PROGNOSTIC TEST OF THE BENEFITS OF A PD-1 ANTIBODIES IN A PATIENT WITH NON-SMALL CELL LUNG CANCER AND METHOD FOR DEVELOPING CLASSIFIERS

Description

Field

[0001] This invention relates to a method for predicting in advance of treatment whether a cancer patient is likely to benefit from administration of immune checkpoint inhibitors, including for example anti-PD-1 and / or anti-CTLA4 agents, allowing the immune system to attack the tumor. This application further relates to methods for developing. i.e., training, a computer-implemented classifier from a development set of samples.Background

[0002] Melanoma is a type of cancer primarily affecting the skin that develops from pigment-containing cells known as melanocytes. The primary cause of melanoma is ultraviolet light (UV) exposure in those with low levels of skin pigment, which causes damage to DNA in skin cells. The UV light may be from either the Sun or from tanning devices. About 25% of melanomas develop from moles. Individuals with many moles, a history of affected family members, or who have poor immune function are all at greater risk of developing a melanoma. A number of rare genetic defects also increase the risk of developing melanoma. Diagnosis of melanoma is typically done by visual inspection of any concerning lesion followed by biopsy.

[0003] Treatment of melanoma is typically removal by surgery. In those with slightly larger cancers nearby lymph nodes may be tested for spread. Most people are cured if spread has not occurred. In those in whom melanoma has spread, immunotherapy, biologic therapy, radiation therapy, or chemotherapy may improve survival. With treatment, the five-year survival rate in the United States is 98% among those with localized disease, but only 17% among those in whom spread has occurred. Melanoma is considered the most dangerous type of skin cancer. Globally, in 2012, it occurred in 232,000 people and resulted in 55,000 deaths.

[0004] Tumor mutations, including mutations associated with melanoma, create specific neoantigens that can be recognized by the immune system. Roughly 50% of melanomas are associated with an endogenous T-cell response. Cytotoxic T-cells (CT cells, or cytotoxic T lymphocytes (CTLs)) are leukocytes which destroy virus-infected cells and tumor cells, and are also implicated in transplant rejection. CTLs that express the CD8 glycoprotein at their surfaces are also known as CD8+ T-cells or CD8+ CTLs. Tumors develop a variety of mechanisms of immune evasion, including local immune suppression in the tumor microenvironment, induction of T-cell tolerance, and immunoediting. As a result, even when T-cells infiltrate the tumor they cannot kill the cancer cells. An example of this immunosuppression in cancer is mediated by a protein known as 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. Production of PD-L1 and PD-L2 is one way that the body naturally regulates the immune system. Many cancer cells make PD-L1, hijacking this natural system and thereby allowing cancer cells to inhibit T-cells from attacking the tumor.

[0005] One approach to the treatment of cancer is to interfere with the inhibitory signals produced by cancer cells, such as PD-L1 and PD-L2, to effectively prevent the tumor cells from putting the brakes on the immune system. Recently, an anti-PD-1 monoclonal antibody, known as nivolumab, marketed as Opdivo ®< , was approved by the Food and Drug Administration for treatment of patients with unresectable or metastatic melanoma who no longer respond to other drugs. In addition, nivolumab was approved for the treatment of squamous and non-squamous non-small cell lung cancer and renal cell carcinoma. Nivolumab has also been approved in melanoma in combination with ipilimumab, an anti-cytotoxic T-lymphocyte-associated protein 4 (CTLA4) antibody. Nivolumab acts as an immunomodulator by blocking ligand activation of the PD-1 receptor on activated T-cells. In contrast to traditional chemotherapies and targeted anticancer therapies, which exert their effects by direct cytotoxic or tumor growth inhibition, nivolumab acts by blocking a negative regulator of T-cell activation and response, thus allowing the immune system to attack the tumor. PD-1 blockers appear to free up the immune system only around the tumor, rather than more generally, which could reduce side effects from these drugs.

[0006] The current clinical results of anti-PD-1 treatment in melanoma patients are encouraging and overall results lead to progression free and overall survival results that are superior to alternative therapies. However, the real promise of these therapies is related to durable responses and long-term clinical benefit seen in a subgroup of around 40% of melanoma patients. Some portion of the other ~60% of patients might do better on alternative therapies. Being able to select which patients derive little benefit from anti-PD-1 treatment from pre-treatment samples would enable better clinical understanding and enhance the development of alternative treatments for these patients. There is also considerable cost related to these therapies, e.g. the recently approved combination of ipilimumab and nivolumab in melanoma while showing spectacular results is only effective in about 55% of patients while costing around $ 295,000 per treatment course. (Leonard Saltz, MD, at ASCO 2015 plenary session: "The Opdivo + Yervoy combo is priced at approximately 4000X the price of gold ($158 / mg)"). This results in a co-pay of around $60,000 for patients on a standard Medicare plan. Avoiding this cost by selecting these treatments only for those patients who are likely to benefit from them would result in substantial savings to the health care system and patients. It is also unclear whether the benefit of the combination of nivolumab and ipilimumab arises from a synergistic effect, or is just the sum of different patient populations responding to either nivolumab or ipilimumab. In any case having a test for nivolumab benefit would shed light on this question.

[0007] Much work has been performed to use the expression of PD-L1 measured by immunohistochemistry (IHC) as a biomarker for selection of anti-PD-1 treatments. Correlations between anti-PD-1 efficacy and outcome have been observed in some studies but not in others. Of particular issue is the current lack of standardization and universally accepted cut-offs in terms of IHC staining, which renders comparison of such data difficult. Of more fundamental issue is the observation that PD-L1 expression appears to be a dynamic marker, i.e. IHC expression changes during tumor evolution and during treatment. If one were to use PD-L1 expression via IHC in a rigorous manner one would require multiple repeat biopsies with a high corresponding risk and cost for patients. In contrast, a serum based test would not suffer from these effects.

[0008] The assignee Biodesix, Inc. has developed classifiers for predicting patient benefit or non-benefit of certain anticancer drugs using mass spectrometry of blood-based 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 publications 2013 / 0344111 and 2011 / 0208433. WO2015 / 039021 describes a method for classifier generation includes a step of obtaining data for classification of a multitude of samples, the data for each of the samples consisting of a multitude of physical measurement feature values and a class label. Individual mini-classifiers were generated using sets of features from the samples. The performance of the mini-classifiers was tested, and those that meet a performance threshold are retained. A master classifier was generated by conducting a regularized ensemble training of the retained / filtered set of mini-classifiers to the classification labels for the samples, e.g., by randomly selecting a small fraction of the filtered mini-classifiers (drop out regularization) and conducting logistical training on such selected mini-classifiers. The set of samples were randomly separated into a test set and a training set. The steps of generating the mini-classifiers, filtering and generating a master classifier are repeated for different realizations of the separation of the set of samples into test and training sets, thereby generating a plurality of master classifiers. A final classifier was defined from one or a combination of more than one of the master classifiers.

[0009] Taguchi et al (2007) "Mass Spectrometry To Classify Non-Small Cell Lung Cancer Patients For Clinical Outcome After Treatment With Epidermal Growth Factor Receptor Tyrosine Kinase Inhibitors: A Multicohort Cross Institutional Study" JNCI Journal of the national cancer institute vol. 99(11) 838-846 describes the VeriStrat 1.0 test which identifies patients who are likely to have no benefit from EFGR inhibitors (like erlotinib or gefitnib).Summary

[0010] The invention is set out in the appended set of claims.

[0011] As will described in more detail in Examples 9 and 10, a practical method of guiding melanoma patient treatment with immunotherapy drugs is disclosed. The method of the invention relates to guiding non-small cell lung cancer (NSCLC) patient treatment with immunotherapy drugs and includes the steps of a) conducting mass spectrometry on a blood-based sample of the patient and obtaining mass spectrometry data; (b) obtaining integrated intensity values in the mass spectrometry data of a multitude of mass-spectral features; and (c) operating on the mass spectral data with a programmed computer implementing a classifier. The classifier is obtained from filtered mini-classifiers combined using a regularized combination method comprising repeatedly conducting logistic regression with extreme dropout on the filtered mini-classifiers. In the operating step the classifier compares the integrated intensity values with feature values of a reference set of class-labeled mass spectral data obtained from blood-based samples obtained from a multitude of other cancer patients treated with an antibody drug blocking ligand activation of programmed cell death 1 (PD-1) with a classification algorithm and generates a class label for the sample. The class label "Good" or the equivalent (e.g., Late in the description of Example 10) predicts the patient is likely to obtain similar benefit from a combination therapy comprising an antibody drug blocking ligand activation of PD-1 and an antibody drug targeting CTLA4 and is therefore guided to a monotherapy of an antibody drug blocking ligand activation of PD-1 (e.g., nivolumab), whereas a class label of "Not Good" or the equivalent (e.g., Early in the description of Example 10) indicates the patient is likely to obtain greater benefit from the combination therapy as compared to the monotherapy of an antibody drug blocking ligand activation of programmed cell death 1 (PD-1) and is therefore guided to the combination therapy. The the mass spectral features include a multitude of features listed in Appendix A, Appendix B or Appendix C.

[0012] A method of treating a melanoma patient is described herein. The method includes performing the method recited above and if the class label is Good or the equivalent the patient is administered a monotherapy of an antibody drug blocking ligand activation of PD-1 (e.g., nivolumab), whereas if the class label of "Not Good" or the equivalent is reported the patient is administered a combination therapy an antibody drug blocking ligand activation of PD-1 and an antibody drug targeting CTLA4, e.g., the combination of nivolumab and ipilimumab.

[0013] The mass spectral features include a multitude of features listed in Appendix A, Appendix B or Appendix C, or features associated with biological functions Acute Response and Wound Healing (see Examples 1, 6 and 10The classifier is obtained from filtered mini-classifiers combined using a regularized combination method, e.g., using the procedure of Figure 8 or Figure 54. The regularized combination method takes the form of repeatedly conducting logistic regression with extreme dropout on the filtered mini-classifiers. In one example the mini-classifiers are filtered in accordance with criteria listed in Table 10. As disclosed in Example 9, the classifier may take the form of an ensemble of tumor classifiers (each having different proportions of patients with large and small tumors) combined in a hierarchical manner. In the illustrated embodiment of Example 9 if any one of the tumor classifiers returns an Early or the equivalent, the label the Not Good or equivalent class label is reported, whereas if all the tumor classifiers return a Late class label the Good or equivalent class label is reported.

[0014] In this method the relatively greater benefit from the combination therapy label means significantly greater (longer) overall survival as compared to monotherapy.

[0015] In another aspect, the reference set takes the form of a set of class-labeled mass spectral data of a development set of samples having either the class label Early or the equivalent or Late or the equivalent, wherein the samples having the class label Early are comprised of samples having relatively shorter overall survival on treatment with nivolumab as compared to samples having the class label Late.

[0016] In embodiments the mass spectral data is acquired from at least 100,000 laser shots performed on the sample using MALDI-TOF mass spectrometry. This methodology is described in Example 1 and in prior patent documents cited in Example 1,

[0017] In one embodiment, as indicated in Examples 6 and 10, the mass-spectral features are selected according to their association with at least one biological function, for example sets of features which are associated with biological functions Acute Response and Wound Healing.

[0018] A practical testing method is described for predicting melanoma patient response to an antibody drug blocking ligand activation of PD-1. The method includes steps of a) conducting mass spectrometry on a blood-based sample of the melanoma patient and obtaining mass spectrometry data; (b) obtaining integrated intensity values in the mass spectral data of a multitude of pre-determined mass-spectral features; and (c) operating on the mass spectral data with a programmed computer implementing a classifier. In the operating step the classifier compares the integrated intensity values obtained in step (b) with feature values of a reference set of class-labeled mass spectral data obtained from a multitude of other melanoma patients treated with the drug with a classification algorithm and generates a class label for the sample. The class label "early" or the equivalent predicts the patient is likely to obtain relatively less benefit from the antibody drug and the class label "late" or the equivalent indicates the patient is likely to obtain relatively greater benefit from the antibody drug. The method of generating the classifier used in this test from a development set of sample data is described in detail in this disclosure.

[0019] A machine is described which is capable of predicting melanoma patient benefit from an antibody drug blocking ligand activation of the programmed cell death 1 (PD-1). The machine includes a memory storing a reference set in the form of feature values for a multitude of mass spectral features obtained from mass spectrometry of blood-based samples from a multitude of melanoma patients treated with the antibody drug. The memory further stores a set of code defining a set of master classifiers each generated from a plurality of filtered mini-classifiers combined using a regularized combination method. The machine further includes a central processing unit operating on the set of code and the reference set and mass spectral data obtained from a blood-based sample of a melanoma patient and responsively generates a class label for the blood-based sample, wherein the class label "early" or the equivalent predicts the patient is likely to obtain relatively less benefit from the antibody drug and the class label "late" or the equivalent indicates the patient is likely to obtain relatively greater benefit from the antibody drug.

[0020] A system is disclosed for predicting patient benefit from an anti-body drug blocking ligand activation of PD-1 in the form of a mass spectrometer for conducting mass spectrometry of the blood-based sample of the patient and the machine as recited in the previous paragraph.

[0021] A a method of generating a classifier for predicting patient benefit from an antibody drug blocking ligand activation of programmed cell death 1 (PD-1) is disclosed. The method includes the steps of: 1) obtaining mass spectrometry data from a development set of blood-based samples obtained from melanoma patients treated with the antibody drug, in which a mass spectrum from at least 100,000 laser shots is acquired from each member of the set; 2) performing spectral pre-processing operations on the mass spectral data from the development sample set, including background estimation and subtraction, alignment, batch correction, and normalization; 3) performing the process of Figure 8 steps 102-150 including generating a master classifier based on a regularized combination of a filtered set of mini-classifiers; 4) evaluating performance of the master classifiers generated in accordance with step 3); and 5) defining a final classifier based on the master classifiers generated in step 3).

[0022] The final classifier includes a training set including feature values for a set of features listed in Appendix A, Appendix B, or Appendix C. In one possible embodiment, the method may include the step of deselecting features from the list of features of Appendix A which are not contributing to classifier performance and performing steps 3), 4), and 5) using a reduced list of features. Such a reduced list of features may take the form of the list of features in one of the sets of Appendix B or the list of features in Appendix C.

[0023] A method of treating a melanoma patient is disclosed. The method includes performing the method recited of predicting whether the patient will benefit from the antibody drug blocking ligand activation of PD-1 as recited above, and if the patient has class label of Late or the equivalent for their blood-based sample then performing a step of administrating the antibody drug to the patient. improved general purpose computer configured as a classifier for classifying a blood-based sample from a human cancer patient to make a prediction about the patient's survival or relative likelihood of obtaining benefit from a drug is disclosed. The improvement is in the form of a memory storing a reference set in the form of feature values for a multitude of mass spectral features obtained from mass spectrometry of blood-based samples from a multitude of melanoma patients treated with an immune checkpoint inhibitor and an associated class label for each of the blood-based samples in the reference set. The data of blood-based samples form a set used for developing the classifier. The memory further stores a set of computer-executable code defining a final classifier based on a multitude of master classifiers, each master classifier generated from a set of filtered mini-classifiers executing a classification algorithm and combined using a regularized combination method, such as extreme dropout and logistic regression. The multitude of master classifiers are obtained from many different realizations of a separation of the development set into classifier training and test sets. The improvement further includes a central processing unit operating on the set of code, the reference set, and mass spectral data obtained from the blood-based sample of the cancer patient to be tested and generating a class label for the blood-based sample.

[0024] The memory stores may feature values of at least 50 of the features listed in Appendix A. The memory stores may feature values for a reduced set of features, such as the features of one of the approaches listed in Appendix B or the list of features of Appendix C. The immune checkpoint inhibitor may comprise an antibody blocking ligand activation of PD-1. The immune checkpoint inhibitor may comprise an antibody blocking ligand activation of CTLA4.

[0025] A laboratory test center is described which includes a mass spectrometer for conducting mass spectrometry of a blood based sample from a cancer patient and a machine configured as a classifier and storing a reference set of mass spectral data as described herein.

[0026] We further describe in Example 9 below a general extension of classifier development to designing the development sets of an ensemble of classifiers to explore different clinical groups, for example different proportions of patients with large and small tumors. A method of generating an ensemble of classifiers from a set of patient samples is disclosed, comprising the steps of: a. defining a plurality of classifier development sample sets from the set of patient samples, each of which have different clinical characteristics (e.g., proportions of patients with large or small tumors, or other relevant clinical groupings); b. conducting mass spectrometry on the set of patient samples and storing mass spectrometry data; c. using a programmed computer, conducting a classifier development exercise using the mass spectral data for each of the development sets defined in step a. and storing in a memory associated with the computer the parameters of the classifiers thus generated, thereby generating an ensemble of classifiers; and d. defining a rule or set of rules for generating a class label for a test sample subject to classification by the ensemble of classifiers generated in step c. A method of testing a sample using the ensemble of classifiers generated in accordance with this method is also disclosed in Examples 8 and 9. In this method, step b. can be performed before or after step a.

[0027] While we describe in the examples details of our discoveries in melanoma and anti-PD-1 and anti-CTLA4 antibody drugs, our studies of protein correlations with classification labels, set forth in great detail below, have allowed us to generalize our discoveries. In particular, we can expect that Example 1, Example 2 and Example 3 classifiers may be relevant / applicable for a broad variety of drugs affecting immunological status of the patient, such as various immune checkpoint inhibitors, high dose IL2, vaccines, and / or combinational therapy, e.g., anti-PD-1 and anti-CTLA4 combination therapy. Furthermore, since effects that are measured in serum reflect the organism status as a whole, and the complement system, found to be relevant in our discoveries, affects innate and adaptive immunity on the global level, not just in a tumor site, the classifiers are expected to have similar performance in different indications (e.g., lung, renal carcinoma), and are not restricted to melanoma.

[0028] A classifier generation method is described, including the steps of: a) obtaining physical measurement data from a development set of samples and supplying the measurement data to a general purpose computer, each of the samples further associated with clinical data; b) identifying a plurality of different clinical sub-groups 1 ... N within the development set based on the clinical data; c) for each of the different clinical sub-groups, conducting a classifier generation process from the measurement data for each of the members of the development set that is associated with such clinical sub-groups, thereby generating clinical sub-group classifiers C1 ... CN; and d) storing in memory of a computer a classification procedure involving all of the classifiers C1 . . . CN developed in step c), each of the classifiers associated with a reference set comprising samples in the development set used to generate the classifier and associated measurement data.

[0029] A multi-stage classifier is disclosed which includes a programmed computer implementing a hierarchical classifier construction operating on mass spectral data of a test sample stored in memory and making use of a reference set of class-labeled mass spectral data stored in the memory. The classifier includes (a) a first stage classifier for stratifying the test mass spectral data into either an Early or Late group (or the equivalent, the moniker not being important); (b) a second stage classifier for further stratifying the Early group of the first stage classifier into Early and Late groups (or Earlier and Later groups, or the equivalent), the second stage implemented if the first stage classifier classifies the test mass spectral data into the Early group and the Early class label produced by the second stage classifier is associated with an exceptionally poor prognosis; and (c) a third stage classifier for further stratifying the Late group of the first stage classifier into Early and Late groups (or Earlier and Later groups, or the equivalent). The third stage classifier is implemented if the first stage classifier classifies the test mass spectral data into the Late group, wherein a Late class label (or Later or the equivalent) produced by the third stage classifier is associated with an exceptionally good prognosis.

[0030] The third stage classifier may comprise one or more classifiers developed from one or more different clinical sub-groups of a classifier development set used to generate the first level classifier. In one example, the third stage classifier includes at least four different classifiers C1, C2, C3, and C4, each developed from different clinical sub-groups. The multi-stage classifier may be configured to predict an ovarian cancer patient as being likely or not likely to benefit from platinum chemotherapy, and wherein the classifiers C1, C2, C3 and C4 are developed from the following clinical subgroups: C1: developed from a subset of patients with non-serous histology or serous histology together with unknown FIGO score; C2: developed from a subset of patients with serous histology not used to develop Classifier C1; C3: developed from a subset of patients with residual tumor after surgery; C4: developed from a subset of patients with no residual tumor after surgery.

[0031] In yet another aspect, we have discovered a method of generating a classifier for classifying a test sample from a development set of samples, each of the samples being associated with clinical data. The method includes the steps of: (a) dividing the development set of samples into different clinical subgroups 1 . . . N based on the clinical data, where N is an integer of at least 2; (b) performing a classifier development process (such as for example the process of Figure 8) for each of the different clinical subgroups 1 . . . N, thereby generating different classifiers C1 . . . CN; and (c) defining a final classification process whereby a patient sample is classified by the classifiers C1 . . . CN.

[0032] We have discovered a method of generating a classifier for classifying a test sample, comprising the steps of: (a) generating a first classifier from measurement data of a development set of samples using a classifier development process; (b) performing a classification of the measurement data of the development set of samples using the first classifier, thereby assigning each member of the development set of samples with a class label in a binary classification scheme (Early / Late, or the equivalent); and (c) generating a second classifier using the classifier development process with an input classifier development set being the members of the development set assigned one of the two class labels in the binary classification scheme by the first classifier (e.g., the Early group), the second classifier thereby stratifying the members of the set with the first class label into two further sub-groups. The method optionally includes the steps (d) dividing the development set of samples into different clinical subgroups 1 . . . N where N is an integer of at least 2; and (e) repeating the classifier development process for each of the different clinical subgroups 1 . . . N, thereby generating different third classifiers C1 . . . CN; and (f) defining a hierarchical classification process whereby: i. a patient sample is classified first by the first classifier generated in step a); ii. if the class label assigned by the first classifier is the class label used to generate the second classifier, then classifying the patient sample with the second classifier; and iii. if the class label assigned by the first classifier is not the class label used to generate the second classifier, then classifying the patient sample with the third classifiers C1 .. . CN; and iv. assigning a final label as a result of classification steps ii or step iii.

[0033] Example 6 below describes our ability to correlate specific mass spectral features with protein functional groups circulating in serum and use such correlations to train a classifier, or to monitor changes in a biological process. A method of training a classifier is disclosed, comprising the steps of: a) obtaining a development set of samples from a population of subjects and optionally a second independent set of samples from a similar, but not necessarily identical population of subjects; b) conducting mass spectrometry on the development set of samples, and optionally on the second set of samples, and identifying mass spectral features present in the mass spectra of the set(s) of samples; c) obtaining protein expression data from a large panel of proteins spanning biological functions of interest for each of the samples in the development set of samples or optionally each of the samples in the second set of samples; d) identifying statistically significant associations of one or more of the mass spectral features with sets of proteins grouped by their biological function using Gene Set Enrichment Analysis methods; and e) with the aid of a computer, training a classifier on the development set of samples using the one or more mass spectral features identified in step d), the classifier in the form of a set of parameters which assigns a class label to a sample of the same type as the development set of samples in accordance with programmed instructions.

[0034] The classifier may be in the form of a combination of filtered mini-classifiers which have been subject to a regularization procedure. The samples in the development set, and optional second sample set, are blood-based samples, e.g., serum or plasma samples from human patients.Brief description of the drawings

[0035] Figure 1 illustrates Kaplan-Meier plots for time-to-progression (TTP)(Fig. 1A) and overall survival (OS)(Fig. 1B) for the cohort of 119 melanoma patients treated with nivolumab with available clinical data and spectra from pre-treatment samples. Figures 2A and 2B are Kaplan-Meier plots of time-to-event data (TTP and OS), respectively, for all 119 patients with available clinical data and spectra from pretreatment samples by prior treatment (no prior ipilimumab, prior ipilimumab). Differences in outcome were not statistically significant. Figures 3A and 3B are Kaplan-Meier plots of time-to-event data (TTP and OS), respectively, for all 119 patients with available clinical data and spectra from pretreatment samples showing relatively good outcomes for patients in cohort 5. Figure 4A and Figure 4B are Kaplan-Meier plots showing time-to-event data for all 119 patients with available clinical data and spectra from pretreatment samples split into development (N=60) and validation (N=59) sets ("DEV1"); Figures 4C and 4D are Kaplan-Meier plots of time-to-event data for all 119 patients with available clinical data and spectra from pretreatment samples split into development (N=60) and validation (N=59) sets for a second split of samples into development and validation sets ("DEV2"). Figure 5 is a plot of bin normalization scalars as a function of disease control (DC); the bin method is used to compare normalization scalars between clinical groups of interest to ensure that windows useful for classification are not used for partial ion current normalization. Figure 6 is a plot of mass spectra from a multitude of samples showing several feature definitions defined within an m / z range of interest; different clinical performance groups are shown in contrasting line conventions. Figure 7 is a plot of bin normalization scalars as a function of DC for a partial ion current normalization performed on the features in the final feature table used for classifier generation. Figures 8A-8B are a flow chart of a classifier development process we used to develop the melanoma / nivolumab classifiers of this disclosure from the sample set of 119 serum samples from melanoma patients in the trial of nivolumab. Figures 9-12 are Kaplan-Meier plots showing classifier performance for the classifiers we developed in Example 1 using the classifier development procedure of Figure 8. Figures 9A and 9B show Kaplan-Meier plots for OS and TTP, respectively, by Early and Late classification groups for "approach 1" (see table 10) for the development set and Figures 9C and 9D show the Kaplan-Meier plots for OS and TTP, respectively, by Early and Late classification groups for approach 1 (see table 10) for the validation set. Figures 10A and 10B show Kaplan-Meier plots for OS and TTP, respectively, by Early and Late classification groups for "approach 2" (see table 10) for the development set, Figures 10C and 10D show the Kaplan-Meier plots for OS and TTP, respectively, by Early and Late classification groups for approach 2 (see table 10) for the validation set. Figures 11A and 11B show Kaplan-Meier plots for OS and TTP, respectively, by Early and Late classification groups for "approach 3" (see table 10) for the development set and Figures 11C and 11D show Kaplan-Meier plots for OS and TTP, respectively, by Early and Late classification groups for approach 3 (see table 10) for the validation set. Figures 12A and 12B show Kaplan-Meier plots for OS and TTP by Early and Late classification groups for "approach 4" (see table 10) for the development set, and Figures 12C and 12D show the Kaplan-Meier plots for OS and TTP by Early and Late classification groups for approach 4 (see table 10) for the validation set. Figures 13A and 13B illustrate Kaplan-Meier plots for OS and TTP, respectively, by Early and Late classification groups for a first approach shown in Table 13 applied to the whole set of 119 samples as development set (100) for classifier generation in accordance with Figure 8. Figures 13C and 13D illustrate Kaplan-Meier plots for OS and TTP, respectively, by Early and Late classification groups for a second approach shown in Table 13 applied to the whole set of 119 samples as development set (100) for classifier generation in accordance with Figure 8. Figure 14 is a Kaplan-Meier plot for the analysis of the Yale cohort of patients treated with anti-PD-1 antibodies, an independent sample set used for validation of the classifiers of Example 1 developed using Figure 8. Figure 15 is an illustration of a laboratory testing center including a mass spectrometer and a machine in the form of a general purpose computer for predicting melanoma patient benefit from antibody drugs blocking ligand activation of PD-1. Figures 16A and 16B are Kaplan-Meier plots of progression free survival (PFS) and overall survival (OS), respectively, for all 173 non-small cell lung cancer (NSCLC) patients of the ACORN NSCLC cohort with available clinical data and spectra from pretreatment samples. The ACORN NSCLC cohort was used to develop and tune the classifier of Example 2 to be predictive of melanoma patient benefit of nivolumab. Figures 17A and 17B are plots of progression free survival PFS and overall survival OS for the three different subsets of the ACORN NSCLC cohort. One subset ("additional filtering" in the figures) was used to filter mini-classifiers in generation of the classifier of Example 2. One subset ("test" in the figures) was used for testing classifier performance together with the melanoma development set samples. The other subset ("validation" in the figures) was used as an internal validation set, in addition to a melanoma subset already held for that purpose. The similarity in the plots of Figures 17A and 17B indicate that the survival data for the "additional filtering" subset was representative of the other subsets. Figures 18A-18H are Kaplan-Meier plots of OS and TTP or PFS for the development and internal validation sets of samples. In particular, Figures 18A-18D are the plots of OS and TTP for the development and validation sets of the melanoma / nivolumab cohort. Note that Figures 18A-18D show the separation in the survival plots of the samples labeled Early and Late by the classifier of Example 2. Figures 18E-18H are the plots of OS and PFS for the development and validation sets of the ACORN NSCLC chemotherapy cohort. Note that the plots in the ACORN NSCLC cohort (Figures 18E-18H) show similar OS and PFS for samples labeled Early and Late by the classifier of Example 2. Figure 19 is a Kaplan-Meier plot for an independent validation cohort of patients treated with anti-PD-1 antibodies (the Yale anti-PD-1 cohort), showing the separation of OS plots for patient samples labeled Early and Late by the classifier of Example 2. Figure 20 is a Kaplan-Meier plot for an independent Yale validation cohort of melanoma patients treated with anti-CTLA4 antibodies, showing the separation of OS plots for patient samples labeled Early and Late by the classifier of Example 2. Figures 21A and 21B are Kaplan-Meier plots of OS (Figure 21A) and disease free survival (DFS, Figure 21B) for the independent validation cohort of patients with ovarian cancer treated with platinum-doublet chemotherapy after surgery. Like the plots of Figures 18E-18G, the plots of Figures 21A and 21B show a lack of separation in OS and DFS for the ovarian cancer patients whose samples are classified as Early or Late by the classifier of Example 2. Figure 22 is a Kaplan-Meier plot of overall survival for the Yale anti-CTLA4 cohort. Figure 23 is a Kaplan-Meier plot for the Yale cohort of melanoma patients treated with anti-CTLA4 antibodies for the "full-set" classifier of Example 1, see Table 13 below. Figure 23 is similar to Figure 20 in that both figures show that the classifiers of Example 1 and Example 2 of this disclosure are able to predict melanoma patients having relatively better or worse outcomes on anti-CTLA4 antibody treatment. Figures 23A and 23B are Kaplan-Meier plots of classifier performance of a second anti-CTLA4 antibody classifier, which was developed from the Yale anti-CTLA4 cohort. Figures 24A and 24B are Kaplan-Meier plots of OS and PFS, respectively, for the ACORN NSCLC cohort used in the development of the classifier of Example 2. Figures 25A and 25B are Kaplan-Meier plots of OS and PFS, respectively, for the ACORN NSCLC cohort, by classification produced by the "full-set" classifier of Example 1. Note the clear separation in the overall survival plot of Figure 25A between the samples classified as Early and Late by the full-set classifier of Example 1. A separation in PFS between the Early and Late classified samples is also shown in Figure 25B. Figures 26A and 26B are Kaplan-Meier plots for the ACORN NSCLC cohort by classification and VeriStrat label (assigned in accordance with the classifier and training set of U.S. Pat. 7,736,905). Note that there is essentially no separation between the VeriStrat Good (VS-G) and Poor (VS-P) samples classified Early by the full-set classifier of Example 1, and the clear separation in the survival plots between those classified Late and those classified Early by the full-set classifier of Example 1. Figures 27A and 27Bare Kaplan-Meier plots of OS and DFS for the ovarian cancer chemotherapy cohort of 138 patients used for internal validation of the classifier of Example 2. Figures 28A and 28B are Kaplan-Meier plots for OS and DFS, respectively, of the ovarian cancer chemotherapy cohort by classification produced by the "full-set" classifier of Example 1. Note the clear separation in the overall survival plot of Figure 28A between the samples classified as Early and Late by the full-set classifier of Example 1. A clear separation in DFS between the Early and Late classified samples is also shown in Figure 28B. Thus, the Figures demonstrate the ability of the full-set classifier of Example 1 to predict ovarian cancer survival on chemotherapy. Figures 29A and 29B are Kaplan-Meier plots of the ovarian cancer chemotherapy cohort by classification and VeriStrat label. Note that there is essentially no separation between the samples classified Early and tested as VeriStrat Good (VS-G) and Poor (VS-P) by the full-set classifier of Example 1, and the clear separation in the survival plots between those classified Late and those classified Early by the full-set classifier of Example 1. As in the NSCLC cohort, it is apparent that outcomes are similar between VeriStrat subgroups within the group of patients classified as Early. Figures 30A-30F are Kaplan-Meier plots of classifications by early and late groups produced by the full set classifier of Example 1 initially (advance of treatment, "baseline" herein) shown in Figure 30A and 30B, after 7 weeks of treatment (WK7) shown in Figures 30C and 30D, and after 13 weeks (WK13), shown in Figures 30E and 30F. Figures 31A and 31B are Kaplan-Meier plots of overall survival and time to progression (TTP), respectively, grouped by the triplet of baseline, WK7, and WK13 classifications produced by the Example 1 full set classifier. There were too few patients with other label combinations for a meaningful analysis. Figures 32A and 32B are Kaplan-Meier plots of overall survival and time to progression (TTP), respectively, which show the outcomes when the patients are grouped according to their triplet of baseline, WK7, and WK13 classifications produced by the Example 2 classifier. (E=Early, L=Late) Figures 33A and 33B are Kaplan-Meier plots of a subset of 104 patients of the development sample set of Example 1, showing the Early and Late labeled patients from the original classifications produced by the Example 1 full set classifier ("Original") and the classifications produced by new "large" and "small" tumor classifiers ("TSAdjusted"). Figures 34A and 34B are Kaplan-Meier plots of overall survival and time to progression (TTP), respectively, of 47 Early patients (classified according to the Example 1 full-set Approach 1 classifier) classified into Earlier and Later groups by a classifier using these 47 samples as its development set. The patients in the Earlier classification group are removed from the development sample set in generating classifiers which take into account tumor size. Figure 35A and 35B are Kaplan-Meier plots of all 119 samples in the original development set of Example 1, with the large tumors classified by a classifier developed using only large tumors, the small tumors classified by a classifier developed using only small tumors and early progressing patients (the Early / Earlier patients of Figure 34) classified as Early (their original Example 1 full-set Approach 1 classifications). These groups are labeled as "Final" and compared with the groups produced by the Example 1 full-set classifier ("Original"). Figures 36A and 36B are waterfall plots showing the class identifications for melanoma patients who had either increasing or decreasing tumor size over the course of treatment with nivolumab. Figure 36A shows the data for the large and small tumor classifiers classifying the members of the development sample set (after removal of the fast progressing patients) and fast progressing patients with available tumor size change data classified as Early ("Final"), whereas Figure 36B shows the data for the full-set classifier of Example 1. Figure 37 is a flow-chart showing the process of development of small and large tumor classifiers from a development sample set. Figure 38 is a flow-chart showing how either the large and small tumor classifier generated in accordance with Figure 37 is used to test a sample of a cancer patient, depending on whether the patient has a large or small tumor. Figure 39 is a flow-chart showing one method for removing from the development set the fast progressing samples, step 3702 of Figure 37. Figures 40A and 40B are Kaplan-Meier plots of overall survival and time to progression, respectively, for classifications generated by an ensemble of seven classifiers, each of which are based on different classifier development sets having different clinical groupings (based on tumor size in this example) and generated in accordance with Figure 8. The ensemble of classifiers is generated from the 119 patient samples described in Example 1. A set of rules define a class label from labels produced by the ensemble of classifiers, such as "Bad", "Good" and "Other," which can be used to guide melanoma patient treatment as explained in Example 9 below. Figure 41 is a Kaplan-Meier plot of overall survival for the classifications obtained by a test composed of the ensemble of seven classifiers of Example 9 for 30 samples in an anti-PD-1 treated validation cohort. Figure 42 is a Kaplan-Meier plot of overall survival by Good and NotGood class labels produced by an ensemble of seven classifiers of Example 8 for melanoma patients treated with anti-PD1 monotherapy (nivolumab) as well as melanoma patients treated with both ipilimumab and nivolumab combination therapy. The plot shows that nivolumab patients having the class label Good have very similar survival as compared to patients with the nivolumab + ipilimumab combination therapy. Figure 43 is an example of a plot of running sum (RS) score (RS(S l,p )) calculated in a gene set enrichment analysis (GSEA) for one protein set S l . Figure 44A and 44B are examples of the null distributions for the two definitions of the enrichment scores (ES) used in a GSEA, showing the calculated ES and the regions assessed to determine the p value. Figures 45A and 45B are "heat maps", namely plots of p values generated by GSEA associating all 351 defined mass spectral features (Appendix A) in our nivolumab study of Examples 1 and 6 with protein functional groups. Figure 45A is the heat map for ES definition 1, and Figure 45B is the heat map for ES definition 2. Only every 5 th< spectral feature is labeled on the x axis. Figure 46 is a schema of a final classification procedure of Example 6. Figures 47A and 47B illustrate Kaplan-Meier plots of overall survival (OS) (Figure 47A) and time to progression (TTP) (Figure 47B) for the melanoma / immunotherapy new classifier development (NCD) cohort by classification group from Classifier 1 of Example 6. Figures 48A and 48B illustrate Kaplan-Meier plots of OS and TTP, respectively, for the 68 samples not classified as "Early" by Classifier 1, by classification group from Classifier 2 of Example 6. Figures 49A and 49B illustrate Kaplan-Meier plots of OS and TTP, respectively, for all 119 samples by overall classification produced by the classifier schema of Figure 46. Figures 50A-50D illustrates Kaplan-Meier plots comparing the performance of the classifier developed in Example 6 ("Current") with those developed in Example 1 (Figure 50A and 50B, OS and TTP, respectively)("IS2") and Example 8 (Figures 50C and 50D, OS and TTP, respectively)("IS6"). Figure 51 is a Kaplan-Meier plot of OS for the validation cohort of Example 6 by classification group. Figure 52 is a block diagram of a classifier training system including a mass spectrometer for conducting mass spectrometry on a development set of samples (not shown, e.g., serum or blood samples), a GSEA platform, including protein assay system and computer with GSEA analysis module, and a computer programmed to conduct a classifier training procedure using sets of mass spectral peaks associated with a particular protein function group of a biological process of interest in the development set of samples. Figures 53A and 53B are Kaplan-Meier plots of time to event data for disease free survival (DFS, Figure 53A) and overall survival (OS, Figure 53B) for a cohort of 138 ovarian cancer patients with available clinical data and mass spectral data, which were used to develop the ovarian classifiers of Example 9. Figures 54A and 54B are a flow chart showing a computer-implemented procedure for developing a classifier from a development sample set. In Example 9, the procedure of Figures 54A and 54B (up to and including step 350) was performed several different times for different configurations or subsets of the development sample set to result in the creation of a tiered or hierarchical series of classifiers (referred to as Classifiers A, B and C), as will be explained in more detail in the description of Example 9. Figures 55A and 55B are Kaplan-Meier plots of time to event data for the 129 patients in the ovarian development sample set of Example 9 with available clinical data, DFS > 1 month, and mass spectral data from pretreatment samples, showing the plots for a split of the sample set into development (N = 65) and validation (N=64) sets. Figure 55A shows the plot of DFS; Figure 55B shows the plot for OS. Note that the plots for the development and validation sample sets are essentially the same. Figures 56A-56D are Kaplan-Meier plots of OS and DFS by Early and Late classification groups produced by the first tier or "Classifier A" classifier of Example 9, for the 129 patients split into development and validation sets. Figure 56A is a plot of OS for the development set; Figure 56B is a plot of DFS for the development set; Figure 56C is a plot of OS for the validation set; Figure 56D is a plot of DFS for the validation set. Figures 57A and 57B are Kaplan-Meier plots of OS and DFS by Early and Late classification groups, for the "Classifier A" of Example 9 run on all 138 samples. Figures 58A and 58B are Kaplan-Meier plots of OS and DFS, respectively, by classification group produced by the Classifier B classifier of Example 9, for the subset of the development set of samples which were used to develop the Classifier B. Figure 59 is a flow chart showing a process for generating a second tier classifier ("Classifier B") from those development set samples that were classified as "Early" by the first tier "Classifier A" classifier of Example 9. Figure 60 is a flow chart showing a process for generating a third tier classifier C of Example 9; in this particular example the third tier consists of a several different classifiers each based on a different and clinically distinct subset of the development sample set. Figure 61 is a diagram showing the construction of a third tier Classifier C, and how it could be used to generate a "Good' class label based on the results of classification by each of the members of the third tier. Figure 62 is a diagram or schema showing the construction of a final classifier composed of a three-stage hierarchical classifier. Figure 63 is a diagram or schema showing the construction of an alternative final classifier in which the third stage of the three-stage hierarchical classifiers is made up of four individual classifiers developed from clinically distinct subgroups. Figures 64A and 64B are Kaplan-Meier plots of OS and DFS, respectively by classification group produced on the development sample set using the final classifier construction of Figure 63 and Example 9. Figures 65A-65C are plots of the average of the percentage of the "PROSE-erlotinib" data set variances explained by each principal component (PC) as a function of the PC index (descending order of variance) from Example 10, for biological functions Acute Response (Figure 65A), Wound Healing (Figure 65B) and Complement system (Figure 65C). Figures 66A-66D are distributions of the Acute Response Score in four sample sets described in Example 10. In the results shown in Figure 66A-66D, twenty nine mass spectrometry features were determined to be correlated with the AR biological function and were used in calculation of the corresponding Score. Figures 67A-67D are distributions of the Acute Response Score in four sample sets: "PROSE-erlotinib" (85 samples with available IS2 labels), Figure 67A; "PROSE-chemo" (122 samples with available IS2 labels), Figure 67B; "Moffitt", Figure 67C; and "Moffitt-Week7", Figure 67D. Figures 67A-67D also shows the Scores split by IS2 classification label for the samples from the Example 1 classifier. Figures 68A-68D are Kaplan-Meier plots for OS, PFS for "PROSE-erlotinib" sample set, Figures 68A-68B, Kaplan-Meier plots for OS and PFS in a "PROSE-chemo" set, Figure 68C-68D, and Kaplan-Meier plots for OS and TTP in "Moffitt" set, Figures 68E-68F, by group defined according to the AR Score threshold defined as illustrated in the Figures. The corresponding numbers of samples in each group, hazard ratios (HRs), log-rank p-values and medians are shown below each plot. Figures 69A-69D are illustrations of the evolution of the Acute Response Score over time for 107 patients in both the Moffitt and Moffitt-Week7 sets, grouped by combination of IS2 label at baseline (before treatment) and week 7 (after treatment). Each line in the plots represent the score of an individual patient. Figure 69A is the plot for patients with class label Early at both baseline and week 7; Figure 69B is the plot for the patients with class labels Late at both baseline and week 7; Figure 69C is the plot for patients with class label Early at baseline and Late at week 7, and Figure 69D is a plot for patients with class label Late at baseline and Early at week 7. Figure 70A-70C are plots of evolution of the Acute Response Score for the 107 patients with samples both in the "Moffitt" and the "Moffitt-Week7" sets, grouped by treatment response. Each line represents one single patient. Figure 70A is the plot for the PD: progressive disease treatment response; Figure 70B is the plot for the PR: partial response to treatment, and Figure 70C is the plot for SD: stable disease treatment response. Figures 71A and 71B are Kaplan-Meier plots of OS and TTP, respectively for the samples both in the "Moffitt" and the "Moffitt-Week7" sets grouped by whether the AR score increased or had a small change or decrease. The plots show the change in AR score has prognostic significance. Figure 72A-72D are the plots of distribution of the Wound Healing Score across four sample sets described in Example 10. Figure 73A-73D are illustrations of the evolution of the Wound Healing Score over time for 107 patients in both the Moffitt and Moffitt-Week7 sets, grouped by combination of IS2 label at baseline (before treatment) and week 7 (after treatment). Each line in the plots represents the score of an individual patient. Figure 73A is the plot for patients with class label Early at both baseline and week 7; Figure 73B is the plot for the patients with class labels Late at both baseline and week 7; Figure 73C is the plot for patients with class label Early at baseline and Late at week 7, and Figure 73D is a plot for patients with class label Late at baseline and Early at week 7. Figure 74A-74C are plots of evolution of the Wound Healing Score for the 107 patients with samples both in the "Moffitt" and the "Moffitt-Week7" sets, grouped by treatment response. Each line represents one single patient. Figure 74A is the plot for the PD: progressive disease treatment response; Figure 74B is the plot for the PR: partial response to treatment, and Figure 74C is the plot for SD: stable disease treatment response. Figures 75A-75D are the plots of distribution of the Complement System Score across four sample sets described in Example 10. Figure 76A-76D are illustrations of the evolution of the Complement System Score over time for 107 patients in both the Moffitt and Moffitt-Week7 sets, grouped by combination of IS2 label at baseline (before treatment) and week 7 (after treatment). Each line in the plots represents the score of an individual patient. Figure 76A is the plot for patients with class label Early at both baseline and week 7; Figure 76B is the plot for the patients with class labels Late at both baseline and week 7; Figure 76C is the plot for patients with class label Early at baseline and Late at week 7, and Figure 76D is a plot for patients with class label Late at baseline and Early at week 7. Figure 77A-77C are plots of evolution of the Wound Healing Score for the 107 patients with samples both in the "Moffitt" and the "Moffitt-Week7" sets, grouped by treatment response. Each line represents one single patient. Figure 77A is the plot for the PD: progressive disease treatment response; Figure 77B is the plot for the PR: partial response to treatment, and Figure 77C is the plot for SD: stable disease treatment response. Figures 78A-78D are Kaplan-Meier plots for OS, PFS ("PROSE-chemo"), Figures 78A- 78B, and OS and TTP ("Moffitt" set), Figures 78C and 78D, by group defined according to the AR score threshold defined by tertiles in the PROSE set. The corresponding number of samples in each group, hazard ratios (HRs), log-rank p-values and medians are shown below each plot. Figures 79A and 79B are Kaplan-Meier plots for classification groups Early and Late, obtained from a classifier developed in accordance with the procedure of Figure 8 but instead of using mass spectral features, the features used for classification are the Biological Function Scores of Example 10, in this example the Acute Response Score, the Wound Healing Score and the Complement Score. Figure 80 is a schematic illustration of system for generating the biological function scores of Example 10 for a sample set. Figure 81 is a flow chart showing the steps for calculating the biological function scores of Example 10 from mass spectrometry data. Figure 82 is an illustration of a partial feature table for sample set ss, F ss< which is used in the calculation of a biological function score of Example 10. Figure 83 is an illustration of an average first principal component vector which is used in the calculation of a biological function score of Example 10. Detailed Description

[0036] A practical test (method) is disclosed in this document for predicting whether a cancer patient is likely to be benefit from administration of immune checkpoint inhibitors such as a monoclonal antibody drug blocking ligand activation of PD-1 on activated T cells, e.g., nivolumab. The method makes use of the mass spectrum of the patient's serum or plasma sample acquired pre-treatment, and a general purpose computer configured as a classifier which assigns a class label to the mass spectrum. The class label can take the form of "early" or the equivalent, or "late" or the equivalent, with the class label "late" indicating that the patient is a member of a class of patients that are likely to obtain relatively greater benefit from the drug than patients that are a member of the class of patients having the class label "early." The particular moniker used for the class label is not particularly important.

[0037] Overall survival is a primary indicator for assessing the benefit of antibody drugs blocking ligand activation of PD-1. Hence, when considering the meaning of the labels Early and Late, in one embodiment the "relatively greater benefit" associated with the Late label means a patient whose sample is assigned the Late label is likely to have significantly greater (longer) overall survival than a patient with the Early class label.

[0038] The term "antibody drug blocking ligand activation of PD-1" is meant to include not only antibodies that bind that 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) would also block ligand activation of PD-1. The term "immune checkpoint inhibitors" is meant to include those that block PD-1 as well as those that block CLTA-4. The term immune checkpoints inhibitors or "checkpoint blockers" is defined as immunomodulatory mAbs that target CTLA4-like receptors and their ligands. See Galluzzi et al., Oncoimmunology 2014; 3:e967147. FDA-approved agents of this type include ipilimumab (anti-CTLA4), nivolumab (anti-PD-1), and pembrolizumab (anti-PD-1). Several publications / abstracts released during the last 13 months reported the results of clinical trials involving additional checkpoint blockers, such as the CTLA4-targeting mAb tremelimumab, the PD-1-targeting mAb pidilizumab, and the PD-L1-targeting mAbs MEDI4736 (durvalumab), MPDL3280A (atezolizumab), and MSB0010718C (avelumab). See Buqué et al. Oncoimmunology. 2015 Mar 2; 4(4).

[0039] Example 1 explains the development of a classifier from a melanoma / nivolumab sample set and provides details of classifier performance.

[0040] Example 2 explains a redevelopment of the classifier of Example 1 that has been tuned to be more predictive and less prognostic for patient benefit from nivolumab.

[0041] Example 3 explains the development of a classifier that is predictive for melanoma patient benefit from anti-CTLA4 antibodies, another immune checkpoint inhibitor.

[0042] Example 4 explains how the classifier developed in accordance with Example 1 is also able to predict whether non-small cell lung cancer (NSCLC) and ovarian cancer patients are likely to have relatively higher or lower overall survival from chemotherapy.

[0043] Example 5, and Figure 15, describes a practical testing environment for conducting the tests of this disclosure on a blood-based sample from a patient in advance of treatment.

[0044] Example 6 describes our studies of proteins which are correlated to the Early and Late class labels in the Example 1 and Example 2 classifiers using Gene Set Enrichment Analysis, which allow us to generalize our discoveries to other immune checkpoint inhibitors and other types of cancers, as well as generate classifiers based on mass spectral features associated with particular protein functional groups.

[0045] Example 7 describes longitudinal studies from patient samples of Example 1 and how changes in class label produced by our classifiers can be used, inter alia, to monitor treatment efficacy and guide treatment.

[0046] Example 8 describes an ensemble of classifiers generated from different clinical subsets of the Example 1 development sample set population, in this example, melanoma patients with large and small tumors.

[0047] Example 9 provides further examples of development of an ensemble of classifiers from clinically different development sets, including a first example of an ensemble of classifiers from the nivolumab / melanoma set of Example and a second example from study of ovarian cancer patients treated with chemotherapy.

[0048] Example 10 describes a methodology for measurement of a biological function score using mass spectrometry data. Example 10 further describes the measurement of biological functions scores in four different sample sets, each blood-based samples from humans with cancer. Example 10 describes how the scores can be used to guide treatment and to build a classifier using biological function scores as features for classifier training, e.g., using the procedure of Figure 8. Example 10 also builds on the discoveries described in Example 6, including correlation of biological functions with mass spectrometry peaks.Example 1Classifier and method for predicting melanoma patient benefit from antibody drug blocking ligand activation of PD-1

[0049] We obtained samples to develop a classifier of Example 1 from a clinical trial of nivolumab in treatment of melanoma. We describe briefly this trial below. We then describe the patient samples which we obtained, the mass spectrometry methods used to obtain spectra from the samples, including spectra processing steps. These mass spectral procedures are preferably in accordance with the so-called "Deep MALDI" method described in U.S. Patent 9,279,798. We then describe in detail a classifier generation process which was used to define a final classifier which can assign class labels to spectra in accordance with the test. We also describe below the results of the classifier generation process and demonstrate its ability to assign class labels to mass spectra from blood-based samples which predict whether the patient providing the sample is likely to obtain relatively greater or lesser benefit from the antibody drug.

[0050] The classifier generation method uses what we have called "combination of mini-classifiers with dropout regularization", or CMC / D, described in U.S. patent application publication 2015 / 0102216. This procedure for developing a classifier is also referred to herein as DIAGNOSTIC CORTEX, a trademark of Biodesix, Inc. See the discussion of Figure 8 below. Applying this procedure to pre-treatment serum spectra from melanoma patients obtained using Deep MALDI spectral acquisition we have identified clinical groups "Early" and "Late." These groups are showing significant differences in outcome (both time to progression (TTP) and overall survival (OS)) following treatment with an anti-PD-1 treatment, nivolumab. See Figures 9-14 and the discussion below. We have presented a test procedure to identify these groups from pre-treatment samples, and validated the results in internal validation sets, and in an external validation set (see Figure 14).

[0051] Patients whose serum classifies as "Early" exhibit significantly faster progression and shorter survival than patients whose serum classifies as "Late" making this test suitable as a biomarker for nivolumab therapy. The clinical groups "Early" and "Late" are not associated with PD-L1 expression, and our classification remains a significant predictor for outcome (both TTP and OS) even when other clinical attributes are included in a multivariate analysis.

[0052] While a correlative approach to test development does not easily lend itself to a deep understanding of the biology underlying the difference between the identified groups, we have done some initial work relating the two different groups to differences in acute phase reactants and the complement system. These studies are set forth in Example 1 and later in this document in Example 6. The success of this project exemplifies the power of the combination of Deep MALDI spectral acquisition and our inventive classifier development method in the construction of clinically useful, practical tests.Clinical trial

[0053] The trial from which samples were available for new classifier development was a study of nivolumab with or without a peptide vaccine in patients with unresectable stage III or stage IV melanoma. The trial is described in the paper of Weber et al., J. Clin. Oncol. vol. 31 pp. 4311-4318 (2013). Patients enrolled in the trial had experienced progression after at least one prior therapy, but no prior PD-1 or PD-L1 treatment. The trial consisted of 6 patient cohorts. Cohorts 1-3 enrolled patients who were ipilimumab-naive, while patients in cohorts 4-6 had progressed after prior ipilimumab therapy. Cohorts 1-5 received the peptide vaccine in addition to nivolumab and cohort 6 received nivolumab alone. Cohort 5 enrolled only patients who had experienced grade 3 dose-limiting toxicities on ipilimumab therapy, while patients in cohorts 4 and 6 could only have experienced at most grade 2 dose-limiting ipilimumab toxicities. Cohorts 1-3 differed in the nivolumab dose (1mg / kg, 3 mg / kg or 10 mg / kg). The number of prior treatment regimens was not restricted. All patients had ECOG performance status (PS) 0-1.

[0054] For the purpose of this new classifier development project, the dose of nivolumab and whether or not peptides were given in addition to nivolumab is not considered to be significant.Samples

[0055] The samples available for this study were pretreatment serum samples. Clinical data and spectra were available from 119 patients. Available outcome data included time-to-progression (TTP), overall survival (OS), and response. The baseline clinical characteristics for patients with available spectra from pretreatment samples are listed in table 1. (Lactate dehydrogenase (LDH) is known to have prognostic significance across many cancer types and is used frequently as an important factor in assessing prognosis for patients with melanoma.) Table 1: Baseline characteristics of patients with available spectraN (%)GenderMale72 (61)Female45 (38)NA2 (2)AgeMedian (Range)61 (16-87)Response*CR0 (0)PR31 (26)SD18 (15)PD70 (59)TTPMedian (days)160OSMedian (weeks)94Prior IpiNo31 (26)Yes88 (74)Cohort19 (8)211 (9)311 (9)410 (8)521 (18)657 (48)PD-L1 expression (5% tumor)Positive8 (7)Negative29 (24)NA82 (69)PD-L1 expression (1% tumor)Positive18 (15)Negative19 (16)NA82(69)PD-L1 expression (1% tumor+immune cells)Positive28 (24)Negative7 (6)NA84 (71)VeriStrat-like classificationGood98 (82)Poor21(18)LDH level x< (IU / L)Median (Range)486 (149-4914)> ULN x x< 100 (85)>2ULN31 (26)*subject to further data review; see details in Appendix D of our prior provisional application, x< Not available for one patient, x x< ULN = upper limit of normal range

[0056] Kaplan-Meier plots for time-to-progression (TTP) and overall survival (OS) for the cohort of 119 patients with baseline samples and acquired spectra from pretreatment samples are shown in Figure 1A and 1B, respectively. Note: Of the 14 patients on the plateau of the TTP Kaplan-Meier plot, 3 (21%) had an objective response of SD rather than PR.

[0057] Figures 2A and 2B illustrate Kaplan-Meier plots of time-to-event data for all 119 patients with available clinical data and spectra from pretreatment samples by prior treatment (no prior ipilimumab, i.e., "Ipi-naïve"; prior ipilimumab). Differences in outcome were not statistically significant.

[0058] Figures 3A and 3B are Kaplan-Meier plots of time-to-event data (TTP and OS), respectively, for all 119 patients with available clinical data and spectra from pretreatment samples showing particularly good outcomes for patients in cohort 5, both in absolute terms and as compared to the ipilimumab-naïve patients and other patients with prior ipilimumab treatment. Recall that Cohort 5 in the nivolumab study involved patients that had progressed after or on prior ipilimumab therapy, and enrolled only patients who had experienced grade 3 dose-limiting toxicities on their prior ipilimumab therapy.

[0059] The relatively large number of samples (119) we obtained from the study allowed for a split of the samples into a development set and an internal validation set for classifier development. Two different splits were studied. The first (referred in the following discussion as "DEV1") was stratified by VeriStrat-like classification, response, censoring of TTP and TTP. The second (referred in to the following discussion as "DEV2"), was stratified by cohort, VeriStrat-like classification, response, censoring of TTP and TTP. (By "VeriStrat-like classification" we mean assignment of a "Good" or "Poor" class label for the mass spectra using the classification algorithm and training set for the VeriStrat test described in US Patent 7,736,905). The assignment of individual samples to either the validation set or the development set is listed in Appendix E of our prior provisional application serial no. 62 / 289,587, from which this case claims priority.. Clinical characteristics are listed for the development and validation split in Tables 2A and 2B and comparison of the time-to-event data between development and validation sets is shown in Figures 4A-4D. Table 2A: Baseline characteristics of patients with available spectra split into development and internal validation sets ("DEV1")Development Set (N=60) n(%)Validation Set (N=59) n(%)GenderMale35 (58)37 (63)Female25 (42)20 (34)NA0 (0)2 (3)AgeMedian (Range)61 (23-86)61 (16-87)ResponseCR0 (0)0 (0)PR16 (27)16 (27)SD9 (15)8 (14)PD35 (58)35 (59)TTPMedian (days)162154OSMedian (weeks)9486Cohort16 (10)3 (5)25 (8)6 (10)38 (13)3 (5)46 (10)4 (7)510 (17)11 (19)625 (42)32 (54)Prior ipiyes41 (68)47 (80)no19 (32)12 (20)VS-likegood10 (17)11 (19)classificationpoor50 (83)48 (81) Table 2B: Baseline characteristics of patients with available spectra split into development and internal validation sets ("DEV2") Development Set (N=60) n(%)Validation Set (N=59) n(%)GenderMale36 (60)36 (61)Female23 (38)22 (37)NA1 (2)1 (2)AgeMedian (Range)60 (16-87)62 (34-85)ResponseCR0 (0)0 (0)PR16 (27)15 (25)SD9 (15)9 (15)PD35 (58)35 (59)TTPMedian (days)162132OSMedian (weeks)9489Cohort14(7)5 (8)26 (10)5 (8)36 (10)5 (8)44 (7)6 (10)511 (18)10 (17)629 (48)28 (47)Prior ipiyes44 (73)44 (75)no16 (27)15 (25)VS-likegood11 (18)10 (17)classificationpoor49 (82)49 (83)

[0060] Kaplan-Meier plots for time-to-progression (TTP) and overall survival (OS) for development and validation sets are shown in Figures 4A, 4B, 4C and 4D. In particular, Figures 4A and 4B show the time-to-event data for all 119 patients with available clinical data and spectra from pretreatment samples split into development (N=60) and validation (N=59) sets for the first split ("DEV1"); Figures 4C and 4D shows the time-to-event data for all 119 patients with available clinical data and spectra from pretreatment samples split into development (N=60) and validation (N=59) sets for the second split ("DEV2").Sample Preparation

[0061] 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 serum of five healthy patients, purchased from ProMedDx, "SerumP3") were spotted onto VeriStrat TM cellulose serum cards (Therapak). The cards were allowed to dry for 1 hour at ambient temperature after which the whole serum spot was punched out with a 6mm skin biopsy punch (Acuderm). Each punch was placed in a centrifugal filter with 0.45 µm nylon membrane (VWR). One hundred µl of HPLC grade water (JT Baker) was added to the centrifugal filter containing the punch. The punches were vortexed gently for 10 minutes then spun down at 14,000 rcf for two minutes. The flow-through was removed and transferred back on to the punch for a second round of extraction. For the second round of extraction, the punches were vortexed gently for three minutes then spun down at 14,000 rcf for two minutes. Twenty microliters of the filtrate from each sample was then transferred to a 0.5 ml eppendorf tube for MALDI analysis.

[0062] All subsequent sample preparation steps were carried out in a custom designed humidity and temperature control chamber (Coy Laboratory). The temperature was set to 30 °C and the relative humidity at 10%.

[0063] 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 20µl serum extract and the mix vortexed for 30 sec. The first three aliquots (2 x 2µl) of sample: matrix mix were discarded into the tube cap. Eight aliquots of 2µl sample:matrix mix were then spotted 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.

[0064] This set of samples was processed for MALDI analysis in three batches. QC samples were added to the beginning (two preparations) and end (two preparations) of each batch run.Spectral Acquisition

[0065] MALDI spectra were obtained using a MALDI-TOF mass spectrometer (SimulTOF 100 s / n: LinearBipolar 11.1024.01 from Virgin Instruments, Sudbury, MA, USA). The instrument was set to operate in positive ion mode, with ions generated using a 349 nm, diode-pumped, frequency-tripled Nd:YLF laser operated at a laser repetition rate of 0.5 kHz. External calibration was performed using a mixture of standard proteins (Bruker Daltonics, Germany) consisting of insulin (m / z 5734.51 Da), ubiquitin (m / z, 8565.76 Da), cytochrome C (m / z 12360.97 Da), and myoglobin (m / z 16952.30 Da).

[0066] Spectra from each MALDI spot (8 spots per sample) were collected as 800 shot spectra that were 'hardware averaged' as the laser fires continuously across the spot while the stage is moving at a speed of 0.25 mm / sec. A minimum intensity threshold of 0.01 V was used to discard any 'flat line' spectra. All 800 shot spectra with intensity above this threshold were acquired without any further processing.

[0067] MALDI-TOF mass spectral data acquisition and processing (both for purposes of acquiring a set of data for classifier development and to perform a test on a sample for patient benefit) is optionally performed in accordance with the so-called "Deep MALDI" method described in U.S. Patent 9,279,798 of H. Röder et al. This "798 patent describes the surprising discovery that collecting and averaging large numbers of laser shots (typically 100,000 to 500,000 or more) from the same MALDI spot or from the combination of accumulated spectra from multiple spots of the same sample, leads to a reduction in the relative level of noise vs. signal and that a significant amount of additional spectral information from mass spectrometry of complex biological samples is revealed. The document also demonstrates that it is possible to run hundreds of thousands of shots on a single spot before the protein content on the spot is completely depleted. Second, the reduction of noise via averaging many shots leads to the appearance of previously invisible peaks (i.e., peaks not apparent at spectra resulting from typical 1,000 laser shots). Even previously visible peaks become better defined and this allows for more reliable measurements of peak intensity and comparisons between samples when the sample is subject to a very large number of shots. The classifier of this disclosure takes advantage of the deep MALDI method to look deep into the proteome of serum samples and uses relatively large numbers of peaks (hundreds) for classification which would not be otherwise observable in conventional "dilute and shoot" spectra obtained from the typical ~ 1000 shot mass spectrum. See e.g. the definition of classification feature values listed in Appendix A.

[0068] The following section of this document describes the spectral processing we used on the raw spectra from the mass spectrometer in order to construct a feature table for use in classifier generation. The following procedures are executed in software in a general purpose computer which receives the spectra from the mass spectrometer. Some of the steps, such as for example defining the features used for classification, may be performed in part or in whole by a human operator by inspection of plots of the mass spectral data.Spectral ProcessingRaster spectra preprocessingRescaling

[0069] Instrument calibration can introduce dramatic drifts in the location of peaks (mass (m) / charge (z) = m / z), most apparent in the high mass region, by batch. This results in an inability to consistently use predefined workflows to process the data that rely on the position of peaks and a set tolerance for alignment. To overcome the problem, rescaling of the m / z data can be performed requiring a standard reference spectrum. The standard is compared to spectra from the current batch to identify if there is a shift in the position of common serum peaks. The m / z position is borrowed from the reference and any 'shift' applied to rescale the spectra. The results are rescaled spectra with comparable m / z across batches. In a sense, this is a batch correction procedure for gross alignment issues.Alignment and filtering

[0070] This workflow performs the ripple filter as it was observed that the resulting averages were improved in terms of noise. The spectra are then background subtracted and peaks are found in order to perform alignment. The spectra that are used in averaging are the aligned ripple filtered spectra without any other preprocessing. The calibration step uses a set of 43 alignment points listed below in table 3. Additional filtering parameters required that the spectra have at least 20 peaks and used at least 5 of the alignment points. Table 3: Alignment points used to align the raster spectram / z m / z m / z m / z m / z 31687614942714040282984153793410739144053350041838034109381512767150479282061152715263577386841217315869580288121257217253643389191286418630663189941355521066720291331376323024756393101388228090 Raster averaging

[0071] Averages were created from the pool of rescaled, aligned, and filtered raster spectra. A random selection of 500 spectra was averaged to create a final sample spectrum of 400,000 shots. We collected multiple 800 shot spectra per spot, so that we end up with a pool in excess of 500 in number of 800 shot raster spectra from the 8 spots from each sample. We randomly select 500 from this pool, which we average together to a final 400,000 shot average deep MALDI spectrum.Deep MALDI average spectra preprocessingBackground estimation and subtraction

[0072] Estimation of background was performed with additional consideration for the high mass region. The two window method of background estimation and subtraction was used (table 4). Table 4: Background estimation windowsWide windowsm / ZWidth300080000300008000031000160000Medium windows300050003000050003100010000Details on background subtraction of mass spectra are known in the art and described in prior US patent no. 7,736,905. Normalization by bin method

[0073] A bin method was used to compare clinical groups of interest to ensure that normalization windows are not selected that are useful for classification. The feature definitions used in this analysis are included in Appendix C of our prior provisional application serial no. , from which this case claims priority.. This method compares feature values by clinical group and calculates the coefficient of variance (CV) of the feature for all samples. A threshold is set for p value and for the CV to remove any region that significantly distinguishes the groups of interest or has intrinsic instability (high CV). We used the clinical group comparisons Progressive Disease (PD) vs Stable Disease (SD) and Partial Response (PR), i.e., disease control (DC) no or yes, to calculate univariate p values. A second comparison was added that used PR vs PD and SD. For both, the p value cutoff was set to 0.22. Feature bins with p values less than 0.22 were not included to calculate the normalization scalars. The CV cutoff that was used was 1.0. Features that had CVs above 1.0 were also excluded. By hand, features above 25kDa were removed and features known to be intrinsically unstable (17kDa region) were also removed. A total of 16 bins were identified to include as normalization windows (see table 5). Table 5: Iteration # 1 normalization binsLeftCenterRight3530.6793657.6683784.6583785.0293931.8844078.7394220.214271.6374323.0654875.5814909.7424943.9035260.6355348.0795435.5245436.475559.4515682.4336050.4216213.6146376.8076510.8526555.9666601.0817751.4147825.127898.82610606.1210751.6610897.210908.6111132.5611356.5112425.2712476.2712527.2617710.3518107.5218504.6919212.9219978.3720743.8222108.9522534.0522959.1523738.524238.7724739.04 A second iteration of normalization by bin was performed on the spectra normalized using iteration # 1 bins. The CVs following the first iteration were in general lower. This allowed a new threshold for CV of 0.68 to be set. The p value cutoff was increased to add stringency to the requirements. The same clinical groups were used in the evaluation. The second iteration resulted in 9 windows for inclusion as normalization bins (see table 6). Table 6: Iteration # 2 normalization bins LeftCenterRight4168.2264194.0334219.8394875.5814909.7424943.9034946.1315011.8545077.5765080.9185170.4055259.8925260.6355348.0795435.5246510.8526555.9666601.0817751.4147825.127898.82610606.1210751.6610897.210908.6111132.5611356.51 The resulting scalars using these windows were found for each spectrum and were compared by disease control groups, i.e., scalars for spectra from patients with disease control were compared with spectra for patients with no disease control to ensure that there were no significant differences in scalars depending on clinical group. The plot of normalization scalars shown in Figure 5 reveals that the distribution of the resulting scalars was not significantly different between the clinical groups, and thus the normalization bins were not useful for classification. The spectra were normalized using partial ion current (PIC) over these windows.Average spectra alignment

[0074] The peak alignment of the average spectra is typically very good; however, a fine-tune alignment step was performed to address minor differences in peak positions in the spectra. A set of alignment points was identified and applied to the analysis spectra (Table 7). Table 7: Alignment points used to align the spectral averagesm / Z m / Z m / Z m / Z 33157934140932106941538916151312116844579423158722808447109714160782829350661286817256671506433137661738366311404518631 Feature Definitions

[0075] After performing all of the above pre-processing steps, the process of classifier development proceeded with the identification and definition of features (m / z regions) that are useful for classification. All 119 average spectra were viewed simultaneously by clinical groups PD, SD, and PR to select features for classification. This method protects that features found only in a subset of the spectra are not missed and that feature definitions are broad enough to cover variations of 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 selected features is shown in Figure 6, with the shaded regions representing the m / z region defined for each feature.Batch correction of analysis spectraReference sample "SerumP3" analysis

[0076] Two preparations of the reference sample, SerumP3, were plated at the beginning (1,2) and end (3,4) of each run of samples through the MALDI-TOF mass spectrometer. The purpose of these samples is to ensure that variations by batch due to slight changes in instrument performance (for example, aging of the detector) can be corrected for. To perform batch correction, one spectrum, which is an average of one of the preparations from the beginning and one from the end of the batch, must serve as the reference for the batch. The procedure used for selecting the pair is described first.

[0077] The reference samples were preprocessed as described above. All 351 features (Appendix A) were used to evaluate the possible combinations (1-3, 1-4, 2-3, 2-4). We compared each possible combination of replicates using the function: A = min abs 1 − ftrval 1 / ftrval 2 , abs 1 − ftrval 2 / ftrval 1 where ftrval1 (ftrval2) is the value of a feature for the first (second) replicate of the replicate pair. This quantity A gives a measure of how similar the replicates of the pair are. For each feature, A is reported. If the value is >0.5, then the feature is determined to be discordant, or 'Bad'. A tally of the bad features is reported for each possible combination. If the value of A is <0.1, then the feature is determined to be concordant and reported as 'Good'. A tally of the Good features is reported for each possible combination. Using the tallies of Bad and Good features from each possible combination, we computed the ratio of Bad / Good. The combination with the lowest ratio was reported as the most similar combination, unlikely to contain any systematic or localized outlier behavior in either of the reference spectra. If no ratio can be found that is less than 0.12, then the batch is declared a failure. Table 8 reports the combinations that were found most similar for each batch. Table 8: SerumP3 preparations found to be most similar by batchBatchCombination12_421_432_4 Batch Correction

[0078] Batch 1 was used as the baseline batch to correct all other batches. The reference sample was used to find the correction coefficients for each of the batches 2 and 3 by the following procedure.

[0079] Within each batch j (2 ≤ j ≤ 3), the ratio r ^ i j = A i j A i 1 and the average amplitude A ¯ i j = 1 2 A i j + A i 1 are defined for each i th< feature centered at (m / z) i , where A i j is the average reference spectra amplitude of feature i in the batch being corrected and A i 1 is the reference spectra amplitude of feature i in batch 1 (the reference standard). It is assumed that the ratio of amplitudes between two batches follows the dependence: r A ¯ m / z = a 0 + a 1 ln A ¯ + b 0 + b 1 ln A ¯ m / z + c 0 m / z 2 .

[0080] On a batch to batch basis, a continuous fit is constructed by minimizing the sum of the square residuals, Δ j = ∑ i r ^ i j − r j a 0 a 1 b 0 b 1 c 0 2 , and using the experimental data of the reference sample. The SerumP3 reference samples are used to calculate the correction function. Steps were taken to not include outlier points in order to avoid bias in the parameter estimates. The values of the coefficients a 0 , a 1 , b 0 , b 1 and c 0 , obtained for the different batches are listed in Appendix B (table B.1) of our prior provisional application serial no. 62 / 289,587 from which this case claims priority.. The projection in the r ^ i j versus (m / z) i plane of the points used to construct the fit for each batch of reference spectra, together with the surface defined by the fit itself, is shown in Figure B.1 of Appendix B of our prior provisional application serial no. 62 / 289,587 from which this case claims priority..

[0081] Once the final fit, r j< (A, (m / z)), is determined for each batch, the next step is to correct, for all the samples, all the features (with amplitude A at (m / z)) according to A corr = A r j A ¯ m / z . After this correction, the corrected ( A ¯ i j , m / z i , r ^ i j ) feature values calculated for reference spectra lie around the horizontal line defined by r = 1, as shown in Figure B.1 of Appendix B of our prior provisional application serial no. 62 / 289,587 from which this case claims priority.. Post-correction coefficients are calculated to compare to quality control thresholds. These coefficients can be found in Appendix B table B.2 and the corresponding plots in Figure B.2 of our prior provisional application serial no. 62 / 289,587 from which this case claims priority..

[0082] Using the 351 features and all SerumP3 samples from all batches, a reproducibility assessment was performed on the feature values before and after batch correction. In summary, the median and average CVs were 14.8% and 18.3% before batch correction. Following batch correction, the median and average CVs were 15.0% and 18.2%. As seen in the plots found in Appendix B of our prior provisional application serial no. 62 / 289,587 from which this case claims priority., the batches were very similar requiring little in correction. This is reflected in the lack of improvement in the CVs by feature over all SerumP3 samples.Partial Ion Current (PIC) Normalization

[0083] We have found it advantageous to perform a normalization of spectra before batch correction (see above). However, we have found that after batch correction we can improve the coefficient of variances (CVs) of features and obtain better results if we do another normalization. This second PIC normalization is based on smaller windows around individual peaks that are identified above.

[0084] The spectra were normalized using a partial ion current (PIC) normalization method. Background information on partial ion current normalization is described in the prior US patent 7,736,905. The full feature table was examined to find regions of intrinsic stability to use as the final normalization windows. First, the univariate p values were found by comparing the DC groups by feature. Features with p values less than 0.15 were excluded from the PIC analysis as these features may contribute meaningful information to the test to be developed. A set of 221 features were used in the PIC analysis, of which 30 features were used for the final normalization (table 9). Table 9: Features used for PIC normalizationm / Zm / Zm / Z3243510489023265540390203420619390383554643810637367965891273839536612127864009665713943440966811409848916732141995068707414255 To normalize, the listed features were summed to find the normalization factor for each sample. All feature values were then divided by the normalization factor to arrive at the final feature table used in the subsequent classifier generation method of Figure 8. The normalization factors were examined by DC groups to test that the calculated factors were not significantly correlated. The plot of Figure 7 illustrates the distribution of the factors. The plots for the two groups are very similar, indicating that the normalization scalars are appropriate to use.Classifier Development using Diagnostic Cortex TM

[0085] After the feature table for features in the mass spectra for the 119 samples was created (as explained above) we proceeded to develop a classifier using the classifier generation method shown in flow-chart form in Figure 8. This method, known as "combination of mini-classifiers with drop-out regularization" or "CMC / D", or DIAGNOSTIC CORTEX TM, is described at length in the pending U.S. patent application publication no. 2015 / 0102216 of H. Röder et al. An overview of the methodology will be provided here first, and then illustrated in detail in conjunction with Figure 8 for the generation of the melanoma / nivolumab classifier.

[0086] In contrast to standard applications of machine learning focusing on developing classifiers when large training data sets are available, the big data challenge, in bio-life-sciences the problem setting is different. Here we have the problem that the number (n) of available samples, arising typically from clinical studies, is often limited, and the number of attributes (measurements) (p) per sample usually exceeds the number of samples. Rather than obtaining information from many instances, in these deep data problems one attempts to gain information from a deep description of individual instances. The present methods take advantage of this insight, and are particularly useful, as here, in problems where p >> n.

[0087] The method includes a first step a) of obtaining measurement data for classification from a multitude of samples, i.e., measurement data reflecting some physical property or characteristic of the samples. The data for each of the samples consists of a multitude of feature values, and a class label. In this example, the data takes the form of mass spectrometry data, in the form of feature values (integrated peak intensity values at a multitude of m / z ranges or peaks, see Appendix A) as well as a label indicating some attribute of the sample (for example, patient Early or Late death or disease progression). In this example, the class labels were assigned by a human operator to each of the samples after investigation of the clinical data associated with the sample. The development sample set is then split into a training set and a test set and the training set is used in the following steps b), c) and d).

[0088] The method continues with a step b) of constructing a multitude of individual mini-classifiers using sets of feature values from the samples up to a pre-selected feature set size s (s = integer 1 . . . n). For example a multiple of individual mini- or atomic classifiers could be constructed using a single feature (s = 1), or pairs of features (s = 2), or three of the features (s = 3), or even higher order combinations containing more than 3 features. The selection of a value of s will normally be small enough to allow the code implementing the method to run in a reasonable amount of time, but could be larger in some circumstances or where longer code run-times are acceptable. The selection of a value of s also may be dictated by the number of measurement data values (p) in the data set, and where p is in the hundreds, thousands or even tens of thousands, s will typically be 1, or 2 or possibly 3, depending on the computing resources available. The mini-classifiers execute a supervised learning classification algorithm, such as k-nearest neighbors (kNN), in which the values for a features, pairs or triplets of features of a sample instance are compared to the values of the same feature or features in a training set and the nearest neighbors (e.g., k=9) in an s-dimensional feature space are identified and by majority vote a class label is assigned to the sample instance for each mini-classifier. In practice, there may be thousands of such mini-classifiers depending on the number of features which are used for classification.

[0089] The method continues with a filtering step c), namely testing the performance, for example the accuracy, of each of the individual mini-classifiers to correctly classify the sample, or measuring the individual mini-classifier performance by some other metric (e.g. the difference between the Hazard Ratios (HRs) obtained between groups defined by the classifications of the individual mini-classifier for the training set samples) and retaining only those mini-classifiers whose classification accuracy, predictive power, or other performance metric, exceeds a pre-defined threshold to arrive at a filtered (pruned) set of mini-classifiers. The class label resulting from the classification operation may be compared with the class label for the sample known in advance if the chosen performance metric for mini-classifier filtering is classification accuracy. However, other performance metrics may be used and evaluated using the class labels resulting from the classification operation. Only those mini-classifiers that perform reasonably well under the chosen performance metric for classification are maintained. Alternative supervised classification algorithms could be used, such as linear discriminants, decision trees, probabilistic classification methods, margin-based classifiers like support vector machines, and any other classification method that trains a classifier from a set of labeled training data.

[0090] To overcome the problem of being biased by some univariate feature selection method depending on subset bias, we take a large proportion of all possible features as candidates for mini-classifiers. We then construct all possible kNN classifiers using feature sets up to a pre-selected size (parameter s). This gives us many "mini-classifiers": e.g. if we start with 100 features for each sample (p = 100), we would get 4950 "mini-classifiers" from all different possible combinations of pairs of these features (s = 2), 161,700 mini-classifiers using all possible combination of three features (s = 3), and so forth. Other methods of exploring the space of possible mini-classifiers and features defining them are of course possible and could be used in place of this hierarchical approach. Of course, many of these "mini-classifiers" will have poor performance, and hence in the filtering step c) we only use those "mini-classifiers" that pass predefined criteria. These filtering criteria are chosen dependent on the particular problem: If one has a two-class classification problem, one would select only those mini-classifiers whose classification accuracy exceeds a pre-defined threshold, i.e., are predictive to some reasonable degree. Even with this filtering of "mini-classifiers" we end up with many thousands of "mini-classifier" candidates with performance spanning the whole range from borderline to decent to excellent performance.

[0091] The method continues with step d) of generating a master classifier (MC) by combining the filtered mini-classifiers using a regularized combination method. In one embodiment, this regularized combination method takes the form of repeatedly conducting a logistic training of the filtered set of mini-classifiers to the class labels for the samples. This is done by randomly selecting a small fraction of the filtered mini-classifiers as a result of carrying out an extreme dropout from the filtered set of mini-classifiers (a technique referred to as drop-out regularization herein), and conducting logistical training on such selected mini-classifiers. While similar in spirit to standard classifier combination methods (see e.g. S. Tulyakov et al., Review of Classifier Combination Methods, Studies in Computational Intelligence, Volume 90, 2008, pp. 361-386), we have the particular problem that some "mini-classifiers" could be artificially perfect just by random chance, and hence would dominate the combinations. To avoid this overfitting to particular dominating "mini-classifiers", we generate many logistic training steps by randomly selecting only a small fraction of the "mini-classifiers" for each of these logistic training steps. This is a regularization of the problem in the spirit of dropout as used in deep learning theory. In this case, where we have many mini-classifiers and a small training set we use extreme dropout, where in excess of 99% of filtered mini-classifiers are dropped out in each iteration.

[0092] In more detail, the result of each mini-classifier is one of two values, either "Early" or "Late" in this example. We can then use logistic regression to combine the results of the mini-classifiers in the spirit of a logistic regression by defining the probability of obtaining an "Early" label via standard logistic regression (see e.g. http: / / en.wikipedia.org / wiki / Logistic_regression) P " Early " feature for a spectrum = exp ∑ w mc mini classifiers I mc feature values Normalization where I(mc(feature values)) = 1, if the mini-classifier mc applied to the feature values of a sample returns "Early", and 0 if the mini-classifier returns "Late". The weights w mc for the mini-classifiers are unknown and need to be determined from a regression fit of the above formula for all samples in the training set using +1 for the left hand side of the formula for the Late-labeled samples in the training set, and 0 for the Early-labeled samples, respectively. As we have many more mini-classifiers, and therefore weights, than samples, typically thousands of mini-classifiers and only tens of samples, such a fit will always lead to nearly perfect classification, and can easily be dominated by a mini-classifier that, possibly by random chance, fits the particular problem very well. We do not want our final test to be dominated by a single special mini-classifier which only performs well on this particular set and is unable to generalize well. Hence we designed a method to regularize such behavior: Instead of one overall regression to fit all the weights for all mini-classifiers to the training data at the same time, we use only a few of the mini-classifiers for a regression, but repeat this process many times in generating the master classifier. For example we randomly pick three of the mini-classifiers, perform a regression for their three weights, pick another set of three mini-classifiers, and determine their weights, and repeat this process many times, generating many random picks, i.e. realizations of three mini-classifiers. The final weights defining the master classifier are then the averages of the weights over all such realizations. The number of realizations should be large enough that each mini-classifier is very likely to be picked at least once during the entire process. This approach is similar in spirit to "drop-out" regularization, a method used in the deep learning community to add noise to neural network training to avoid being trapped in local minima of the objective function.

[0093] Other methods for performing the regularized combination method in step (d) that could be used include: Logistic regression with a penalty function like ridge regression (based on Tikhonov regularization, Tikhonov, Andrey Nikolayevich (1943). Doklady Akademii Nauk SSSR 39 (5):195-198.) The Lasso method (Tibshirani (1996). 58 (1):267-288). Neural networks regularized by drop-out (Shrivastava, "Improving Neural Networks with Dropout", Master's Thesis, Graduate Department of Computer Science, University of Toronto), available from the website of the University of Toronto Computer Science department. General regularized neural networks (Girosiet al, Neural Computation, (7), 219 (1995)).

[0094] Our approach of using drop-out regularization has shown promise in avoiding overfitting, and increasing the likelihood of generating generalizable tests, i.e. tests that can be validated in independent sample sets. The performance of the master classifier is then evaluated by how well it classifies the subset of samples forming the test set.

[0095] In step e), steps b)-d) are repeated in the programmed computer for different realizations of the separation of the set of samples into test and training sets, thereby generating a plurality of master classifiers, one for each realization of the separation of the set of samples into training and test sets. The performance of the classifier is evaluated for all the realizations of the separation of the development set of samples into training and test sets. If there are some samples which persistently misclassify when in the test set, the process optionally loops back and steps b), c) and d) and e) are repeated with flipped class labels for such misclassified samples.

[0096] The method continues with step f) of defining a final classifier from one or a combination of more than one of the plurality of master classifiers. In the present example, the final classifier is defined as a majority vote of all the master classifiers resulting from each separation of the sample set into training and test sets, or alternatively by an average probability cutoff.

[0097] Turning now to Figure 8, the classifier development process will be described in further detail in the context of the melanoma / nivolumab classifier.

[0098] The set of 119 samples we had available was initially randomly divided into two subsets, a set of 59 samples to be used for validation of the classifier we generated, and a development set (100) of the remaining 60 samples. This split was performed twice (see the discussion of DEV1 and DEV2 above) with stratification as described previously.

[0099] At step 102, a definition of the two class labels (or groups) for the samples in the development set 100 was performed. While some preliminary approaches used for classifier development employed well-defined class labels, such as response categories, these proved to be unsuccessful. All approaches discussed in this application make use of time-to-event data for classifier training. In this situation, the initial class label definition (Early, Late) is not obvious and, as shown in Figure 8, the process uses an iterative method to refine class labels at the same time as 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 on either TTP or OS and half of the samples with the lowest time-to-event outcome are assigned the "Early" class label (early death or progression, i.e. poor outcome) while the other half are assigned the "Late" class label (late death or progression, i.e. good outcome). A classifier is then constructed using the outcome data and these class labels. This classifier can then be used to generate classifications for all of the development set samples and these are then used as the new class labels for a second iteration of the classifier construction step. This process is iterated until convergence.

[0100] While one could define "Early" and Late" by setting a cutoff based on clinical data, we started with an initial assignment on training labels using TTP (time to progression) data, with 30 patients with lowest TTP assigned the class label "Early", and the 30 patients with highest TTP assigned the class label "Late". It will be noted that later in the process we have a procedure for flipping class labels for samples which persistently misclassify, so these initial class label assignments are not necessarily fixed. After this initial class label definition is arrived at the samples are then assigned to the Early and Late classes based on outcome data as indicated by the two groups 104 and 106 in Figure 8.

[0101] At step 108, the Early and Late samples of the development set (100) are then divided randomly into training (112) and test sets (110), 30 patients each. This division is performed in a stratified manner. Twenty samples from each class were generally assigned to the training set and the remainder to the test set. Occasionally, during the class label flip refinement process, the number of samples in the Early group dropped too low to allow 20 samples to be assigned to the training set and still have a reasonable number of samples (e.g. more than 7 or 8) in the test set. In these cases, a smaller number of samples, for example 17 or 18, were assigned to the training set from each class. In all cases the number of samples assigned to training from each class was the same. The training set (112) is then subject to steps 120, 126 and 130. In step 120, many k-nearest neighbor (kNN) mini-classifiers (mCs) that use the training set as their reference set are constructed (defined) using subsets of features from the 351 mass spectral features identified (see Appendix A). For many of the investigations we performed, all possible single features and pairs of features were examined (s = 2); however, when fewer features were used, triplets were also sometimes considered (s = 3). Although different values of k in the kNN algorithm were tried in preliminary investigations, the approaches described in Example 1 all use k=9. 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 set of 351 features. This was done using the bagged feature selection approach outlined in Appendix F of our prior provisional application serial no. 62 / 289,587 (from which this case claims priority).. Further details on this methodology, its rationale and benefits, are described in the pending U.S. patent application of J. Roder et al., serial no 15 / 091,417 filed April 5, 2016 (Published as US 2016 / 0321561) and in U.S provisional application serial no. 62 / 319,958 filed April 8, 2016 from which this case claims priority.. A reduced, selected list of features for different approaches for classification is listed in Appendix B.

[0102] In step 126 a filtering process was used to select only those mini-classifiers (mC) that had useful or good performance characteristics. This can be understood in Figure 8 by the spectra 124 containing many individual features (shown by the hatched regions) and the features alone and in pairs are indicated in the feature space 122. For some of the kNN mini-classifiers, the features (singly or in pairs) perform well for classification of the samples and such mini-classifiers are retained (indicated by the "+" sign in Figure 8 at 128) whereas others indicated by the "-" sign are not retained.

[0103] To target a final classifier that has certain performance characteristics, these mCs were filtered as follows. Each mC is applied to its training set and performance metrics are calculated from the resulting classifications of the training set. Only mCs that satisfy thresholds on these performance metrics pass filtering to be used further in the process. The mCs that fail filtering are discarded. For this project, accuracy filtering and hazard ratio filtering were used. Sometimes, a simple single filter was used, and sometimes a compound filter was constructed by combining multiple single filters with a logical "AND" operation. For accuracy filtering, the classifier was applied to the training set of samples, or a subset of the training set of samples, and the accuracy of the resulting classification had to lie within a preset range for the mC to pass filtering. For hazard ratio filtering, the classifier was applied to the training set, or a subset thereof. The hazard ratio for a specified outcome (TTP or OS) was then calculated between the group classified as Early and the rest classified as Late. The hazard ratio had to lie within specified bounds for the mC to pass filtering. In this particular classifier exercise, we filtered the mini-classifiers by Hazard ratio (HR) between early / late, and by accuracy for early class label (with TTP defined as < 100 days) and late (with TTP defined as > 365 days). Table 10: Parameters used in mini-classifiers and filteringApproachDepth s (max # features per mC)Development / Validation SplitFilter13DEV1HR on TTP and classification accuracy on TTP < 100 days and TTP > 365 days22DEV1HR on OS32DEV2HR on TTP and classification accuracy on TTP < 100 days and TTP > 365 days42DEV2HR on OS Here, "approach" means different classifier development exercises. In essence, the process of Figure 8 was repeated a number of different times, with each iteration a different value of the depth parameter s was used and different mini-classifier filtering criteria were used in step 126 of Figure 8 as indicated by Table 10. Note that this exercise was done twice for each of the two separations of the entire set of 119 samples into development and validation sets (DEV1 and DEV 2).

[0104] At step 130, we generated a master classifier (MC) for each realization of the separation of the development set into training and test sets at step 108. Once the filtering of the mCs was complete, at step 132 the mCs were combined in one master classifier (MC) using a logistic regression trained using the training set class labels, step 132. See the previous discussion of drop-out regularization. To help avoid overfitting the regression is regularized using extreme drop out with only a small number of the mCs chosen randomly for inclusion in each of the logistic regression iterations. The number of dropout iterations was selected based on the typical number of mCs passing filtering to ensure that each mC was likely to be included within the drop out process multiple times. All approaches outlined in Example 1 left in 10 randomly selected mCs per drop out iteration. Approaches using only single and pairs of features used 10,000 drop out iterations; approaches using single, pairs and triplets of features used 100,000 drop out iterations.

[0105] At step 134, we evaluated the performance of the MC arrived at in step 132 and its ability to classify the test set of samples (110). With each iteration of step 120, 126, 130, 134 we evaluate the performance of the resulting MC on its ability to classify the members of the test set 110.

[0106] After the evaluation step 134, the process looped back via loop 135 to step 108 and the generation of a different realization of the separation of the development set into training and test sets. The process of steps 108, 120, 126, 130, 132, 134 and looping back at 135 to a new separation of the development set into training and test sets (step 108) was performed many times, and in this project six hundred and twenty five different realizations (loops) were used. The methodology of Figure 8 works best when the training set classes have the same number of samples. Hence, if classes had different numbers of members, they were split in different ratios into test and training. The use of multiple training / test splits (loop 135 and the subsequent performance of steps 120, 126, 130, 132 and 134) avoids selection of a single, particularly advantageous or difficult, training set split for classifier creation and avoids bias in performance assessment from testing on a test set that could be especially easy or difficult to classify.

[0107] At step 136, there is an optional procedure of analyzing the data from the training and test splits, and as shown by block 138 obtaining the performance characteristics of the MCs from each training / test set split and their classification results. Optional steps 136 and 138 were not performed in this project.

[0108] At step 144, we determine if there are samples which are persistently misclassified when they are present in the test set 110 during the many iterations of loop 135. If so, we flip the class label of such misclassified samples and loop back in loop 146 to the beginning of the process at step 102 and repeat the methodology shown in Figure 8.

[0109] If at step 144 we do not have samples that persistently misclassify, we then proceed to step 150 and define a final classifier in one of several ways, including (i) a majority vote of each master classifier (MC) for each of the realizations of the separation 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) is a probability of being in one of the two training classes (Early or Late). These MC probabilities can be averaged to yield one average probability for a sample. When working with the development set (100), this approach is adjusted to average over MCs for which a given sample is not included in the training set ("out-of-bag" estimate). These average probabilities can be converted into a binary classification by applying a threshold (cutoff). During the iterative classifier construction and label refinement process, classifications were assigned by majority vote of the individual MC labels obtained with a cutoff of 0.5. This process was modified to incorporate only MCs where the sample was not in the training set for samples in the development set (modified, or "out-of-bag" majority vote). This procedure gives very similar classifications to using a cutoff of 0.5 on the average probabilities across MCs.

[0110] After the final classifier is defined at step 150, the process optionally continues with a validation step 152 in which the final classifier defined at step 150 is tested on an internal validation set of samples, if it is available. In the present example, the initial set of 119 samples was divided into development set (100) and a separate internal validation set, and so this validation set existed and was subject to the validation step 152. Ideally, in step 154 this final classifier as defined at step 150 is also validated on an independent sample set. We describe the validation on an independent set of samples later in this Example 1.Results of Example 1 classifier

[0111] The goal of this exercise was to determine if the final classifier defined in accordance with Figure 8 could demonstrate a separation between the Early and Late class labels, in other words, predict from a mass spectrum of a pre-treatment serum sample whether the patient providing the sample was likely to obtain benefit from administration of nivolumab in treatment of cancer. We achieved this goal, as demonstrated by the classifier performance data in this section of the document. The performance of the classifiers was assessed using Kaplan-Meier plots of TTP and OS between samples classified as Early and Late, together with corresponding hazard ratios (HRs) and log-rank p values. The results for the four approaches of table 10 are summarized in table 11. The table lists the classifier performance for both the Development Set (100 in Figure 8), and the other half of the set of 119 samples, i.e., the Validation Set. Note in table 11 the substantial difference in median OS and TTP between the Early and Late groups as identified by the classifier developed in Figure 8, indicating the ability of the classifier to identify clinically useful groups. Table 11: Final Classifier performance summary for the four approaches of table 10Development Setapproach#Early / #LateOS HR (95% CI)OS log-rank pOS Median (Early, Late)TTP HR (95% CI)TTP log-rank pTTP Median (Early, Late)126 / 340.38 (0.17-0.73)0.00667, not reached (weeks)0.55 (0.28-0.96)0.04083,254 (days)226 / 340.30 (0.13-0.57)0.00165, not reached (weeks)0.53 (0.27-0.94)0.03283,317 (days)325 / 350.32 (0.13-0.58)0.00161, not reached (weeks)0.43 (0.19-0.71)0.00382, 490 (days)428 / 320.36 (0.16-0.71)0.00573, not reached (weeks)0.48 (0.24-0.84)0.01383, 490 (days)Validation Set123 / 360.47 (0.20-0.91)0.02960, 101 (weeks)0.51 (0.25-0.88)0.021132,181 (days)225 / 340.51 (0.23-1.01)0.05373, 101 (weeks)0.46 (0.22-0.76)0.00690,273 (days)316 / 430.49 (0.18-0.95)0.04055, 99 (weeks)0.50 (0.20-0.86)0.02186, 179 (days)418 / 410.43 (0.16-0.80)0.01355, 101 (weeks)0.48 (0.20-0.80)0.01287, 183 (days) Kaplan-Meier plots corresponding to the data in table 11 are shown in Figures 9-12 for the four approaches. The classifications per sample for each approach are listed in Appendix G of our prior provisional application serial no. 62 / 289,587 from which this case claims priority.. Baseline clinical characteristics are summarized by classification group for each of the four approaches in table 12. Table 12: Clinical characteristic by classification group Approach 1Approach 2Approach 3Approach 4Early (N=49)Late (N=70)Early (N=51)Late (N=68)Early (N=41)Late (N=78)Early (N=46)Late (N=73)GenderMale3042314123492745Female1728182716291728AgeMedian (Range)63 (23-86)60 (16-87)61 (23-86)61 (16-87)63 (23-86)60 (16-87)62 (23-86)60 (16-87)ResponsePR922823625724SD414414315414PD3634393132383535Cohort136362727256565656329382938473736473561561551661562631273021362334Prior IpiYes102111209221021No3949404832563652VS-like classificationgood2870306820782573poor210210210210PD-L1 expression (5% tumor)Positive35Negative1415NA3250PD-L1 expression (1% tumor)Positive810Negative910NA3250PD-L1 expression (1% tumor / immune cells)Positive1216Negative43NA3351 The data in tables 11-12 and Figures 9-12 demonstrate that it is possible to build classifiers able to identify patients with better and worse outcomes on nivolumab therapy from mass spectra generated from pre-treatment serum samples. Patients classified as Late have better OS and TTP than patients classified as Early. In addition, patients classified as Late are more likely than patients classified as Early to have a partial response to or stable disease on nivolumab therapy. Table 12 illustrates that the proportions of patients PD-L1 positive is similar in Early and Late classification groups, so the two quantities are not significantly correlated and the serum classifier of Example 1 would provide additional information to any provided by PD-L1 expression levels.

[0112] These results shows a consistency in the ability to identify patients with better and worse outcomes on nivolumab therapy across development / validation set splits, with good generalization between validation and development set performance. Having demonstrated that there is an appreciable, validating performance, to create the most robust classifier, the two mini-classifier filtering methods used were applied to the whole set of 119 samples as the development set 100 of Figure 8. The results are summarized in Table 13 and the Kaplan-Meier plots are shown in Figure 13. (A classifier constructed in accordance with Figure 8 using all 119 samples in the development set as explained here and in Table 13 is referred to as "the full-set classifier of Example 1" later in this document.) Table 13: Performance of the two classifiers built on the whole set of 119 samples as development setApproach#Early / #LateOS HR (95% CI)OS log-rank pOS Median (Early,Late)TTP HR (95% CI)TTP log-rank pTTP Median (Early, Late)2 (compound TTP filtering)47 / 720.42 (0.22-0.63)<0.00161, 113 (weeks)0.55 (0.33-0.80)0.00485, 200 (days)1 (simple OS filtering)47 / 720.38 (0.19-0.55)< 0.00161, not reached (weeks)0.50 (0.29-0.71)0.00184, 230 (days) These two approaches show very similar performance, both being very consistent with the classifiers built on a development / validation split of the 119 samples. This is a further indication that the procedure of Figure 8 produces reliable classifiers with good generalization.

[0113] Clinical characteristics are summarized by the classification groups given by approach 1 (simple OS filtering) in Table 14. There is no hint of association of Early and Late classification with PD-L1 expression with any cutoff or with gender or age. More than 70% of patients in cohort 5 are classified as Late, while patients in cohorts 4 and 6 are split roughly evenly between the two classification groups. LDH is significantly higher in the Early group than in the Late group (Mann-Whitney p value = 0.003). Table 14: Baseline characteristics by classification group for the full-set classifier approach 1Early (N=47)Late (N=72)GenderMale28 (60)44 (61)Female17 (36)28 (39)AgeMedian (Range)61 (23-86)60 (16-87)ResponsePR7 (15)24 (33)SD4 (9)14 (19)PD36 (77)34 (47)Cohort12 (4)7 (10)25 (11)6 (8)33 (6)8 (11)47 (15)3 (4)56 (13)15 (21)624 (51)33 (46)Prior IpiYes10 (21)21 (29)No37 (79)51 (71)VS-like classificationgood26 (55)72 (100)poor21 (45)0 (0)PD-L1 expression (5% tumor)Positive3 (6)5 (7)Negative15 (32)14 (19)NA29 (62)53 (74)PD-L1 expression (1% tumor)Positive9 (19)9 (13)Negative9 (19)10 (14)NA29 (62)53 (74)PD-L1 expression (1% tumor / immune cells)Positive13 (28)15 (21)Negative4 (9)3 (4)NA30 (64)54 (75)LDH level x< Median (Range)655 (417-1130)469 (351-583)(IU / L)>ULN x x< 43 (91)57 (80)>2ULN23 (49)8 (11) x< Missing for one patient, x x< ULN = upper limit of normal range

[0114] Multivariate analysis shows that Early / Late classification remains independently significant when adjusted for other clinical factors. Table 15: Multivariate analysis of OS and TTPOSTTPCovariateHR (95% CI)P valueHR (95% CI)P valueLate vs Early2.86 (1.69-5.00)<0.0012.22 (1.41-3.45)<0.001Male vs Female1.66 (0.96-2.88)0.0691.73 (1.09-2.75)0.020Prior Ipi (no vs yes)0.63 (0.35-1.12)0.1120.70 (0.42-1.16)0.168PD-L1 (5%) -ve / NA vs +ve0.81 (0.27-2.41)0.7041.25 (0.53-2.98)0.613PD-L1 (5%) -ve / +ve vs NA1.00 (0.55-1.81)0.9871.18 (0.69-2.03)0.546LDH (IU / L) / 10001.77 (1.26-2.48)<0.0011.59 (1.16-2.17)0.004 This table shows that the serum test adds information additional to the other available clinical characteristics and that the serum classification label remains a significant predictor of both TTP and OS even when adjusting for other available characteristics, including PD-L1 expression level and LDH level. In particular, it is of note that even though LDH level and test classification are associated with each other, they are both simultaneous independently significant predictors of OS and TTP. In addition, we have examined the dependence of the Early / Late classification on tumor size. Classification was significantly associated with tumor size (Mann-Whitney p < 0.001), with the median tumor size in the Early classification group being 53 cm, compared with 16 cm in the Late classification group. However, while tumor size only showed a trend to significance as a predictor of OS and TTP (p = 0.065 and p=0.07, respectively) in univariate analysis, Early and Late classification retained its highly significant predictive power of OS and TTP when adjusted for tumor size (p<0.001 for OS and p=0.004 for TTP).

[0115] Given the magnitude of the difference in outcomes between the classification groups Early and Late and its independence from other clinical information, the classifiers developed can provide additional information to physicians and patients to inform the decision whether nivolumab (or another anti-PD-1 antibody), or an alternative therapy is an appropriate treatment for the patient. A final classifier which would be considered preferred would be one generated from all the 119 patient samples in the development set with OS filtering (approach 1 in Table 13, Kaplan-Meier plots of Figure 13Cand 13D). This classifier is referred to as the full-set classifier or Example 1 or "IS2" later in this document.Independent validation of the classifier generated using Figure 8 with a second set of samples

[0116] A set of 30 pretreatment samples from patients treated with anti-PD-1 antibodies at Yale University were available as an independent validation cohort.

[0117] Deep MALDI spectra were generated from these samples and processed using identical procedures to those used in classifier development of Figure 8 and described previously. The classifier of "approach 1" for the whole set (table 13 approach 1) was applied to the resulting feature table, yielding a classification of "Early" or "Late" for each sample. Ten samples were classified as "Early" and the remaining 20 as "Late". The Kaplan-Meier plot of overall survival for the cohort is shown in Figure 14 and a summary of the analysis of OS is given in Table 16. As was the case with Figure 13, note the clear separation between the Early and Late groups in Figure 14, indicating the ability of the final classifier of Table 13 and Figure 13 to correctly classify the samples in the independent validation cohort. Table 16: Summary of the performance of the classifier on the Yale anti-PD-1 antibody-treated cohort#Early / #LateOS HR (95% CI)OS log-rank pOS Median (Early,Late)10 / 200.270.0024221, 1471(0.05-0.52)(days)

[0118] The classifier of Example 1 also validated well on an independent cohort of 48 melanoma patients treated with the anti-CTLA4 antibody ipilimumab. See the Example 3 section below.Protein Identification

[0119] Our approach to generating a classifier is based on a correlational analysis relating peak intensities to clinical outcome using the CMC / D process described above in conjunction with Figure 8. As such, the proteins underlying the feature definitions used in the classifications may not be causally related to the outcome to treatment. It is also not trivial to relate peaks measured in a MALDI-TOF experiment to previously identified proteins and their functions. However, by studying the literature of observed serum proteins in MALDI-TOF studies it is still possible to give names to some of the peaks used in our classification. A tentative list includes: c1 inhibitor (4133Da), ITIH4, C1 fragment (4264 Da), Beta-defensin 4A (4381 Da), Leukocyte-specific transcript 1 protein (5867 Da), Leukocyte-specific transcript 1 protein (5889 Da), Gamma sectretase C-terminal fragment 50 of Amyloid beta A4 protein (5911 Da), Granulin-3 (Granulin-B) (5997 Da), C-C motif chemokine 20 (7318 Da), complement C3a (8413 Da), Apolipoprotein C3 (9109 Da), CDC42 small effector protein 2 (9226 Da), B melanoma antigen 3 (10285 Da), Serum Amyloid A (SAA) (11686 Da),C-reactive protein (CRP) (23469 Da).

[0120] While the appearance of acute phase reactants like SAA and CRP was expected, we were surprised to find some evidence of proteins related to the complement system (complement C3a, c1 inhibitor, C1 fragment). The complement system is involved in the immune response and cancer immunotherapy (Markiewski et al., Trends Immunol. 30:286 (2009)), and has recently been suggested to have a role in PD-1 inhibition (Woo et al.,Annual Review of Immunology, 33:445 (2015).

[0121] It appears that at least some part of the classification of patients into Early and Late groups is related to an interplay between acute phase reactants and the activation of the complement system. A more detailed explanation of the relationship between the features for classification, and the biological functions of the Early and Late class labels is set forth later in this document in Example 6.Conclusions on the classifiers we developed in Example 1

[0122] Applying the DIAGNOSTIC CORTEX TM procedure (Figure 8) to pre-treatment serum spectra from melanoma patients obtained using Deep MALDI spectral acquisition we have identified clinical groups "Early" and "Late". These groups are showing significant differences in outcome (both TTP and OS) following treatment with an anti-PD-1 treatment, nivolumab. We have presented a test procedure to identify these groups from pre-treatment samples, and validated the results in internal validation sets, and in an external validation set.

[0123] Patients whose serum classifies as "Early" exhibit significantly faster progression and shorter survival than patients whose serum classifies as "Late" making this test suitable as a biomarker for nivolumab therapy. The clinical groups "Early" and "Late" are not associated with PD-L1 expression, and our classification remains a significant predictor for outcome (both TTP and OS) even when other clinical attributes are included in a multivariate analysis.

[0124] While a correlative approach to test development does not easily lend itself to a deep understanding of the biology underlying the difference between the identified groups, we have done some initial work relating the two different groups to differences in acute phase reactants and the complement system. The success of this project exemplifies the power of the combination of Deep MALDI spectral acquisition and our classifier development methods in the construction of clinically useful tests.Example 2 (reference)Predictive classifier for melanoma patient benefit from immune checkpoint inhibitors

[0125] The classifier of Example 1, including the classifiers built from one half of the sample set and the classifier built using the whole set of samples, showed similar performance and validated well on two independent sample sets. It was discovered that the classifiers of Example 1 also split a cohort of 173 first line, advanced non-small cell lung cancer (NSCLC) patients treated with platinum-doublet + cetuximab, and yet another cohort of 138 ovarian cancer patients treated with platinum-doublet after surgery into groups with better and worse OS and progression-free survival (PFS). For further details on this discovery, see Example 4 below.

[0126] It is possible to argue that the classifiers of Example 1 above have a strong prognostic component, since they seem to stratify patients according to their outcomes regardless of whether treatment was immunotherapy or chemotherapy. While clinically this might or might not matter, Example 2 describes the development of a classifier and test capable of splitting patients treated with immune checkpoint inhibitors according to their outcome, while not stratifying outcomes of patients treated with chemotherapy - i.e., a predictive test between immune checkpoint inhibitors and chemotherapy with less prognostic component.Patient Samples

[0127] The samples we used to develop the classifier of Example 2 were pretreatment serum samples from two different cohorts. The first cohort was the 119 samples from melanoma patients treated with nivolumab, and discussed at length in Example 1, see Table 1 above for the baseline clinical characteristics, Figure 1A and 1B, etc. The set was split into development and validation sets as explained above in Example 1.

[0128] A second cohort, referred to as the ACORN NSCLC cohort, was a set of 173 pretreatment serum samples, clinical data and associated mass spectra from non-small cell lung cancer (NSCLC) patients treated with platinum-doublet plus cetuximab. The purpose of the second cohort was to tune the development of a classifier such that it was (a) predictive for patient benefit on nivolumab but also (b) not outcome predictive in NSCLC patients treated with chemotherapy. Available outcome data included progression-free-survival (PFS) and overall survival (OS). Selected clinical characteristics for patients with available spectra from pretreatment samples are listed in table 18. Table 18 Selected baseline characteristics of NSCLC patients with available spectraN (%)GenderMale108 (62)Female65 (38)AgeMedian (Range)66.4 (35.4 - 86.3)PFSMedian (months)4.3OSMedian (months)9.5 Kaplan-Meier plots for 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 16A and 16B.

[0129] The ACORN NSCLC cohort was split into 3 subsets: 58 samples (referred to as "AddFilt" subset herein) were assigned for additional filtering of mini-classifiers as explained in detail below; 58 samples (the "Test" subset herein) were used for testing the classifier performance together with the melanoma development set samples; and 57 samples (the "Validation" subset herein) were used as part of the internal validation set, in addition to the melanoma validation subset already held for that purpose. A method was implemented in order to choose this split while ensuring the 3 subsets were balanced. The details of how we generated these splits are not particularly important. Suffice it to say that we analyzed mass spectral features which were correlated with overall survival, and chose the subsets such that it minimized the difference in such features across the three subsets. The clinical characteristics are listed for each of the subsets in table 19 and comparison of the time-to-event data is shown in Figures 17A and 17B. Table 19. Baseline characteristics of NSCLC patients with available spectra for each of the 3 created subsets"AddFilt" subset (N=58) n (%)"Test" subset (N=58) n (%)"Validation" subset (N=57) n (%)GenderMale19 (33)22 (38)24 (42)Female39 (67)36 (62)33 (58)AgeMedian (Range)68.3 (35.4 - 86.3)66.8 (42.3 - 85.0)62.0 (44.9 - 76.7)PFSMedian (months)5.04.14.4OSMedian (months)12.47.59.1 Sample preparation

[0130] The sample preparation for the serum samples of both the ACORN NSCLC cohort was the same as the nivolumab cohort and described in detail in Example 1.Spectral acquisition

[0131] The acquisition of mass spectral data for the ACORN NSCLC cohort was the same as the nivolumab cohort and described in detail in Example 1.Spectral processing

[0132] The mass spectral data processing, including raster spectra preprocessing, Deep MALDI average spectra preprocessing (background subtraction, normalization, alignment) and batch correction, was the same as described in Example 1. In the project of Example 2 we did not use feature 9109 in the development of the classifier; however, we used all other 350 features listed in Appendix A.Classifier development

[0133] We used the classifier development process of Figure 8 (described at length in Example 1) in developing the classifier of Example 2.Definition of class labels (step 102, Figure 8A)

[0134] The classifier development of Example 2 makes use of OS data from the melanoma / nivolumab development set for initial assignment of classification group labels to the samples in the development set. In this situation, class labels are not obvious and, as shown in Figure 8, the classifier development process uses an iterative method to refine class labels at the same time as creating the classifier. See Figure 8B, step 144 and loop 146. An initial guess is made for the class labels in the initial iteration of the method, at step 102. The samples were sorted on OS and half of the samples with the lowest time-to-event outcome were assigned the "Early" class label (early death, i.e. poor outcome) while the other half were assigned the "Late" class label (late death, i.e. good outcome). A classifier was then constructed in accordance with Figure 8 using the outcome data and these class labels. This classifier was used to generate classifications for all of the development set samples. The labels of samples that persistently misclassified across the ensemble of master classifiers created for the many splits into training and test sets (loop 146) were flipped. This refined set of class labels were then used as the new class labels for a second iteration of the method at step 102. This process was iterated until convergence, in exactly the same way as for Example 1.Select Training and Test sets (108, Figure 8A)

[0135] The development set samples were split into training and test sets (step 108) in multiple different random realizations. Six hundred and twenty five realizations were used (i.e. iterations through loop 135). The methodology of Figure 8 works best when the training classes Early and Late have the same number of samples in each realization or iteration through loop 135. Hence, if the training classes had different numbers of members, they were split in different ratios into test and training.Creation and Filtering of Mini-Classifiers (120, 126, Figure 8A)

[0136] Many k-nearest neighbor (kNN) mini-classifiers (mCs) that use the training set as their reference set were constructed in step 120 using subsets of features. In this project we used k=9. To be able to consider subsets of single and two features and improve classifier performance, we deselected features that were not useful for classification from the set of 350. This was done using a bagged feature selection approach as outlined in Appendix F of our prior provisional application serial no. 62 / 289,587 from which this case claims priority. and in U.S. patent application of J. Roder et al. serial no. 15 / 091,417.

[0137] To target a final classifier that has certain performance characteristics, these mCs were filtered in step 126 of Figure 8A as follows. Each mC is applied to its training set and to the "AddFilt" NSCLC subset, and performance metrics are calculated from the resulting classification. Only mCs that satisfy thresholds on these performance metrics pass filtering to be used further in the process. The mCs that fail filtering are discarded. For this project hazard ratio based on OS filtering was used. Table 20 shows the filtering criteria at step 126 used for deselecting mini-classifiers in each label-flip iteration (step 146) of the classifier development. In every iteration a compound filter was constructed by combining two single filters with an "AND" operation, see Table 21. Such filtering criteria were designed so that patients treated with immunotherapy would split according to their OS outcomes but patients treated with chemotherapy would not. Table 20: Criteria used in the filtering stepIterationHR range (melanoma samples, training subset)HR range (NSCLC samples, "AddFilt" subset)02.0 - 10.00.9 - 1.11112.0 - 10.00.8 - 1.2522.0 - 10.00.8 - 1.2532.0 - 10.00.8 - 1.2542.0 - 10.00.8 - 1.25 Table 21: Parameters used in mini-classifiers and filtering Depth (parameter s, max # features per mC)Filter2HR on OS (melanoma samples) and HR on OS (NSCLC samples, see Table 20 for HR values. Combination of mini-classifiers using logistic regression with dropout (steps 130, 132)

[0138] Once the filtering of the mCs was complete, the mCs were combined in one master classifier (MC) using a logistic regression trained using the training set class labels. To help avoid overfitting the regression is regularized using extreme drop out with only a small number of the mCs chosen randomly for inclusion in each of the logistic regression iterations. The number of dropout iterations was selected based on the typical number of mCs passing filtering to ensure that each mC was likely to be included within the drop out process multiple times. In all label-flip iterations, ten randomly selected mCs were taken in each of the 10,000 drop out iterations.

[0139] Although the ACORN NSCLC "AddFilt" subset of samples has been used in the filtering step as described above, such samples were not included in the training of the logistic regression that combines the mCs (or the reference sets of the kNN mini-classifiers). Only the training subset drawn from the development set (melanoma samples) was used for that purpose.Training / Test splits (loop 135 of Figure 8A)

[0140] The use of multiple training / test splits in loop 135, and in this Example 625 different training / test set realizations, avoids selection of a single, particularly advantageous or difficult, training set for classifier creation and avoids bias in performance assessment from testing on a test set that could be especially easy or difficult to classify.Definition of final test in step 150

[0141] The output of the logistic regression that defines each MC is a probability of being in one of the two training classes (Early or Late). These MC probabilities over all the 625 training and test set splits can be averaged to yield one average probability for a sample. When working with the development set, this approach is adjusted to average over MCs for which a given sample is not included in the training set ("out-of-bag" estimate). These average probabilities can be converted into a binary classification by applying a threshold (cutoff). During the iterative classifier construction and label refinement process, classifications were assigned by majority vote of the individual MC labels obtained with a cutoff of 0.5. This process was modified to incorporate only MCs where the sample was not in the training set for samples in the development set (modified, or "out-of-bag" majority vote). This procedure gives very similar classifications to using a cutoff of 0.5 on the average probabilities across MCs.

[0142] Results. The performance of the classifiers developed in Example 2 was assessed using Kaplan-Meier plots of OS between samples classified as Early and Late, together with corresponding hazard ratios (HRs) and log-rank p values. This performance estimation was performed separately for melanoma patients treated with nivolumab and for NSCLC samples treated with chemotherapy. The results are summarized in table 22 for the time endpoints available for each sample set. Kaplan-Meier plots corresponding to the data in table 22 are shown in Figures 18A-18H. The plots of Figure 18A-H and the corresponding results of table 22, in the case of the ACORN NSCLC set, did not consider any of the "AddFilt" subset samples since they were directly used in the mC filtering step in all of the 625 training / test realizations. The classifications of all 119 samples in the melanoma cohort are listed in Appendix I of our prior provisional application serial no. 62 / 289,587 from which this case claims priority., together with their classifications obtained from Example 1 for comparison. The features used in the final label flip iteration of this classifier development are given in Appendix C. Table 22: Performance summary#Early / #LateOS HR (95% CI)OS log-rank pOS Median (Early,Late)TTP HR (95% CI)TTP log-rank pTTP Median (Early,Late)Development (melanoma nivo)25 / 350.42 (0.18-0.81)0.01461, 145 (weeks)0.53 (0.27-0.94)0.03484, 285 (days)Validation (melanoma nivo)18 / 410.52 (0.21-1.02)0.05855, 99 (weeks)0.43 (0.17-0.70)0.00480, 183 (days)PFS HR (95% CI)PFS log-rank pPFS Median (Early,Late)Development (NSCLC chemo)35 / 230.60 (0.34-1.07)0.0897.0, 10.1 (months)0.81 (0.48-1.37)0.4343.9, 4.2 (months)Validation (NSCLC chemo)32 / 250.99 (0.54-1.81)0.9629.0, 9.5 (months)0.98 (0.58-1.68)0.9524.2, 5.1 (months) Note that in Figures 18A-18D, for the melanoma / nivolumab cohort there is a clear separation in the OS and TTP curves between the classes Early and Late in both the development and validation sets, whereas in the NSCLC / chemotherapy cohort there is a little or no separation between the Early and Late classes in the development and validation sets. This demonstrates that our use of the NSCLC cohort for filtering mini-classifiers in the development of the Master Classifiers and in the definition of the final classifier was useful in creating a classifier that is predictive of benefit of nivolumab in melanoma but does not stratify outcomes of NSCLC patients treated with chemotherapy.

[0143] Baseline clinical characteristics of the melanoma / nivolumab sample set (development + validation) are summarized by classification group in table 23. Clinical characteristics of the ACORN NSCLC sample set ("Test"+"Validation") are also summarized by classification group in table 24. Table 23: Clinical characteristic by classification group (melanoma sample set)Early (N=43) n (%)Late (N=76) n (%)GenderMale28 (65)44 (58)Female14(33)31(41)NA1 (2)1 (1)AgeMedian (Range)63 (23-87)60.5 (16-79)ResponsePR7 (16)24 (32)SD3 (7)15 (20)PD33 (77)37 (49)Cohort11 (2)8 (11)24 (9)7 (9)33 (7)8 (11)45 (12)5 (7)55 (12)16 (21)625 (58)32 (42)Prior IpiNo8 (19)23 (30)Yes35 (81)53 (70)VS-like classificationgood24 (56)74 (97)poor19 (44)2 (3)PD-L1 expression (5% tumor)Positive3 (7)5 (7)Negative11 (26)18 (24)NA29 (67)53 (70)PD-L1 expression (1% tumor)Positive7 (16)11 (14)Negative7 (16)12 (16)NA29 (67)53 (70)PD-L1 expression (1% tumor / immune cells)Positive10 (23)18 (24)Negative3 (7)4 (5)NA30 (70)54 (71)LDH level x< Median (Range)655 (414-1292)469 (353-583)(IU / L)>ULN x x< 40 (93)60 (80)>2ULN21 (49)10 (13) x< Missing for one patient, x x< ULN = upper limit of normal range Table 24: Clinical characteristic by classification group (ACORN NSCLC sample set) Early (N=67) n (%)Late (N=48) n (%)GenderMale37 (55)32 (67)Female30 (45)16 (33)AgeMedian (Range)66.5 (42.3-85.0)63.4 (46.3-80.7)

[0144] The reader will note that the data in Table 14 differs slightly from the data in Table 23. In particular, the classifications we obtain from the classifier of Example 2 are not exactly the same as the classifications we get from the Example 1 classifier; although not surprisingly a sizeable proportion of samples get the same label for both classifiers. See Appendix I of our prior provisional application serial no. 62 / 289,587 from which this case claims priority. for the labels produced from the Example 1 and Example 2 classifiers for all 119 samples. The reader will also note that Appendix C lists the subset of the 350 features that we used for the final classifier of Example 2.

[0145] The idea behind the reduced features set of Appendix C for the classifier of Example 2 is that instead of just taking all of the mCs (and associated features) that give a prognostic behavior for treatment with nivolumab, in the Example 2 classifier we only use the subset of mCs that, in addition to being useful to predict benefit / nonbenefit from nivolumab treatment, show no separation on the ACORN NSCLC cohort. So, we use a smaller subset of all mCs as we have an additional constraint on their behavior. Then, when we combine these mC, we obtain a different classifier with different sample classifications and different behavior (at least on the ACORN NSCLC set). To get the different behavior on the ACORN set, we have to get some different labels on the melanoma set - but not enough to destroy the nice separation between the Early and Late populations that we had with the Example 1 classifier. A priori it is not clear that it is possible to do this, but the results of Example 2 demonstrate that it is possible to generate such a classifier.

[0146] The results of a multivariate analysis of the melanoma sample set (development + validation) are shown in table 25. Two samples, for which gender was not available, and one sample without LDH level were not considered in such analysis. Table 25: Multivariate analysis of OS and TTP for the melanoma set (Development+Validation)OSTTPCovariateHR (95% CI)P valueHR (95% CI)P valueLate vs Early2.38 (1.37-4.17)0.0022.22 (1.41-3.57)<0.001Male vs Female1.73 (1.00-3.01)0.0521.76 (1.11-2.79)0.017Prior Ipi (no vs yes)0.63 (0.35-1.13)0.1220.66 (0.40-1.11)0.115PD-L1 (5%) -ve vs +ve0.69 (0.23-2.06)0.5061.10 (0.46-2.64)0.825PD-L1 (5%) -ve vs NA0.81 (0.44-1.48)0.4911.04 (0.61-1.79)0.848LDH (IU / L) / 10001.74 (1.25-2.44)0.0011.54 (1.12-2.11)0.009

[0147] Even though the classification of Example 2 is significantly associated with LDH level (table 23), both quantities are independently significant predictors of TTP and OS in multivariate analysis. In addition, analysis of tumor size showed that although strongly associated with Early and Late classification (Mann-Whitney p < 0.001), with the median tumor size being 48 cm in the Early group and 16 cm in the Late group, Early / Late classification retained its significance as a predictor of OS and TTP when adjusted for tumor size (p=0.014 for OS and p=0.005 for TTP). Thus the classification has independent predictive power in addition to other prognostic factors such as LDH and tumor size.Independent validation of Example 2 Classifier

[0148] The developed classifier was applied to several sample sets from two different cancer types (melanoma and ovarian) and therapies (immune checkpoint inhibitors and chemotherapies). For all samples Deep MALDI spectra were generated and processed using identical procedures to those used in development. For each sample a classification of "Early" or "Late" was obtained. For each cohort, Kaplan-Meier plots of the available time endpoints are shown and a summary of the analysis of time-to-event is given.A. Yale anti-PD-1 cohort

[0149] The following results refer to a set of 30 pretreatment samples from patients with advanced unresectable melanoma treated with anti-PD-1 antibodies at Yale University. Figure 19 is a Kaplan-Meier plot of overall survival for the Yale cohort of patients. It shows a clear separation of the survival curves between the two classes Early and Late. The statistics for the results are shown in Table 26. Table 26: Summary of the performance of the classifier on the Yale anti-PD-1 antibody-treated cohort#Early / #LateOS HR (95% CI)OS log-rank pOS Median (Early,Late)7 / 230.37 (0.07-0.89)0.034221, 832 (days) These results show that the classifier of Example 2 generalized well to an independent sample set. The baseline clinical data available for this cohort of patients is summarized in table 27. Table 27 : Baseline characteristics of the cohort n (%)GenderMale19 (63)Female11 (37)AgeMedian (Range)55.5 (26-83)RaceWhite29 (97)Black1 (3)VeriStrat LabelGood24 (80)Poor6 (20)

[0150] We obtained qualitatively similar results for this cohort of anti-PD-1 antibody treated patients for the full-set classifier developed in accordance with Example 1, Table 13, OS filtering.B. Yale Anti-CTLA4 cohort

[0151] The following results refer to a set of 48 pretreatment samples from advanced, unresectable melanoma patients treated with anti-CTLA-4 antibodies (ipilimumab or other similar antibodies) at Yale University. The small amount of baseline clinical data available for this cohort of patients is summarized in table 28. Table 28: Baseline characteristics of the Yale anti-CTLA4 cohortn (%)GenderMale31(65)Female17 (35)RaceWhite48 (100)VeriStrat LabelGood40 (83)Poor8 (17)

[0152] Figure 20 is a Kaplan-Meier plot of overall survival for the Yale anti-CTLA4 cohort by classification produced by the classifier of Example 2. Note the clear separation of the Early and Late classes produced by the classifier of Example 2 on this cohort, indicating the classifier's ability to predict melanoma patient benefit from administration of anti-CTLA4 antibodies. The statistics for the performance of the classifier on this cohort are set forth in table 29. Table 29: Summary of the performance of the classifier on the Yale cohort treated with ipilimumab#Early / #LateOS HR (95% CI)OS log-rank pOS Median (Early, Late)16 / 320.27 (0.05-0.28)< 0.0001156, 782 (days) B. Ovarian Chemotherapy

[0153] The following results refer to a set of 138 pretreatment samples from patients with ovarian cancer treated with platinum-doublet chemotherapy after surgery. Two of these patients have no disease-free survival (DFS) data. The Kaplan-Meier plots for the ovarian cancer cohort by classification produced by the classifier of Example 2 are shown in Figures 21A (overall survival) and Figure 21B (disease free survival). Note that the classifiers did not produce a stratification of the Early and Late classification groups in this cohort. This is consistent with the classifier not stratifying the NSCLC patients treated with chemotherapy. Statistics for the classifier performance shown in Figures 21A and 21B are set forth in Table 30. Table 30: Summary of the performance of the classifier on the cohort of patients with ovarian cancer treated with platinum-doublet chemotherapy after surgery#Early / #LateOS HR (95% CI)OS log-rank pOS Median (Early,Late)DFS HR (95% CI)DFS log-rank pDFS Median (Early, Late)77 / 610.98 (0.61-1.56)0.92241, 41 (months)0.94 (0.60-1.47)0.78725, 27 (months) Reproducibility

[0154] Two sample sets (the melanoma / nivolumab and Yale anti-CTLA4 cohorts) were rerun to assess the reproducibility of the classifier of Example 2. Spectra were acquired in completely separate batches from the original runs. In all cases, the mass spectrometer used for the original mass spectral acquisition and the rerun mass spectral acquisition had been used on other projects of the assignee in the interim. Reproducibility of the classifier labels is summarized in table 31.A. The melanoma / nivolumab cohort

[0155] The classifier was concordant between the original run and rerun in 113 of the 119 samples, for an overall concordance of 95%. Of the 43 samples originally classified as Early, 38 were classified as Early and 5 as Late in the rerun. Of the 76 samples originally classified as Late, 75 were classified as Late and 1 classified as Early in the rerun.B. Yale Anti-CTLA4 cohort

[0156] The reproducibility of the classifier was evaluated only in a subset of 43 samples of the cohort. The classifier was concordant between the original run and rerun in 40 of those 43 samples, for an overall concordance of 93%. Within the subset of 43 samples, of the 13 originally classified as Early, 12 were classified as Early and 1 as Late in the rerun. Of the 30 samples originally classified as Late, 28 were classified as Late and 2 classified as Early in the rerun. Table 31: Reproducibility of the classifications assigned by the predictive classifier across sample setsSample SetLabel concordancenivolumab113 / 119 (95%)anti-CTLA440 / 43 (93%) Example 2 Conclusions

[0157] We 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 not separating patients treated with chemotherapies according to their outcome (OS and PFS or DFS). The classifier was constructed using deep MALDI mass spectra generated from pretreatment serum samples, and was trained using half of the available 119 melanoma samples treated with nivolumab. One third of the 173 available NSCLC samples treated with platinum-doublet plus cetuximab chemotherapy was used to tune the classifier to be predictive and not just prognostic.

[0158] The classifier generalized well by stratifying by outcome two independent cohorts of melanoma patients treated with immune checkpoint inhibitors and not stratifying by outcome a third cohort of ovarian cancer patients treated with post-surgery platinum-doublet chemotherapy.

[0159] The classifier demonstrated acceptable reproducibility on two separate cohorts, with concordance of 93% or higher.Example 3 (reference)Classifier for predicting melanoma patient benefit from anti-CTLA4 antibodies

[0160] Figure 20 and Table 29 above demonstrate that the classifier of Example 2 was able to predict melanoma patient benefit from administration of anti-CTLA4 antibodies. Ipilimumab is a specific example of such a drug. It is known that the majority of the patients in the Yale cohort received ipilimumab. We are not certain that they all received ipilimumab and not some other anti-CTLA4 antibody under development. These results demonstrate that it is possible to predict from a blood-based sample in advance of treatment whether a melanoma patient is likely to benefit from anti-CTLA4 antibody drugs. Note further that this classifier for predicting patient benefit for anti-CTLA4 antibody drugs was developed from a sample set of patients who were treated with nivolumab, which is an anti-PD-1 antibody.

[0161] We further found that the full-set classifier of Example 1 also validated well on this cohort. (The term "full-set classifier of Example 1" means the classifier developed from all 119 patient samples in the nivolumab cohort in accordance with Figure 8 as explained above in Example 1, see discussion of Table 13, Approach 2, with OS mini-classifier filtering.) Figure 23 shows the Kaplan-Meier curves for the Early and Late groups from the full set classifier of Example 1 applied to this cohort. Table 32 is a summary of the performance of the classifier on the Yale anti-CTLA4 antibody-treated cohort for the full-set classifier of Example 1. Table 32: Summary of the performance of the classifier#Early / #LateOS HR (95% CI)OS log-rank pOS Median (Early, Late)20 / 280.33 (0.10-0.47)0.0002156, 804 (days)22, 115 (weeks) Table 33: Baseline characteristics of the cohort by classification group Early (N=20)Late (N=28)n(%)n(%)GenderMale12 (60)19 (68)Female8 (40)9 (32)VeriStrat LabelGood12 (60)28 (100)Poor8 (40)0 (0) A subset of 43 samples from this cohort (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 in the original run (95% concordance). Two samples initially classified as Early were classified as Late on the rerun.

[0162] We also developed a classifier for identification of patients with better and worse outcomes on anti-CTLA4 therapy using pre-treatment serum samples from patients subsequently treated with anti-CTLA4 agents. We will now describe the pertinent aspects of this classifier development exercise and the performance of the classifier developed from this sample set.

[0163] In this classifier development exercise, we had 48 pretreatment serum samples available from a cohort of patients, subsequently treated with anti-CTLA4 agents along with associated clinical data. Most patients are known to have received ipilimumab, but some may have received an alternative anti-CTLA4 antibody. These are the same 48 patients that we ran Example 1 and Example 2 validation tests on which are already described above in Example 3. Overall survival (OS) was the only outcome endpoint available.

[0164] For this classifier development, we used the same spectra and the same spectral processing and features, i.e. identical feature table, as we did for the Examples 1 and 2. The only difference was that feature 9109 was dropped from the feature table, as we are concerned that it has reproducibility issues and little value for classification.

[0165] We used the same Diagnostic Cortex method of Figure 8 in classifier development, as detailed throughout the description of Examples 1 and 2, with label flip iterations. The initial class definitions were based on shorter or longer OS. We used mini-classifier filtering based on hazard ratio for OS between the classification groups of the training set.

[0166] The resulting classifier assigned 16 patients to the Early group and 32 to the Late group. (This compares with 20 in the Early group and 28 in the Late group from the Example 1 full-set classifier.) The Kaplan-Meier plot of classifier performance is shown in Figure 23A for the groups defined by the new classifier trained on the anti-CTLA4 cohort. Note the clear separation in the overall survival between the Early and Late groups in the plot of Figure 23A. None of the patients in the Early group survived more than 3 years. The statistics for the classifier performance are as follows: #Early / #LateHR (95% CI)log rank pMedian OS16 / 320.24 (0.04-0.23)<0.0001Early: 155 days, Late: 804 days A comparison of this result with the classifier performance we obtained for the Example 1 full-set classifier operating on the Yale cohort is shown in Figure 23B. Note in Figure 23B the almost perfect overlap between the Early and Late groups in the classifiers developed from the melanoma / nivolumab sample set ("example 1 full-set classifier" and the classifiers developed from the Yale anti-CTLA4 cohort ("Early" and "Late").

[0167] So, comparing the Kaplan-Meier plots of Figure 23 and the statistics of the two classifiers, the classifier performance results do not seem to be significantly better with the classifier developed on the anti-CTLA4 cohort than with the classifier developed on the nivolumab-treated cohort. The fact that the two classifiers produce quite similar results using completely different development sets supports our assertions that the Diagnostic Cortex method of Figure 8 produces tests that do not overfit to development sample sets, but rather extract the information that will generalize to other sample sets.

[0168] As described herein, the testing method of making a prediction of whether a melanoma patient is likely to benefit from anti-CTLA4 antibody drug involves the following steps: (a) conducting mass spectrometry on a blood-based sample of the patient and obtaining mass spectrometry data; (b) obtaining integrated intensity values in the mass spectrometry data of a multitude of pre-determined mass-spectral features (such as for example the features listed in Appendix A or some subset thereof, e.g., after a deselection of noisy features that do not significantly contribute to classifier performance such as the features of one of the sets of Appendix B or Appendix C); and (c) operating on the mass spectral data with a programmed computer implementing a classifier (e.g., a classifier generated in accordance with Figure 8 as explained in Example 2 or the full-set classifier of Example 1).

[0169] In the operating step the classifier compares the integrated intensity values with feature values of a training set of class-labeled mass spectral data obtained from blood-based samples obtained from a multitude of other melanoma patients. This training set could be of one of two types, namely class labeled mass-spectral data from a set of samples from patients treated with an antibody drug blocking ligand activation of programmed cell death 1 (PD-1), e.g., nivolumab. Alternatively, this training set could be class-labeled mass spectral data from a set of samples from patients treated with an anti-CTLA4 antibody. The classifier performs this comparison with a classification algorithm. The classifier generates a class label for the sample, wherein the class label "early" or the equivalent predicts the patient is likely to obtain relatively less benefit from the anti-CTLA4 antibody drug and the class label "late" or the equivalent indicates the patient is likely to obtain relatively greater benefit from the anti-CTLA4 antibody drug.

[0170] Additionally, preferably the mass spectral data is acquired from at least 100,000 laser shots performed on the blood-based sample using MALDI-TOF mass spectrometry.

[0171] The classifier is obtained from a combination of filtered mini-classifiers using a regularized combination method. The mini-classifiers may be filtered as explained in Example 2 by retaining mini-classifiers defined during classifier generation so that mass spectra from patients treated with immunotherapy would split according to their OS outcomes, but when such mini-classifiers are applied to mass spectra from patients treated with chemotherapy they would not split. This is considered optional, as the "full-set" classifier of Example 1, which can be used in this test, was developed without using any filtering of mini-classifiers on a set of samples from a chemotherapy cohort.

[0172] A testing environment for conducting the test of Example 3 on a blood-based sample of a melanoma patient can take the form of the system shown in Figure 15 and described in detail below in Example 5.Example 4Classifier for predicting better or worse survival in ovarian and NSCLC patients treated with chemotherapy

[0173] We discovered that the classifier of Example 1 split a cohort of 173 first line, advanced non-small cell lung cancer (NSCLC) patients treated with platinum-doublet + cetuximab, and yet another cohort of 138 ovarian cancer patients treated with platinum-doublet after surgery into groups with better and worse OS and progression-free (or disease-free) survival (PFS or DFS). Practical tests for predicting better or worse survival in ovarian and NSCLC patients will be described in this section. Further details on the ACORN NSCLC and ovarian cancer cohorts will also be described in this Example 4. Note that the classifier of Example 2 (with the use of the ACORN NSCLC cohort for mini-classifier filtering) does not identify those ovarian and NSCLC patients which are likely to benefit from chemotherapy, hence the classifier and test for predicting better or worse survival in ovarian and NSCLC on platinum chemotherapy is constructed in accordance with Example 1.ACORN NSCLC cohort

[0174] A set of 173 pretreatment blood-based samples from patients with previously untreated advanced non-small cell lung cancer (NSCLC) were available. Patients received platinum-based chemotherapy with cetuximab as part of a clinical trial. The most important baseline clinical data available for this cohort of patients are summarized in table 34 and OS and progression-free survival (PFS) for the whole cohort is shown in Figures 24A and 24B, respectively. Table 34: Baseline characteristics of the ACORN NSCLC cohortn (%)GenderMale108 (62)Female65 (38)RaceWhite133 (77)Black23 (13)Other17 (10)Histologysquamous63 (36)non-squamous110 (64)VeriStrat LabelGood122 (71)Poor51 (29)Performance Status061 (35)1112 (65)Disease StageIIIB8 (5)IV165 (95)TreatmentCarboplatin / Paclitaxel / Cetuximab68 (39)Carbo- or Cisplatin / Gemcitabine / Cetuximab69 (40)Carbo- or Cisplatin / Pemetrexed / Cetuximab36 (21)AgeMedian (range)66 (35-86)

[0175] Deep MALDI spectra had been generated from these samples for a prior project, but in an identical way to that described in Example 1, and these were processed using identical procedures to those used in development of the classifiers of Example 1. The full-set classifier of Example 1 was applied to the resulting feature table, yielding a classification of "Early" or "Late" for each sample. One hundred sixteen samples were classified as Early and the remaining 57 as Late. The Kaplan-Meier plot of overall survival for the cohort by classification group is shown in Figures 25A and 25B and a summary of the analysis of OS and PPFS is given in table 35. Patient baseline characteristics are summarized by classification group in table 36. Table 35: Summary of the performance of the classifier on the ACORN NSCLC cohort#Early / #LateEndpointHR (95% CI)log-rank pMedian (Early,Late)116 / 57OS0.36 (0.28-0.55)<0.00017.0, 20.1 (months)116 / 57PFS0.60 (0.44-0.82)0.00173.9, 5.9 (months) Table 36: Baseline characteristics by classification of full-set classifier of Example 1 EarlyLaten (%)n (%)GenderMale74 (64)34 (60)Female42 (36)23 (40)RaceWhite86 (74)47 (82)Black19 (16)4 (7)Other11 (9)6(11)Histologysquamous45 (39)18 (32)non-squamous71 (61)39 (68)VeriStrat LabelGood65 (56)57 (100)Poor51 (44)0 (0)Performance Status034 (29)27 (47)182 (71)30 (53)Disease StageIIIB5 (4)3 (5)IV111 (96)54 (95)TreatmentCarboplatin / Paclitaxel / Cetuximab46 (40)22 (39)Carbo- or Cisplatin / Gemcitabine / Cetuximab46 (40)23 (40)Carbo- or Cisplatin / Pemetrexed / Cetuximab24 (21)12 (21)AgeMedian (range)66.5 (35-85)64 (46-86) Figures 26A and 26B shows the time-to-event outcomes broken down by classification and VeriStrat label (using the classification algorithm, feature definitions, and NSCLC training set described in U.S. Patent 7,736,905). It is apparent that both VeriStrat groups within the Early classification group have similarly poor outcomes.Ovarian cancer cohort

[0176] A set of 165 samples from an observation trial of patients with ovarian cancer were available. Patients underwent surgery followed by platinum-based chemotherapy. Samples were taken at the time of surgery. Of the 165 patients, 23 did actually not start chemotherapy, were not newly diagnosed, or had received prior therapy for ovarian cancer. Outcome data was not available for an additional four patients. Data are presented here for the remaining 138 patients. The most important baseline clinical data available for these patients are summarized in table 37 and OS and disease-free survival (DFS) are shown in Figures 27A and 27B. Note, two patients of the 138 did not have DFS available. Table 37: Baseline characteristics of the ovarian cohortn (%)Histologyserous100 (72)non-serous38 (28)VeriStrat LabelGood110 (80)Poor27 (20)Indeterminate1 (1)FIGO113 (9)23 (2)354 (39)429 (21)NA39 (28)Histologic GradeNA2 (1)17 (5)253 (38)376 (55)Metastatic Diseaseyes20 (14)no118 (86)AgeMedian (range)59 (18-88)

[0177] Deep MALDI spectra had been generated from these samples for a prior project, in an identical manner as outlined in Example 1, and these were processed using identical procedures to those used in development of the nivolumab test described in Example 1. The Example 1 full-set classifier was applied to the resulting feature table, yielding a classification of "Early" or "Late" for each sample. Seventy six samples were classified as Early and the remaining 62 as Late. The Kaplan-Meier plots of overall and disease-free survival by classification are shown in Figures 28A and 28B, and a summary of the analysis of OS and PFS is given in table 38. Patient baseline characteristics are summarized by classification group in table 39. Table 38: Summary of the performance of the full-set classifier of Example 1 on the ovarian cancer cohort#Early / #LateEndpointHR (95% CI)log-rank pMedian (Early, Late)76 / 62OS0.48 (0.30-0.76)0.002135, not reached (months)74 / 62PFS0.48 (0.30-0.73)0.001117, not reached (months) Table 39: Baseline characteristics by classification produced from full-set classifier of Example 1 EarlyLaten(%)n(%)Histologyserous61 (80)39 (63)non-serous15 (20)23 (37)VeriStrat LabelGood48 (63)62 (100)Poor27 (36)0 (0)Indeterminate1 (1)0 (0)FIGO12 (3)11 (18)22 (3)1 (2)332(42)22(35)422 (29)7 (11)NA18 (24)21 (34)Histologic GradeNA0 (0)2 (3)11 (1)6 (10)232 (42)21 (34)343 (57)33 (53)Metastatic Diseaseyes15 (20)5 (8)no61 (80)57 (92)AgeMedian (range)60 (35-88)57.5 (18-83)

[0178] Figures 29A and 29B show the time-to-event outcomes broken down by classification and VeriStrat label (testing in accordance with U.S. Patent 7,736,905). It is apparent that both VeriStrat groups Good and Poor within the Early classification produced by the full-set classifier of Example 1 have similarly poor outcomes.

[0179] In summary, a testing method of making a prediction of whether an ovarian or NSCLC patient is likely to benefit from chemotherapy, e.g., platinum doublet chemotherapy, involves the following steps: a) conducting mass spectrometry on a blood-based sample of the patient and obtaining mass spectrometry data; (b) obtaining integrated intensity values in the mass spectrometry data of a multitude of pre-determined mass-spectral features (such as for examples the features listed in Appendix A or some subset thereof, e.g., after a deselection of noisy features that do not significantly contribute to classifier performance such as the features of one of the sets of Appendix B); and (c) operating on the mass spectral data with a programmed computer implementing a classifier (e.g., a classifier generated in accordance with Figure 8, the full-set classifier of Example 1).

[0180] In the operating step the classifier compares the integrated intensity values with feature values of a training set of class-labeled mass spectral data obtained from blood-based samples obtained from a multitude of melanoma patients treated with an antibody drug blocking ligand activation of programmed cell death 1 (PD-1), e.g., nivolumab, with a classification algorithm. The classifier generates a class label for the sample, wherein the class label "early" or the equivalent predicts the patient is likely to obtain relatively less benefit and / or have worse outcome from the chemotherapy and the class label "late" or the equivalent indicates the patient is likely to obtain relatively greater benefit and / or have better outcome from the chemotherapy. In one embodiment the chemotherapy is platinum-doublet chemotherapy.

[0181] Additionally, preferably the mass spectral data is acquired from at least 100,000 laser shots performed on the blood-based sample using MALDI-TOF mass spectrometry.

[0182] The classifier is preferably obtained from a combination of filtered mini-classifiers using a regularized combination method.

[0183] A practical testing environment for conducting the test on ovarian and NSCLC cancer patients is described in Figure 15 and the following section.EXAMPLE 5Laboratory test center and computer configured as classifier

[0184] Once the classifier as described in conjunction with Examples 1, 2, 3, 9 (or the other Examples) has been developed, its parameters and reference set can now be stored and implemented in a general purpose computer and used to generate a class label for a blood-based sample, e.g., in accordance with the tests described in Examples 1, 2, 3 and 4. Depending on the particular clinical question being asked (and the type of patient the sample is obtained from), the class label can predict in advance whether a melanoma or other cancer patient is likely to benefit from immune checkpoint inhibitors such as antibodies blocking ligand activation of PD-1, such as nivolumab, or for example predict melanoma or other cancer patient benefit from antibodies blocking CTLA4, or if developed in accordance with Example 1, predict whether an ovarian or NSCLC cancer patient is likely to have better or worse overall survival on chemotherapy.

[0185] Figure 15 is an illustration of a laboratory testing center or system for processing a test sample (in this example, a blood-based sample from a melanoma, ovarian or NSCLC patient) using a classifier generated in accordance with Figure 8. The system includes a mass spectrometer 1506 and a general purpose computer 1510 having CPU 1512 implementing a classifier 1520 coded as machine-readable instructions and a memory 1514 storing reference mass spectral data set including a feature table 1522 of class-labeled mass spectrometry data. This reference mass spectral data set forming the feature table 1522 will be understood to be the mass spectral data (integrated intensity values of predefined features, see Appendix A or Appendix B), associated with a development sample set to create the classifier of Figure 8 and Examples 1-4. This data set could be from all the samples, e.g., for the full-set classifier of Example 1 or a subset of the samples (e.g., development set of one half the samples) plus a set of mass spectral data from NSCLC patients used to develop the classifier of Example 2. It will be appreciated that the mass spectrometer 1506 and computer 1510 of Figure 15 could be used to generate the classifier 1520 in accordance with the process of Figure 8.

[0186] The operation of the system of Figure 15 will be described in the context of conducting a predictive test for predicting patient benefit or non-benefit from antibodies blocking ligand activation of PD-1 as explained above. The following discussion assumes that the classifier 1520 is already generated at the time of use of the classifier to generate a class label (Early or Late, or the equivalent) for a test sample. The method of operation of Figure 15 for the other tests (benefit from anti-CTLA4 drugs, overall survival prediction on chemotherapy in ovarian and NSCLC, etc.) is the same.

[0187] The system of Figure 15 obtains a multitude of samples 1500, e.g., blood-based samples (serum or plasma) from diverse cancer (e.g., melanoma) patients and generates a class label for the sample as a fee-for-service. The samples 1500 are used by the classifier 1520 (implemented in the computer 1510) to make predictions as to whether the patient providing a particular sample is likely or not likely to benefit from immune checkpoint inhibitor therapy. The outcome of the test is a binary class label such as Early or Late or the like which is assigned to the patient blood-based sample. The particular moniker for the class label is not particularly important and could be generic such as "class 1", "class 2" or the like, but as noted earlier the class label is associated with some clinical attribute relevant to the question being answered by the classifier. As noted earlier, in the present context the Early class label is associated with a prediction of relatively poor overall survival, and the Late class label is associated with a prediction of relatively better (longer) overall survival on the immune checkpoint inhibitor.

[0188] The samples may be obtained on serum cards or the like in which the blood-based sample is blotted onto a cellulose or other type card. Aliquots of the sample are spotted onto one or several spots of a MALDI-TOF sample "plate" 1502 and the plate inserted into a MALDI-TOF mass spectrometer 1506. The mass spectrometer 1506 acquires mass spectra 1508 from each of the spots of the sample. The mass spectra are represented in digital form and supplied to a programmed general purpose computer 1510. The computer 1510 includes a central processing unit 1512 executing programmed instructions. The memory 1514 stores the data representing the mass spectra 1508. Ideally, the sample preparation, spotting and mass spectrometry steps are the same as those used to generate the classifier in accordance with Figure 8 and Examples 1 and 2.

[0189] The memory 1514 also stores a data set representing classifier 1520, which includes a) a reference mass spectral data set 1522 in the form of a feature table of N class-labeled spectra, where N is some integer number, in this example a development sample set of spectra used to develop the classifier as explained above or some sub-set of the development sample set (e.g., DEV1 or DEV2 above in Example 1, or all of the 119 samples). The classifier 1520 includes b) code 1524 representing a kNN classification algorithm (which is implemented in the mini-classifiers as explained above), including the features and depth of the kNN algorithm (parameter s) and identification of all the mini-classifiers passing filtering, c) program code 1526 for executing the final classifier generated in accordance with Figure 8 on the mass spectra of patients, including logistic regression weights and data representing master classifier(s) forming the final classifier, including probability cutoff parameter, mini-classifier parameters for each mini-classifier that passed filtering, etc., and d) a data structure 1528 for storing classification results, including a final class label for the test sample. The memory 1514 also stores program code 1530 for implementing the processing shown at 1550, including code (not shown) for acquiring the mass spectral data from the mass spectrometer in step 1552; a preprocessing routine 1532 for implementing the background subtraction, normalization and alignment step 1554 (details explained above), filtering and averaging of the 800 shot spectra at multiple locations per spot and over multiple MALDI spots to make a single 100,000 + shot average spectrum (as explained above) a module (not shown) for calculating integrated intensity values at predefined m / z positions in the background subtracted, normalized and aligned spectrum (step 1556), and a code routine 1538 for implementing the final classifier 1520 using the reference dataset feature table 1522 on the values obtained at step 1556. The process 1558 produces a class label at step 1560. The module 1540 reports the class label as indicated at 1560 (i.e., "Early" or "Late" or the equivalent).

[0190] The program code 1530 can include additional and optional modules, for example a feature correction function code 1536 (described in U.S. patent application publication 2015 / 0102216) for correcting fluctuations in performance of the mass spectrometer, a set of routines for processing the spectrum from a reference sample to define a feature correction function, a module storing feature dependent noise characteristics and generating noisy feature value realizations and classifying such noisy feature value realizations, modules storing statistical algorithms for obtaining statistical data on the performance of the classifier on the noisy feature value realizations, or modules to combine class labels defined from multiple individual replicate testing of a sample to produce a single class label for that sample. Still other optional software modules could be included as will be apparent to persons skilled in the art.EXAMPLE 6 (reference)Correlation of Protein Functional Groups with Classification Groups and Mass Spectral Features

[0191] When building tests using the procedure of Figure 8, it is not essential to be able to identify which proteins correspond to which mass spectral features in the MALDI TOF spectrum or to understand the function of proteins correlated with these features. Whether the process produces a useful classifier depends entirely on classifier performance on the development set and how well the classifier performs when classifying new sample sets. However, once a classifier has been developed it may be of interest to investigate the proteins or function of proteins which directly contribute to, or are correlated with, the mass spectral features used in the classifier. In addition, it may be informative to explore protein expression or function of proteins, measured by other platforms, that are correlated with the test classification groups.

[0192] Appendix K to our prior provisional application serial no. 62 / 289,587 from which this case claims priority. sets forth the results of an analysis aimed at associating protein function with classification groups of the Example 1 and Example 2 classifiers and the mass spectral features measured in Deep MALDI spectra. In Appendix K, the nomenclature "IS2" corresponds to the full-set approach 1 classifier of Example 1, and "IS4" corresponds to the classifier of Example 2. A summary of the pertinent details of the methods we used and the results we found are set forth in this Example. The discoveries we made lead to new examples of how the classifiers of this disclosure can be characterized and generalization of the classifiers to other immune checkpoint inhibitors and other cancer indications beyond melanoma.

[0193] The data we used for the study include the feature table created during the application of the Example 1 full-set approach 1 classifier on a set of 49 serum samples ("Analysis Set") composed of patients with cancer and some donors without cancer. This is a table of feature values for each of the 59 features used in the Example 1 whole set approach 1 classifier (see Appendix B) and 292 other features (see Appendix A) not used in the Example 1 full-set approach 1 classifier. The feature values were obtained from MALDI-TOF mass spectra obtained using the fully specified spectral acquisition and spectral processing processes defined in the Example 1 description for each of the 49 samples. We also used the list of classifications (Early / Late) obtained for the 49 samples in the Analysis Set produced by the Example 1 full-set approach 1 classifier. We also used the list of classifications (Early / Late) obtained for the 49 samples in the Analysis Set produced by the Example 2 classifier. We also used a table of 1129 protein / peptide expression measurements obtained from running a SomaLogic 1129 protein / peptide panel on the Analysis Set.

[0194] We used a method known as Gene Set Enrichment Analysis (GSEA) applied to protein expression data. Background information on this method is set forth in Mootha et al.,Nat Genet. 2003; 34(3):267 and Subramanian et al., PNAS USA 2005; 102(43):15545. Specific protein sets were created based on the intersection of the list of SomaLogic 1129 panel targets and results of queries from GeneOntology / AmiGO2 and UniProt databases.

[0195] The implementation of the GSEA method was performed using Matlab. Basically, in our method we evaluated the correlation r between individual proteins and group labels (Early and Late). Once these correlations had been calculated for each protein, we ranked the proteins by value of r from largest to smallest, with larger values of r indicating greater correlation, and a value of r = 0 meaning no correlation was found. We then calculated an enrichment score (ES) (as explained in the Subramanian et al. paper above), which is designed to reflect the degree to which elements of a particular protein set are over-represented at the top or bottom of the ranked list of proteins. We considered two possible definitions for the enrichment score, the details of which are set forth at page 5 of Appendix K of our prior provisional application serial no. 62 / 289,587 from which this case claims priority.. We also calculated the corresponding p value for the proteins, in order to assess the significance of the deviation of the calculated enrichment score from its average value for a random distribution. We also calculated a running sum (RS), as part of the calculation of the enrichment score, the details of which are explained in Appendix K of our prior provisional application serial no. 62 / 289,587 from which this case claims priority..

[0196] The results for the correlation of the protein sets with the Example 1 full-set approach 1 classifier class labels (Early or Late) of the 49 samples are shown in table 40. Table 40: Results of GSEA applied to protein sets and Example 1 class labels of the Analysis SetDefinition 1 of enrichment scoreDefinition 2 of enrichment scoreES definition 1p valueES definition2p valueAcute inflammatory response0.4240.020.4800.03Activation of innate immune response0.4120.550.5180.56Regulation of adaptive immune response-0.2340.900.3380.95Positive regulation of glycolytic process-0.4950.290.6730.21Immune T-cells-0.1560.970.2740.96Immune B-cells0.2130.910.3120.95Cell cycle regulation-0.2070.810.3710.50Natural killer regulation-0.4060.390.4290.68Complement system0.5520.010.5650.02Acute response0.5390.100.7000.02Cytokine activity-0.23 10.680.3420.74Wound healing-0.3730.110.4760.11Interferon-0.1780.940.3300.84Interleukin-100.1900.770.3320.64Growth factor receptor signaling-0.2210.450.3090.84Immune Response Type 1-0.4020.560.5060.71Immune Response Type 20.5110.480.5520.84Acute phase0.5720.010.693<0.01Hypoxia-0.2470.650.3630.71Cancer0.1530.960.2980.80 There are correlations at the p < 0.05 level of the class labels with the protein sets corresponding to the following biological processes: acute inflammatory response, acute phase, and complement system. Correlations with p values around 0.1 were found for the wound healing protein set. The correlation for the acute response has a p value of 0.02. We then used statistical methods (see Appendix K of our prior provisional application serial no. 62 / 289,587 from which this case claims priority.) for identifying subsets of proteins of the complement system, acute phase, acute response, and acute inflammatory response protein sets that are most important for these correlations, the results of which are shown in Tables 41A, 41B, 41C and 41D, respectively. Table 41A: Proteins included in the extended leading edge set for complement (Amigo9). * indicates proteins to the right of the minimum of RS and † indicates proteins with anti-correlations of at least as great magnitude as that at the maximum of RS. UniProtIDProtein NameCorrelationP valueP01024Complement C3b0.626<0.01P02741C-reactive protein0.582<0.01P02748Complement C90.559<0.01P01024Complement C3a anaphylatoxin0.556<0.01P01024Complement C30.525<0.01P11226Mannose-binding protein C0.461<0.01P06681Complement C20.4180.01P12956ATP-dependent DNA helicase II 70 kDa subunit0.4120.01P01031Complement C5a0.3940.02P02743Serum amyloid P0.3910.02P07357P07358P07360Complement C80.3770.03P01031P13671Complement C5b,6 Complex0.3500.04P01031Complement C50.3370.05P00751Complement factor B0.3230.05P01024Complement C3b, inactivated0.3230.05P05155C1-Esterase Inhibitor0.3130.06P00736Complement C1r0.3100.07P13671Complement C60.3100.07P48740Mannan-binding lectin serine peptidase 10.2830.09P16109P-Selectin-0.475*†<0.01Q6YHK3CD109-0.364†0.03 Table 41B: Proteins included in the extended leading edge set for acute phase (UniProt1). * indicates proteins to the right of the minimum of RS and † indicates proteins with anti-correlations of at least as great magnitude as that at the maximum of RS. UniProtIDProtein NameCorrelationP valueP01009alpha1-Antitrypsin0.801<0.01P0DJI8Serum amyloid A0.704<0.01P18428Lipopolysaccharide-binding protein0.640<0.01Q14624Inter-alpha-trypsin inhibitor heavy chain H40.603<0.01P02741C-reactive protein0.582<0.01P02671P02675P02679D-dimer0.529<0.01P11226Mannose-binding protein C0.461<0.01P00738Haptoglobin0.455<0.01P02743Serum amyloid P0.3910.02P02765alpha2-HS-Glycoprotein-0.593*t<0.01P02787Transferrin-0.502*t<0.01P08697alpha2-Antiplasmin-0.347*†0.04P08887Interleukin-6 receptor alpha chain-0.347*†0.04 Table 41C: Proteins included in the extended leading edge set for acute response (Amigo11). * indicates proteins to the right of the minimum of RS and † indicates proteins with anti-correlations of at least as great magnitude as that at the maximum of RS. UniProtIDProtein NameCorrelationP valueP18428Lipopolysaccharide-binding protein0.640<0.01Q14624Inter-alpha-trypsin inhibitor heavy chain H40.603<0.01P05155C1-Esterase Inhibitor0.3130.06P13726Tissue Factor0.3060.07P48740Mannan-binding lectin serine peptidase 10.2830.09P05231Interleukin-60.2660.11P02765alpha2-HS-Glycoprotein-0.593*t<0.01 Table 41 D: Proteins included in the extended leading edge set for acute inflammatory response (Amigo1). * indicates proteins to the right of the minimum of RS and † indicates proteins with anti-correlations of at least as great magnitude as that at the maximum of RS. UniProtIDProtein NameCorrelationP valueP01009alpha1-Antitrypsin0.801<0.01Q14624Inter-alpha-trypsin inhibitor heavy chain H40.603<0.01P02741C-reactive protein0.582<0.01P01024Complement C3a anaphylatoxin0.556<0.01P01024Complement C30.525<0.01P10600Transforming growth factor beta-30.498<0.01Q00535Cyclin-dependent kinase 5:activator p35 complexQ150780.492<0.01P07951Tropomyosin beta chain0.478<0.01P02679Fibrinogen gamma chain dimer0.475<0.01P11226Mannose-binding protein C0.461<0.01P00738Haptoglobin0.455<0.01P12956ATP-dependent DNA helicase II 70 kDa subunit0.4120.01P02743Serum amyloid P0.3910.02P07357P07358P07360Complement C80.3770.03P06744Glucose phosphate isomerase0.3640.03P06400Retinoblastoma 10.3400.04P01031Complement C50.3370.05P08107Hsp700.3060.07Q9Y5S2Myotonic dystrophy protein kinase-like beta0.2900.09Q8NEV9Q14213Interleukin-270.2900.09P05231Interleukin-60.2660.11P01019Angiotensinogen0.2630.12P02765alpha2-HS-Glycoprotein-0.593*t<0.01O00626Macrophage-derived chemokine-0.535*†<0.01P02649Apolipoprotein E-0.421†0.01P08697alpha2-Antiplasmin-0.347†0.04P08887Interleukin-6 receptor alpha chain-0.347†0.04P08514P05106Integrin alpha-IIb: beta-3 complex-0.303†0.07Q9BZR6Nogo Receptor / reticulon 4 receptor-0.300†0.08P00747Angiostatin-0.276†0.10

[0197] We further investigated the interaction of the processes related to the ontologies common to the groups of proteins we identified from RS plots. We found that the processes are related to the complement activation, activation and regulation of the immune system, as well as innate immune response and inflammatory response. Charts showing these relationships are found in Appendix K of our prior provisional application serial no. 62 / 289,587 from which this case claims priority.. Appendix K, Figure 8 also shows "heat maps" (i.e., plots of p values generated by GSEA analysis associating mass spectral features with protein expression values for all the m / z features used by the classifier of Example 1, Appendix B), which demonstrate the results of the correlation of the protein sets with the mass spectral features used for the Example 1 full-set approach 1 classifier. See also Figures 45A and 45B and the discussion thereof later in this document. We discovered that many of the mass spectral features used in the Example 1 classifier are related to the following biological processes: (1) acute phase, (2) acute response, (3) complement system, and (4) acute inflammatory response. Very few of the mass spectral features used are associated with the specific immune-related protein functions we investigated (i.e., "activation of innate immune response", "regulation of adaptive immune response", "immune T-cells", "immune B-cells", "interferon", "interleukin-10"). These relationships are demonstrated by the heat maps of Figure 8 of Appendix K of our prior provisional application serial no. 62 / 289,587 from which this case claims priority., with darker areas associated with lower p values and thus higher correlation between protein and class label. Figure 8 shows the heat maps for two different definitions of the enrichment score, but the plots are similar. See also Figures 45A and 45B of this document and the associated discussion below.

[0198] We performed the same analysis of the correlation between the protein sets with the class labels produced by the classifier of Example 2. The results are shown in Table 42. Note that the proteins associated with the following biological processes are strongly correlated with the class labels: acute inflammatory response, complement system, acute response and acute phase. In addition, proteins associated with Immune Response Type 2 were also strongly correlated with the class labels. Table 42: Results of GSEA applied to protein sets and Example 2 class labels of the Analysis SetDefinition 1Definition2ES definition1p valueES definition2p valueAcute inflammatory response0.451<0.010.4870.03Activation of innate immune response0.5110.330.5110.58Regulation of adaptive immune response0.1730.990.3320.96Positive regulation of glycolytic process-0.3280.800.4460.91Immune T-cells-0.1810.910.3230.81Immune B-cells0.4040.220.4180.65Cell cycle regulation-0.1780.950.3380.75Natural killer regulation-0.2410.870.2550.10Complement system0.629<0.010.633<0.01Acute response0.5350.120.6880.03Cytokine activity-0.2240.710.4010.35Wound healing-0.3240.230.4990.06Interferon0.2010.860.3190.88Interleukin-100.3060.050.3690.31Growth factor receptor signaling-0.1990.660.3450.55Immune Response Type 1-0.2270.980.3860.97Immune Response Type 20.852<0.010.8710.04Acute phase0.659<0.010.752<0.01Hypoxia0.2420.680.4080.45Cancer0.1540.950.2730.96 There are correlations at the p < 0.05 level of the class labels with the protein sets corresponding to acute inflammatory response, acute response, acute phase, complement system, interleukin-10 (with ES definition 1), and immune response type 2. Correlations with p values below 0.1 were also found for the wound healing protein set.

[0199] We further identified proteins from the complement system, acute phase, acute response, acute inflammatory response, and interleukin-10, and immune response type 2 processes important for the Example 2 classifier. The results are listed in Tables 11A-11F of Appendix K of our prior provisional application serial no. 62 / 289,587 from which this case claims priority.. Many of the proteins listed in Tables 11A-11D are found in the Tables 41A-41D. We also calculated the correlations of the protein sets with the mass spectral features used in Example 2 classifier and produced heat maps, see Figures 10 and 11 of Appendix K of our prior provisional application serial no. 62 / 289,587 from which this case claims priority. Many mass spectral features used in the Example 2 classification are related to the complement system and acute inflammation / acute phase reaction. However, these biological processes are somewhat less dominant as compared to the mass spectral features used in the Example 1 full-set approach 1 classifier. Table 12 of Appendix K of our prior provisional application serial no. 62 / 289,587 from which this case claims priority shows the number of mass spectral features (out of the 351 listed in Appendix A) which are associated with the protein sets at various significance levels (p values). In general, there are more mass spectral features associated with the proteins of the acute inflammation, complement system, acute response, acute phase, immune response type 2, and wound healing processes as compared to other processes.

[0200] The above discussion, and the report of Appendix K of our prior provisional application serial no. 62 / 289,587, from which this case claims priority, demonstrates that the combination of Deep MALDI measurements with simple modifications of GSEA shows promise for extracting useful information on the biological functions related to our test labels of this disclosure. It also allows us to gain insight into the functions related to the mass spectral features (Deep MALDI peaks) that we measure from serum samples.

[0201] The protein functions associated with the mass spectral features used in the Example 1 full-set approach 1 classifier were consistent with the protein functions associated with the class labels, namely acute phase reactants and the 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 plausible biological functions did not show any significant association with the class labels. This does not imply that these other functions are not relevant for the biology of immunotherapies; it just means that we have no evidence that they play a major role in the classifications produced by the full-set approach 1 classifier of Example 1. However, although we measured features (proteins) related to most of these other functions, they were not used in our tests, and the test classifications were not significantly associated with these protein functions.

[0202] In the case of the Example 2 classifier, we saw that the strength of association of classification with acute phase and complement functions increased quite significantly compared with Example 1, which may be an indication that the classifier of Example 2 is a "cleaner" test, less confounded by prognostic effects. We also observed that IL-10 related functions are associated with the Example 2 class groups at the p=0.05 significance level.

[0203] There are limitations with these data caused mainly by the limited size of the Analysis Set of samples, resulting in fairly wide null distributions, and hence limited statistical power. The simplest way to improve on this would be to have paired Deep MALDI / Somalogic data on more samples. This would also allow us to have an independent validation of these present results. While the number of proteins in the Somalogic panel is rather large, one could also consider using additional or extended panels.

[0204] Due to computer resource limitations we did not perform a false discovery rate analysis wrapped around the protein sets. Such an analysis would also require further theoretical work to assess the applicability of, and possibly work to improve on, the suggested set normalizations in the Subramanian paper. While this should in principle be done, the observed effects, especially for some of the mass spectral features, are so clear and large that we do not expect any substantial qualitative changes to the main conclusions of this analysis.

[0205] Because it is well-known that many proteins in circulation are related to acute phase reactants and the complement system, it may be not surprising that these two functions appear associated with many of the mass spectral features we measure with Deep MALDI. However, we did see other significant correlations, especially in the case of the Example 2 classifier, indicating that our results are not a trivial reflection of abundance of circulating proteins. In addition, within the subset of features used in the full-set approach 1 Example 1 classifier (see Appendix B), the proportion of features associated with acute phase and the complement system was substantially higher than that observed in the whole set of 351 Deep MALDI features listed in Appendix A.

[0206] The above discoveries can be used to build a classifier used in guiding immune checkpoint inhibitor treatment for a cancer patient. In particular, such a classifier includes a memory storing a reference set of class-labelled mass spectral data obtained from blood-based samples of melanoma patients treated with an immune checkpoint inhibitor agent. The mass spectral data is in the form of feature values for a multitude of mass spectral features, wherein the mass spectral features are identified 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. See the above discussion and the heat maps of Figures 8, 10 and 11 of Appendix K of our prior provisional application serial no. 62 / 289,587 from which this case claims priority. The classifier further includes a programmed computer (see Figure 15) implementing a classification algorithm on a set of mass spectral data including feature values for the multitude of mass spectral features obtained from a test blood-based sample and the reference set and generating a class label for the test blood-based sample. Alternatively, the mass spectral features may further include features associated with immune response type 2 and interleukin-10 processes. In one embodiment, the features include the features listed in Appendix A, Appendix B or Appendix C.

[0207] Furthermore, application of GSEA methods to the data obtained using the SomaLogic panel in combination with the results of full-set approach 1 classifier of Example 1 and further analysis allowed us to correlate up- and down-regulation of proteins with Early and Late classifications. In particular, proteins in Tables 41 A-D with positive correlation coefficient, are correlated with up-regulation in samples classified as Early.

[0208] Further analysis of the running sum of the Complement protein set allowed identifying proteins that have the biggest impact on the correlation of the corresponding proteins sets with the classification results (Table 5, Appendix K of our prior provisional application serial no. 62 / 289,587 from which this case claims priority). Table 43 below lists proteins which have P-value for correlation with classification labels ≤ 0.05 and are included in Group1 (leading edge) of the Complement protein set. Table 43: Proteins associated with gene ontology "Complement system" and correlated with full- set approach 1 classifier of Example 1.UniProt IDFull NameCorrelated Expression in the "Early" groupP02741C-reactive proteinUpP00736Complement C1rUpP01024Complement C3UpP01024Complement C3a anaphylatoxinUpP01024Complement C3bUpP01031Complement C5UpP01031Complement C5aUpP01031Complement C5b,6 ComplexUpP07357Complement C8UpP02748Complement C9UpP00751Complement factor BUpP01024Complement C3b, inactivatedUpP16109P-SelectinDownP02743Serum amyloid PUpP06681Complement C2UpP12956ATP-dependent DNA helicase II 70 kDa subunitUpP11226Mannose-binding protein CUpQ6YHK3CD109Down One can see that the group classified as Early is characterized by up-regulation of most of the components of the complement system, which raises the question of possible biological relationship between this up-regulation and unfavorable prognosis of patients classified as Early, as well as their little benefit from PD-1 inhibitors and, possibly, other types of immunotherapy.

[0209] A growing body of evidence suggests a complex role for the complement system in tumorigenesis, which, depending on an intricate balance of multiple factors, can be pro- or antitumor. Neoplastic transformation is accompanied by an increased capacity to activate complement. Pio et al., Adv Exp Med Biol 772:229 (2014). Activated complement proteins have been shown to inhibit tumor growth by promoting complement-dependent cytotoxicity (Janelle & Lamarre; Oncoimmunology 3, e27897 (2014)) and inhibition of Treg function. Mathern & Heeger, Clin J Am Soc Nephrol 10:1636 (2015).

[0210] On the other hand, recent data have demonstrated that activated complement proteins, interacting with the components of the innate and adaptive immune systems, can promote carcinogenesis. For example, in mouse models of melanoma it was shown that "decomplementation" led to a robust antitumor CD8+ response and improved cytotoxic activity of NK cells, while activated complement system resulted in limited accumulation of tumorspecific cytotoxic T cells (CTLs), and, at the same time, promoted tumor infiltration with immunosuppressive myeloid-derived suppressor cells (MDSc), which suppress NK-and T-cell functions. Janelle et al., Cancer Immunol Res 2, 200-6 (2014). Similarly, complement activation and C5a signaling were shown to be associated with recruitment of MDSCs into tumors, suppression of effector CD8+ and CD4+ T cells, generation of regulatory T cells (Tregs), Th2 predominant immune responses, and facilitation of lung and liver metastasis in models of breast and cervical cancer. Markiewski et al.,Nat Immunol 9:1225 (2008);

[0211] Vadrevu et al., Cancer Res 74:3454 (2014). C5a was shown to promote differentiation of Tregs, causing inhibition on antitumor activity. Gunn et al., J Immunol 189:2985 (2012). Complement can assist the escape of tumor cells from immunosurveillance, support chronic inflammation, promote angiogenesis, activate mitogenic signaling pathways, sustain cell proliferation and insensitivity to apoptosis, and participate in tumor invasion and migration. Pio et al., supra. In lung cancer models, blockade of C5a signaling led to the inhibition of key immunosuppressive molecules within the tumor. These molecules included IL-10, IL-6, CTLA4, LAF3, and PDL18.

[0212] The latter findings have direct implications for the activity of the immune checkpoint inhibitors in cancer patients, and, consequently, for the role of our Example 1 and Example 2 classifiers. In particular, the observed upregulation of the complement system proteins in the group classified as Early may indicate that these patients have higher levels of immunosuppression, and / or higher levels of pro-tumor inflammation, related to the activation of the corresponding immune checkpoints, and as a result are less responsive to such drugs as nivolumab, ipilimumab, pembrolizumab, or other agents targeting these pathways. Interestingly, it has been shown that the complement protein C5a promotes the expression of the PD-1 ligands, PD-L1 and PD-L2. Zhang, J. Immunol. 2009; 182:5123. In this scenario one could envision that excessive complement upregulation might compete with efforts to inhibit PD-1. On the other hand, the results of recent clinical trials suggest that patients with tumor microenvironment characterized by high expression of PDL1 and presence of Tregs are more likely to respond to anti-PD-1, anti-CTLA4, or high dose IL-2 therapy. Though we do not know how exactly upregulation of the complement system is correlated with Example 1 and Example 2 classifications, this connection is in line with the biological effects discussed above.

[0213] Consequently, we can expect that Example 1 and Example 2 classifiers may be relevant for the broad variety of drugs affecting the immunological status of the patient, such as various immune checkpoint inhibitors, high dose IL-2, vaccines, and / or combinational therapy. Furthermore, since effects that are measured in serum reflect the organism status as a whole, and the complement system affects innate and adaptive immunity on the global level, not just in a tumor site, the classifiers are expected to have similar performance in different indications in cancer (e.g., lung, renal carcinoma), and are not restricted to melanoma.

[0214] Another large gene ontology protein set correlated with Example 1 and Example 2 Early and Late classifications is "Acute inflammatory response". Proteins correlated with these classification groups from this set (p≤0.05) are presented in Table 44. Table 44. Proteins associated with gene ontology "Acute Inflammatory response" and correlated with full set approach 1 Example 1 classification.UniProtIDProtein NameCorrelated Expression in the "Early" groupP01009alpha1-AntitrypsinUpQ14624Inter-alpha-trypsin inhibitor heavy chain H4UpP02741C-reactive proteinUpP01024Complement C3a anaphylatoxinUpP01024Complement C3UpP10600Transforming growth factor beta-3UpQ0053, Q15078Cyclin-dependent kinase 5:activator p35 complexUpP07951Tropomyosin beta chainUpP02679Fibrinogen gamma chain dimerUpP11226Mannose-binding protein CUpP00738HaptoglobinUpP12956ATP-dependent DNA helicase II 70 kDa subunitUpP02743Serum amyloid PUpP07357, P07358, P07360, P06744Complement C8UpGlucose phosphate isomeraseUpP06400Retinoblastoma 1UpP01031Complement C5UpP01019AngiotensinogenDownP02765alpha2-HS-GlycoproteinDownO00626Macrophage-derived chemokineDownP02649Apolipoprotein EDownP08697alpha2-AntiplasminDownP08887Interleukin-6 receptor alpha chainDown It is generally accepted that cancer triggers an intrinsic inflammatory response that creates a protumorigenic microenvironment (Mantovani et al., Nature 454:436 (2008)); "smouldering" inflammation is associated with most, if not all, tumors and supports their progression. Porta et al., Immunobiology 214:761 (2009). Inflammation is intrinsically associated with the complement system, and complement system promotes tumor growth in the context of inflammation. Hence, it seems logical that both systems came out as significantly correlated with Example 1 and Example 2 classifications.

[0215] Tumor associated inflammatory response can be initiated and / or modulated by cancer therapy. On one hand, it can have tumor-promoting functions, but on the other hand it can enhance presentation of tumor-antigens and subsequent induction of anti-tumor immune response. Grivennikov et al., Cell 140:883 (2010). While a T-cell inflamed microenvironment, characterized by recruiting of CD8+ and CD4+ lymphocytes to the tumor, is considered a necessary condition for effective immunotherapeutic treatment, activation of the elements of the acute inflammatory pathway is likely correlated with the negative prognosis. The exact mechanisms of action remain poorly understood, but our data on upregulation of this system in the Early group seem to be consistent with the existing clinical data.

[0216] The present disclosure demonstrates how it is possible to incorporate biological insight into a classifier development exercise, such as for example the approach of Figures 8A and 8B, using mass spectrometry data. The present disclosure also demonstrates how it is possible to test biologically motivated hypotheses about the relevance of biological functions for certain disease states using mass spectral data. This is achieved by using gene set enrichment analysis to associate biological functions, via subsets of proteins related to these functions, with features (peaks) in mass spectra obtained from serum samples, and using such identified features to train a classifier with the aid of a computer.

[0217] In particular, classifier development and training methods are disclosed which make use of mass spectrometry of a development set of samples. Protein expression data are obtained from a large panel of proteins spanning biological functions of interest either for each of the samples in the development set of samples or, alternatively and more typically, for each of the samples in another sample cohort for which mass spectrometry data are also available. The latter case is preferred because the measurement of abundance of many proteins via a protein assay requires a large amount of sample and is expensive and time consuming. It is also not necessary to construct the relation between mass spectral peaks and biological function for every development project because we can infer the correlation of mass spectral features to function from any reference set that has sufficient protein coverage. In the following we exemplify our methods using these two options interchangeably.

[0218] In our method, we identify statistically significant associations of one or more of the mass spectral features with sets of proteins grouped by their biological function. With the aid of a computer, we then use these one or more mass spectral features that were identified (and typically 10-50 of such features) to train a classifier. This training may take the form of a classifier development exercise, one example of which is shown in Figure 8 and described in detail previously. The classifier is in the form of a set of parameters and associated program instructions or code, which when executed by a computer assigns a class label to mass spectrometry data of a sample of the same type as the development set of samples in accordance with the programmed instructions.

[0219] As described in this Example, using Gene Set Enrichment Analysis (GSEA) methods, it is possible to look for statistically significant associations of mass spectral features with sets of proteins grouped by their biological function ("protein functional groups") avoiding direct protein identification, and taking advantage of the high-throughput aspect of mass spectrometry. A system or set of components that conducts GSEA and identifies such mass spectral features with a particular protein functional group or subset is referred to herein as a "platform" or "GSEA platform." Such a platform consists of both a known, conventional protein expression assay system (e.g., the SOMAscan assay provided by SomaLogic of Boulder CO) and a computer for implementing GSEA analytical procedures to identify the mass spectra features associated with functional groups of proteins as described in this document.

[0220] It is necessary to have matched mass spectral data (preferably from a high sensitivity method such as "Deep MALDI" see U.S. Patent 9,279,798) and protein expression data from a large panel of proteins spanning biological functions of interest on a single set of serum samples. Using well-known protein databases, such as UniProt or GeneOntology / AmiGO2, subsets of proteins from the universe of measured proteins can be defined based on their biological functions. The entire list of measured proteins is first ranked according to the correlation of each protein with the mass spectral feature of interest. The GSEA method then looks for over- or under-representation of the proteins included in a particular protein functional subset as a function of rank in this ranked list of all measured proteins and provides a way of assessing its statistical significance. Thus, the association of the mass spectral feature of interest with different protein functional groups can be assessed. This procedure can be repeated for as many spectral features and as many protein functional groups as desired. Setting a cutoff on the degree of association or the p value for significance of the association, all mass spectral features associated with a particular protein functional subset can be identified. This set of mass spectral features is then used to train a classifier, such as a k-Nearest Neighbor (kNN) classifier.

[0221] One method we prefer for classifier training and development is known as Diagnostic Cortex, which is described at length previously in the context of Figure 8. It has been demonstrated that using the Diagnostic Cortex classifier development procedure, clinical data and Deep MALDI mass spectrometry data can be combined to produce clinically useful molecular diagnostic tests. One advantage of this method is that it allows for the design and tuning of tests to meet required standards of clinical utility. The use of subsets of functionally related mass spectral features (identified from the procedures mentioned in the previous paragraph) instead of all available mass spectral data provides an additional option for test design and optimization. The mass spectral features associated with one or more protein functional subsets can be selected and combined with the clinical data of a set of samples to create a new classifier and associated test, allowing for the investigation of the relevance of individual biological functions or groups of biological functions for the required classification task.

[0222] We also describe creation of a multitude of different classifiers using different feature subsets related to different protein functional groups and combine them, for example by simple majority vote, more complex ensemble averaging, or some rule-based system, to produce an overall classification that combines the information content across various biological functions. In one variation of this, we can look at the functional subsets becoming relevant on the groups defined by the classifier after taking out, or taking care of the main effects, by using the peaks related to the functional groups in the classifier, and using these newly relevant peaks for building a new classifier in terms of a hierarchy of biological functions. For example, we can use peaks associated with acute response function to train a first level classifier. We train a second classifier using a set of peaks associated with a wound healing protein function. A sample which tests Late (or the equivalent) on the first level classifier is then classified by the second level classifier. If one has a big enough set, it is possible to iterate this process further and define a third level classifier on a set of peaks associated with a third protein function and use that for a group classified by the second level classifiers, etc. As there are often multiple protein functional groups associated with a given peak this approach attempts to disentangle compound effects.

[0223] Thus, this disclosure describes a method of generating a classifier including the steps of: a) obtaining a development set of samples from a population of subjects; b) conducting mass spectrometry on the development set of samples and identifying mass spectral features present the mass spectra of the development set of samples; c) obtaining protein expression data from a large panel of proteins spanning biological functions of interest for each of the samples in the development set of samples, or, alternatively, for each of the samples in an additional cohort of samples with associated mass spectral data; d) identifying statistically significant associations of one or more of the mass spectral features with sets of proteins grouped by their biological function using the set of samples with matched mass spectral and protein expression data; and e) with the aid of a computer, using the one or more mass spectral features identified in step d) and clinical data from the development set of samples to train a classifier, the classifier in the form of a set of parameters which assigns a class label to a sample of the same type as the development set of samples in accordance with programmed instructions.

[0224] A programmed computer configured as a classifier generated in accordance with the method of the previous paragraph is also described.

[0225] A method of testing a sample is disclosed, which includes the steps of a) training a classifier using a set of mass spectra features that have been determined to have statistically significant associations with sets of proteins grouped by their biological function; b) storing the parameters of the classifier including a feature table of the set of mass spectral features in a memory; c) conducting mass spectrometry on a test sample; and d) classifying the test sample with the trained classifier with the aid of the computer. In one variation, the steps include training two classifiers using different subsets of features associated with different functional groups of proteins, storing the parameters of the first and second classifiers including a feature table of the sets of mass spectral features in a memory, and logical instructions for combining the first and second classifiers into a final classifier; c) conducting mass spectrometry on a test sample; and d) classifying the test sample with the final classifier with the aid of the computer.

[0226] A computer configured as a classifier is disclosed including 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 spectra features have been determined to have statistically significant associations with sets of proteins grouped by their biological function, and a set of parameters defining a classifier including a classification algorithm operating on mass spectral data from a test sample and the feature table.

[0227] In another aspect, a classifier development system is disclosed, including a mass spectrometer for conducting mass spectrometry on a development set of samples to generate mass spectral data, said data including a multitude of mass spectral features; a platform for conducting a gene set enrichment analysis on the development set of samples or, more typically, another set of samples with associated mass spectral data, and identifying statistically significant associations of one or more of the mass spectral features with sets of proteins grouped by their biological function; and a computer programmed to train a classifier using the one or more mass spectral features identified by the platform, the classifier in the form of a set of parameters which assigns a class label to a sample of the same type as the development set of samples in accordance with programmed instructions.

[0228] In the above methods and systems, the development set of samples can take the form of blood-based samples (serum or plasma) from humans, for example humans enrolled in a clinical trial of a drug or combination of drugs. Such humans can be cancer patients. We describe the inventive methods and systems below in the context of a development set of blood-based samples obtained from melanoma patients treated with an immunotherapy drug, namely a programmed cell death 1 (PD-1) checkpoint inhibitor.

[0229] This document demonstrates that it is possible to associate features in mass spectra with biological functions without direct identification of the proteins or peptides producing the mass spectral feature, and incorporate biological insights into the choice of mass spectral features for use in reliable classifier training or development, e.g., using the Diagnostic Cortex platform.

[0230] Association of mass spectral features directly with biological processes is important as it is often difficult and time-consuming, and sometimes impossible, to identify the proteins or peptides producing individual peaks in mass spectra. This method circumvents the need for thousands of protein identification studies, which even when successful, do not always allow the matching of biological processes to mass spectral peaks (the specific functions of many peptides and protein fragments remain to be determined).

[0231] The ability to find mass spectral features generated from human serum reliably associated with biological processes provides a new way to monitor these processes in a longitudinal manner in a minimally-invasive, high throughput manner. Serum samples could be collected from patients at many time points during the course of a therapy or disease and changes in specific biological processes could potentially be inferred from the analysis of mass spectra generated from the serum samples. Such changes can be due to an intervention (e.g., treatment), or to the natural evolution of the disease. While the study considered an application in oncology, this could be of interest across many disease areas.

[0232] The incorporation of biological insights into classifier training for reliable molecular diagnostic test development provides another avenue for the design and tuning of the tests that can be created. Experience with the Diagnostic Cortex platform has shown that in some situations the ability to tune tests to meet clinical needs is reduced and similar tests are produced despite attempts to tune towards other performance goals. It was believed that this was due to the dominance of certain mass spectral features and correlations between features, but previous attempts to remove these dominating effects to allow investigation of other subsidiary, but potentially important, effects had proved quite unsuccessful. The ability to determine which features are associated with individual biological processes provides a new way to look at the universe of mass spectral features that can be used in classification, allowing us to attempt to separate out effects that might confound each other or to remove processes that dominate classification to reveal other processes that can improve test performance and test biological hypotheses.

[0233] The application of GSEA-based feature selection could be very broad and potentially could lead to the extension of our understanding of the role of specific biological processes and related treatment in any disease. As an example, type 2 diabetes is known to be a metabolic disorder. However, it is also known that inflammation plays an important role (see Kinget al., J Periodontol. 2008 Aug;79(8 Suppl): 1527-34. doi: 10. 1902 / jop.2008.080246 and Odegaard et al., Cardiovasc Diabetol. 2016 Mar 24;15(1):51. doi: 10.1186 / s12933-016-0369-6). If we decide to build a prognostic test for diabetes, we might consider selecting separate feature sets: one associated with the insulin pathway, and another one with inflammation. Using the approach outlined in this report, we can attempt to separate out the effects of these two broad biological processes on prognosis, and even estimate the relative effect of each of them on the prognostic classification. Furthermore, if we try to find a predictive test for a novel drug, we might even be able to better understand the mechanism of action of the therapy, for example, if only one of the hypothetically relevant feature sets would work well.

[0234] Since almost no disease is defined just by a single process disruption, the application of the methods of this disclosure is very broad, and is limited only by the adequate measurement of the related proteins in the sample of choice and by (in)sufficient understanding of the roles of these proteins in particular biological processes. So, theoretically, this method could allow researchers to separate and test effects of multiple biological processes associated with practically any disorder, as well as to better understand the mechanism of action of treatments. While this study involves MALDI mass spectrometry of serum, and so is limited to processes that can be explored via circulatory proteins and peptides, the method per se does not depend on sample type and so can be extended even beyond this already wide regime of applicability.

[0235] Further details and an example of GSEA-based feature selection with an application in classifier development will now be described with particularity.

[0236] Two sample sets were used in this study: 1. A set of 49 serum samples ("the GSEA cohort", or "analysis" set or cohort) from 45 patients with non-small cell lung cancer and 4 subjects without cancer, for which mass spectral data were collected and protein expression data were generated using the 1129 protein SOMAscan ®< aptamer panel (SomaLogic, Boulder, CO). We used this set to determine peaks which were used for classifier training from a GSEA analysis. 2. A set of 119 pretreatment serum samples from 119 patients with advanced melanoma who were treated with the anti-programmed cell death-1 (PD-1) therapy, nivolumab, with or without the addition of a multi-peptide vaccine as part of a clinical trial ("the NCD cohort", also called "Moffitt" in this disclosure) (Details of the trial can be found in Weber et al., J Clin Oncol 2013, 31(34) 4311). Outcome data were available for patients in this cohort and mass spectral data were collected from these samples. Samples from this set were used in classifier training. This sample set is described in Example 1. Generation of Protein Expression Data and GSEA platform

[0237] Protein expression data were collected from the GSEA cohort using the 1129 protein SOMAscan aptamer panel by SomaLogic at their laboratory in Boulder, CO. A list of the 1129 proteins contained in the assay is contained in Appendix A of our prior provisional application serial no. 62 / 340,727 filed May 24, 2016 (from which this case claims priority). Further details on the identification of protein groups associated with mass spectral features are set forth later on in this document.

[0238] The generation and processing of mass spectral data from both the GSEA cohort and the NCD cohort was performed as explained in great detail previously in Example 1.Application of Gene Set Enrichment Analysis (GSEA) Methods

[0239] GSEA (see the Mootha et al. and Subramanian et al. papers cited previously) was introduced as a method to help deal with or try to minimize some essential problems in gene expression analysis studies: identification of gene sets and resulting tests in development sample sets that cannot generalize to other sample sets (overfitting), the multiple testing problem, and the inability to identify smaller expression changes consistent across multiple related genes that might be swamped by larger randomly occurring expression changes in a dataset. These are problems inherent in dealing with "p>n" datasets, i.e. where the number of measured expression values greatly exceeds the number of samples for which the measurements are available. Instead of looking at expression differences feature by feature (gene by gene), the method looks for expression differences that are consistent across pre-specified groups or sets of features. The feature sets can be created based on biological insight or one can use feature sets that have been defined by prior hypothesis-free studies. Correlating with sets of features rather than single features provides some protection against identifying isolated features that are randomly correlated with study groups and would not generalize to other sample sets. Typically, the number of feature sets that are tested for correlation is smaller than the number of single features in a typical gene expression study, so this reduces somewhat the multiple testing problem. In addition, because the method looks for consistent correlations across a group of features, it is possible to identify a significant effect that is smaller in magnitude (per feature) than that which could be identified for a single feature.Definition of protein sets

[0240] Specific protein sets were created based on the intersection of the list of SOMAscan 1129 panel proteins and results of queries from GeneOntology / AmiGO2 and UniProt databases. The AmiGO2 queries were filtered by: document category: annotation taxon: H sapiens evidence type: experimental The individual filters used are listed in Table 45. Table 45: Filters used in the AmiGO2 database Protein Set NameKeywordAmigo 1Acute inflammatory responseAmigo 2Activation of innate immune responseAmigo 3Regulation of adaptive immune responseAmigo 4Positive regulation of glycolytic processAmigo 5Immune T-cellsAmigo 6Immune B-cellsAmigo 7Cell cycle regulationAmigo 8Natural killer regulationAmigo 9Complement systemAmigo 11Acute responseAmigo 14Cytokine activityAmigo 16Wound healingAmigo 17InterferonAmigo 18Interleukin-10Amigo 20Growth factor receptor signalingAmigo 21Immune ResponseAmigo 22Immune Response Type 1Amigo 23Immune Response Type 2 The UniProt queries were filtered by: Organism: H. sapiens DB: reviewed (SwissProt) The individual filters used in the UniProt database are listed in table 46. Table 46: Filters used in the UniProt database Uniprot 1Acute phaseUniprot 2HypoxiaUniprot 4Cancer The proteins included in each protein set are listed in the Appendix D of our prior provisional application serial no. 62 / 340,727 (from which this case claims priority). (Uniprot 4 Cancer is not included in the listing as it contains more than 400 proteins.)Implementation of GSEA Method

[0241] The implementation of the GSEA method was done on a general purpose computer using Matlab (version R2015a). The process can be decomposed into several steps.

[0242] Let us assume that we have data from a set of N s samples and for each sample, i, we are given a continuous variable Q i and the expression values of N f proteins, F i j< , where i runs over the samples (1≤i≤ N s ) and j runs over the proteins (1≤j≤ N f ). We have k predefined protein sets S l (1≤l≤k) that we are interested in correlating with the mass spectral data. Each protein set S l consists of N h l< members (1<N h l< < N f ) and is a subset of the complete set of N f proteins.1. Evaluation of the correlation between individual proteins and a continuous variable (the mass spectral feature value)

[0243] The strength of the correlation, r j , between each individual protein j in the full protein set and the continuous variable associated with the samples is calculated. Spearman's rank correlation was used to assess the degree of correlation. Once a correlation, r j , had been calculated for each protein, j, the N f proteins were ranked by r j , from largest to smallest.2. Calculation of an Enrichment Score

[0244] As explained in the Subramanian et al. paper cited previously, the enrichment score, ES l , is designed to reflect the degree to which elements of a particular protein set, S l , are over-represented at the top or bottom of the ranked list of proteins. We start at the top of the rank list and construct a running sum, RS(S l, p), at item p on the ranked list by starting at zero and adding a term |r j | / N norm for the jth item in the ranked list if protein j is contained in S l and subtracting a term 1 / (N-N h l< ) for the jth item in the ranked list if protein j is not contained in S l until one reaches item p. N norm is a normalization coefficient defined by N norm = Σ|r j |, where the sum runs over all proteins j contained in S l . An example of a calculated RS(S l ,p) is shown in Figure 43.

[0245] We consider two possible definitions for ES. First, ES l is defined in terms of the largest positive value of RS(S l ,p) as a function of p, RS max , and the smallest value of RS(S l ,p), RS min . These are illustrated in Figure 43. If RS max ≥ |RS min |, ES = RS max ; if RS max < |RS min |, ES = RS min (This is the definition used in Subramanian et al. with their exponent p set to 1.) To be able to take account of protein sets containing mixtures of over- and under-expressed proteins by group or mixture or proteins meaningfully correlated and anti-correlated with the continuous variable, we also consider an alternative definition of ES as RS max + |RS min |. (If all proteins in the protein set are over-expressed (positively correlated), the two definitions are identical.)3. Calculation of the corresponding p value

[0246] To assess the significance of the deviation of the calculated ES from its average value for a random distribution, the null distribution of ES is calculated by generating many realizations of a random association between the continuous variable and the protein expressions and evaluating ES for each. These realizations are created by permuting the values of the continuous variable assigned to each sample. Note that this maintains the correlation structure within the protein expression values for each sample. Once the null distribution has been generated, the p value for the calculated ES can be read off as the proportion of random permutations generating an ES further from random (more extreme) than the calculated ES. (Note that the first definition of ES requires an assessment of positive and negative ES separately). Examples of the null and calculated ES for one protein set are shown in Figures 44A and 44B. The null distribution has to be evaluated separately for each comparison (each individual continuous variable (i.e., each mass spectral feature) and protein set pair). For the correlation with mass spectral feature value, 2000 realizations were generated.4. Corrections for multiple testing

[0247] The p values produced by the method outlined above do not take into account multiple testing. It is possible to extend the analysis to take account of multiple testing either by a very conservative Bonferroni correction or by generating many permutations of the random permutations over the continuous variable also over the ranked protein list for all protein sets and computing the ES for each realization. This latter method also requires a normalization of the ES to allow the combination of results across different protein sets. At present neither of these approaches has been implemented and the results in this report have not been corrected for multiple testing.GSEA Results

[0248] Using the methodology described above, a GSEA p value was obtained for each mass spectral feature for each protein functional group. The results for the correlation of the protein functional groups with the all 351 defined mass spectral features (Appendix A) for the 49 samples of the GSEA cohort are shown in the heat maps of Figure 45A and 45B. In particular, Figures 45A and 45B show the p values generated by the GSEA analysis associating all 351 defined mass spectral features with different protein functional groups (biological processes). Figure 45(a) shows the p values for ES definition1 and Figure 45(b) for ES definition 2. Note: Mass spectral features are ordered in increasing m / z and only every 5 th< spectral feature is labeled on the x axis.

[0249] Table 47 shows the number of mass spectral features (out of the 351 defined) associated with the protein functional groups with p values below a variety of thresholds for each of the protein functional subsets for definition 1 and definition 2 of the enrichment score. Table 47: Number of mass spectral features associated with protein functional groups for p values below a variety of thresholds Protein Set Function ES Definition 1 p value ES Definition 2 p value < 0.001* <0.01 <0.03 <0.05 <0.1 < 0.0005* <0.01 <0.03 <0.05 <0.1 Acute Inflammation13365931295285779112Innate Immune Response03612290051027Adaptive Immune Response011250141012Glycolytic Process33817361391735Immune T-cells00381300244Immune B-cells02281500227Cell cycle0146101381520NK Regulation0002400127Complement1562951221578447793125Acute Response0922336408304785Cytokine Activityo151218002213Wound Healing215324678116406077Interferon0146170581420Interleukin-100251025o171325Growth Factor Receptor Signaling002618012413Immune Response1831558016234681Immune Response Type 11259162231423Immune Response Type 21510163707121834Acute phase841678411166690110149Hypoxia002102603142440Cancer03111732118303446*Indicates that ES was greater (or smaller) than that obtained in any of the realizations generated to assess the null distribution

[0250] It is clear that many of the 351 mass spectral features are associated with acute phase reactants (acute response, acute phase, acute inflammation), the complement system, or wound healing. However, there also exist mass spectral features that are associated with other quite distinct protein functional groups, such as glycolytic process, cell cycle, or cancer. Hence, it is potentially possible to use measurements of mass spectral features that have been determined to be associated with a particular biological function from serum samples from a patient in order to monitor the particular biological function in the patient.

[0251] A cutoff of p = 0.05 was chosen for the first definition of enrichment score (ES definition 1) and for each protein functional subset, so that the mass spectral features with GSEA p < 0.05 were taken to be associated with the biological function. The mass spectral features associated with several of the protein functional subsets (Acute Response, Wound Healing, Immune Response) investigated are tabulated in Appendix E of our prior provisional application 62 / 340,727.

[0252] We then proceeded to develop classifiers using the Figure 8A and 8B methodology and peaks associated with particular protein functional groups. First, classifier development using the procedure of Figure 8, steps 102-150 was performed on all 119 samples in the new classifier development (NCD) cohort, with the subset of 33 mass spectral features associated with the acute response protein functional group. Using the procedure of Figure 8, we created a classifier, referred to below as "Classifier 1" which 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 at each step of refinement of the class labels. Fifty one samples were assigned to the poor performing group and these were given an "Early" classification label. The particular choice of moniker for the class label generated by the classifier is not particularly important.

[0253] The remaining 68 samples, assigned to the good performing group, were used as the development set for a second classifier generated in accordance with Figure 8 steps 102-150, referred to as "Classifier 2". This classifier was trained on the subset of 26 mass spectral features which had been identified as being associated with wound healing, but not associated with acute response or immune response. The second classifier again used no feature deselection and stratified patients well into groups with better or worse TTP. Samples in the good TTP group were assigned a "Late" classification and samples in the poor TTP group were assigned an "Early" classification.

[0254] We then defined a final classifier as a hierarchical combination of classifiers 1 and 2. The resulting final classifier (i.e., a combination of two Diagnostic Cortex classifiers with logical instructions for use in a hierarchical manner) uses a total of 59 features, listed in Appendix D of this document. Figure 46 illustrates schematically how a classification is assigned to a test sample by the combination of Classifier 1, based on mass spectral features associated with acute response, and Classifier 2, based on mass spectral features associated with wound healing but not with acute response or immune response. In particular, the mass spectrum is obtained for a test sample ("test spectrum" in Figure 46) and in particular feature values for the 59 features of Appendix D are obtained. This data is supplied to Classifier 1 and Classifier 1 then produces a label for the sample. If the label reported is Early (or the equivalent) the sample is assigned the Early classification label. If Classifier 1 does not produce the Early label, the spectral data is supplied to Classifier 2. If Classifier 2 produces the Early label, then the sample is assigned the Early classification label. If Classifier 2 does not produce the Early label, then the sample is assigned the Late classification label. The Early and Late labels have the same clinical meaning as explained in Example 1 previously.

[0255] Note that this is an example of the creation of multiple different classifiers using different feature subsets related to different protein functional groups and the combination of them, by a rule-based system, to produce an overall classification that combines the information content across various biological functions. In this particular example we look at the functional groups becoming relevant on the groups defined by the classifier after taking out, or taking care of the main effects, by using the peaks related to the functional groups used in the first classifier, and using newly relevant peaks for building a new or second classifier in terms of a hierarchy of biological functions. If one has a big enough set, it is possible to iterate this process further and define a third level classifier on a set of peaks associated with a third protein function and use that for a group classified by the second level classifiers, etc. As there are often multiple protein functional groups associated with a given peak this approach attempts to disentangle compound effects.Results1. Classifier 1 alone

[0256] Classifier 1 assigned 51 of 119 samples an "Early" classification. Figures 47A and 47B show the Kaplan-Meier plots of OS and TTP for the classifications provided by Classifier 1 for the NCD cohort. The classifier based on acute response achieves a clear separation between the good and poor prognosis groups.2. Classifier 2 alone

[0257] Classifier 2 assigned 35 of the 68 samples not classified as "Early" by Classifier 1 an "Early" classification and the remaining 33 a "Late" classification. Kaplan-Meier plots of OS and TTP for the classifications provided by Classifier 2 of these 68 samples are shown in Figures 48A and 48B.

[0258] Classifier 2, using features associated with wound healing, but not those associated with acute response or immune response, further stratifies the 68 samples not classified as "Early" by classifier 1.2. Final classifier defined as a combination of classifiers 1 and 2

[0259] Combining the classifiers in a hierarchical manner as shown in Figure 46, one obtains a superior binary stratification of the whole set of 119 samples. Thirty three (28%) samples were classified as "Late" and 88 (72%) as "Early". This is illustrated in the Kaplan-Meier plots of OS and TTP for the whole NCD cohort of 119 samples by overall classification in Figures 49A and 49B. Associated statistics characterizing the clear stratification of the cohort are given in table 48. Table 48: Statistics related to the Kaplan-Meier plots of Figures 49A and 49B (CPH = Cox Proportional Hazard)OSTTPlog-rank pCPH pHR (95% CI)log-rank pCPH pHR (95% CI)Late vs Early0.0060.0080.42 (0.22-0.80)< 0.0010.0010.40 (0.23-0.68)Median (95% CI) in weeksMedian (95% CI) in daysEarly80 (59-99)92 (82-162)LateNot reached (78-undefined)541 (163-undefined)

[0260] Table 49 shows some landmark survival and progression-free statistics and table 50 summarizes the response data. Table 49: Proportions still alive and progression-free at key time pointsEarlyLate% alive at 1 year6388% alive at 2 years3461% progression-free at 6 months3167% progression-free at 1 year2560 Table 50: Response by test classification Early (n=86)Late (n=33)PR19 (22%)12 (36%)SD9 (10%)9 (27%)PD58 (67%)12 (36%) Table 51 shows the baseline patient characteristic by classification group. Table 51: Baseline patient characteristics by test classification Early (N=86) n (%)Late (N=33) n (%)GenderMale52 (60)20 (61)Female32 (37)13 (39)NA2 (2)0 (0)AgeMedian (Range)62 (16-87)60 (27-76)Cohort16 (7)3 (9)210 (12)1 (3)36 (7)5 (15)48 (9)2 (6)513 (15)8 (24)643 (50)14 (42)Prior IpiNo22 (26)9 (27)Yes64 (74)24(73)PD-L1 expression (5% tumor)Positive6 (7)2 (6)Negative20 (23)9 (27)NA60 (70)22 (67)PD-L1 expression (1% tumor)Positive14 (16)4 (12)Negative12 (14)7 (21)NA60 (70)22 (67)PD-L1 expression (1% tumor / immune cells)Positive20 (23)8 (24)Negative5 (6)2 (6)NA61 (71)23 (70)Serum LDH levels< ULN12 (4)6 (18)< 2ULN55 (64)32 (97)median range496 (174-4914)472 (149-789)Baseline tumor sizemedian31.05 (1.50-259.03)11.43 (0.88-1.13)range Fisher's exact test shows a significant correlation of serum LDH level < 2ULN with classification (p < 0.001) and Mann-Whitney p = 0.070 for association of LDH level with classification. Baseline tumor size was greater in the Early group than in the Late group (Mann-Whitney p < 0.001). Classification was not associated with PD-L1 expression at any available cutoff, however.

[0261] Multivariate analysis of the time-to-event outcomes allows the adjustment of the effect sizes (hazard ratios) for other known prognostic characteristics, such as serum LDH level. The results of this analysis are given in table 52. Classification remains a significant predictor of both OS and TTP, in addition to serum LDH level, indicating that the classification is providing supplementary information on outcome following nivolumab therapy. Table 52: Multivariate Analysis of Time-to-Event EndpointsOSTTPCPH pHR (95% CI)CPH pHR (95% CI)Late vs Early0.0230.46 (0.23-0.90)0.0020.43 (0.25-0.74)Female vs Male0.0791.63 (0.95-2.82)0.0321.65 (1.04-2.61)LDH / 10000.0041.66 (1.18-2.33)0.0151.49 (1.08-2.06)PD-L1 5% + vs -0.6960.81 (0.29-2.30)0.7501.14 (0.52-2.50)PD-L1 5% NA vs -0.5691.19 (0.65-2.20)0.8640.95 (0.54-1.67)Prior Ipilimumab Yes vs No0.3830.77 (0.44-1.37)0.3690.80 (0.48-1.31)

[0262] It is interesting to note that the performance of the overall classification obtained using these feature subsets selected using biological hypotheses may be superior to that obtained previously on this sample cohort. Figures 50A-50D compare the Kaplan-Meier plots for the present classification with those obtained for two previously developed classifiers, one (IS2 = full-set classifier of Example 1) developed using all mass spectral features simultaneously and the other (IS6) using an ensemble of classifiers with clinically different development subsets again using all mass spectral features.(These classifiers are described in Example 1, and Example 8, respectively).

[0263] Samples from a cohort of 30 patients also treated with anti-PD 1 therapy were available for independent validation of the classifier. Twenty-one patients (70%) were classified as "Early" and 9 (30%) as "Late". The Kaplan-Meier plot for OS is shown in Figure 51 and associated statistics in table 53. (TTP was not available for this observational cohort.) Table 53: Statistics related to the Kaplan-Meier plots of Figure 51OSlog-rank pCPH pHR (95% CI)Late vs Early0.0160.0300.20 (0.05-0.86)Median (95% CI) in weeks1 year survival2 year survivalEarly37 (21-68)38%27%Late210 (20-undefined)89%89% Example 6 Conclusions and Discussion

[0264] This study of Example 6 has demonstrated that it is possible to: 1) associate features in mass spectra with biological functions without direct identification of the proteins or peptides producing the mass spectral feature, and 2) incorporate biological insights into the choice of mass spectral features for use in reliable classifier development, e.g., using the Diagnostic Cortex platform of Figure 8.

[0265] It will be further appreciated that once a classifier has been developed as explained above, it is then stored as a set parameters in memory of a computer (e.g., feature table of mass spectral features used for classifications, identification of mini-classifiers, logistic regression weights, kNN parameters, program code for executing one or more master classifiers and logic defining a final classifier, as per Figure 8 step 150 or Figure 46, etc.). A laboratory test center, for example as described in Figure 15, includes such a computer as well as a mass spectrometer to conduct mass spectrometry on a blood-based sample. The resulting mass spectrum is subject to pre-processing steps (same as performed on the samples of the classifier development set) and then the classifier is applied to the mass spectral data of the sample. The classifier then generates a class label, e.g., Early or Late, and provides the class label to a requesting physician or clinic as a fee for service.

[0266] With reference to Figure 52, it will be further appreciated that a classifier development system 5200 has been disclosed which includes a mass spectrometer 5202 for conducting mass spectrometry on a development set of samples, or, alternatively and more typically, another independent set of samples, to generate mass spectral data. The data includes intensity data for a multitude of mass spectral features. The system includes a platform 5204 for conducting a gene set enrichment analysis on the development set of samples, or, more typically, the other independent sample set, including a protein assay system such as the SOMAscan system of SomaLogic or the equivalent, and a computer for identifying statistically significant associations of one or more of the mass spectral features with sets of proteins grouped by their biological function. The system further includes a computer 5206 programmed to train a classifier on the development set of samples using the one or more mass spectral features identified by the GSEA platform, e.g., using the procedure of Figure 8. The classifier is in the form of a set of parameters and programmed instructions which assign a class label to a sample of the same type as the development set of samples in accordance with the programmed instructions.

[0267] In this document we use the terms classifier training and classifier development interchangeably, to mean a process of constructing a classifier in a computer (i.e., specifying the parameters for such a classifier) and testing its ability to classify a set of samples (the development set of samples or some subset thereof). Typically, this process occurs in an iterative manner to tweak the parameters to optimize classifier performance, such as by refining class labels assigned to members of the development set, refining filtering parameters, feature deselection, varying the parameter k, etc. It will also be noted that while the present example describes the use of k-nearest neighbor with majority vote as a classification algorithm, in principle the invention can use other supervised learning classification algorithms, such as margin-based classifiers, support vector machine, decision trees, etc., or a classifier configured as a multitude of filtered mini-classifiers combined using a regularization procedure, for example as generated using the procedure of Figure 8.Example 7 (reference)Longitudinal studies

[0268] We conducted an analysis of samples collected during treatment of the nivolumab study (described in Example 1), and specifically at weeks 7 ("WK7") and weeks 13 ("WK13") of the trial. We explored how the classifications for a given patient changed over time, using the full-set classifier of Example 1 and the classifier of Example 2. We found that in some patients the labels changed e.g., an initial class label of Late at the commencement of treatment, followed by Late at week 7 and Early at week 13. As another example, some patients had the class label of Early at commencement of treatment, followed by Early at week 7 and Late at week 13.

[0269] The results for the longitudinal studies using the Example 1 full-set classifier are shown in Figures 30A - 30F. Figure 30A and 30B are Kaplan-Meier plots for overall survival (figure 30A) and time to progression (TTP)(Figure 30B) for Early and Late groups as defined by the baseline classifications. Figures 30C and 30D are Kaplan-Meier plots for overall survival and TTP, respectively for Early and Late groups as defined by the week 7 classifications. Figures 30E and 30F are Kaplan-Meier plots for overall survival and TTP, respectively for Early and Late groups as defined by the week 13 classifications, for the 90 patients for which we had class labels at all three time points. Table 54 is a table of the survival analysis for the plots of Figure 30A-30F. Table 54HR (95% CI)log-rank p valueMediansBaseline OS0.45 (0.21-0.78)0.008Early: 84 weeks, Late: Not reachedWK7 OS0.39 (0.14-0.64)0.002Early: 60 weeks, Late: 113 weeksWK13 OS0.33 (0.14-0.54)< 0.001Early: 61 weeks, Late: Not reachedBaseline TTP0.53 (0.24-1.01)0.055Early: 91 days, Late: 457 daysWK7 TTP0.37 (0.10-0.61)0.003Early: 91 days, Late: 782 daysWK13 TTP0.58 (0.25-1.15)0.112Early: 112 days, Late: 457 days Table 55 shows the distribution of the classifications across the three time points for all samples. Table 55: Distribution of classifications across the three time points: Baseline, WK7, WK13. Missing classifications are denoted by "-". ClassificationsnEarly Early Early15Early Early -8Early - -8Early Early Late1Early Late Early1Early Late Late14Late Early Early3Late Early Late3Late Early -3Late Late Early12Late Late Late41Late Late -6Late - -4

[0270] The majority of classifications remain the same across the available time points (82 / 119 = 69%). There are proportionately more changes from Early to Late than from Late to Early and most of the changes from Early to Late occurred at WK7 and remained Late at WK13. It is possible that this is due to the onset of the immunotherapy treatment. Half of the patients with a change from Late to Early at WK7 reverted back to Late at WK13, when the sample was available. Twelve (16%) of the patients classified as Late at WK7 changed to Early at WK13 and half of these progressed between 70 and 93 days, although three of the others experienced progression-free intervals in excess of 1000 days.

[0271] Figures 31A and 31B show Kaplan-Meier curves that plot the outcomes when the patients are grouped according to their triplet of baseline, WK7, and WK13 classifications. Figure 31A is a plot of overall survival; Figure 31B is a plot of time to progression. In these figures, the groups are labeled by the baseline classification first, the WK7 classification second, and the WK13 classification last (i.e. "Early Late Early" indicates baseline classification of Early, WK7 classification of Late, and WK13 classification of Early). Repeated Early classifications mark particularly poor OS and TTP and, at least in OS, having an Early label at WK 13 indicates poorer prognosis, even when the previous two classifications were Late. However, a Late label at WK13 and WK7 corresponded to better outcomes, even if the baseline classification had been Early. Note: there were too few patients with other label sequences for a meaningful analysis.

[0272] Table 56 shows the medians for the plots of Figures 31A and 31B. Table 56Median OS (weeks)Median TTP (days) from day 84Early Early Early(N=15)4184Early Late Late (N=14)94789Late Late Early (N=12)78Not reachedLate Late Late (N=41)Not reached457

[0273] We repeated this analysis for the classifications produced by the Example 2 classifier over time. The results are generally similar to those presented here for the Example 1 full set classifier. The majority of classifications remain the same across the available time points (59 / 90 = 66% across all three time points, 83 / 107 = 78% across the first two time points). There are proportionately more changes from Early to Late than from Late to Early and most of the changes from Early to Late occurred at WK7 and remained Late at WK13. It is possible that this is due to the onset of the immunotherapy treatment. Most of the patients with a change from Late to Early at WK7 reverted back to Late at WK13, when the sample was available.

[0274] Figures 32A and Figure 32B are Kaplan-Meier plots which show the outcomes when the patients are grouped according to their triplet of baseline, WK7, and WK13 classifications produced by the Example 2 classifier. Figure 32A is a plot of overall survival; Figure 32B is a plot of time to progression. The groups are labeled by the baseline classification first, the WK7 classification second, and the WK13 classification last. Repeated Early classifications mark particularly poor OS and TTP, while repeated Late classifications indicate particularly good OS and TTP. Changing from Early to Late at WK7 and staying Late at WK13 leads to similar outcomes to having a Late classification at all three time points.

[0275] The statistics for Figure 32A are set forth in table 57. Table 57: Medians for the OS plots of Figure 32AMedian OS (weeks)Early Early Early(N=9)47Early Early Late (N=5)99Early Late Late (N=11)Not reachedLate Early Late (N=5)99Late Late Early (N=6)78Late Late Late (N=50)113

[0276] The statistics for Figure 32B are set forth in table 58. Table 58: Medians for the TTP plots of Figure 32BMedian TTP (days) from day 84Early Early Early(N=5)78Early Late Late (N=9)789Late Late Late (N=38)782

[0277] We have some theories for why the class labels changed over time. It is possible that the class label changes were induced by biological changes in the patients caused by the commencement of the nivolumab therapy. It is possible that the changes were due to the influence on tumor size on classification labels (see the discussion below), and large tumor shrinkage that some patients achieve. Whatever the origin of the changes, we do observe that most patients keep their baseline label. Of those patients whose class label changes over time, we observe that when the label changes from Early at baseline to Late later on (week 7 or 13) these patients have relatively good outcome, similar to those patients having a Late baseline class label. Accordingly, in one embodiment, the test of Example 1 or 2 can be conducted periodically over the course of treatment, e.g. every 4, 6 or 8 weeks. By comparing the results and the progression of class labels over time during treatment it may be possible to monitor the therapeutic effect of the nivolumab treatment, or predict the patient's prognosis or overall survival. This treatment monitoring can be direct, i.e., direct changing of some immune status measured by the class label, or indirect, i.e., the change in class label is a proxy or approximation of measurement of tumor shrinkage / expansion. How often one would want to conduct the test and determine the patient's class label during the course of treatment might also depend on whether the change in class label is due to direct action of the drug on the patient's immune system or whether one has to wait for an indirect effect of the treatment on shrinkage (or lack thereof) of the tumor.

[0278] A change of the patient's label over time from Late to Early may be an early indication of lack of efficacy of the drug. This lack of efficacy could be optionally confirmed by conducting radiological studies of the patient, e.g., CT scan to determine tumor size and change as compared to baseline. Potentially, if the changes from Late to Early are due to the drug changing the immune state of the patient directly and if this happens in a relatively short space of time (say 4 weeks or so) the baseline Late label could be an indication to commence treatment with nivolumab and the subsequent Early label could be used to either stop treatment, change treatment to a different treatment (such as combination nivolumab and ipilimumab) or take other action. Another possibility would be to conduct a monitoring test periodically during the first few weeks of treatment and use the class labels to indicate how long the patient needs to take the nivolumab. Currently, patients take the drug until disease progression. This can be a long time and the drug is very expensive. So, if there is a way to tell within the first few months whether a patient could stop nivolumab treatment without detriment to outcome it could result in some savings to health care costs. In any event, in one possible embodiment, the tests of Examples 1 and 2 are conducted periodically over the course of treatment. The class labels are compared over the course of treatment. The status of the class label over the course of treatment can be used to guide treatment or predict the patient's prognosis, either maintain the treatment, stop the treatment, or change the treatment in some fashion such as by combining nivolumab with another drug in a combination treatment regime.

[0279] In one specific example of how the monitoring tests can be done, the initial class label is determined in accordance with Example 1 or Example 2 using the system of Figure 15 (described above), at least once again within the first four weeks of treatment, and at least once again after the first four weeks of treatment.EXAMPLE 8 (reference)Classifiers trained from tumor size information

[0280] We have discovered a method for generating a classifier that takes into account tumor size at baseline which improves the classifier performance. This method and how it is used in practice will be explained in this section. Note that the studies described below used the same 119 samples of Example 1, tumor size data was provided for all patients, and we used the same sample feature table data (mass spectral data for features in Appendix A) as we did for Example 1. The only difference was that feature m / z 9109 was dropped from the feature table as it has possible reproducibility issues and little value for classification.

[0281] Initial attempts at taking account of tumor size in the assignment of a prognostic label for melanoma patients treated with nivolumab indicated that for patients who had available tumor size follow data on treatment, there was a definite influence of tumor size at baseline on the classification we should assign to the samples in the development set. We noticed this by taking the data of the 104 patients for whom we initially had baseline and follow up tumor size data and splitting the set into two: one half with smaller tumors at baseline and the other half with larger tumors at baseline. We then used the classifier development method of Figure 8 as we had done to make the classifier of Example 1, and made separate classifiers, one for patients with smaller tumors and one for patients with larger tumors. We then proceeded to classify the samples in the development set using either the large or small tumor classifiers, depending in the size of the tumor at baseline.

[0282] We noticed that some of the smaller tumors classified by the Example 1 full-set classifier as Late now got an Early classification and some of the larger tumors that had been classified by the Example 1 full set classifier as Early were now classified as Late. In particular, some of the small tumors that previously were classified as Late and which had demonstrated huge tumor growth on treatment were now classified as Early and some of the large tumors that previously had been classified as Early and which had shrunk considerably in the first 26 weeks of treatment were now classified as Late. Plotting the Kaplan-Meier plots for this subset of 104 patients, taking the classifications from the two separate classifiers, as defined by pretreatment tumor size, increased the hazard ratio between the new Early and Late groups, as shown in Figures 33A and 33B. In these figures, "Original Early / Late" are the classification groups defined for the 104 patient subset using Example 1 full-set classifier, approach 1, and "TSAdjusted Early / Late" are the classifications generated by the new classifiers for the smaller and larger tumors, each applied to the samples with smaller and larger tumors, respectively.

[0283] The results of the survival analysis comparison between the Early and Late groups adjusted for tumor size are given in table 58. Table 58: Performance of classifications obtained for the subset of 104 patients when adjusted for tumor size# Early / # LateHR OS (95% CI)Log-rank pMedian OS Earlier / Later (weeks)HR TTP (95% CI)Log-rank pMedian TTP Earlier / Later (days)TS Adjusted44 / 600.24 (0.11-0.37)<0.00169 / not reached0.36 (0.18-0.49)<0.00188 / 541Original36 / 680.42 (0.20-0.68)0.00273 / not reached0.57 (0.32-0.88)0.015157 / 362 These results indicated to us that we were not making optimal decisions on classification for some of the samples with the smallest and largest tumors, and that we could improve by taking tumor size into account when designing and generating the classifier of Example 1.

[0284] We then obtained baseline tumor size data for the remaining 15 samples. When we applied the new classifiers to these samples (depending on whether they were in the small tumor size group or large tumor size group), and added these patients into the Kaplan-Meier analysis, we noticed that the improved separation almost disappeared. Apparently, we were doing a worse job of classifying these 15 samples than we had done before. We also tried to carry out the same approach as above training on all 119 samples, but again the result was more or less no improvement from our initial classification. These observations led us to the conclusion that these 15 patients whom we had initially omitted were essentially different - indeed they were omitted because they did not reach a follow up tumor size assessment, having very early progression (all 15 patients progressed before 78 days). We hypothesized that for patients who progress very quickly, tumor size plays a much weaker role than it does for the patients who remain progression free for a longer period of time. To try to keep the improvement noted above for the 104 patients reaching the 26 week assessment and classify the other 15 samples correctly, we decided to first find a classifier to remove the patients progressing the fastest, and then repeat classifier development by tumor size for the remaining samples. That is, we wanted to remove from classifier development those samples with the patients progressing fastest, and then conduct a new classifier development taking into account tumor size, and generate a "small tumor" classifier and a "large tumor" classifier. These new classifiers are designed for later use in testing a patient for immune checkpoint inhibitor benefit, with the additional input at the time of testing data on whether the patient has a "large" or "small" tumor and then using the appropriate large or small tumor classifier. It will be apparent from the following discussion that the methodology we describe below may be useful generally in generating "small tumor" and "large tumor" classifiers in the oncology setting.

[0285] In order to remove the patients progressing the fastest, we returned to the full set Example 1 classifier. Using this classifier, we divided the samples in the development set into Early (N=47) and Late (N=72) groups. We took the Early group of 47 samples and used the same methodology of Figure 8 as detailed above, using the Early group of 47 samples as the input development sample set. In performing the new classifier development we performed label flips for misclassified samples in an iterative manner until convergence to make a classifier that splits these 47 samples into two sub-groups, which we called "Earlier" and "Later". The initial class definitions (Figure 8, step 102) were based on shorter and longer OS and we used filtering (Figure 8 step 126) based on hazard ratio for OS between the classification groups of the training set. This produced a classifier that split the 47 Early patients into two groups with better ("Later") and worse ("Earlier") outcomes.

[0286] The Kaplan-Meier plots for OS and TTP for the groups generated by this classifier are shown in Figures 34A and 34B. Twenty two patients were assigned to the Earlier group and 25 to the Later group. The results of the survival analysis comparison between the Earlier and Later groups are given in table 59. Table 59: Performance of classifier developed on only full-set classifier "Early" samples# Earlier / # LaterHR OS (95% CI)Log-rank pMedian OS Earlier / Later (weeks)HR TTP (95% CI)Log-rank pMedian TTP Earlier / Later (days)22 / 250.57 (0.27-1.10)0.09426 / 730.60 (0.31-1.07)0.08577 / 132

[0287] It is apparent that the Earlier group has particularly poor outcomes in terms of TTP and OS. We decided to remove these 22 samples from further analysis and leave them their already assigned "Early" classification. We then conducted two new classifier developments, again using the procedure of Figure 8, one for large tumors and one for small tumors, using the remaining 97 samples of the initial development sample set. This set of 97 samples was split into two groups depending on tumor size: the smallest 49 samples were used to generate one classifier (small tumor classifier) and the largest 48 samples were used to generate another classifier (large tumor classifier). Both classifiers were trained as before using the procedure of Figure 8 in an iterative manner, with label flips for misclassified samples until convergence, with the initial class assignments of "Early" and "Late" (Figure 8, step 102) defined based on duration of OS.

[0288] When the data for the patients with larger tumors, patients with smaller tumors and quickly progressing patients (Early / Earlier classification) were combined, the Kaplan-Meier plots of Figures 35A and 35B were obtained. The results of the survival analysis comparison between the Early and Late groups are given in table 60. Table 60: Performance of classifications obtained for all 119 patients when adjusted for tumor size, first removing the 22 patients classified as having especially poor outcomes# Early / # LateHR OS (95% CI)Log-rank pMedian OS Earlier / Later (weeks)HR TTP (95% CI)Log-rank pMedian TTP Earlier / Later (days)Final60 / 590.32 (0.19-0.51)<0.00161 / not reached0.40 (0.24-0.57)<0.00183 / 490Original47 / 720.38 (0.19-0.55)0.00261 / not reached0.50 (0.29-0.71)0.00184 / 230 The final Late group has slightly better outcomes than the original Late group and is composed of significantly fewer patients. Outcomes in the final Early group are quite similar to those of the original Early group, although its size has increased by 28%. The hazard ratios between the groups are slightly better for both endpoints than they were for the original Example 1 classifications.

[0289] Investigation of classifier performance in this context can also be made by plotting the percent change in tumor size from baseline to a later point in time (e.g., 26 weeks after commencement of treatment) for each member of the development set, and indicating in such a plot whether the data points represent Early or Late classified patients. Such plots, known as "waterfall plots," are shown in Figures 36A and 36B. The waterfall plots show the percentage reduction in tumor size for the 104 patients assessable at the 26 week evaluation. Figure 36A is the plot for the "final" classifiers (i.e., taking into account tumor size and using either the large or small tumor classifier). Figure 36B is the plot for the original Example 1 full-set classifier.

[0290] What is noticeable from comparing the plots of Figures 36A and 36B is that the new final classifications using tumor size classifiers is considerably better at classifying the patients with tumor growth as Early. That is, the majority of patients having significant tumor growth over the course of treatment were classified as Early when the tumor size classifiers were used, as would be expected given the clinical meaning the Early class label. Moreover, the majority of the patients with significantly diminished tumor size over the course of treatment were classified as Late, which is also expected. However, comparing the right hand side of Figures 36A and 36B, the new classifiers using tumor size data performed slightly worse in identifying the patients with tumor shrinkage as Late as compared to the Example 1 full set classifier.

[0291] A preferred method for generating classifiers using tumor size data in a development sample set can be summarized and explained in flow chart form. Referring now to Figure 37, the classifier development process is shown at 3700 and includes a first step 3702 of removing the samples of patients from the development sample set who progressed fastest after commencement of treatment. This step 3702 is shown in detail in Figure 39 and will be explained in detail below.

[0292] At step 3704, once these samples are removed from the development set, the samples remaining in the development set are sorted based on tumor size at baseline into small tumor and large tumor groups, 3706 and 3707, respectively.

[0293] At step 3708, the small tumor samples (associated mass spectral data, feature values of features listed in Appendix A) are used as the development set for generating a small tumor classifier using the procedure of Figure 8.

[0294] At step 3710, the performance of the classifier developed at step 3708 is then verified by using the classifier to classify the small tumor samples 3706.

[0295] At step 3712, the parameters of the small tumor classifier generated at step 3708 are then stored for later use in classifying test samples for patients with small tumors. These parameters include, inter alia, the data identifying the small tumor sample mass spectra data sets forming the reference set for classification; the feature values at predefined mass spectral features for the reference set; identification and kNN parameters of the mini-classifiers passing filtering; logistic regression weights derived from the combination of mini-classifiers with drop-out regularization; and the definition of the final classifier from the master classifiers generated during the performance of Figure 8 on the small tumor development sample set (Figure 8B, step 150).

[0296] The steps 3714, 3716 and 3718 are performed on the large tumor samples 3707, exactly the same as for steps 3708, 3710 and 3712 described above.

[0297] The manner of use of the classifiers generated in accordance with Figure 37 is shown in Figure 38. The classifiers are used for conducting a test of blood-based sample of a melanoma or other cancer patient to determine whether they are likely or not to obtain benefit from an immune checkpoint inhibitor in treatment of cancer, such as anti-PD-1 antibody or anti-CTLA4 antibody. This process is shown at 3800. At step 3802, tumor size data for the patient is obtained. Such data could be obtained from CT or PET scan data of the patient. The tumor size data ideally will accompany a blood-based sample provided for testing. Alternatively, some surrogate or proxy for tumor size data could be used (for example, some combination of mass spectral features alone or combined with other serum proteins measured by alternative methods, such as ELISA). At step 3804, the determination is made as to whether the tumor is "large" or "small", again using this data. The criteria for this determination could take the form of criteria used to sort samples in step 3704 of Figure 37.

[0298] If the tumor size is "small", then at step 3806 the sample is classified using the small tumor classifier generated and stored at step 3712 of Figure 37, using the system of Figure 15. That is, the blood-based sample is subject to mass spectrometry, and the mass spectrometry data is subject to the steps shown in Figure 15 with the classifier used for classification being the small tumor classifier generated in Figure 37. At step 3808 the class label Early or Late is assigned to the sample. If the patient is identified as Late, the patient is predicted to obtain benefit from the immune checkpoint inhibitor and have improved overall survival as compared to a class label of Early.

[0299] If at step 3804 the tumor size is "large", the large tumor classifier of Figure 37 is then used to classify the blood-based sample of the patient using the system of Figure 15. At step 3812 the class label Early or Late from the classifier is reported. The Early and Late labels have the same meaning as explained in the previous paragraph.

[0300] Figure 39 is a flow chart showing a procedure 3702 of removing the fastest progressing samples from a development set as a preliminary step in generating the large and small tumor classifiers of Figure 37. At step 3904, a classifier is generated over all the samples in a development sample set using the procedure of Figure 8. An example of this is the full set classifier of Example 1, described above.

[0301] At step 3906, this classifier is then used to classify all the samples in the development sample set. Each member of the development sample set is then classified as either Early or Late. The Early and Late patients are grouped into two groups shown at 3908 and 3910.

[0302] At step 3912, a new classifier is generated using the process of Figure 8, with the Early patient group 3908 forming the development sample set of Figure 8. The process Figure 8 is performed in an effort to split the Early patients into Earlier and Later sub-groups. An example of this was described at the beginning of this section of the document and the results shown in Figures 34A and 34B. At step 3914, after the classifier of 3912 has been generated it is applied to all the Early samples (3908) and the resulting classifications of the Early patients into "Earlier" and "Later" sub-groups 3916 and 3918 is performed. The Earlier patients 3916 are then identified and removed from the development sample set. The process of Figure 37 then proceeds at step 3704 with the development sample set minus this "Earlier" sub-group of patients to produce the small tumor and large tumor classifiers.

[0303] From the above discussion, it will be apparent that the disclosure describes a machine (e.g., Figure 15, 1510) programmed as a classifier for classifying a cancer patient as likely or not likely to benefit from an immune checkpoint inhibitor. The machine 1510 includes a memory 1514 storing parameters of a small tumor classifier and a large tumor classifier, and a reference set of class-labeled mass spectral data for each of the small tumor classifier and the large tumor classifier. The reference sets are obtained from blood-based samples of other cancer patients treated with the immune checkpoint inhibitor, as explained above. The machine further includes a processing unit (Figure 15, 1512) executing a classifier defined by the parameters stored in the memory. The parameters defining the classifier for each of the small tumor classifier and large tumor classifier include parameters defining a classifier configured as a combination of filtered mini-classifiers with drop out regularization, e.g., resulting from the procedure of Figure 8 steps 102-150. The mass spectral data can be obtained from performing MALDI-TOF mass spectrometry on the blood-based samples and wherein each of the samples is subject to at least 100,000 laser shots, e.g. using the so-called Deep MALDI methods described in Example 1.

[0304] A method of generating a classifier for classifying cancer patients as likely or not likely to benefit from a drug has been described, comprising the steps of: 1) obtaining a development sample set (Fig. 8, 100) in the form of a multitude of blood-based samples; 2) conducting mass spectrometry on the development sample set (see Example 1); 3) removing from the development sample set samples from patients with a relatively fast progression of disease after commencement of treatment by the drug; (step 3702, Figure 37) 4) sorting the remaining samples based on tumor size at baseline (commencement of treatment) into large and small tumor groups; (Figure 37 step 3704); 5) for the small tumor group: a) generating a small tumor classifier using the small tumor group as an input development sample set in a classifier development exercise; (Figure 37, step 3708) b) verifying the performance of the small tumor classifier in classification of the members of the small tumor sample group; (Figure 37, step 3710) and c) storing the parameters of the small tumor classifier; (Figure 37, step 3712); and 6) for the large tumor group: a) generating a large tumor classifier using the large tumor group as an input development sample set in a classifier development exercise; (Figure 37 step 3714) b) verifying the performance of the large tumor classifier in classification of the members of the large tumor sample group; (Figure 37 step 3716) and c) storing the parameters of the large tumor classifier (Figure 37, 3718).

[0305] Preferably, as explained above in this section, the classifier development exercise of step 5a) and step 6a) takes the form of implementing the procedure of Figure 8 steps 102-150.

[0306] A method of classifying a cancer patient as likely or not likely to benefit from a drug is contemplated. The method includes the steps of a) determining whether the patient has a large or small tumor; (Figure 38, 3802); b) if the patient has a large tumor, using the large tumor classifier generated in the method described above to classify a blood-based sample of the patient as likely or not likely to benefit from the drug, (Figure 38, 3806) and c) if the patient has a small tumor, using the small tumor classifier generated as described above to classify a blood-based sample of the patient as likely or not likely to benefit from the drug.

[0307] In still another aspect, a method of classifying a cancer patient as likely or not likely to benefit from a drug is contemplated. The method includes the step of a) conducting two classifier generation exercises on a development set of samples which are sorted into small tumor and large tumor groups, resulting in the generation of a large tumor classifier and a small tumor classifier and storing the large tumor and small tumor classifiers in a programmed computer (Figure 37). The development set of samples are blood-based samples which have been subject to mass spectrometry. The method includes a step b) of determining whether the patient has a large or small tumor, either directly from tumor measurement data or indirectly using a surrogate for tumor measurement data (Figure 38 step 3084). The method further includes a step c) of conducting mass spectrometry on the blood-based sample of the cancer patient (Figure 15, 1506, 1508). If the patient has a large tumor, the method includes a step of using the large tumor classifier generated in step a) and the mass spectrometry data obtained in step c) with the programmed computer to classify the patient as likely or not likely to benefit from the drug, and if the patient has a small tumor, using the small tumor classifier generated in step a) and the mass spectrometry data obtained in step c) with the programmed computer to classify the patient as likely or not likely to benefit from the drug.

[0308] In one example, the cancer patient is a melanoma patient, and the drug is an antibody drug targeting programmed cell death 1 (PD-1). However the methods described in Figures 37-39 are applicable to other types of cancer patients and drugs.Example 9Development of an ensemble of classifiers from clinically different classifier development sets and use thereof to guide treatment

[0309] Our work described in Example 8 made use of the development of different classifiers using clinically different development sets, i.e., with one set from "small tumor" patients and another set from "large tumor" patients. In this Example, we extend this method of developing classifiers more generally and describe an ensemble of different classifiers, each derived from clinically different development sets. In one implementation of this Example, each development set represents different tumor sizes or different proportions of tumor sizes in a population of melanoma patients. From this approach, we have discovered a reproducible ternary classification method and system which is better able to identify patients who do so badly on the immune checkpoint inhibitor anti-PD-1 that they might be better not taking it at all and, perhaps more importantly, others that do so well on anti-PD-1 monotherapy that they might be just as well off taking anti-PD-1 monotherapy rather than undergoing anti-PD1 / anti-CTLA4 combination therapies, which incur tremendous addition expense and can have severe toxicity side effects. Later in this Example, we describe the development and implementation of an ensemble of classifiers to predict survival of ovarian cancer patients on chemotherapy.

[0310] Accordingly, in this section we describe a different approach to classifier development that we have not considered before, namely designing the development sets of a set of classifiers to explore different clinical groups, and using an ensemble of classifiers obtained from such development sets to result in a, for example, ternary (three-level) classification scheme. We further describe how a class label produced from this ensemble of classifiers can be used to guide treatment of a cancer patient or predict survival of a cancer patient. Those skilled in the art will appreciate that the present example of designing the development set of a set of classifiers with different clinical groups is offered by way of example and not limitation, and that this methodology can be extended to other classifier development scenarios generally, including in particular other classifier developments to predict patient benefit or survival from treatment with drugs.A. Ensemble of classifiers for melanoma patient benefit from nivolumab

[0311] In the melanoma / nivolumab portion of this Example, the deep MALDI feature table for the pretreatment serum samples from patients treated with nivolumab at the Moffitt Cancer Center (see Example 1 and Appendix A) was used for classifier development. For classifier development, the 104 samples for the patients who had tumor size follow up data were used. These 104 samples were split into two groups according to baseline tumor size: the 50 patients with smallest tumors and the 54 patients with largest tumors. Each of these subsets was used as the development set to develop a classifier using the process of Figure 8, with bagged feature deselection and filtering of mini-classifiers on overall survival. These aspects have been described previously in this document in Example 8 and Appendix F of our prior provisional application serial no. 62 / 289,587 from which this case claims priority.

[0312] In addition, five other subsets of the 104 sample classifier development set were defined as additional or alternative development sets. The first of these took the set of 50 patients with smallest tumors, dropped 10 of them, and replaced these with 10 patients from the set of 54 with the larger tumors. The second of these took the set of 50 patients with smallest tumors, dropped 20 of them, and replaced these with 20 patients from the set of 54. Three other development sets were defined extending this approach further. The fifth classifier was accordingly a subset of the original 54 large tumor size set. In this way, 5 development sets of 50 patient samples were generated that contained different proportions of patients with smaller and larger tumor sizes (80%-20%, 60%-40%, 40%-60%, 20%-80%, and 0%-100%, respectively). For each of these 5 development sets, classifiers were generated using the same procedure of Figure 8 described in detail above, i.e., each classifier was defined as a final classifier (Fig. 8, step 150) as an ensemble average over 625 master classifiers generated from 625 test / training splits of the development set used for that classifier, and each master classifier is a logistic regression combination of a multitude of mini-classifiers that pass overall survival performance filtering criteria, and regularized by extreme drop out. Each classifier produces a binary class label for a sample, either Early or Late, and Early and Late have the same clinical meaning as explained in Example 1. Hence, we obtained an ensemble of 7 different classifiers (the 5 developed as described here, plus the "large" and "small" tumor classifiers described in the "Classifiers incorporating tumor size information" section, Example 8), each of which was developed on a clinically different classifier development set. It will be noted that the "large" tumor classifier described in the "Classifiers incorporating tumor size information" section and the fifth of the new classifiers generated from 50 "large" tumor patients are similar, but distinct in that they were formed from different sets of patients. This ensemble of seven classifiers is referred to herein as "IS6" or "the IS6 classifier."

[0313] An alternative method for defining the classifier development sets with different clinical groupings is as follows: 1. Order the 104 samples by tumor size. 2. Take the 50 samples with the smallest tumor size for one classifier development and the remaining 54 samples with the largest tumor for another, just as here. 3. Define 5 other classifier development sets by a. Dropping the 10 samples with the smallest tumor size and taking the next 50 samples for a classifier development set. b. Dropping the 20 samples with the smallest tumor size and taking the next 50 samples for a second classifier development set. c. Dropping the 30 samples with the smallest tumor size and taking the next 50 samples for a third classifier development set. d. Dropping the 40 samples with the smallest tumor size and taking the next 50 samples for a forth classifier development set. e. Dropping the 50 samples with the smallest tumor size and taking the next 50 samples for a fifth classifier development set.

[0314] Classifiers are then developed from each of these seven classifier development sets using the procedure of Figure 8 steps 102-150. One then establishes rules to combine the classification results from these seven classifiers, e.g., as explained below. This method of designing classifier development sets may have similar performance as the classifiers produced from the development sets described in the previous paragraphs, but may be more reproducible, for example in a rerunning of the samples or have better performance in identifying patients with particularly good or poor outcomes.

[0315] To conduct a test on a patient's blood-based sample, the sample is subject to mass spectrometry as described above in the description of Figu...

Claims

1. A method of guiding non-small cell lung cancer (NSCLC) patient treatment with immunotherapy drugs, comprising: a) conducting mass spectrometry on a blood-based sample of the patient and obtaining mass spectrometry data; b) obtaining integrated intensity values in the mass spectrometry data of a multitude of mass-spectral features; and c) operating on the mass spectral data with a programmed computer implementing a classifier, wherein the classifier is obtained from filtered mini-classifiers combined using a regularized combination method comprising repeatedly conducting logistic regression with extreme dropout on the filtered mini-classifiers; wherein in the operating step the classifier compares the integrated intensity values with feature values of a reference set of class-labeled mass spectral data obtained from blood-based samples obtained from a multitude of other cancer patients treated with an antibody drug blocking ligand activation of programmed cell death 1 (PD-1) with a classification algorithm and generates a class label for the sample, wherein the class label "Good" or the equivalent predicts the patient is likely to obtain similar benefit from a combination therapy comprising an antibody drug blocking ligand activation of PD-1 and an antibody drug targeting CTLA4 and is therefore guided to a monotherapy of an antibody drug blocking ligand activation of PD-1, whereas a class label of "Not Good" or the equivalent indicates the patient is likely to obtain greater benefit from the combination therapy as compared to the monotherapy of an antibody drug blocking ligand activation of programmed cell death 1 (PD-1) and is therefore guided to the combination therapy, wherein the mass spectral features include a multitude of features listed in Appendix A, Appendix B or Appendix C.

2. The method of claim 1, wherein the mini-classifiers are filtered in accordance with criteria listed in Table 10.

3. The method of claim 1, wherein the classifier comprises an ensemble of tumor classifiers combined in a hierarchical manner.

4. The method of claim 1, wherein the relatively greater benefit from the combination therapy label means significantly greater (longer) overall survival as compared to monotherapy.

5. The method of claim 1, wherein the reference set comprise a set of class-labeled mass spectral data of a development set of samples having either the class label Early or the equivalent or Late or the equivalent, wherein the samples having the class label Early are comprised of samples having relatively shorter overall survival on treatment with nivolumab as compared to samples having the class label Late.

6. The method of claim 1, wherein the mass spectral data is acquired from at least 100,000 laser shots performed on the sample using MALDI-TOF mass spectrometry.

7. The method of claim 1, wherein the mass-spectral features are selected according to their association with the biological functions Acute Response and Wound Healing.