition, analysis of tumor size showed that, despite a strong correlation with early and late categories (Mann-Whitney p < 0.001), the median tumor size of the early and late groups was 48 cm and 16 cm, respectively. When adjusting for tumor size, the early / late category retained its significance as a predictor of OS and TTP (p = 0.014 for OS and p = 0.005 for TTP). Therefore, in addition to other prognostic factors such as LDH and tumor size, the classification also has independent predictive power.
[0464] Example 2 Independent Validation of the Classifier
[0465] The developed classifier was applied to several sample sets from two different cancer types (melanoma and ovarian) and treatments (immune checkpoint inhibitors and chemotherapy). For all samples, deep MALDI spectra were generated and processed using the same program used in development. For each sample, a classification of "early" or "late" was obtained. For each group, a Kaplan-Meier plot of the available time endpoints was shown, and an analysis summary of the event time was given.
[0466] A. Yale Anti-PD-1 Group
[0467] The following results involve a cohort of 30 pre-treatment samples from patients with advanced unresectable melanoma who were treated with an anti-PD-1 antibody at Yale University. Figure 19 Figure 26 shows the Kaplan-Meier curve for overall survival of the Yale cohort. It shows a clear separation of the survival curves between the early and late stages. Table 26 shows the statistical data of the results.
[0468] Table 26: Summary of Classifier Performance for the Yale Anti-PD-1 Antibody Treatment Cohort
[0469]
[0470] These results show that the classifier of Example 2 generalizes well to independent sample sets.
[0471] Table 27 summarizes the baseline clinical data of this cohort of patients.
[0472] Table 27: Baseline characteristics of the cohort
[0473]
[0474] Qualitatively similar results were obtained for the full set classifier developed according to Example 1, Table 13, OS filtering in this anti-PD-1 antibody treated patient cohort.
[0475] B. Yale Anti-CTLA4 Group
[0476] The following results relate to a cohort of 48 pre-treatment samples from patients with advanced unresectable melanoma who were treated with anti-CTLA-4 antibodies (ipilimumab or other similar antibodies) at Yale University. Table 28 summarizes the limited baseline clinical data available for this cohort of patients.
[0477] Table 28: Baseline Characteristics of the Yale Anti-CTLA4 Cohort
[0478]
[0479] Figure 20 2 is a Kaplan-Meier plot of overall survival for the Yale anti-CTLA4 cohort classified by the classifier of Example 2. Note the clear separation of the early and late classes generated by the classifier of Example 2 in this cohort, demonstrating the ability of the classifier to predict benefit for melanoma patients administered anti-CTLA4 antibodies. Table 29 lists the statistics for the performance of the classifier for this cohort.
[0480] Table 29: Summary of classifier performance for the Yale cohort treated with Ipilimumab
[0481]
[0482] B. Ovarian chemotherapy
[0483] The following results relate to a cohort of 138 pre-treatment samples from ovarian cancer patients who were treated with platinum-doublet chemotherapy after surgery. Disease-free survival (DFS) data were not available for two of these patients. Figure 21A (overall survival) and Figure 21B (Disease-free survival) shows a Kaplan-Meier plot of the ovarian cancer cohort classified by the classifier of Example 2. Note that the classifier does not stratify early and late stage categories in this cohort. This is consistent with the classifier not stratifying NSCLC patients treated with chemotherapy. Table 30 lists the Figure 21A and 21BStatistics of the classifier performance are shown in .
[0484] Table 30: Summary of classifier performance for the cohort of ovarian cancer patients treated with platinum doublet chemotherapy following surgery
[0485]
[0486] Reproducibility
[0487] Two more sample sets (melanoma / nivolumab and Yale anti-CTLA4 cohorts) were run to assess the reproducibility of the classifier of Example 2. Spectra were acquired in batches completely separate from the initial run. In all cases, the mass spectrometer used for the initial and rerun mass spectra acquisitions was used during this time on other projects of the assignee. The reproducibility of the classifier signatures is summarized in Table 31.
[0488] A. Melanoma / Nivolumab Group
[0489] In 113 of the 119 samples, the classifier was consistent between the initial run and the rerun, with an overall agreement of 95%. In the rerun, of the 43 samples initially classified as early, 38 were classified as early and 5 as late. In the rerun, of the 76 samples initially classified as late, 75 were classified as late and 1 as early.
[0490] B. Yale Anti-CTLA4 Group
[0491] The reproducibility of the classifier was assessed only on a subset of 43 samples from the cohort. In 40 of these 43 samples, the classifier was consistent between the initial run and the rerun, with an overall agreement of 93%. In the subset of 43 samples, of the 13 samples initially classified as early, 12 were classified as early and 1 as late. In the rerun, of the 30 samples initially classified as late, 28 were classified as late and 2 as early.
[0492] Table 31: Reproducibility of the prediction classifier across different sample sets
[0493] Sample Set Markup consistency Nivolumab 113 / 119(95%) Anti-CTLA4 40 / 43(93%)
[0494] Example 2 Conclusion
[0495] They were able to construct a classifier that could separate patients treated with immune checkpoint inhibitors into groups with better and worse outcomes (TTP, OS), while failing to separate patients treated with chemotherapy based on outcomes (OS and PFS or DFS). The classifier was constructed using deep MALDI mass spectra generated from pre-treatment serum samples and trained on 119 melanoma samples, half of which were treated with nivolumab. A third of 173 NSCLC samples treated with platinum doublet plus cetuximab chemotherapy were used to adjust the classifier for predictive, rather than just prognostic, results.
[0496] The classifier generalized well to stratify two independent cohorts of melanoma patients treated with immune checkpoint inhibitors by outcome and did not stratify a third cohort of ovarian cancer patients treated with postoperative platinum-doublet chemotherapy by outcome.
[0497] The classifier showed acceptable reproducibility in two independent cohorts, with a concordance of 93% or higher.
[0498] The following clauses are provided as further description of the invention disclosed in Example 2:
[0499] 1. An improved general-purpose computer configured as a classifier for classifying a blood sample from a human cancer patient to predict the patient's relative likelihood of survival or benefit from a drug, comprising:
[0500] a memory storing a reference set in the form of feature values for a plurality of mass spectral features derived from mass spectrometric analysis of blood samples from a plurality of melanoma patients treated with an immune checkpoint inhibitor, and an associated class label for each blood sample in a training set, the blood samples forming a classifier development set;
[0501] The memory further stores a set of computer executable codes for defining a final classifier based on a plurality of master classifiers, each master classifier being generated from a set of filtered small classifiers that perform a classification algorithm and combined using a regularized combination method; wherein the plurality of master classifiers are obtained from a number of different implementations of splitting the development set into classifier training and test sets; and
[0502] a central processing unit that operates on the code set, the reference set, and mass spectrometry data obtained from a blood sample of a cancer patient to be tested, and generates a class label for the blood sample; and
[0503] The small classifiers are filtered, in part, by their classifier performance on feature values of a set of mass spectrometric data obtained from blood samples of non-small cell lung cancer (NSCLC) patients.
[0504] 2. An improved computer as described in claim 1, wherein the code set is programmed to generate a category marker in the form of early or equivalent or late or equivalent for the sample, wherein the category marker early or equivalent predicts that the patient is likely to obtain relatively less benefit from the immune checkpoint inhibitor drug, and the category marker late or equivalent indicates that the patient is likely to obtain relatively greater benefit from the immune checkpoint inhibitor.
[0505] 3. The improved computer of clause 2, wherein the immune checkpoint inhibitor comprises a monoclonal antibody that blocks ligand activation of programmed cell death 1 (PD-1).
[0506] 4. The improved computer of clause 2, wherein the cancer patient providing the blood sample to be tested has been diagnosed with lung cancer, ovarian cancer, or melanoma.
[0507] 5. The improved computer of clause 2, wherein the relatively greater benefit associated with late-stage markers means significantly greater (longer) overall survival compared to early-stage markers.
[0508] Example 3
[0509] A classifier for predicting melanoma patients who would benefit from anti-CTLA4 antibodies
[0510] Above Figure 20 Table 29 shows that the classifier of Example 2 is able to predict whether the administration of anti-CTLA4 antibodies will benefit melanoma patients. Ipilimumab is a specific example of such a drug. It is understood that most patients in the Yale cohort received ipilimumab treatment. It is uncertain whether they all received ipilimumab rather than other anti-CTLA4 antibodies under development. These results indicate that it is possible to predict from a blood sample whether a melanoma patient is likely to benefit from an anti-CTLA4 antibody drug before treatment. It is further noted that the classifier for predicting the benefit of a patient from an anti-CTLA4 antibody drug was developed from a sample set of patients treated with the anti-PD-1 antibody nivolumab.
[0511] It was further found that the full set classifier of Example 1 was also well validated in this group. (The term "full set classifier of Example 1" refers to the Figure 8 Classifier developed on all 119 patient samples in the nivolumab cohort, as explained in Example 1 above, see discussion of Method 2 in Table 13, OS mini-classifier filtering.) Figure 23 Kaplan-Meier curves for the early and late groups of the ensemble classifier from Example 1 applied to this cohort are shown. Table 32 is a summary of the classifier performance for the ensemble classifier of Example 1 for the Yale anti-CTLA4 antibody treated cohort.
[0512] Table 32: Classifier performance summary
[0513]
[0514] Table 33: Baseline characteristics of the cohort according to classification groups
[0515]
[0516] A subset of 43 samples from this group (16 classified as early and 27 classified as late) was rerun with independent sample preparation and spectral acquisition. Of the 43 samples, 41 were assigned the same class label as the initial run (95% consistency). Two samples that were initially classified as early were classified as late in the rerun.
[0517] A classifier was also developed to identify patients with better and worse outcomes on anti-CTLA4 therapy using pre-treatment serum samples from patients subsequently treated with an anti-CTLA4 agent. Relevant aspects of this classifier development exercise and the performance of the classifier developed from this sample set will now be described.
[0518] In this classifier development exercise, 48 pre-treatment serum samples and associated clinical data were obtained from a cohort of patients who were subsequently treated with an anti-CTLA4 agent. Most patients were known to have received ipilimumab, but some may have received another anti-CTLA4 antibody. These were the same 48 patients who ran the validation tests for Examples 1 and 2, which are described in Example 3 above. Overall survival (OS) was the only available endpoint.
[0519] For the development of this classifier, the same spectra and the same spectral processing and features, i.e., the same feature table, were used as in Examples 1 and 2. The only difference was that feature 9109 was deleted from the feature table due to concerns that it had reproducibility issues and was of little value for classification.
[0520] In classifier development, we used Figure 8 The same Diagnostic Cortex method was used as described in detail in the description of Examples 1 and 2, with label flipping iterations. Initial category definitions were based on shorter or longer OS. A small classifier based on the hazard ratio of OS was used to filter between the classification groups in the training set.
[0521] The resulting classifier assigned 16 patients to the early group and 32 to the late group. (This compares to 20 patients in the early group and 28 patients in the late group generated by the full set classifier in Example 1.) The Kaplan-Meier curves of classifier performance for the groups defined by the new classifier trained on the anti-CTLA4 cohort are shown in Figure 2. Figure 23A Note Figure 23A The figure shows a clear separation in overall survival between the early and late groups. No patients in the early group survived beyond 3 years. The classifier performance statistics are as follows:
[0522]
[0523] Figure 23B The comparison of this result with the classifier performance obtained by running the full set classifier of Example 1 on the Yale cohort is shown in . Figure 23B There is almost complete overlap between the early and late groups in the classifier developed from the melanoma, nivolumab sample set ("Example 1 Full Set Classifier") and the classifier developed from the Yale anti-CTLA4 cohort ("Early" and "Late").
[0524] Therefore, comparison Figure 23 The Kaplan-Meier plots and statistics of the two classifiers show that the classifier performance results for the classifier developed on the anti-CTLA4 cohort do not appear to be significantly better than the classifier developed on the nivolumab-treated cohort. The fact that the two classifiers produced very similar results using completely different development sets supports the assertion that Figure 8 The Diagnostic Cortex method produces tests that are not tailored to the development sample set, but rather extract information that will generalize to other sample sets.
[0525] In one embodiment, a test method for predicting whether a melanoma patient is likely to benefit from an anti-CTLA4 antibody drug involves the following steps:
[0526] (a) performing mass spectrometry analysis on a patient's blood sample and obtaining mass spectrometry data;
[0527] (b) obtaining integrated intensity values in the mass spectral data for a plurality of predetermined mass spectral features (e.g., the features listed in Appendix A or a subset thereof, e.g., after deselecting noise features that do not significantly contribute to classifier performance, e.g., the features of the set in Appendix B or Appendix C); and
[0528] (c) using a program-controlled computer implementing a classifier (e.g., as described in Example 2) Figure 8 The generated classifier or the full set classifier of Example 1 operates on the mass spectrometry data.
[0529] In the operation step, the classifier compares the integrated intensity value with the feature value of a training set of class-labeled mass spectral data obtained from blood samples obtained from many other melanoma patients. The training set can be one of two types, namely, class-labeled mass spectral data from a sample set of patients treated with an antibody drug (e.g., nivolumab) that blocks ligand activation of programmed cell death 1 (PD-1). Alternatively, the training set can be class-labeled mass spectral data from a sample set of patients treated with anti-CTLA4 antibodies. The classifier performs this comparison using a classification algorithm. The classifier generates a class label for the sample, where the class label "early" or equivalent predicts that the patient may obtain relatively less benefit from the anti-CTLA4 antibody drug, and the class label "late" or equivalent indicates that the patient may obtain a relatively large benefit from the anti-CTLA4 antibody drug.
[0530] Additionally, preferably, mass spectral data is acquired from at least 100,000 laser shots performed on the blood sample using MALDI-TOF mass spectrometry.
[0531] The classifier is preferably obtained from a combination of filtered small classifiers using a regularized combination method. As described in Example 2, the small classifiers can be filtered by retaining the small classifiers defined during classifier generation, so that the mass spectra of patients treated with immunotherapy are divided according to their OS classification results, but when such a small classifier is applied to the mass spectra of patients receiving chemotherapy, these mass spectra will not be divided. This is considered optional because the "full set" classifier of Example 1 that can be used in this test was developed without any filtering of any small classifiers on the sample set from the chemotherapy group.
[0532] The test environment for performing the test of Example 3 on the blood sample of a melanoma patient can be Figure 15 The form of the system is shown and described in detail in Example 5 below.
[0533] The following clauses are provided as further description of the disclosed invention of Example 3.
[0534] 1. A method for predicting the response of a melanoma patient to an antibody drug targeting CTLA-4, comprising:
[0535] a) performing mass spectrometry analysis on a patient's blood sample and obtaining mass spectrometry data;
[0536] (b) obtaining integrated intensity values of a plurality of predetermined mass spectral features in the mass spectral data; and
[0537] (c) operating the mass spectrometry data using a programmable computer implementing the classifier;
[0538] Wherein, in the operation step, the classifier uses a classification algorithm to compare the integrated intensity value with feature values of a reference set of class-labeled mass spectrometry data obtained from blood samples obtained from multiple other melanoma patients, wherein the other melanoma patients are treated with (1) an antibody drug targeting programmed cell death 1 (PD-1) or (2) an antibody drug targeting CTLA4, and generates class labels for the samples, wherein the class label "early" or an equivalent predicts that the patient may obtain relatively less benefit from the antibody drug targeting CTLA-4, and the class label "late" or an equivalent indicates that the patient may obtain relatively greater benefit from the antibody drug targeting CTLA4.
[0539] 2. The method of clause 1, wherein the predetermined mass spectral characteristics include a plurality of characteristics listed in Appendix A, Appendix B, or Appendix C.
[0540] 3. The method of clause 1, wherein the classifier is configured as a combination of filtered small classifiers using a regularized combination method.
[0541] 4. The method of clause 1, wherein the mass spectral data is acquired from at least 100,000 laser shots performed on the sample using MALDI-TOF mass spectrometry.
[0542] 5. The method of clause 1, wherein the small classifiers are filtered according to any one of the criteria listed in Table 10.
[0543] 6. The method of clause 1, wherein the classifier is defined based on a plurality of main classifiers generated by splitting a development sample set multiple times to generate a training set and a test set.
[0544] 7. The method of clause 1, wherein the reference set is in the form of class-labeled mass spectrometry data obtained from blood samples obtained from a plurality of melanoma patients treated with an antibody drug targeting programmed cell death 1 (PD-1).
[0545] 8. Machines that predict whether antibody drugs targeting CTLA-4 will benefit melanoma patients, including:
[0546] a memory storing a reference set in the form of feature values for a plurality of mass spectral features obtained from mass spectra of blood samples of a plurality of melanoma patients who were (1) treated with an antibody drug that blocks ligand activation of programmed cell death 1 (PD-1) or (2) treated with an antibody drug that targets CTLA4;
[0547] The memory further stores a set of codes for defining a final classifier based on a plurality of main classifiers, each main classifier being generated from filtered mini-classifiers combined using a regularized combination method;
[0548] a central processing unit that operates on the code set and the reference set and mass spectrometry data obtained from a blood sample of a melanoma patient to be tested, and responsively generates a class label for the blood sample, wherein the class label "early stage" or an equivalent predicts that the patient is likely to derive relatively little benefit from an antibody drug targeting CTLA4, and the class label "late stage" or an equivalent indicates that the patient is likely to derive relatively greater benefit from an antibody drug targeting CTLA4.
[0549] 9. A machine as described in clause 8, wherein the memory stores integrated intensity values of multiple features listed in Appendix A or Appendix B or Appendix C.
[0550] 10. A method for treating a melanoma patient, comprising the step of administering to the patient an antibody drug targeting CTLA-4,
[0551] wherein the patient has been preselected for such administration by performing the method of any one of clauses 1 to 7 on a blood sample of the patient and the patient has been assigned a late or equivalent class label.
[0552] Example 4
[0553] Classifier for predicting improvement or worsening in patients with ovarian cancer and NSCLC receiving chemotherapy
[0554] It was found that the classifier of Example 1 divided a group of 173 first-line, advanced non-small cell lung cancer (NSCLC) patients treated with platinum doublets + cetuximab, and another group of 138 ovarian cancer patients treated with platinum doublets after surgery into groups with better or worse OS and progression-free (or disease-free) survival (PFS or DFS). This section will describe the actual test for predicting better or worse survival rates for ovarian cancer and NSCLC patients. This Example 4 will also describe further details of the ACORN NSCLC and ovarian cancer groups. Please note that the classifier of Example 2 (using the ACORN NSCLC group filtered for small classifiers) cannot identify those ovarian cancer and NSCLC patients who may benefit from chemotherapy, so the classifier and test for predicting better or worse survival of ovarian and NSCLC with platinum chemotherapy are constructed according to Example 1.
[0555] ACORN NSCLC Cohort
[0556] A cohort of 173 pre-treatment blood samples was available from patients with previously untreated advanced non-small cell lung cancer (NSCLC). The patients received platinum-based chemotherapy with cetuximab as part of a clinical trial. The most important baseline clinical data available for this patient cohort are summarized in Table 34, with OS and progression-free survival (PFS) for the entire cohort shown in Table 34, respectively. Figure 24A and 24B middle.
[0557] Table 34: Baseline Characteristics of the ACORN NSCLC Cohort
[0558]
[0559]
[0560] Deep MALDI spectra were generated from these samples, but in the same manner as described in Example 1, and these samples were processed using the same procedure used in the development of the classifier of Example 1. The full set classifier of Example 1 was applied to the resulting feature table, generating a classification of "early" or "late" for each sample. One hundred and sixteen samples were classified as early and the remaining 57 samples as late. Figure 25A and 25B Kaplan-Meier curves for overall survival of the cohort by classification group are shown, and a summary of the analysis of OS and PPFS is given in Table 35. The baseline characteristics of the patients by classification group are summarized in Table 36.
[0561] Table 35: Summary of classifier performance in the ACORN NSCLC cohort
[0562]
[0563] Table 36: Baseline characteristics of classification according to the ensemble classifier of Example 1
[0564]
[0565]
[0566] Figure 26A and 26B Shown are time-to-event outcomes broken down by classification and VeriStrat signature (using the classification algorithm, signature definition, and NSCLC training set described in US Patent 7,736,905). Clearly, both VeriStrat arms in the early category group had similarly poor outcomes.
[0567] Ovarian cancer group
[0568] There are 165 samples available from an observational trial of patients with ovarian cancer. Patients who underwent surgery received platinum-based chemotherapy. Samples were taken at the time of surgery. Of these 165 patients, 23 patients did not actually start chemotherapy, were not newly diagnosed, or had received prior treatment for ovarian cancer. An additional 4 patients had no outcome data. Data for the remaining 138 patients are presented here. The most important baseline clinical data obtained for these patients are summarized in Table 37, with OS and disease-free survival (DFS) shown in Table 37. Figure 27A and27B Note that 2 of the 138 patients had no DFS available.
[0569] Table 37: Baseline Characteristics of the Ovarian Group
[0570]
[0571]
[0572] Deep MALDI spectra were generated from these samples in the same manner as described in Example 1, and these samples were processed using the same procedures used in the development of the nivolumab test described in Example 1. The Example 1 full set classifier was applied to the resulting feature table, generating a classification of "early" or "late" for each sample. 76 samples were classified as early, and the remaining 62 were classified as late. Figure 28A and 28B Kaplan-Meier plots of overall and disease-free survival by category are shown in Table 3, and a summary of the OS and PFS analyses is given in Table 38. The baseline characteristics of the patients by category group are summarized in Table 39.
[0573] Table 38: Summary of the performance of the ensemble classifier of Example 1 in the ovarian cancer cohort
[0574]
[0575] Table 39: Baseline characteristics of the classifications produced by the ensemble classifier of Example 1
[0576]
[0577]
[0578] Figure 29A and 29B Shown are times to event outcomes broken down by classification and VeriStrat signatures (tested according to US Patent 7,736,905).It is clear that both the good and poor VeriStrat cohorts in the early categories generated by the ensemble classifier of Example 1 also had poor outcomes.
[0579] In summary, a test to predict whether a patient with ovarian cancer or NSCLC is likely to benefit from chemotherapy, such as a platinum-doublet chemotherapy, includes the following steps:
[0580] a) performing mass spectrometry analysis on a patient's blood sample and obtaining mass spectrometry data;
[0581] (b) obtaining integrated intensity values in the mass spectral data for a plurality of predetermined mass spectral features (e.g., the features listed in Appendix A or a subset thereof, e.g., after deselecting noise features that do not significantly contribute to classifier performance, e.g., the features of the set in Appendix B); and
[0582] (c) using a programmed computer implementing a classifier (e.g., based on Figure 8 The resulting classifier, the ensemble classifier of Example 1 ), operates on mass spectrometry data.
[0583] In the operation step, the classifier uses a classification algorithm to compare the integrated intensity value with the feature value of the training set of class-labeled mass spectrometry data obtained from blood samples obtained from multiple melanoma patients treated with an antibody drug that blocks ligand activation of programmed cell death 1 (PD-1), such as nivolumab. The classifier generates a class label for the sample, wherein the class label "early stage" or an equivalent predicts that the patient is likely to obtain relatively less benefit and / or have a poorer outcome from chemotherapy, and the class label "late stage" or an equivalent indicates that the patient is likely to obtain relatively greater benefit and / or obtain a better outcome from chemotherapy. In one embodiment, the chemotherapy is platinum doublet chemotherapy.
[0584] Additionally, preferably, mass spectral data is acquired from at least 100,000 laser shots performed on the blood sample using MALDI-TOF mass spectrometry.
[0585] The classifier is preferably obtained from an ensemble of filtered small classifiers using a regularized ensemble method.
[0586] The actual testing environment for testing patients with ovarian cancer and NSCLC is as follows Figure 15 and described in the next section.
[0587] The following clauses are provided as further examples of the invention disclosed in Example 4.
[0588] 1. A method for predicting overall survival in patients with non-small cell lung cancer (NSCLC) or ovarian cancer treated with chemotherapy, comprising:
[0589] a) performing mass spectrometry analysis on a patient's blood sample and obtaining mass spectrometry data;
[0590] (b) obtaining integrated intensity values of a plurality of predetermined mass spectral features in the mass spectral data; and
[0591] (c) operating the mass spectrometry data using a programmable computer implementing the classifier;
[0592] Wherein, in the operation step, the classifier uses a classification algorithm to compare the integrated intensity value with the feature value of a reference set of class-labeled mass spectrometry data obtained from blood samples obtained from multiple melanoma patients treated with an antibody drug targeting programmed cell death 1 (PD-1), and generates a class label for the sample, wherein the class label "early" or equivalent predicts that the patient may have relatively less benefit and / or a poor outcome (survival), and the class label "late" or equivalent indicates that the patient may obtain a relatively greater benefit and / or a better outcome (survival) from chemotherapy.
[0593] 2. The method of clause 1, wherein the chemotherapy comprises platinum doublet chemotherapy.
[0594] 3. The method of clause 1, wherein the chemotherapy comprises a combination of platinum doublet plus cetuximab chemotherapy.
[0595] 4. A method as described in clause 1, wherein the predetermined mass spectral characteristics include multiple characteristics listed in Appendix A or Appendix B.
[0596] 5. The method of any one of clauses 1 to 4, wherein the mass spectral data is acquired from at least 100,000 laser shots performed on the sample using MALDI-TOF mass spectrometry.
[0597] 6. The method of any of clauses 1 to 5, wherein the classifier is obtained from a filtered combination of small classifiers combined using a regularizer.
[0598] 7. The method of clause 6, wherein the small classifiers are filtered according to any one of the criteria listed in Table 10.
[0599] 8. The method of clause 6, wherein the classifier is obtained from a plurality of main classifiers generated by splitting the development sample set multiple times to generate a training set and a test set.
[0600] 9. A machine for predicting overall survival in patients with non-small cell lung cancer (NSCLC) or ovarian cancer treated with chemotherapy, comprising:
[0601] a memory storing a reference set in the form of characteristic values of a plurality of mass spectral features obtained from mass spectra of blood samples of a plurality of melanoma patients treated with an antibody drug that blocks ligand activation of programmed cell death 1 (PD-1);
[0602] The memory further stores a set of codes for defining a final classifier based on a plurality of main classifiers, each main classifier being generated from filtered mini-classifiers combined using a regularized combination method;
[0603] a central processing unit that operates on the code set and the reference set and mass spectrometry data obtained from a blood sample of a NSCLC or ovarian cancer patient to be tested, and responsively generates a class label for the blood sample, wherein the class label "early stage" or an equivalent predicts that the patient is likely to have relatively less benefit and / or poorer survival, and the class label "advanced stage" or an equivalent indicates that the patient is likely to have relatively greater benefit from chemotherapy and / or better survival.
[0604] 10. A laboratory testing center comprising the machine of clause 9 and a MALDI-TOF mass spectrometer configured for performing mass spectrometric analysis of blood samples from patients with NSCLC or ovarian cancer.
[0605] 11. A method for treating a patient with NSCLC or ovarian cancer, comprising:
[0606] Administering chemotherapy to patients with NSCLC or ovarian cancer,
[0607] Wherein the method of any one of clauses 1 to 9 is performed on a blood sample of the patient, the patient having previously been selected for chemotherapy, and the sample being assigned a late stage or equivalent class label.
[0608] 12. The method of clause 11, wherein the chemotherapy comprises platinum doublet chemotherapy.
[0609] 13. The method of clause 11, wherein the cancer patient is an NSCLC patient, and wherein the chemotherapy comprises a platinum doublet plus cetuximab.
[0610] Example 5
[0611] Laboratory testing center and computer configured as a classifier
[0612] Once a classifier has been developed in conjunction with the classifiers described in Examples 1, 2, 3, 9 (or other examples), its parameters and reference set can now be stored and implemented in a general purpose computer and used to generate class signatures for blood samples, for example, according to the tests described in Examples 1, 2, 3, and 4. Depending on the specific clinical question being asked (and the type of patient from whom the sample was obtained), the class signature can predict in advance whether a melanoma or other cancer patient is likely to benefit from an immune checkpoint inhibitor, such as an antibody that blocks ligand activation of PD-1, or, for example, predict whether a melanoma or other cancer patient will benefit from an antibody that blocks CTLA4, such as nivolumab, or, if developed according to Example 1, predict whether an ovarian or NSCLC cancer patient is likely to have better or worse overall survival with chemotherapy.
[0613] Figure 15 Is used based on Figure 815. A diagram of a laboratory testing center or system that processes a test sample (in this example, a blood sample from a melanoma, ovarian, or NSCLC patient) using a generated classifier. The system includes a mass spectrometer 1506 and a general purpose computer 1510 having a CPU 1512 that implements a classifier 1520 encoded as machine-readable instructions and a memory 1514 that stores a reference mass spectral data set including a feature table 1522 of mass spectral data labeled for class. This reference mass spectral data set forming the feature table 1522 will be understood to be the same as that used to create the mass spectral data set. Figure 8 Mass spectrometry data associated with the development sample set of the classifiers of Examples 1-4 (integrated intensity values of predefined features, see Appendix A or Appendix B). This dataset can be from all samples, for example, for the full set of classifiers for Example 1 or a subset of samples (e.g., a development set of half of the samples) plus a set of mass spectrometry data from NSCLC patients from which the classifier of Example 2 was developed. It should be understood that Figure 15 The mass spectrometer 1506 and computer 1510 can be used according to Figure 8 The process generates a classifier 1520.
[0614] Figure 15 The operation of the system will be described in the context of performing a predictive test as described above to predict whether a patient will benefit or not benefit from an antibody that blocks ligand activation of PD-1. The following discussion assumes that the classifier 1520 has already been generated when using the classifier to generate a class label (early or late, or equivalent) for a test sample. The classifier 1520 used for other tests (benefit from anti-CTLA4 drugs, overall survival prediction for ovarian and NSCLC chemotherapy, etc.) Figure 15 The operation method is the same.
[0615] Figure 15 The system obtains multiple samples 1500, such as blood samples (serum or plasma) from patients with different cancers (e.g., melanoma), and generates class labels for the samples on a fee-for-service basis. A classifier 1520 (implemented in computer 1510) uses the samples 1500 to predict whether the patient providing the particular sample is likely to benefit from immune checkpoint inhibitor therapy. The result of the test is a binary class label assigned to the patient's blood sample, such as early or late stage, etc. The specific name of the class label is not particularly important and can be general, such as "Class 1", "Class 2", etc., but as previously described, the class label is associated with a clinical attribute that is related to the question being answered by the classifier. As previously described, in this article, the early stage class label is associated with a prediction of relatively poor overall survival, and the late stage class label is associated with a prediction of relatively better (longer) overall survival with immune checkpoint inhibitors.
[0616] The sample can be obtained on a serum card or the like, where a blood sample is applied to a cellulose or other type of card. Aliquots of the sample are dropped onto one or more spots on a MALDI-TOF sample "plate" 1502, and the plate is inserted into a MALDI-TOF mass spectrometer 1506. The mass spectrometer 1506 acquires a mass spectrum 1508 from each spot of the sample. The mass spectrum is represented in digital form and provided to a programmed general purpose computer 1510. The computer 1510 includes a central processing unit 1512 that executes the programmed instructions. A memory 1514 stores data representing the mass spectrum 1508. Ideally, the sample preparation, spotting, and mass spectrometry steps are similar to those used for the analysis of the mass spectrum according to the method described in the accompanying drawings. Figure 8 The steps for generating the classifier are the same as those in Examples 1 and 2.
[0617] The memory 1514 also stores a data set representing a classifier 1520, which includes a) a reference mass spectrum data set 1522 having a feature table of spectra labeled for N classes, where N is some integer, in this example a development sample set of spectra used to develop the classifier as explained above, or some subset of the development sample set (e.g., DEV1 or DEV2 in Example 1 above, or all 119 samples). The classifier 1520 includes b) code 1524 representing a kNN classification algorithm (which is implemented in mini-classifiers, as described above), including features and depth (parameter s) of the kNN algorithm and identification of all mini-classifiers that pass filtering, and c) program code 1526 for performing a classification on the patient's mass spectra according to Figure 8 The final classifier generated in includes the logarithmic regression weights and data representing the main classifiers that form the final classifier, including probability cutoff parameters, small classifier parameters of each small classifier that passes filtering, etc., and d) a data structure 1528 for storing classification results, the classification results including the final category label for the test sample. Memory 1514 also stores program code 1530 for implementing the process shown at 1550, including code (not shown) for acquiring mass spectral data from a mass spectrometer in step 1552; a pre-processing routine 1532 for implementing background subtraction, normalization, and alignment steps 1554 (details detailed above); a module for filtering and averaging 800 emission spectra at multiple positions across each spot and multiple MALDI spots to form a single 100,000+ emission average spectrum (as explained above); a module for calculating integrated intensity values at predefined m / z positions in the background-subtracted, normalized, and aligned spectra (step 1556); and a code routine 1538 for implementing a final classifier 1520 using a reference data set signature table 1522 for the values obtained at step 1556. Process 1558 generates a class label at step 1560. Module 1540 reports the class label indicated at 1560 (i.e., "early" or "late," or equivalent).
[0618] Program code 1530 may include additional and optional modules, such as feature correction function code 1536 (described in U.S. Patent Application Publication No. 2015 / 0102216) for correcting for performance fluctuations in a mass spectrometer, a set of routines for processing spectra from a reference sample to define a feature correction function, a module for storing noise signatures associated with a feature and generating noise signature realizations and classifying these noise signature realizations, a module for storing a statistical algorithm for obtaining statistics on the performance of a classifier for noise signature realizations, or a module for combining class labels defined from multiple individual replicates of a sample to generate a single class label for the sample. It will be apparent to one skilled in the art that other optional software modules may also be included.
[0619] Figure 15 The system can be implemented as a laboratory test processing center that obtains multiple patient samples from oncologists, patients, clinics, etc. and generates class labels for the patient samples on a fee-for-service basis. The mass spectrometer 1506 need not be physically located at the laboratory test center, but rather the computer 1510 can obtain data representing the mass spectrum of the test samples over a computer network.
[0620] Example 6
[0621] Correlation between protein functional groups, taxonomic groups, and mass spectrometric features
[0622] When using Figure 8 When constructing a test program for a classifier, it is not necessary to be able to determine which proteins correspond to which mass spectral features in the MALDI TOF spectrum, or to understand the protein functions associated with these features. Whether the process produces a useful classifier depends entirely on the classifier's performance on the development set and how well the classifier performs when classifying a new set of samples. However, once a classifier is developed, it may be of interest to investigate the functions of proteins or proteins that are directly related to or correlated with the mass spectral features used in the classifier. Additionally, it may be beneficial to measure protein expression or protein function using other platforms related to the test classifier set.
[0623] Appendix K of prior provisional application serial number 62 / 289,587 sets forth the results of analyses designed to correlate protein function with the classification panels of the classifiers of Examples 1 and 2 and the mass spectral features measured in the deep MALDI spectra. In Appendix K, the term "IS2" corresponds to the full set Method 1 classifier of Example 1 and "IS4" corresponds to the classifier of Example 2. Details of the methods used and a summary of the results found are listed in this Example. The discoveries made lead to new examples of how the classifiers of the present disclosure can be characterized and generalized to other immune checkpoint inhibitors and other cancer indications beyond melanoma.
[0624] The data used for the study consisted of a table of features created when the full set method 1 classifier of Example 1 was applied to a set of 49 serum samples (the "Analysis Set") consisting of cancer patients and some donors without cancer. This is a table of feature values for each of the 59 features used in the full set method 1 classifier of Example 1 (see Appendix B) and the 292 other features not used in the full set method 1 classifier of Example 1 (see Appendix A). The feature values were obtained from MALDI-TOF mass spectra obtained for each of the 49 samples using the fully specified spectral acquisition and spectral processing procedures defined in the description of Example 1. The classification (early / late) list obtained for the 49 samples was also used in the analysis set generated by the full set method 1 classifier of Example 1. The classification list (early / late) obtained for the 49 samples was also used in the analysis set generated by the classifier of Example 2. A table of 1129 protein / peptide expression measurements obtained by running the SomaLogic 1129 protein / peptide panel on the analysis set was also used.
[0625] A method called gene set enrichment analysis (GSEA) was applied to protein expression data. Background information on the method is provided in Mootha et al., PGC-1α-responsive genes involved in oxidative phosphorylation are coordinately downregulated in human diabetes. Nat Genet. 2003; 34(3): 267-73 and Subramanian et al., Gene set enrichment analysis: A knowledge-based approach for interpreting genome-wide expression profiles. Proc Natl Acad Sci USA 2005; 102(43): 15545-50, the contents of which are incorporated herein by reference. Specific protein groups were created based on the intersection of the SomaLogic 1129 group target list and the GeneOntology / AmiGO2 and UniProt database query results.
[0626] The implementation of the GSEA method was performed using Matlab. Basically, in this method, the correlation r between a single protein and a group marker (early and late) was evaluated. Once these correlations have been calculated for each protein, the proteins are sorted according to the r value from largest to smallest, where a larger r value indicates a larger correlation, and a value of r=0 means that no correlation was found. Then, the enrichment score (ES) is calculated (as explained in the paper by Subramanian et al. above), which is designed to reflect the extent to which the elements of a particular protein group are overrepresented at the top or bottom of the sorted protein list. Two possible definitions of the enrichment score are considered, the details of which are listed on page 5 of Appendix K of the previous provisional application serial number 62 / 289,587. The p-value of the corresponding protein is also calculated to evaluate the significance of the deviation of the calculated enrichment score from the mean of its random distribution. A running sum (RS) is also calculated as part of the calculation of the enrichment score, the details of which are explained in Appendix K of the previous provisional application serial number 62 / 289,587.
[0627] Table 40 shows the results of the correlation of the proteome with the class label (early or late) of the method 1 classifier of Example 1 full set for 49 samples.
[0628] Table 40: Results of GSEA applied to the proteomes of the analysis set and the class labels of Example 1
[0629]
[0630]
[0631] At the p < 0.05 level, there was a correlation between the class markers and the protein groups corresponding to the following biological processes: acute inflammatory response, acute phase, and complement system. For the wound healing protein group, a correlation with a p value of approximately 0.1 was found. The correlation for the acute response had a p value of 0.02. Statistical methods (see Appendix K of previous provisional application serial number 62 / 289,587) were then used to identify subsets of proteins from the complement system, acute phase, acute response, and acute inflammatory response protein groups that were most important for these correlations, and the results are shown in Tables 41A, 41B, 41C, and 41D, respectively.
[0632] Table 41A: Proteins included in the extended leading edge group of complement (Amigo9). * indicates proteins located to the right of the RS minimum, Proteins with anti-correlations of at least the same magnitude as the RS maximum are indicated.
[0633]
[0634]
[0635] Table 41B: Proteins included in the extended front group of the acute phase (UniProt1). * indicates proteins located to the right of the RS minimum, Proteins with anti-correlations of at least the same magnitude as the RS maximum are indicated.
[0636]
[0637] Table 41C: Proteins included in the extended front group of the acute response (Amigol1). * indicates proteins located to the right of the RS minimum, Proteins with anti-correlations of at least the same magnitude as the RS maximum are indicated.
[0638]
[0639]
[0640] Table 41D: Proteins included in the extended front set of acute inflammatory response (Amigo1). * indicates proteins located to the right of the RS minimum, Proteins with anti-correlations of at least the same magnitude as the RS maximum are indicated.
[0641]
[0642]
[0643] The interactions of ontologically related processes associated with the protein groups identified from the RS graph were further investigated. These processes were found to be related to complement activation, activation and regulation of the immune system, and innate immune responses and inflammatory responses. A chart showing these relationships is found in Appendix K of the previous provisional application serial number 62 / 289,587. In Appendix K, Figure 8 Also shown is a "heat map" (i.e., a plot of p-values generated by GSEA analysis correlating mass spectral features with protein expression values for all m / z features used by the classifiers of Example 1, Appendix B), illustrating the results of the correlation of the proteome with the mass spectral features used in the full set of Method 1 classifiers of Example 1. See also Figure 45A and 45Band as discussed later in this article. Many of the mass spectral features used in the classifier of Example 1 were found to be associated with the following biological processes: (1) acute phase, (2) acute response, (3) complement system, and (4) acute inflammatory response. Few of the mass spectral features used were associated with the specific immune-related protein functions studied (i.e., "activation of innate immune response," "regulation of adaptive immune response," "immune T-cell," "immune B-cell," "interferon," "interleukin-10"). These relationships were established by the previous provisional application serial number 62 / 289,587 in Appendix K. Figure 8 The heatmap indicates that darker regions are associated with lower p-values and therefore higher correlations between the protein and the class label. Figure 8 Heatmaps are shown for two different definitions of enrichment score, but the plots are similar. See also the Figure 45A and 45B And the related discussion below.
[0644] The same analysis was performed on the correlation between the proteome and the class signature generated by the classifier of Example 2. The results are shown in Table 42. Note that proteins associated with the following biological processes are strongly correlated with the class signature: acute inflammatory response, complement system, acute response, and acute phase. In addition, proteins associated with immune response type 2 are also strongly correlated with the class signature.
[0645] Table 42: Results of GSEA applied to the proteomes of the analysis set and the class labels of Example 2
[0646]
[0647]
[0648] At the p < 0.05 level, there were correlations of the category markers with the protein groups corresponding to acute inflammatory response, acute response, acute phase, complement system, interleukin-10 (ES definition 1), and immune response type 2. For the wound healing proteome, correlations with p values below 0.1 were found.
[0649] Proteins from the complement system, acute phase, acute response, acute inflammatory response, and interleukin-10 were further identified, as well as immune response type 2 processes important for the classifier of Example 2. The results are listed in Tables 11A-11F of Appendix K of the previous provisional application serial number 62 / 289,587. Many of the proteins listed in Tables 11A-11D are shown in Tables 41A-41D. The correlation of the proteome with the mass spectral features used in the classifier of Example 2 and the resulting heat maps was also calculated, see Figures 10 and 11 of Appendix K of the previous provisional application serial number 62 / 289,587. Many of the mass spectral features used in the classification of Example 2 are related to the complement system and acute inflammation / acute phase response. However, these biological processes are somewhat less dominant than the mass spectral features used in the classifier of the complete set of Example 1, Method 1. Table 12 of Appendix K of prior provisional application serial number 62 / 289,587 shows the number of mass spectrometry features (out of the 351 features listed in Appendix A) associated with the proteome at various significance levels (p-values). In general, there were more mass spectrometry features associated with proteins involved in acute inflammation, the complement system, acute response, acute phase, type 2 immune response, and wound healing processes than with other processes.
[0650] The above discussion, as well as the report in Appendix K of the previous provisional application serial number 62 / 289,587, demonstrates that the combination of deep MALDI measurements with a simple modification of GSEA shows promise in extracting useful information about the biological functions associated with the test markers of the present disclosure. It also enables a deeper understanding of the functions associated with the mass spectral features (deep MALDI peaks) measured from serum samples.
[0651] The protein functions associated with the mass spectrometric features used in the full set method 1 classifier of Example 1 are consistent with the protein functions associated with the class markers (i.e., acute phase reactants and complement system). These functions are also consistent with the functions of the available protein IDs of the features used in the same Example 1 classifier (Table 17, Example 1). Other reasonable biological functions did not show any significant association with the class markers. This does not mean that these other functions are unrelated to the biology of immunotherapy; it simply means that there is no evidence that they play a major role in the classification produced by the full set method 1 classifier of Example 1. However, although features (proteins) associated with most of these other functions were measured, they were not used in these tests and the test classifications were not significantly correlated with these protein functions.
[0652] In the case of the Example 2 classifier, it was found that the strength of association of the classification with the acute phase and complement function was significantly increased compared to Example 1, which may indicate that the Example 2 classifier is a "purer" test that is less confounded by prognostic effects. IL-10-related functions were also observed to be associated with the Example 2 category group at a significance level of p = 0.05.
[0653] These data have limitations, primarily due to the limited size of the analyzed sample set, which results in a rather wide distribution of null points and, therefore, limited statistical power. The simplest approach would be to acquire paired deep MALDI / Somalogic data on a larger number of samples. This would also enable independent validation of these results. While the number of proteins in the Somalogic panel is quite large, the use of additional or expanded panels could also be considered.
[0654] Due to limited computer resources, false discovery rate analysis was not performed for the proteome. Such an analysis would require further theoretical work to assess the applicability of the ensemble regularization suggested in the Subramanian paper, and there may be room for improvement. While this should in principle be done, the observed effects, particularly for certain mass spectrometric features, are so clear and large that we do not expect any substantial qualitative change in the main conclusions of this analysis.
[0655] Because it is well known that many proteins in the circulation are associated with acute phase reactants and the complement system, it may not be surprising that these two functions appear to correlate with many of the mass spectral features measured using deep MALDI. However, other significant correlations were found, particularly in the case of the Example 2 classifier, suggesting that these results are not a trivial reflection of circulating protein abundance. Furthermore, in the subset of features used in the full set Method 1 Example 1 classifier (see Appendix B), the proportion of features associated with the acute phase and complement systems was significantly higher than that observed in the entire 351 deep MALDI feature set listed in Appendix A.
[0656] The above findings can be used to construct a classifier for guiding immune checkpoint inhibitor therapy in cancer patients. Of particular note, such a classifier includes a memory that stores a reference set of class-labeled mass spectral data obtained from blood samples of melanoma patients treated with immune checkpoint inhibitors. The mass spectral data is in the form of characteristic values of multiple mass spectral features, wherein the mass spectral features are consistent with at least proteins circulating in serum associated with the following biological processes: (1) acute phase, (2) acute response, (3) complement system, and (4) acute inflammatory response. See the discussion above and in Appendix K of the previous provisional application serial number 62 / 289,587. Figure 8 , 10 and 11. The classifier also includes a program-controlled computer that performs a classification algorithm on a set of mass spectrometry data (see Figure 15), the set of mass spectral data includes characteristic values of multiple mass spectral features obtained from the test blood sample and a reference set, and generates a class label for the test blood sample. Alternatively, the mass spectral features may also include features related to immune response type 2 and interleukin-10 processes. In one embodiment, these features include the features listed in Appendix A, Appendix B, or Appendix C.
[0657] Furthermore, applying the GSEA method to the data obtained using the SomaLogic panel, combined with the results and further analysis of the full set Method 1 classifier of Example 1, enabled the up- and down-regulation of proteins to be associated with the early and late categories. Of particular note, proteins with positive correlation coefficients in Tables 41A-D were associated with up-regulation in samples classified as early.
[0658] Further analysis of the running sum of the complement proteomes allowed identification of proteins that had the greatest impact on the correlation between the corresponding proteome and the classification results (Table 5, Appendix K of the previous provisional application serial number 62 / 289,587). Table 43 below lists proteins that had a P value of ≤0.05 for correlation with the class marker and were included in group 1 (leading edge) of the complement proteome.
[0659] Table 43: Proteins associated with the gene ontology "Complement system" and associated with the Ensemble Method 1 classifier of Example 1.
[0660]
[0661]
[0662] It can be seen that the group classified as early is characterized by upregulation of most components of the complement system, raising the question of a possible biological relationship between this upregulation and the poor prognosis of patients classified as early, and the minimal benefit these patients receive from PD-1 inhibitors and possibly other types of immunotherapy.
[0663] There is increasing evidence that the complement system plays a complex role in tumorigenesis, which can be either pro-tumor or anti-tumor depending on the complex balance of multiple factors. Tumor transformation is accompanied by an increase in the ability to activate complement. Pio, R., Corrales, L. & Lambris, JD The role of complement in tumor growth. Adv Exp Med Biol 772, 229-62 (2014). Activated complement proteins have been shown to inhibit tumor growth by promoting complement-dependent cytotoxicity (Janelle, V. & Lamarre, A. Role of the complement system in NK cell-mediated antitumor T-cell responses. Oncoimmunology 3, e27897 (2014)) and inhibit Treg function. Mathern, DR & Heeger, PSMolecules Great and Small: The Complement System. Clin J Am Soc Nephrol 10, 1636-50 (2015).
[0664] On the other hand, recent data show that activated complement proteins interact with components of the innate and adaptive immune systems and promote carcinogenesis. For example, in a mouse model of melanoma, it was shown that "decomplementation" leads to a powerful anti-tumor CD8+ reaction and improved cytotoxic activity of NK cells, while the activated complement system leads to a limited accumulation of tumor-specific cytotoxic T cells (CTLs), while promoting tumor infiltration by utilizing immunosuppressive bone marrow-derived suppressor cells (MDSc) that inhibit NK cell and T cell function. Janelle, V. et al. Transient complement inhibition promotes atumor-specific immune response through the implication of natural killer cells. Cancer Immunol Res 2, 200-6 (2014). Similarly, complement activation and C5a signaling have been shown to be associated with the recruitment of MDSCs to tumors, the suppression of effector CD8+ and CD4+ T cells, the generation of regulatory T cells (Treg), the dominant immune response of Th2, and the promotion of lung and liver metastasis in breast cancer and cervical cancer models. Markiewski, MM et al. Modulation of the antitumor immune response by complement. Nat Immunol 9, 1225-35 (2008); Vadrevu, SK et al. Complement c5a receptor facilitates cancer metastasis by altering T-cell responses in the metastatic niche. Cancer Res 74, 3454-65 (2014). C5a was shown to promote Treg differentiation, leading to inhibition of anti-tumor activity. Gunn, L. et al. Opposing roles for complement component C5a in tumor progression and the tumor microenvironment. J Immunol 189, 2985-94 (2012). Complement can help tumor cells evade immune surveillance, support chronic inflammation, promote angiogenesis, activate mitogenic signaling pathways, maintain cell proliferation and insensitivity to apoptosis, and participate in tumor invasion and migration. Pio et al., supra. In a lung cancer model, blocking C5a signaling led to the inhibition of key immunosuppressive molecules within the tumor. These molecules include IL-10, IL-6, CTLA4, LAF3, and PDL18.
[0665] The latter finding has direct implications for the activity of immune checkpoint inhibitors in cancer patients, and therefore for the role of the classifiers in Examples 1 and 2. Specifically, the observed upregulation of complement system proteins classified as early may indicate that these patients have higher levels of immunosuppression associated with activation of the corresponding immune checkpoints, and / or higher levels of pro-tumor inflammation, and are therefore less responsive to drugs such as nivolumab, ipilimumab, pembrolizumab, or other drugs targeting these pathways. Interestingly, the complement protein C5a has been shown to promote the expression of the PD-1 ligands, PD-L1 and PD-L2. Zhang, J. Immunol. 2009; 182: 5123-5130. In this context, one can imagine that excessive complement upregulation may compete with efforts to inhibit PD-1. On the other hand, recent clinical trial results suggest that patients with tumor microenvironments characterized by high PDL1 expression and the presence of Tregs are more likely to respond to anti-PD-1, anti-CTLA4, or high-dose IL-2 therapy. Although it is not known how exactly the upregulation of the complement system is related to the classification of Example 1 and Example 2, this connection is consistent with the biological effects discussed above. Therefore, it can be expected that the classifiers of Example 1 and Example 2 may be related to a variety of drugs such as various immune checkpoint inhibitors, high-dose IL-2, vaccines and / or combination treatments that affect the patient's immunological state. In addition, since the effect measured in serum reflects the state of the entire organism, and the complement system affects innate and adaptive immunity at a global scale, not only at the tumor site, it is expected that the classifier has similar performance (such as lung cancer, kidney cancer) in different indications in cancer, and is not limited to melanoma.
[0666] Another large gene ontology protein group associated with the early and late categories of Examples 1 and 2 was "acute inflammatory response." Proteins associated with these classification groups from this set (p≤0.05) are listed in Table 44.
[0667] Table 44. Proteins associated with the gene ontology "acute inflammatory response" and classified according to the comprehensive method 1 Example 1.
[0668]
[0669]
[0670] It is generally believed that cancer triggers an intrinsic inflammatory response that creates a tumor-promoting microenvironment (Mantovani, A., Allavena, P., Sica, A. & Balkwill, F. Cancer-related inflammation. Nature 454, 436-44 (2008)); "smoldering" inflammation is associated with most, if not all, tumors and supports their progression. Porta, C. et al. Cellular and molecular pathways linking inflammation and cancer. Immunobiology 214, 761-77 (2009). Inflammation is intrinsically linked to the complement system, which promotes tumor growth in the context of inflammation. Therefore, it seems reasonable that these two systems are significantly associated with the classification of Examples 1 and 2.
[0671] Tumor-related inflammatory responses can be initiated and / or regulated by cancer treatment. On the one hand, it has a tumor-promoting function, and on the other hand, it can enhance the presentation of tumor antigens, thereby inducing an anti-tumor immune response. Grivennikov, SI, Greten, FR & Karin, M. Immunity, inflammation, and cancer. Cell 140, 883-99 (2010). Although a T cell inflamed microenvironment characterized by the recruitment of CD8+ and CD4+ lymphocytes to the tumor is considered a necessary condition for effective immunotherapy, the activation of acute inflammatory pathway components may be associated with a negative prognosis. The exact mechanism of action is still poorly understood, but the data on the upregulation of this system in early-stage groups seem consistent with existing clinical data.
[0672] This disclosure demonstrates how mass spectrometry data can be used to incorporate biological insights into classifier development exercises (e.g. Figure 8A and 8B The present disclosure also illustrates how mass spectrometry data can be used to test biologically motivated hypotheses about the relevance of biological functions to certain disease states. This is achieved by using gene set enrichment analysis to correlate biological functions with features (peaks) in mass spectra obtained from serum samples via subsets of proteins associated with these functions, and using these identifying features to train a classifier with the aid of a computer.
[0673] Specifically, a method for developing and training a classifier using mass spectra from a development sample set is disclosed. Protein expression data is obtained from a larger set of proteins that cover the biological function of interest for each sample in the development sample set, or more generally, for each sample in another sample group for which mass spectrometry data is available. The latter case is preferred because measuring the abundance of many proteins by protein determination requires a large number of samples and is expensive and time-consuming. Because the correlation between mass spectrometry features and functions can be inferred from any reference set with sufficient protein coverage, it is not necessary to construct a relationship between mass spectrometry peaks and biological functions for each development project. These methods are illustrated alternately with two options below.
[0674] In this approach, one or more mass spectral features are identified that are statistically significantly associated with a group of proteins grouped according to their biological function. These identified mass spectral features (typically 10-50 such features) are used to train a classifier with the aid of a computer. This training can take the form of a classifier development exercise, an example of which is Figure 8 As shown and described in detail above, a classifier is in the form of a set of parameters and associated program instructions or code that, when executed by a computer, assigns class labels to mass spectrometry data of samples of the same type as the development sample set according to programmed instructions.
[0675] As described in this embodiment, using the gene set enrichment analysis (GSEA) method, it is possible to find statistically significant associations of mass spectral features with groups of proteins grouped according to their biological function ("protein functional groups"), avoiding direct protein identification and taking advantage of the high throughput of mass spectra. The system or set of components that performs GSEA and associates such mass spectral features with specific protein functional groups or subsets is referred to herein as a "platform" or "GSEA platform". Such a platform consists of a known conventional protein expression assay system (e.g., the SOMAscan assay provided by SomaLogic of Boulder, CO) and a computer for implementing the GSEA analysis program, which is used to identify mass spectral features associated with protein functional groups, as described in this document.
[0676] Mass spectrometry data (preferably from a highly sensitive method such as "deep MALDI", see U.S. Patent 9,279,798, the contents of which are incorporated herein by reference) must be matched with protein expression data from a larger set of proteins covering the biological function of interest in a group of serum samples. Using well-known protein databases such as UniProt or GeneOntology / AmiGO2, protein subsets of all measured proteins can be defined according to their biological function. First, the entire list of measured proteins is sorted according to the correlation between each protein and the mass spectrometry feature of interest. Then, in the sorted list of all measured proteins, the GSEA method searches for proteins included in a specific protein function subset that are over- or under-represented as determined by the sorting and provides a way to assess their statistical significance. Thus, the association of the mass spectrometry feature of interest with different protein function groups can be assessed. The procedure can be repeated as needed for as many spectral features and as many protein function groups as possible. By setting a cutoff value for the degree of association or a p-value for the significance of the association, all mass spectrometry features associated with a specific protein function subset can be determined. This set of mass spectrometry features is then used to train a classifier, such as a k-nearest neighbor (kNN) classifier.
[0677] A preferred method for classifier training and development is called Diagnostic Cortex, which is Figure 8 It has been described in detail in the context of . It has been demonstrated that using the Diagnostic Cortex classifier development program, clinical data and deep MALDI mass spectrometry data can be combined to produce clinically useful molecular diagnostic tests. One advantage of this approach is that it allows tests to be designed and adjusted to meet the required clinical utility criteria. Using a subset of functionally relevant mass spectrometry features (determined from the procedures mentioned in the previous paragraph) instead of all available mass spectrometry data provides additional options for test design and optimization. Mass spectrometry features associated with one or more subsets of protein functions can be selected and combined with clinical data from a sample set to create new classifiers and related tests, allowing investigation of the relevance of individual biological functions or groups of biological functions for the desired classification task.
[0678] Also described are methods for creating multiple different classifiers using different feature subsets associated with different protein functional groups, and combining these classifiers, for example, by simple majority voting, more complex cluster averaging, or some rule-based system, to produce an overall classification that encompasses the information content of various biological functions. In one variation of this classification, after removing or accounting for major effects, by using peaks associated with functional groups in the classifiers and constructing new classifiers based on the hierarchy of biological functions, it is possible to find functional subsets that become relevant within the various groups defined by the classifiers. For example, a first-level classifier can be trained using peaks associated with acute response functions. A second classifier can be trained using a set of peaks associated with wound healing protein functions. Then, samples from the late stage (or equivalent) are tested on the first-level classifier and classified by the second-level classifier. If a sufficiently large set is available, this process can be further iterated, and a third-level classifier can be defined on a set of peaks associated with a third protein function and applied to the group classified by the second-level classifier, etc. Since multiple protein functional groups are often associated with a given peak, this approach attempts to disentangle complex effects.
[0679] Thus, in one aspect of the present disclosure, a method for generating a classifier is described, comprising the steps of: a) obtaining a development sample set from a subject population; b) performing mass spectrometry analysis on the development sample set and identifying mass spectral features present in the mass spectra of the development sample set; c) obtaining protein expression data from a larger set of proteins, wherein for each sample in the development sample set or for each sample in another sample group with associated mass spectrometry data, the larger set of proteins encompasses a biological function of interest; d) using the sample set with matched mass spectrometry and protein expression data, identifying one or more mass spectral features that are statistically significantly associated with groups of proteins grouped according to their biological function; and e) using the one or more mass spectral features identified in step d) and clinical data from the development sample set, with computer assistance, to train a classifier, wherein the classifier is in the form of a set of parameters that, according to program instructions, assign class labels to samples of the same type as the development sample set.
[0680] In another aspect, the invention may take the form of a programmed computer configured to generate a classifier according to the method of the preceding paragraph.
[0681] In another aspect, a method for testing a sample is disclosed, comprising the steps of: a) training a classifier using a set of mass spectral features that have been determined to have statistically significant associations with groups of proteins grouped according to their biological functions; b) storing parameters for the classifier comprising a table of features for the set of mass spectral features in a memory; c) performing mass spectrometry analysis on the test sample; and d) classifying the test sample using the trained classifier with the aid of a computer. In one variation, the steps include training two classifiers using different subsets of features associated with different functional groups of proteins, storing parameters for the first and second classifiers comprising a table of features for the set of mass spectral features in a memory, and logic instructions for combining the first and second classifiers into a final classifier; c) performing mass spectrometry analysis on the test sample; and d) classifying the test sample using the final classifier with the aid of a computer.
[0682] In another aspect, a computer configured as a classifier is disclosed, comprising a memory storing a feature table in the form of intensity data for a set of mass spectral features obtained from a biological sample, wherein the set of mass spectral features has been determined to have a statistically significant association with a group of proteins grouped according to their biological function, and a set of parameters defining a classifier, the classifier comprising a classification algorithm that operates on the mass spectral data from a test sample and the feature table.
[0683] In another aspect, a classifier development system is disclosed that includes a mass spectrometer for performing mass spectrometry analysis on a development sample set to generate mass spectrometry data, the data comprising a plurality of mass spectrometry features; a platform for performing gene set enrichment analysis on the development sample set, or more typically, on another sample set having associated mass spectrometry data, and identifying statistically significant associations of one or more mass spectrometry features with groups of proteins grouped by their biological function; and a computer programmed to use the one or more mass spectrometry features identified by the platform to train a classifier that assigns class labels to samples of the same type as the development sample set as programmed in the form of a set of parameters.
[0684] In the above methods and systems, the development sample set can take the form of a blood sample (serum or plasma) from a human, such as a human being enrolled in a clinical trial of a drug or drug combination. Such a human can be a cancer patient. The methods and systems of the present invention are described in the context of a development set of blood samples obtained from melanoma patients treated with immunotherapy drugs (i.e., programmed cell death 1 (PD-1) checkpoint inhibitors).
[0685] This paper demonstrates that it is possible to associate features in mass spectra with biological function without directly identifying the protein or peptide that generates the mass spectrometric feature and to incorporate biological insights into the selection of mass spectrometric features for reliable classifier training or development, for example, using the Diagnostic Cortex platform.
[0686] Correlating mass spectrometric features directly with biological processes is important because identifying the protein or peptide that produces a single peak in a mass spectrum is often difficult, time-consuming, and sometimes impossible. This approach avoids the need for thousands of protein identification studies, which, even if successful, do not always allow for the matching of biological processes to mass spectrometric peaks (the specific functions of many peptides and protein fragments remain to be determined).
[0687] The ability to discover mass spectrometric signatures generated by human serum that reliably correlate with biological processes provides a new approach to longitudinally monitor these processes in a minimally invasive, high-throughput manner. Serum samples can be collected from patients at many time points during treatment or disease, and changes in specific biological processes can be inferred from analysis of the mass spectra generated by the serum samples. Such changes may be due to intervention (e.g., treatment) or the natural evolution of the disease. While this study considered applications in oncology, this could be of interest across many disease areas.
[0688] Incorporating biological insights into classifier training for the development of reliable molecular diagnostic tests provides another avenue for designing and tuning the tests that can be created. Experience with the Diagnostic Cortex platform has shown that in some cases, the ability to tune a test to meet clinical needs is reduced and produces similar tests despite attempts to tune for other performance targets. This is thought to be due to the dominance of certain mass spectral features and correlations between features, but previous attempts to remove these dominant effects to allow investigation of other ancillary but potentially important effects have proven unsuccessful. The ability to determine which features are associated with a single biological process provides a new way to view the universe of mass spectral features that can be used for classification, allowing attempts to separate effects that may confound one another or remove processes that dominate classification in order to reveal other processes that can improve test performance and test biological hypotheses.
[0689] The application of feature selection based on GSEA may be very extensive and may lead to an extension of the understanding of the role of specific biological processes and related treatments in any disease. For example, type 2 diabetes is known to be a metabolic disorder. However, inflammation is also known to play an important role (see GL King et al., The role of inflammatory cytokines indiabetes and its complications.J Periodontol.2008Aug;79(8Suppl):1527-34.doi:10.1902 / jop.2008.080246and AO Odegaard et al., Oxidative stress, inflammation, endothelial dysfunction and incidence of type 2diabetes.CardiovascDiabetol.2016Mar 24;15(1):51.doi:10.1186 / s12933-016-0369-6). If you decide to establish a prognostic test for diabetes, you may consider selecting different feature sets: one related to the insulin pathway and the other related to inflammation. Using the methods outlined in this report, one can attempt to separate the prognostic impact of these two broad biological processes and even estimate the relative influence of each on prognostic classification. Furthermore, if one is trying to discover a predictive test for a new drug, one may even be able to better understand the mechanism of action of that treatment, for example, as long as a hypothesized set of relevant features works well.
[0690] Since few diseases are defined by the disruption of a single process, the application of the disclosed method is very broad and is limited only by the appropriate measurement of the relevant proteins in the selected sample and the (in)adequate understanding of the role of these proteins in the specific biological process. Therefore, in theory, this method can allow researchers to separate and test the impact of multiple biological processes related to almost any disease, as well as better understand the mechanism of action of treatments. Although the present study involved MALDI mass spectrometry of serum and was limited to methods that can be explored by circulating proteins and peptides, the method itself is not dependent on the sample type and can therefore be extended even beyond its already broad range of applicability.
[0691] Further details and examples of the application of GSEA-based feature selection and classifier development will now be described in detail.
[0692] Two sample sets were used in this study:
[0693] 1. A set of 49 serum samples ("GSEA cohort" or "analysis" set or cohort) collected from 45 patients with non-small cell lung cancer and 4 cancer-free subjects, for which mass spectrometry data were collected and 1129 proteins were used. The protein expression data were generated using the aptamer set (SomaLogic, Boulder, CO). This set was used to identify peaks for classifier training from GSEA analysis.
[0694] 2. A cohort of 119 pre-treatment serum samples from 119 patients with advanced melanoma treated with the anti-programmed cell death-1 (PD-1) therapy nivolumab, with or without the addition of a peptide vaccine as part of a clinical trial (the "NCD cohort," also referred to in this disclosure as "Moffitt") (details of the trial can be found in J. Weber et al., Safety, Efficacy, and biomarkers with Vaccine in Ipilimumab-Refractory or Melanoma, J Clin Oncol 2013 Dec 1; 31(34):4311). Result data were available for this cohort of patients, and mass spectrometry data were collected from these samples. This sample set was used for classifier training. This sample set is described in Example 1.
[0695] Generation of protein expression data and GSEA platform
[0696] Protein expression data were collected from the GSEA cohort using the 1129-protein SOMAscan aptamer panel from SomaLogic, Inc. at its laboratory in Boulder, CO. A list of the 1129 proteins included in the assay is included in Appendix A of prior provisional application serial number 62 / 340,727, filed May 24, 2016. Further details regarding identifying protein groups associated with mass spectrometry signatures are presented later in this document.
[0697] Generation and processing of mass spectrometry data from the GSEA cohort and the NCD cohort were performed as previously explained in detail in Example 1.
[0698] Application of Gene Set Enrichment Analysis (GSEA)
[0699] GSEA (see the previously cited papers by Mootha et al. and Subramanian et al.) was introduced as a method to help address or attempt to minimize some fundamental problems in gene expression analysis studies: the identification of gene sets and the inability of tests derived in the development sample set to generalize to other sample sets (overfitting), the multiple testing problem, and the inability to identify small, consistent expression changes among multiple correlated genes that might be overwhelmed by larger, randomly occurring expression changes in the dataset. These problems are inherent in dealing with "p>n" datasets, i.e., where the number of measured expression values greatly exceeds the number of samples available for measurement. Rather than looking for expression differences on a feature-by-feature (gene-by-gene) basis, the method looks for consistent expression differences across pre-specified groups or sets of features. Feature sets can be created based on biological insights or defined using prior, non-hypothetical studies. Correlating with feature sets rather than individual features provides some protection against identifying isolated features that are randomly correlated with the study group and that do not generalize to other sample sets. Typically, the number of feature sets tested for association is smaller than the number of individual features in a typical gene expression study, thus reducing the multiple testing problem. Additionally, because the method looks for consistent correlations across a set of features, significant effects can be identified that are smaller in magnitude (per feature) than those that can be identified for a single feature.
[0700] Definition of proteome
[0701] Specific protein groups were created based on the SOMAscan1129 protein list and query results from the GeneOntology / AmiGO2 and UniProt databases.
[0702] AmiGO2 queries are filtered by:
[0703] - Document Category: Annotation - Taxonomic Group: Homo sapiens
[0704] -Evidence Type: Experimental
[0705] Table 45 lists the individual filters used.
[0706] Table 45: Filters used in the AmiGO2 database
[0707] Proteome name Keywords Amigo 1 Acute inflammatory response Amigo 2 Activate innate immune response Amigo 3 Regulation of the adaptive immune response Amigo 4 Positive regulation of glycolysis Amigo 5 immune T cells Amigo 6 Immune B cells Amigo 7 Cell cycle regulation Amigo 8 Natural Killer Regulation Amigo 9 complement system Amigo 11 Acute reactions Amigo 14 Cytokine activity Amigo 16 Wound healing Amigo 17 interferon Amigo 18 Interleukin-10 Amigo 20 Growth factor receptor signaling Amigo 21 immune response Amigo 22 Immune response type 1 Amigo 23 Immune response type 2
[0708] UniProt queries are filtered by:
[0709] Organism: Homo sapiens
[0710] DB: Review (SwissProt)
[0711] Table 46 lists the various filters used in the UniProt database.
[0712] Table 46: Filters used in the UniProt database
[0713] Uniprot 1 Acute phase Uniprot 2 Hypoxia Uniprot 4 cancer
[0714] The proteins included in each proteome are listed in Appendix D of prior provisional application serial number 62 / 340,727. (Uniprot 4 cancer is not included in the list because it contains more than 400 proteins.)
[0715] Implementation of the GSEA method
[0716] The GSEA method was implemented using Matlab (version R2015a) on a general-purpose computer. The process can be broken down into several steps.
[0717] Suppose there is a set N s The data of the sample, and for each sample i, give the continuous variable Q i , and N f Protein expression value Where i is in the sample (1≤i≤N s ) and j is run on proteins (1≤j≤N f ). There are k predefined protein groups S l (1≤l≤k), and the focus is on the association of these groups with mass spectrometry data. Each protein group S l Depend on member Composed and is N f A subset of the complete set of proteins.
[0718] 1. Evaluate the correlation between individual proteins and continuous variables (mass spectrometry feature values)
[0719] Calculate the correlation r between each individual protein j in the complete proteome and the continuous variable associated with the sample j The Spearman rank correlation was used to assess the strength of the correlation. Once the correlation r for each protein j has been calculated, j , then N f Protein by r j Arrange from largest to smallest.
[0720] 2. Calculate the enrichment score
[0721] As explained in the previously cited paper by Subramanian et al., the enrichment score ES l Designed to reflect specific proteome S lThe degree to which an element is over-represented at the top or bottom of the ranked protein list. Starting from the top of the ranked list, a running sum RS(S l,P ), starting from zero, if protein j is included in S l In the sorted list, add the item j | r j | / N norm , and if protein j is not included in S l , then subtract the item j from the sorted list Until item p is reached. N norm By N norm =Σ|r j | regularization coefficient defined, where the sum is in S l The calculated RS(S l,P ) is an example of Figure 43 shown.
[0722] Consider two possible definitions of ES. First, in terms of RS(S l,P )'s maximum positive value RS 最大 , and RS(S l,P ) minimum value RS 最小 To define ES l These are Figure 43 If RS 最大 ≥|RS 最小 |, then ES=RS 最大 ; If RS 最大 <|RS 最小 |, then ES=RS 最小 (This is the definition used by Subramanian et al., with the exponent p set to 1.) To be able to consider protein groups containing a mixture of over- and under-expressed proteins, or mixtures or proteins that are meaningfully correlated and anti-correlated with the continuous variable, we also considered defining ES as RS 最大 +|RS 最小 |. (If all proteins in a proteome are overexpressed (positively correlated), the two definitions are identical.)
[0723] 3. Calculate the corresponding p-value
[0724] To assess the significance of the deviation of the calculated ES from the mean of its random distribution, a null distribution of the ES is calculated by generating many realizations of the random association between the continuous variable and protein expression and evaluating the ES for each realization. These realizations are created by permuting the values of the continuous variable assigned to each sample. Note that this will maintain the correlation structure within the protein expression values of each sample. Once the null distribution is generated, the p-value of the calculated ES can be read as the proportion of random permutations that produce an ES that is far less random (more extreme) than the calculated ES. (Note that the first definition of ES requires the evaluation of positive and negative ES separately). Figure 44A and 44B An example of the zero-point distribution and the calculated ES for one proteome is shown. The zero-point distribution had to be evaluated separately for each comparison (each individual continuous variable (i.e., each mass spectral feature) and proteome pair). For the correlation with the mass spectral feature value, 2000 realizations were generated.
[0725] 4. Correction for Multiple Testing
[0726] The p-values generated by the above approach do not account for multiple testing. It would be possible to extend the analysis to account for multiple testing, either by using a very conservative Bonferroni correction or by generating many random permutations of the protein lists across all proteomes on the continuous variable and calculating an ES for each realization. The latter approach would also require regularizing the ES to allow for combining results across different proteomes. Neither approach has been implemented at this time, and the results reported here have not been corrected for multiple testing.
[0727] GSEA results
[0728] Using the above method, GSEA p-values were obtained for each mass spectral feature of each protein functional group. Figure 45A and 45B The heatmap shows the correlation results of protein functional groups of the 49 samples in the GSEA group with all 351 identified mass spectrometric features (Appendix A). Specifically, Figure 45A and 45B The p-values generated by GSEA analysis correlating all 351 defined mass spectral features with different protein functional groups (biological processes) are shown. Figure 45(a) shows the p-values for ES definition 1, and Figure 45(b) shows the p-values for ES definition 2. Note: The mass spectral features are arranged in order of increasing m / z, and only the fifth spectral feature is labeled on the x-axis.
[0729] Table 47 shows the number of mass spectral features (out of 351 defined mass spectral features) associated with protein functional groups with p-values below various thresholds for each protein functional subset for Definition 1 and Definition 2 of the enrichment score.
[0730] Table 47: Number of mass spectrometric features associated with protein functional groups for p-values below various thresholds
[0731]
[0732]
[0733] *Indicates that the ES is greater (or less) than the ES obtained in any realization used to evaluate the null distribution
[0734] It is clear that many of the 351 mass spectrometric features are associated with acute phase reactants (acute response, acute phase, acute inflammation), the complement system, or wound healing. However, there are also mass spectrometric features associated with other completely different functional groups of proteins, such as glycolysis, the cell cycle, or cancer. Therefore, it is potentially possible to use the measurement results of mass spectrometric features that have been determined to be associated with specific biological functions in serum samples from patients to monitor specific biological functions in patients.
[0735] For the first definition of enrichment score (ES definition 1) and each protein functional subset, a cutoff value of p = 0.05 was selected so that mass spectrometric features with GSEA p < 0.05 were associated with biological function. Appendix E of the previous provisional application 62 / 340,727 lists mass spectrometric features associated with several protein functional subsets studied (acute response, wound healing, immune response).
[0736] Then, continue using Figure 8A and 8B First, all 119 samples in the new classifier development (NCD) cohort were subjected to the Figure 8 The classifier was developed in steps 102-150, where a subset of 33 mass spectral features were associated with the acute response protein functional group. Figure 8 Using the procedure described above, a classifier was created, hereafter referred to as "Classifier 1," that was able to stratify melanoma / nivolumab patients into two groups with better and worse prognosis in terms of OS and TTP (Classifier 1). No feature deselection was used, i.e., all 33 mass spectral features associated with the acute-response protein functional group were used in each optimization step of the class labels. 51 samples were assigned to the poorly performing group and were given the "early" class label. The specific choice of the names of the class labels generated by the classifier is not particularly important.
[0737] The remaining 68 samples assigned to the good performance group are used as the basis Figure 8A development set of a second classifier (referred to as "Classifier 2") generated by steps 102-150 of the present invention was used. This classifier was trained on a subset of 26 mass spectral features that had been identified as being associated with wound healing but not with acute or immune responses. The second classifier again did not use feature deselection and stratified patients into groups with better or worse TTP. Samples in the good TTP group were assigned the "late" classification, and samples in the poor TTP group were assigned the "early" classification.
[0738] The final classifier is then defined as a hierarchical combination of Classifier 1 and Classifier 2. The resulting final classifier (ie, two Diagnostic Cortex classifiers combined by logical instructions used in a hierarchical manner) uses a total of 59 features listed in Appendix D of this document. Figure 46 Schematically illustrates how a classifier 1 based on a mass spectral signature associated with an acute response and a classifier 2 based on a mass spectral signature associated with wound healing but not with an acute response or an immune response are combined to assign a classification to a test sample. Specifically, a mass spectrum ( Figure 46 The "test spectrum" in Appendix D is obtained, and in particular the feature values of the 59 features of Appendix D are obtained. This data is provided to Classifier 1 and Classifier 1 then generates a label for the sample. If the reported label is early (or equivalent), the sample is assigned an early category label. If Classifier 1 does not produce an early label, the spectral data is provided to Classifier 2. If Classifier 2 generates an early label, the sample will be assigned an early category label. If Classifier 2 does not generate an early label, the sample will be assigned a late category label. The early and late labels have the same clinical significance as previously explained in Example 1.
[0739] Note that this is an example of creating multiple distinct classifiers using distinct feature subsets associated with different protein functional groups, and combining these classifiers through a rule-based system to produce an overall classification that encompasses the information content of various biological functions. In this particular example, after removing or accounting for major effects, by using the peaks associated with the functional groups used in the first classifier and constructing a new or second classifier based on the hierarchy of biological functions, it is found that within each of the groups defined by the classifiers, subsets of functions become correlated. If a large enough set is available, this process can be further iterated, and a third-level classifier can be defined on a set of peaks associated with a third protein function, applied to the group classified by the second-level classifier, and so on. Since multiple protein functional groups are often associated with a given peak, this approach attempts to disentangle compound effects.
[0740] result
[0741] 1. Single classifier 1
[0742] Classifier 1 assigned the “early” classification to 51 of the 119 samples. Figure 47A and 47B Kaplan-Meier plots are shown for OS and TTP for the NCD cohort as classified by classifier 1. The acute response-based classifier achieved a clear separation between good and poor prognosis groups.
[0743] 2. Single classifier 2
[0744] Classifier 2 assigned the “early” classification to 35 of the 68 samples that were not classified as “early” by classifier 1, and the remaining 33 were assigned the “late” classification. Figure 48A and 48B Kaplan-Meier plots of OS and TTP for the classification of these 68 samples provided by classifier 2 are shown.
[0745] Classifier 2 further stratified the 68 samples that were not classified as “early” by classifier 1, using features related to wound healing but not including features related to acute or immune responses.
[0746] 2. The final classifier defined as the combination of classifiers 1 and 2
[0747] like Figure 46 As shown, combining the classifiers in a hierarchical manner can achieve excellent binary stratification for the entire set of 119 samples. Thirty-three (28%) of the samples were classified as "late stage" and 88 (72%) were classified as "early stage". Figure 49A and 49B The OS and TTP of the entire NCD cohort of 119 samples stratified by population are illustrated in the Kaplan-Meier plots. Table 48 gives the relevant statistics that characterize the clear stratification of the cohort.
[0748] Table 48: Figure 49A and 49B Kaplan-Meier plot related statistics (CPH = Cox proportional hazards)
[0749]
[0750] Table 49 shows some milestone survival and progression-free statistics, and Table 50 summarizes response data.
[0751] Table 49: Proportions of those still alive and progression-free at key time points
[0752] Early Late stage % survival within 1 year 63 88 % survival within 2 years 34 61 % progression-free within 6 months 31 67 % progression-free within 1 year 25 60
[0753] Table 50: Responses by test
[0754] Early stage (n=86) Late stage (n=33) PR 19(22%) 12(36%) SD 9(10%) 9(27%) PD 58(67%) 12(36%)
[0755] Table 51 shows the baseline patient characteristics by classification group.
[0756] Table 51: Baseline Patient Characteristics by Test
[0757]
[0758] Fisher's exact test showed that serum LDH levels <2 ULN were significantly associated with classification (p < 0.001), and the Mann-Whitney p for the association between LDH levels and classification was 0.070. The baseline tumors in the early-stage group were larger than those in the late-stage group (Mann-Whitney p < 0.001). However, classification was not associated with PD-L1 expression at any available cutoff value.
[0759] Multivariate analysis of time-to-event outcomes adjusted for the effect size (hazard ratio) of other known prognostic features, such as serum LDH level. The results of this analysis are presented in Table 52. In addition to serum LDH level, the classification remained a significant predictor of OS and TTP, indicating that the classification provides complementary information about outcomes after nivolumab treatment.
[0760] Table 52: Multivariate Analysis of Time-to-Event Endpoints
[0761]
[0762] Notably, the overall classification performance obtained using these feature subsets selected by the biological hypothesis may be better than that previously achieved in this sample cohort. Figures 50A-50D The Kaplan-Meier plots for this classification were compared with those of two previously developed classifiers, one classifier (IS2 = the full set classifier of Example 1) developed using all mass spectral features simultaneously, and another classifier (IS6) using a cluster of classifiers with clinically different development subsets that again used all mass spectral features. (These classifiers are described in Examples 1 and 8, respectively.)
[0763] Samples from a cohort of 30 patients also treated with anti-PD1 therapy were used for independent validation of the classifier: 21 patients (70%) were classified as "early stage" and 9 (30%) were classified as "late stage." Figure 51 Kaplan-Meier plots for OS are shown, and related statistics are shown in Table 53. (TTP was not available for this observational cohort.)
[0764] Table 53: Figure 51Kaplan-Meier plot related statistics
[0765]
[0766] Example 6 Conclusion and Discussion
[0767] This study of Example 6 has demonstrated that:
[0768] 1) Correlate features in mass spectra with biological function without directly identifying the protein or peptide that produces the mass spectrometric feature, and
[0769] 2) Incorporate biological insights into the selection of mass spectrometric features for robust classifier development, e.g., using Figure 8 Diagnostic Cortex platform.
[0770] It will be further appreciated that once a classifier has been developed as described above, it is stored in the memory of a computer as setup parameters (e.g., a signature table of mass spectral features for classification, identification of mini-classifiers, logistic regression weights, kNN parameters, program code for executing one or more main classifiers, and logic defining the final classifier, such as Figure 8 Step 150 or Figure 46 etc.). For example, Figure 15 The laboratory testing center includes this computer and a mass spectrometer for performing mass spectrometry analysis on blood samples. The resulting mass spectra undergo preprocessing steps (the same steps performed on the samples in the classifier development set), and then the classifier is applied to the sample's mass spectral data. The classifier then generates a class label, such as early or late stage, which is provided to the requesting physician or clinic on a fee-for-service basis.
[0771] refer to Figure 52 , it will be further appreciated that a classifier development system 5200 has been disclosed that includes a mass spectrometer 5202 for performing mass spectrometry analysis on a development sample set or, alternatively and more typically, another independent sample set, to generate mass spectrometry data. The data includes intensity data for a number of mass spectrometry features. The system includes a platform 5204 for performing gene set enrichment analysis on the development sample set, or, more typically, the other independent sample set, including a protein measurement system, such as SomaLogic's SOMAscan system or equivalent, and a computer for identifying statistically significant associations of one or more mass spectrometry features with groups of proteins grouped according to their biological function. The system also includes a computer 5206 that is programmed to use one or more mass spectrometry features identified by the GSEA platform (e.g., using Figure 8A classifier is trained on the development sample set using a program (a set of parameters and program instructions). A classifier is a set of parameters and program instructions that assigns class labels to samples of the same type as the development sample set according to the programmed instructions.
[0772] In this document, the terms classifier training and classifier development are used interchangeably to refer to the process of building a classifier in a computer (i.e., specifying the parameters of such a classifier) and testing its ability to classify a set of samples (either the development sample set or some subset thereof). Typically, this process occurs in an iterative manner to adjust the parameters to optimize classifier performance, such as by improving the class labels assigned to members of the development set, improving filtering parameters, feature deselection, changing the parameter k, etc. It will also be noted that although the present example describes the use of k-nearest neighbors with majority voting as the classification algorithm, in principle the present invention can use other supervised learning classification algorithms, such as marginal-based classifiers, support vector machines, decision trees, etc., or classifiers that are configured as multiple filtered small classifiers combined using a regularization procedure, such as using Figure 8 generated by the program.
[0773] The following clauses are provided as further description of the invention disclosed in Example 6.
[0774] 1. A classifier for guiding immune checkpoint inhibitor therapy in cancer patients, comprising:
[0775] a memory storing a reference set of class-labeled mass spectral data obtained from a blood sample of a melanoma patient treated with an immune checkpoint inhibitor, the mass spectral data in the form of feature values for at least 50 mass spectral features, wherein the mass spectral features are consistent with proteins circulating in serum associated with at least the following biological processes: (1) acute phase, (2) acute response, (3) complement system, and (4) acute inflammatory response; and
[0776] A programmed computer executes a classification algorithm on a set of mass spectral data, the set of mass spectral data including feature values of a plurality of mass spectral features obtained from a tested blood sample and a reference set, and generates a class label for the tested blood sample.
[0777] 2. A classifier as described in claim 1, wherein the mass spectral features include features listed in Appendix A, Appendix B, Appendix C or Appendix D.
[0778] 3. The classifier of clause 1, wherein the mass spectrometric signature further comprises signatures associated with the following additional biological processes: type 2 immune response and interleukin-10.
[0779] 4. The classifier of clause 1, wherein the mass spectral data for the test blood sample and the reference sample set are acquired using at least 100,000 laser shots performed on the sample using MALDI-TOF mass spectrometry.
[0780] 5. The classifier of clause 1, wherein the test blood sample is obtained from a melanoma patient.
[0781] 6. The classifier of clause 1, wherein the immunotherapy comprises an antibody drug targeting programmed cell death 1 (PD-1).
[0782] 7. The classifier of clause 1, wherein the immunotherapy comprises an antibody drug targeting CLTA4.
[0783] 8. A method for training a classifier, comprising the following steps:
[0784] a) obtaining a development sample set from a subject population, and optionally obtaining a second independent sample set from a similar, but not necessarily identical, subject population;
[0785] b) performing mass spectrometry analysis on the development sample set and optionally on the second sample set, and identifying mass spectrometry features present in the mass spectra of the sample sets;
[0786] c) for each sample in the development sample set or each sample in the second sample set, obtaining protein expression data from a larger set of proteins covering the biological function of interest;
[0787] d) identifying statistically significant associations between one or more of said mass spectral features and combinations of proteins grouped according to their biological function; and
[0788] e) training a classifier based on the development sample set using one or more mass spectral features identified in step d) with computer assistance, wherein the classifier is in the form of a set of parameters that assign class labels to samples of the same type as the development sample set according to program instructions.
[0789] 9. The method of clause 8, wherein step d) further comprises the step of performing a gene set enrichment analysis.
[0790] 10. The method of clause 8 or clause 9, wherein the classifier is in the form of an ensemble of filtered small classifiers that are subjected to a regularization procedure.
[0791] 11. The method of any of clauses 8-10, wherein the samples in the development set and optionally the second sample set are blood samples.
[0792] 12. The method of any one of clauses 8 to 11, wherein step b) comprises performing at least 100,000 laser shots on each of the samples in the sample set in MALDI-TOF mass spectrometry.
[0793] 13. A method as described in clause 8, wherein the classifier trained in step e) is considered to be a first classifier, and the method further comprises repeating step e) for a second set of one or more mass spectral features, wherein the mass spectral features are associated with a group of different proteins related to different biological functions, thereby training a second classifier.
[0794] 14. The method of clause 13, further comprising the step of defining a final classifier from a combination of the first classifier and the second classifier.
[0795] 15. The method of clause 13, wherein the second classifier is used to further stratify the members of the classification group assigned by the first classifier.
[0796] 16. A computer configured as a classifier, the classifier being trained according to any one of clauses 8-15.
[0797] 17. A method for testing a sample, comprising the following steps:
[0798] a) training a classifier according to any of clauses 8-11;
[0799] b) storing parameters of a classifier including the signature table of the set of mass spectral signatures in a memory;
[0800] c) performing mass spectrometry analysis on the test sample; and
[0801] d) Using a computer and a trained classifier, classify the test samples.
[0802] 18. A method for testing a sample, comprising the following steps:
[0803] a) training the first classifier and the second classifier according to clause 13;
[0804] b) storing parameters of the first and second classifiers (including a feature table of the mass spectral feature set) in a memory, and logic instructions for combining the first and second classifiers into a final classifier;
[0805] c) performing mass spectrometry analysis on the test sample; and
[0806] d) classifying the test sample using the final classifier defined in step b) with the help of a computer.
[0807] 19. A computer configured as a classifier, comprising:
[0808] a memory storing a signature table in the form of intensity data for a set of mass spectral features obtained from a biological development sample set, wherein the set of mass spectral features has been determined to have a statistically significant association with a combination of proteins present in the biological sample grouped according to biological function;
[0809] A set of parameters is defined for a classifier comprising a classification algorithm that operates on mass spectral data from a test sample and a signature table.
[0810] 20. A classifier development system comprising:
[0811] a mass spectrometer for performing mass spectrometric analysis on the development sample set and, optionally, a second independent sample set to generate mass spectrometric data comprising a plurality of mass spectrometric features;
[0812] A platform for performing gene set enrichment analysis on the development sample set or, optionally, a second independent sample set, and identifying statistically significant associations of one or more mass spectrometric features with groups of proteins grouped according to their biological function; and
[0813] A computer programmed to train a classifier for the development sample set using one or more mass spectral features identified by the platform, the classifier being in the form of a set of parameters that assign class labels to samples of the same type as the development sample set according to program instructions.
[0814] 21. The system of clause 20, wherein the development sample set and the optional second independent sample set are blood samples from humans.
[0815] 22. The system of clause 21, wherein the blood sample used to develop the sample set is obtained from a melanoma patient being treated with an immunotherapy drug.
[0816] 23. A classifier training method comprising the following steps:
[0817] a) Obtaining a development sample set, and optionally a second independent sample set, from a subject population:
[0818] b) performing mass spectrometry analysis on the development sample set and the optional second sample set to identify mass spectrometry features present in the mass spectra of the sample sets;
[0819] c) obtaining protein expression data from a larger set of proteins that encompasses the biological function of interest for each sample in the development sample set or, optionally, each sample in the second sample set;
[0820] d) identifying statistically significant associations between one or more mass spectrometric features and groups of proteins grouped according to their biological function;
[0821] e) using the one or more mass spectral features identified in step d) to train a first classifier on samples from the development sample set with the aid of a computer, wherein the classifier assigns class labels to samples of the same type as the development sample set according to programming instructions in the form of a set of parameters, the classifier generating at least a first class label and a second class label, and
[0822] f) training a second classifier with computer assistance using one or more mass spectrometric features associated with the different groups of proteins identified in step d) that are grouped by different biological functions, and
[0823] g) classifying the samples with the first classifier, wherein if the first classifier generates a first class label, reporting the class label and if the first classifier generates a second class label, using the second classifier to further stratify the samples.
[0824] 24. A classifier training method comprising the following steps:
[0825] (a) performing mass spectrometry analysis and gene set enrichment analysis on a development blood sample set, or performing mass spectrometry analysis on a development set of blood samples and a second independent set of blood samples, and performing gene set enrichment analysis on the second set of samples;
[0826] (b) identifying groups of mass spectrometric peaks that are statistically significantly associated with groups of proteins grouped according to their biological functions;
[0827] (c) executing a classifier training procedure in a computer using one of the peak sets associated with the first protein functional group identified in step b), the classifier training procedure classifying the mass spectrometry data of the development sample set or a subset thereof.
[0828] 25. The method of clause 24, further comprising repeating step c) for a second set of peaks identified in step b) that are associated with a second functional group of proteins different from the first functional group of proteins.
[0829] 26. The method of clause 25, further comprising repeating step c) for a third set of peaks identified in step b) for a third functional group of proteins that is different from the first and second functional groups of proteins.
[0830] 27. A method of testing a subject, comprising:
[0831] Train the classifier according to Item 24;
[0832] classifying samples from the subject using a classifier at a first time point;
[0833] A second sample obtained from the subject at a later time point is classified using the classifier.
[0834] 28. The method of clause 27, wherein the sample is provided by a patient enrolled in a clinical trial of a drug, wherein the first time point is prior to treatment with the drug, and wherein the later time point is after treatment has begun and the patient is still participating in the clinical trial.
[0835] 29. A method for evaluating a biological process in a human body, comprising the steps of:
[0836] a) training a classifier in accordance with Clause 24;
[0837] b) performing mass spectrometry analysis on blood samples from humans;
[0838] c) classifying the sample using the data obtained from step b) and the classifier trained in step a), thereby obtaining a first category label of the sample;
[0839] d) performing mass spectrometry analysis on a second blood sample collected from the human at a later time point after the sample obtained in step b);
[0840] e) classifying the second blood sample using the data obtained from step 3) and the classifier trained in step a), thereby obtaining a second class label;
[0841] f) comparing the first and second categories of markers, wherein the comparison provides information about biological processes occurring in the human body.
[0842] 30. A method for evaluating a biological process in a human body, comprising the steps of:
[0843] a) obtaining a development set of blood samples from a subject population, and optionally obtaining a second independent set of blood samples from a similar, but not necessarily identical, subject population;
[0844] b) performing mass spectrometry analysis on the development blood sample set and optionally on the second blood sample set, and identifying mass spectrometry features present in the mass spectra of the blood sample sets;
[0845] c) obtaining protein expression data for each blood sample in the development sample set or each sample in the second blood sample set from a larger set of proteins encompassing a biological function of interest;
[0846] d) identifying statistically significant associations between one or more mass spectrometric features and groups of proteins grouped according to their biological function;
[0847] e) performing mass spectrometry analysis on the blood sample from the human, comprising obtaining mass spectrometry feature values of the one or more mass spectrometry features identified in step d).
[0848] 31. The method of clause 30, further comprising the steps of obtaining a second blood sample from the human, and performing mass spectrometry analysis on the second blood sample from the human, comprising obtaining mass spectrometry characteristic values for the one or more mass spectrometry characteristics determined in step d).
[0849] 32. The method of clause 31, wherein the human is participating in a clinical trial of the drug or drug combination.
[0850] 33. A method as described in claim 30, wherein the person is participating in a clinical trial of a drug or drug combination, and wherein the method further comprises the following steps: repeatedly obtaining blood samples from the person during the process of the person being enrolled in the clinical trial, and performing mass spectrometry analysis on the blood samples, including obtaining characteristic values in the mass spectrometry of one or more mass spectrometry features identified in step d) together with the protein group grouped according to their biological function.
[0851] 33. A method of monitoring a group of patients enrolled in a clinical trial, comprising performing the method of clause 30 on each patient participating in the clinical trial.
[0852] 34. The method of claim 33, further comprising repeatedly obtaining blood samples from patients enrolled in the clinical trial during the course of the trial, and performing mass spectrometry analysis on the blood samples, comprising obtaining in the mass spectrometry characteristic values of one or more mass spectrometry features identified in step d) of claim 30 as being associated with the protein group grouped according to its biological function.
[0853] Example 7
[0854] Longitudinal study
[0855] Samples collected during treatment of the nivolumab study (described in Example 1), specifically at week 7 ("WK7") and week 13 ("WK13") of the trial, were analyzed. Using the full set classifier of Example 1 and the classifier of Example 2, we explored how the classification of a given patient changed over time. We found that in some patients, the labeling changed, such as an initial class label of late at the start of treatment, then late at week 7 and early at week 13. As another example, some patients had an early class label at the start of treatment, then early at week 7 and late at week 13.
[0856] Figures 30A-30F The results of a longitudinal study using the full set classifier of Example 1 are shown in . Figure 30A and 30B is the overall survival of the early and late groups defined by baseline categories ( Figure 30A ) and time to progression (TTP) ( Figure 30B ) Kaplan-Meier plot. Figure 30C and 30DKaplan–Meier plots of overall survival and TTP for the early and late groups defined by week 7 categories, respectively. Figure 30E and 30F Kaplan-Meier plots of overall survival and TTP for the early and late groups defined by the week 13 categories, respectively, for the 90 patients with category labels at all three time points. Figures 30A-30F Table of plots for survival analysis.
[0857] Table 54
[0858]
[0859]
[0860] Table 55 shows the classification distribution of all samples at the three time points.
[0861] Table 55: Distribution of categories at three time points: baseline, WK7, and WK13. Missing categories are indicated by “-”.
[0862]
[0863] Most classifications remained unchanged across the available time points (82 / 119 = 69%). There were correspondingly more changes from early to late than from late to early, with most changes from early to late occurring at WK7 and remaining late at WK13. This may be due to the start of immunotherapy treatment. Where samples were available, half of the patients who changed from late to early at WK7 reverted to late at WK13. Twelve (16%) of the patients classified as late at WK7 changed to early during WK13, half of whom progressed between 70 and 93 days, although three of them experienced a progression-free interval of more than 1000 days.
[0864] Figure 31A and 31B Shown are Kaplan-Meier curves that plot the results when patients are grouped according to the ternary classification of baseline, WK7, and WK13. Figure 31A is a graph of overall survival; Figure 31B Is a graph of time to progression. In these figures, the groups are first labeled by baseline category, followed by WK7 classification and finally WK13 classification (i.e., "early late early" represents early baseline classification, early WK7 classification and early WK13 classification). Repeating the early category marks particularly poor OS and TTP, and at least in OS, having early marking at WK 13 indicates a poor prognosis, even if the first two classifications are late. However, late marking of WK13 and WK7 corresponds to better results, even if the baseline classification is early. Note: There are too few patients in the other marker sequences for meaningful analysis.
[0865] Table 56 shows Figure 31A and 31B The median of the curve.
[0866] Table 56
[0867]
[0868] This analysis was repeated for the classifications produced by the Example 2 classifier over time. The results are generally similar to those presented here for the full set classifier of Example 1. Most classifications remained unchanged at all time points (59 / 90=66% for three time points, 83 / 107=78% for the first two time points). There were correspondingly more changes from early to late than from late to early, with most of the changes from early to late occurring at WK7 and remaining late at WK13. This may be due to the start of immunotherapy treatment. Where samples were available, most of the patients who changed from late to early at WK7 returned to late stage at WK13.
[0869] Figure 32A and Figure 32B is a Kaplan-Meier curve graph showing the results when the patients were grouped according to the baseline, WK7 and WK13 ternary classification generated by the classifier of Example 2. Figure 32A is a graph of overall survival; Figure 32B This is a plot of time to progression. The groups are labeled first by baseline category, then by WK7 category, and finally by WK13 category. Repeating the early category indicates particularly poor OS and TTP, while repeating the late category indicates particularly good OS and TTP. Changing from early to late at WK7 and maintaining late at WK13 resulted in similar outcomes to having the late category at all three time points.
[0870] Figure 32A The statistical data are listed in Table 57.
[0871] Table 57: Figure 32A Median OS
[0872] Median OS (weeks) Early Early Early (N=9) 47 Early Early Late (N=5) 99 Early Late Late (N=11) Not reached Late Early Late (N=5) 99 Late Late Early (N=6) 78 Late Late Late Late (N=50) 113
[0873] Figure 32B The statistical data are listed in Table 58.
[0874] Table 58: Figure 32B The median of the TTP graph
[0875]
[0876] There are some theories as to why the class markers change over time. Changes in the patient's biology after the start of nivolumab treatment may induce changes in the class markers. This change may be due to the effect of tumor size on the class marker (see discussion below), as well as the larger tumor shrinkage achieved by some patients. Regardless of the cause of the change, it was indeed observed that most patients maintained their baseline markers. In patients whose class markers changed over time, it was observed that when the markers changed from early at baseline to late at a later time (week 7 or 13), these patients had relatively good outcomes, similar to patients with late baseline class markers. Therefore, in one embodiment, the tests of Example 1 or 2 can be performed regularly throughout the course of treatment, for example, every 4, 6 or 8 weeks. By comparing the results over time during treatment and the progression of the class markers, the therapeutic effect of nivolumab treatment can be monitored, or the prognosis or overall survival of the patient can be predicted. This treatment monitoring can be direct, i.e., a direct change in certain immune status measured by the class marker, or indirect, i.e., the change in the class marker is a substitute or approximation of the tumor shrinkage / expansion measurement. How frequently one wants to test and determine a patient's class marker during treatment may also depend on whether changes in the class marker are due to a direct effect of the drug on the patient's immune system, or whether one must await an indirect effect of the treatment on tumor shrinkage (or lack thereof).
[0877] Over time, a change in a patient's marker from late to early disease could be an early sign of a drug's lack of efficacy. This lack of efficacy can optionally be confirmed by performing radiological studies on the patient, such as CT scans, to determine tumor size and changes compared to baseline. Potentially, if the change from late to early disease is due to the drug directly altering the patient's immune status, and if this occurs within a relatively short time interval (e.g., around 4 weeks), the baseline late marker could be an indication to start treatment with nivolumab, and subsequent early markers could be used to stop treatment, change treatment to a different treatment (e.g., combining nivolumab and ipilimumab), or take other actions. Another possibility is to monitor patients regularly during the first few weeks of treatment and use the class marker to indicate how long the patient needs to take nivolumab. Currently, patients take the drug until the disease progresses. This can be a long time, and the drug is very expensive. Therefore, if there were a way to tell patients whether nivolumab treatment can be stopped within the first few months without compromising outcomes, some healthcare costs could be saved. In any case, in one possible embodiment, the tests of Examples 1 and 2 are performed periodically during treatment. The class markers are compared during treatment. The status of the class marker during treatment can be used to guide treatment or predict a patient's prognosis, maintain treatment, stop treatment, or change treatment in some way, such as combining nivolumab with another drug in a combination therapy regimen.
[0878] In one specific example of how monitoring testing may be performed, using Figure 15 The system (described above) according to Example 1 or Example 2 re-determines the initial class label at least within the first four weeks of treatment and again after at least the first four weeks of treatment.
[0879] The following clauses are provided as further description of the invention described in Example 7.
[0880] 1. First, classify the patient sample according to the method of Example 1, 2, 3, 4 or 6, and repeatedly obtain samples from the patient during the treatment process or during the patient's enrollment in a clinical trial, perform mass spectrometry analysis on the sample, and classify the sample using a classifier, thereby repeatedly generating category labels.
[0881] 2. The method of clause 1 further comprising the steps of determining whether the category marker changes during treatment and using the change in the category marker to guide the patient's treatment.
[0882] 3. The method of clause 1 further comprising the step of determining whether the category signature has not changed during treatment, and using the lack of change in the category signature to guide treatment of the patient.
[0883] 4. The method of clause 1, wherein the repeating step is performed first before treatment, again within at least the first four weeks of treatment, and again after at least the first four weeks of treatment.
[0884] 5. As shown in Example 6, a method for determining an association between a biological function and a mass spectrometric feature obtained from a blood sample (e.g., using GSEA), by repeatedly obtaining blood samples from a patient and performing mass spectrometry on the samples, and analyzing the mass spectrometric data from the samples over a period of time to observe or understand changes in the biological function over time, such as upregulation or downregulation of a particular protein associated with the biological function, such as during treatment or during the patient's participation in a clinical trial.
[0885] 6. The invention of clause 5, further comprising the step of performing classification of each blood sample using a computer-implemented classifier trained on the development sample set and a set of mass spectrometric features associated with biological functions.
[0886] Example 8
[0887] Training a classifier based on tumor size information
[0888] A method has been discovered to generate a classifier that takes into account the tumor size at baseline, which improves the performance of the classifier. This method and how it is used in practice will be explained in this section. Note that the study described below used the same 119 samples as Example 1, providing tumor size data for all patients, and used the same sample feature table data as Example 1 (mass spectrometry data for the features in Appendix A). The only difference is that the feature m / z 9109 was deleted from the feature table because it has possible reproducibility issues and is of little value in classification.
[0889] Initial attempts to consider tumor size in assigning prognostic signatures to melanoma patients treated with nivolumab showed that, for patients for whom follow-up tumor size data were available at the time of treatment, baseline tumor size had a definite influence on the class to which samples in the development set should be assigned. This was discovered by obtaining data from 104 patients for whom both baseline and follow-up tumor size data were initially available, and dividing the set into two parts: one with smaller tumors at baseline and one with larger tumors at baseline. Figure 8 The classifier development method in [ 0 ] was used as for the classifier in Example 1, and separate classifiers were generated, one for patients with smaller tumors and another for patients with larger tumors. Then, samples in the development set were classified using either the large tumor classifier or the small tumor classifier, depending on the size of the tumor at baseline.
[0890] Note that some smaller tumors that were classified as late stage by the Example 1 ensemble classifier now receive the early stage category, and some larger tumors that were classified as early stage by the Example 1 ensemble classifier are now classified as late stage. Of particular note, some small tumors that were previously classified as late stage and that showed substantial tumor growth on treatment are now classified as early stage, and some large tumors that were previously classified as early stage and that showed significant shrinkage in the first 26 weeks of treatment are now classified as late stage. Figure 33A and 33B As shown, Kaplan-Meier plots were drawn for this subset of 104 patients, and the classifications by the two classifiers defined according to pre-treatment tumor size added hazard ratios between the new early and late groups. In these figures, "initial early / late" are the classification groups defined for the subset of 104 patients using the full set classifier of Example 1 (Method 1), while "TS-adjusted early / late" are the classifications generated by the new classifiers for smaller and larger tumors, which were applied to samples of smaller and larger tumors, respectively.
[0891] Table 58 presents the results of the survival analysis comparison between the early and late groups adjusted for tumor size.
[0892] Table 58: Classification performance of the 104 patient subset adjusted for tumor size
[0893]
[0894] These results indicate that suboptimal decisions were made when classifying some samples of the smallest and largest tumors and that this could be improved by taking tumor size into account when designing and generating the classifier of Example 1.
[0895] Baseline tumor size data were then obtained for the remaining 15 samples. When the new classifiers were applied to these samples (depending on whether they were in the small or large tumor size group) and these patients were added to the Kaplan-Meier analysis, it was noted that the improved separation had almost disappeared. Clearly, it was doing a worse job of classifying these 15 samples than before. An attempt was also made to perform the same method as described above for all 119 samples, but the result was more or less no improvement compared to the initial classification. These observations led to the conclusion that the 15 patients initially omitted were inherently different—in effect, they were omitted because they had not reached follow-up tumor size assessment and had very early progression (all 15 patients progressed before 78 days). It was hypothesized that for patients with very rapid progression, tumor size plays a much smaller role than for patients with a long-term progression-free period. In order to maintain the above improvements for the 104 patients who reached the 26-week assessment and correctly classify the other 15 samples, it was decided to first develop a classifier that removed the patients with the fastest progression and then repeat the classifier development process for the remaining samples, tailored to tumor size. That is, it is desirable to remove from classifier development those samples of the patients who have progressed the fastest, and then develop new classifiers based on tumor size, and generate a "small tumor" classifier and a "large tumor" classifier. These new classifiers are designed to be subsequently used to test patients for the benefit of immune checkpoint inhibitors, with the additional input of data on whether the patient has a "large" or "small" tumor, and then use the appropriate large or small tumor classifier. As will be apparent from the discussion below, the approach described below may generally be helpful in generating "small tumor" and "large tumor" classifiers in an oncology setting.
[0896] In order to exclude the patients with the fastest progression, we returned to the fu...
Claims
1. A machine that predicts which lung cancer patients will benefit from immunotherapy, including: a memory storing a reference set in the form of feature values of a plurality of mass spectral features obtained from mass spectra of blood samples of a plurality of lung cancer patients treated with immunotherapy, wherein the plurality of mass spectral features include a plurality of features listed in Appendix A, Appendix B, or Appendix C; The memory further stores a set of codes for defining a final classifier based on a plurality of main classifiers, each main classifier being generated from filtered mini-classifiers combined using a regularized combination method; A central processing unit operates on the code set and the reference set and mass spectrometry data obtained from a blood sample of a cancer patient to be tested, and responsively generates a class label for the blood sample, wherein the class label "good" or an equivalent predicts that the patient is likely to obtain similar benefit from a combination therapy comprising an antibody drug that blocks ligand activation of programmed cell death 1 (PD-1) and an antibody drug that targets cytotoxic T-lymphocyte-associated protein 4 (CTLA4) compared to a monotherapy of an antibody drug that blocks ligand activation of programmed cell death 1 (PD-1), and the class label "bad" or an equivalent indicates that the patient is likely to obtain greater benefit from the combination therapy.
2. The machine of claim 1, wherein the mass spectral data for the reference set and the samples are obtained from at least 100,000 laser shots using MALDI-TOF mass spectrometry on the samples forming the reference set and the samples to be tested.
3. The machine of claim 1, wherein the mini-classifiers are filtered according to any one of the following criteria: (a) Filters with depth s of 3, 1 for the development and internal validation sets, and classification accuracy for hazard ratio for time to progression and for time to progression less than 100 days and time to progression greater than 365 days; (b) Depth s of 2, 1 for the development and internal validation sets, and a filter for the hazard ratio for overall survival; (c) filters with depth s of 2, development and internal validation sets of 2, and hazard ratio for time to progression and classification accuracy for time to progression less than 100 days and time to progression greater than 365 days; (d) Filter with depth s of 2, development and internal validation sets of 2, and hazard ratio for overall survival.
4. The machine of claim 1, wherein the final classifier is defined from a plurality of master classifiers, each master classifier being generated from separating a classifier development sample set into a training set and a test set.
5. The machine of claim 1, wherein: The reference set included mass spectrometry data from patients treated with nivolumab and patients treated with both ipilimumab and nivolumab.
6. The machine of claim 1, wherein the regularized combination method comprises logistic regression with extreme value dropout.
7. The machine of claim 1, wherein the lung cancer is non-small cell lung cancer.
8. A method for generating a classifier, the method comprising: a) performing mass spectrometry analysis on a blood sample of a lung cancer patient and obtaining mass spectrometry data; (b) obtaining integrated intensity values in mass spectral data of a plurality of mass spectral features, wherein the mass spectral features include a plurality of features listed in Appendix A, Appendix B, or Appendix C; and (c) operating the mass spectrometry data using a program-controlled computer implementing a classifier; Wherein in step (c), the classifier uses a classification algorithm to compare the integrated intensity value with the feature values of a reference set of class-labeled mass spectrometry data obtained from blood samples obtained from multiple other cancer patients treated with antibody drugs that block ligand activation of programmed cell death 1 (PD-1), and detects a class label for the sample.
9. The method of claim 8, wherein the classifier is obtained from filtered small classifiers combined using a regularized combination method.
10. The method of claim 9, wherein the regularized combination method comprises repeatedly performing logistic regression with extreme value dropout on the filtered small classifiers.
11. The method of claim 8, wherein the mass spectral data is acquired from at least 100,000 laser shots performed on the sample using MALDI-TOF mass spectrometry.
12. The method of claim 8, wherein the lung cancer is non-small cell lung cancer.
13. A method for training a classifier, the method comprising the following steps: a) obtaining a development sample set from a lung cancer subject population; b) performing mass spectrometry analysis on the development sample set, and identifying mass spectrometry features present in the mass spectra of the development sample set; c) for each sample in the development sample set, obtaining protein expression data from a large set of proteins covering a biological function of interest; d) identifying a statistically significant association between one or more of the mass spectral features and a combination of proteins grouped according to their biological function; wherein the mass spectral features include a plurality of features listed in Appendix A, Appendix B or Appendix C, and e) Training a classifier based on the development sample set using one or more mass spectral features identified in step d) with computer assistance, wherein the classifier is in the form of a set of parameters that assign class labels to lung cancer samples according to program instructions.
14. The method of claim 13, wherein step d) further comprises the step of performing gene set enrichment analysis.
15. The method of claim 13, wherein the classifier is in the form of a combination of filtered small classifiers that are subjected to a regularization procedure. The method of claim 13 , wherein the samples in the development sample set are blood samples.
17. The method of claim 13, wherein step b) comprises performing at least 100,000 laser shots on each of the samples in the sample set in MALDI-TOF mass spectrometry.
18. A method as claimed in claim 13, wherein the classifier trained in step e) is considered to be a first classifier, and the method further comprises repeating step e) for one or more mass spectral features of a second group, wherein the mass spectral features are associated with a group of different proteins related to different biological functions, thereby training a second classifier.
19. The method of claim 13, wherein the lung cancer is non-small cell lung cancer.
20. The method of claim 13, wherein step a) further comprises obtaining a second independent sample set from a similar population of lung cancer subjects, Step b) further comprises performing mass spectrometry analysis on the second independent sample set and identifying mass spectrometry features present in the mass spectra of the second independent sample set, and Step c) further comprises obtaining protein expression data from a large set of proteins covering the biological function of interest for each sample in the second independent sample set.
21. The method of claim 20, wherein the samples in the second independent sample set are blood samples.
