JP2021501422A - Temozolomide reaction predictors and methods - Google Patents

Abstract

The system and method intended uses a reaction prediction model for temozolomide, which is based on RNAseq information, protein quantification information, and methylation information and has at least 85% predictive accuracy.

Description

This application is a co-pending US provisional application with application numbers 62 / 579,127 filed on October 30, 2017 and application numbers 62/727,245 filed on September 5, 2018, both of which are herein in their entirety. Claim priority based on (incorporated).

The field of the present invention is a system and method for predicting a patient's drug response to temozolomide, especially when the patient is diagnosed with cancer.

The background description includes information that may be helpful in understanding the present invention. None of the information provided herein has been found to be prior art or related art of the claimed invention, and any specific or implied reference publication is in prior art. It was not recognized as being.

All publications and patent applications described herein are incorporated by reference to the same extent that each individual publication or patent application is specifically and individually indicated and incorporated by reference. If the definition or use of a term in an incorporated reference conflicts with or conflicts with the definition of that term provided herein, the definition of that term provided herein applies and is incorporated into the reference. The definition of the term does not apply.

Temozolomide (TMZ) is a chemotherapeutic agent used as the standard treatment for glioblastoma and melanoma, and has recently shown limited but promising activity in patients with metastatic colorectal cancer (mCRC). Was done. TMZ is an agent having an alkylating / methylating activity on the N-7 or O-6 position of a guanine residue in DNA, and often triggers cell death of susceptible cells. However, various DNA damage repair enzymes, especially O-6-methylguanine-DNA methyltransferase (MGMT), can counteract the effects of temozolomide in at least some of the tumor cells. More recently, epigenetic silencing of the MGMT gene has been reported, and tumor cells with epigenetic silencing of the MGMT gene have been found to be more susceptible to TMZ killing.

Therefore, MGMT is considered a TMZ resistance marker. Usually, the MGMT expression status of tumors can be evaluated by the digital polymerase chain reaction (PCR) method known as methyl-BEAMing (MB), and a cutoff of> 60% MGMT methylation predicts TMZ response. In another approach, mass spectrometry (MS) proteomic analysis can objectively quantify MGMT proteins and other available protein biomarkers in formalin-fixed paraffin-embedded (FFPE) tissue sections. In this case, a 200 amol / μg MGMT protein cutoff (predetermined based on the detection limit of the assay) is a predictor of response in TMZ-treated mCRC patients. The amount of MGMT protein can also correlate with MGMT methylation status.

Among mCRC patients treated with TMZ, those whose tumors expressed low or undetectable levels of MGMT protein had longer mPFS than their counterparts with higher MGMT protein levels. 80% between MGMT protein expression quantified by mass spectrometry and MGMT methylation status by MB, as presented in Abstract # 11601 of ASCO Annual Meeting, Chicago, Illinois (June 2-6, 2017). Correlation was observed. Quantitative proteomics objectively measured MGMT protein in FFPE tumor samples and retrospectively identified 9 of 9 responders to TMZ. The digital PCR methylation assay (methyl-BEAMing) retrospectively identified 7 of 8 responders to TMZ. Therefore, the researchers concluded that quantitative proteomic analysis of MGMT could potentially be used to select patients with TMZ therapy mCRC.

However, such an approach was considered proteomic and methylation analysis only in the retrospective manner. In addition, the detection limits of mass spectrometry and the potential for deficiencies in methylation detection potentially further reduce the accuracy of responder analysis. In fact, it was observed that there was only about 80% agreement on the reaction association with MGMT between mass spectrometry and methylation analysis. In addition, the authors did not present any analytical options that could support or suggest response prediction with clinically useful predictive accuracy.

Therefore, even if various systems and methods for predicting a specific drug response are known in the art, it enables highly reliable, simple and robust treatment prediction of the drug, and predicts the treatment response in a patient-specific manner. There remains a need for systems and methods that enable it.

The subject of the present invention relates to various devices, systems, and methods for predicting a therapeutic response to temozolomide in the treatment of solid tumors in patients. In one aspect of the subject matter of the present invention, the inventors contemplate a method of predicting a therapeutic response to temozolomide in a patient. The method includes a step of providing RNAseq information, protein quantification information, and methylation information from a patient's tumor and another step of calculating a reaction prediction for temozolomide by a reaction prediction model. RNAseq information, protein quantification information, and methylation information are used.

Most preferably, the reaction prediction model uses the K-nearest neighbor approach, and RNAseq information, protein quantification information, and methylation information are subgrouped. For example, RNAseq information can be subgrouped using a log2 (TPM + 1) cutoff value of 3.5, protein quantification information can be subgrouped using a cutoff value of 200 amol / mL, and / or methylation information can be subgrouped. It is subgrouped using the cutoff value of 60% promoter CpG methylation. Most typically, the reaction prediction model has a prediction accuracy of at least 80% or at least 85%.

RNAseq information, protein quantification information, and methylation information are provided from FFPE or fresh tumor samples, and the tumor may be a solid tumor (eg, metastatic colon cancer, glioblastoma, or melanoma). It is planned.

Another aspect of the subject matter of the present invention includes a method of treating a patient having a tumor. The method includes a step of providing RNAseq information, protein quantification information, and methylation information from a patient's tumor (for example, a solid tumor) and a step of calculating the reaction probability to temozolomide by a reaction prediction model, and includes a reaction prediction model. Uses RNAseq information, protein quantification information, and methylation information. The method is then followed by the step of administering temozolomide to a patient with a temozolomide reaction probability of> 0.5. Preferably, the reaction prediction model uses the K-nearest neighbor approach and / or the reaction prediction model has a prediction accuracy of at least 85%.

In some embodiments, RNAseq information, protein quantification information, and methylation information are subgrouped. Preferably, RNAseq information is subgrouped with a log2 (TPM + 1) cutoff value of 3.5 and / or protein quantification information is subgrouped with a cutoff value of 200 amol / mL and / or. Methylation information is subgrouped using the 60% promoter CpG methylation cutoff value.

In some embodiments, RNAseq information, protein quantification information, and methylation information are provided from FFPE samples or fresh tumor samples. In some embodiments, the tumor is metastatic colon cancer, glioblastoma, or melanoma.

The various objectives, features, aspects, and advantages of the subject matter of the present invention will become apparent from the following detailed description of preferred embodiments in conjunction with the accompanying drawings.

The samples and assays used in the experimental studies are shown. FIG. 2A shows a graph of the percentage change (from baseline) in tumor volume between TMZ-treated patients (n = 41) by MGMT protein status and (FIG. 2B) MGMT promoter hypermethylation status. FIG. 3 shows a graph of percentage changes (from baseline) in tumor volume between TMZ treated patients. FIG. 3 shows graphs of progressionless survival (PFS, FIG. 4A) and overall survival (OS, FIG. 4B) of patients with metastatic colorectal cancer who received TMZ treatment by MGMT protein expression level. FIG. 5 shows a graph of PFS (FIG. 5A) and OS (FIG. 5B) of patients with metastatic colorectal cancer who received TMZ treatment stratified by MGMT methylation status. The graph of PFS (FIG. 6A) and OS (FIG. 6B) of the metastatic colorectal cancer patient who received TMZ treatment by RNA-seq analysis is shown. The graph of the metastatic colorectal cancer patient who received TMZ treatment by the MGMT protein expression level is shown. It is a graph which schematically exemplifies the cutoff value of RNAseq data. It is a graph which schematically exemplifies the agreement between the RNAseq value and the proteomic value. It is a graph which schematically exemplifies the agreement between the RNAseq value and the proteomic value. A bar graph of the match between the amount of MGMT protein and MGMT methylation is shown. FIG. 5 is a graph depicting PFS (progressive survival) and OS (survival) for the MGMT protein level subgroup. FIG. 5 is a graph depicting PFS (progressive survival) for the MGMT RNAseq level subgroup. It is a graph which drew PFS (progressive survival) for one combination of MGMT subgroups. FIG. 5 is a graph depicting PFS (progressive survival) and OS (survival) for the other combination of MGMT subgroups. It is a graph which drew the temozolomide reaction prediction accuracy based on various input variables for various classifiers. The graph of the average accuracy of the prediction model by the leave pair-out cross-validation of Example 2 is shown. A graph of the mean predictive accuracy of unknown samples for 58 predictive modeling strategies by MGMT assessment method group is shown. The groups are ordered from left to right by the average accuracy of Example 2. A schematic diagram of machine learning for predicting drug reaction from various types of data is shown. The flow chart of the regression and classification pipeline of the prediction model is shown. The graph of the relationship between the total expression and the MGMT protein determined by various regression models is shown. The graph of the relationship between MGMT protein and MGMT gene determined by various regression models is shown. The graph of the accuracy value using the methylation value is shown. Another graph of the accuracy value using the methylation value is shown. The heat map of the training data set and the test data set of the prediction model is shown.

The inventor has now used a model based on a combination of tumor (preferably subgrouped) RNAseq information, protein quantification information and promoter methylation information to provide unexpectedly high accuracy and clinically useful temozolomide. We found that a predictive model could be established. In a particularly preferred embodiment, the model is based on the K-nearest neighbor approach.

Although posterior associations between MGMT protein levels or promoter methylation and temozolomide reactions are generally known, such associations are not always predicted and, of course, not with high accuracy. You should be aware. The present inventor has now quantified and preferably subgrouped parameters from a patient's tumor: (1) RNAseq information, especially those measured in TPM units and (2) protein quantification information, especially measured by mass analysis. It was observed that it is possible to establish an extremely accurate prediction model for predicting the patient's therapeutic response by combining (3) methylation information, especially that measured in the MGMT promoter region.

Therefore, the present inventor contemplates a method of predicting a therapeutic response to temozolomide in patients with tumors. The method comprises providing RNAseq information, protein quantification information and methylation information from a patient's tumor. Next, a reaction prediction model that takes into account RNAseq information, protein quantification information, and methylation information is used to establish a reaction prediction for temozolomide. As will be readily apparent, the type of information will determine the nature of the sample, at least to some extent.

As used herein, the term "tumor" refers to one or more cancer cells, cancer tissues, malignant tumor cells, or malignant tumors that can be located or found in one or more anatomical locations in the human body. Means tissues and is used synonymously with them. As used herein, the term "patient" includes both those who have been diagnosed with a condition (eg, cancer) and who are being tested and / or tested for the purpose of detecting or identifying the condition. Should be noted. Therefore, a patient with a tumor means both those who have been diagnosed with cancer and those who are suspected of having cancer. As used herein, the terms "provide" or "providing" make, produce, place, make available, move, or make ready to use. Meaning and including any act.

Therefore, in most aspects of the subject matter of the invention, tumor samples will be used to obtain all relevant information. Any suitable method of obtaining a patient's tumor sample (tumor cells or tumor tissue) (or, by comparison, healthy tissue of the patient or healthy person) is contemplated. Most typically, tumor samples can be obtained from the patient via biopsy (including liquid biopsy or via tissue resection during surgery, independent biopsy procedures, etc.). It can be in a fresh state or in a state that has been processed to further processes to obtain tissue omics data (eg, frozen formalin-fixed paraffin-embedded (FFPE) samples, etc.). For example, tumor cells or tissue can be fresh or frozen. In another example, the tumor cell or tissue can be in the form of a cell / tissue extract. In some embodiments, the tumor sample can be obtained from a single tissue or anatomical region or from multiple different tissues or anatomical regions. For example, metastatic breast cancer tissue can be obtained from the patient's breast and, for metastatic breast cancer tissue, from other organs (eg, liver, brain, lymph nodes, blood, lungs, etc.). Preferably, it is possible to obtain a healthy or corresponding normal tissue of the patient (eg, the non-cancerous breast tissue of the patient), or a healthy tissue of a healthy person (other than the patient) through a similar method for comparison. It is also possible to obtain.

In some embodiments, the tumor sample is available from the patient at multiple time points to determine any changes in the tumor sample over the relevant period. For example, a tumor sample (or a sample suspected of having a tumor) can be obtained before and after the sample is determined or diagnosed as cancerous. In another example, a tumor sample (or a sample suspected of having a tumor) is before, then, and / or after (eg, termination, eg, radiation therapy, chemotherapy, immunotherapy, etc.) a single or series of antitumor treatments. Can be obtained at times). In yet another example, a tumor sample (or a sample suspected to be a tumor) may be obtained during tumor progression in identifying new metastatic tissue or cells.

From the resulting tumor cells or tissues, DNA (eg, genomic DNA, extrachromosomal DNA, etc.), RNA (eg, mRNA, miRNA, siRNA, shRNA, etc.), and / or proteins (eg, membrane proteins, cytoplasmic proteins, etc.) Omics data can be obtained by isolating (such as nuclear proteins) and further analyzing. Alternatively and / or additionally, the step of obtaining omics data may include receiving omics data from a database that stores omics information of one or more patients and / or healthy individuals. For example, omics data for a patient's tumor can be obtained from DNA, RNA, and / or proteins isolated from the patient's tumor tissue, and the omics data obtained can be for tumors of the same type or different types. It can be stored in a database (eg, cloud database, server, etc.) along with other omics datasets of other patients who have it. Correspondence of a healthy person or patient Omics data obtained from a normal tissue (or healthy tissue) can also be stored in a database so that relevant datasets can be retrieved from the database during analysis. Similarly, when obtaining protein data, these data may also include protein activity, especially if the protein has enzymatic activity (eg, polymerase, kinase, hydrolase, lyase, ligase, oxidoreductase, etc.).

As used herein, omics data is, but is not limited to, information related to genomics, proteomics, and transcriptomics, as well as specific gene expression or transcript analysis, as well as other cells. Includes properties and biological functions of. With respect to genomics data, suitable genomics data include whole-genome sequencing and / or exome sequencing (typically at least 10x, more typically at least 20x) coverage of both tumor and corresponding normal samples. Depth) includes DNA sequence analysis information that can be obtained. Alternatively, DNA data can also be provided from sequence records already established from the preceding sequencing (eg, SAM, BAM, FASTA, FASTQ, or VCF files). Thus, datasets may include unprocessed or processed datasets, and exemplary datasets include those having BAM format, SAM format, FASTQ format, or FASTA format. However, it is particularly preferred that the dataset be provided in BAM format or as a BAMBAM difference object (eg, US Patent Application Publication No. 2012/0059670A1 and US Patent Application Publication No. 2012/0066001A1). Omics data can be derived from whole-genome sequencing, exome sequencing, transcriptome sequencing (eg RNA-seq), or gene-specific analysis (eg PCR, qPCR, hybridization, LCR, etc.). Similarly, computer analysis of sequence data can be performed in many ways. However, in the most preferred method, the analysis uses a BAM file and a BAM server, as disclosed, for example, in US Patent Application Publication No. 2012/0059670A1 and US Patent Application Publication No. 2012/0066001A1. Performed in silico by position-guided synchronous alignment of tumor and normal samples. Such an analysis favorably reduces false-positive neoepitope and significantly reduces memory and computer resource requirements.

Alternatively or additionally, the relevant information is preferably obtained directly from the tumor, but one or more of the data can also be obtained from the database. For example, when using fresh tumor samples to achieve whole-genome sequencing and RNA analysis, relevant information may be provided from the database or optimally from the sequencing center. Proteomics analysis can be performed from FFPE samples using a laser microdissection, and mass spectrometry can be performed using such samples. Therefore, sources do not necessarily have to come from a single source and can be assembled from a variety of sources. Similarly, the intended analysis may utilize data at different time points, such as before surgery and before administration of temozolomide, or after surgery and before administration of temozolomide.

Thus, preferably, suitable genomic information is RNA of MGMT that provides, for example, mutations, duplications, or deletions of the MGMT gene, as well as quantitative information on transcription (and splice mutations or other mutations, if any). Includes whole-genome sequencing or exome sequencing that can identify sequence information, especially RNAseq information. Alternatively, quantitative information can also be obtained by hybridization and / or other PCR-based methods. Similarly, protein information is obtained, preferably using mass spectrometry, including selected reaction monitoring methods, antibody-based information, and / or staining methods.

The DNA damage alkylating agent temozolomide (TMZ) is approved for the treatment of glioblastoma, melanoma, and lymphoma. The MGMT enzyme is involved in repairing damage from alkylating agents. MGMT epigenetic silencing is associated with TMZ resistance in melanoma studies and occurs in about one-third of colorectal cancer (CRC). Unfortunately, however, previous studies have demonstrated that low MGMT protein expression can increase the prediction of TMZ response and even more than MGMT methylation in mCRC patients. The relationship between various MGMT assays and outcomes remains unclear.

Therefore, we intend to use two or more types of datasets for a single molecule entity in the construction of predictive models. Advantageously, suitable types of data sets include DNA copy count data, DNA mutation data, RNA spice variant data, RNA expression level data, promoter methylation data, epigenetic modification data, protein data, and protein activity data. including. Most typically, such data are readily available and / or inferrable from various pathway models (eg, PARADIGM). It is also contemplated that if more than one type of dataset is used, then at least three different types of datasets will be used. As will be readily apparent, the choice of classification algorithm will be, at least to some extent, a function of the type of dataset, and one of ordinary skill in the art will be able to determine the appropriate classification algorithm for a given dataset. Let's do it. In addition, the cutoff value can be predetermined or independently learned using additional machine learning, as illustrated below.

Regardless of the particular method used to obtain the quantitative metric, it is generally preferred that one or more of the RNAseq information, protein quantitative information, and methylation information be subgrouped using one or more thresholds. Is intended. For example, RNAseq information can be subgrouped by TPM (transcription parts per million) threshold, and protein quantification information can be subgrouped by specific values such as detection threshold or 200amol, as described in more detail below. And the methylation information can be subgrouped by a threshold determined by methyl-BEAMing (eg, 60% methylated MGMT promoter sequence).

We further train predictive models by using one or more of such RNAseq information, protein quantification information, and methylation information, and / or one or more thresholds of subgrouped information. It is intended that it is possible to verify the accuracy of the prediction model. There are many model-building methods known in the art for reaction prediction models, and the model intended is to group or group one or more of RNAseq information, protein quantification information, and methylation information. It should be noted that it can be used without or in combination with any of them. However, as described in more detail below, the model preferably uses RNAseq information, protein quantification information, and methylation information subgrouped.

Similarly, there are various multivariate classification algorithms known in the art for predictive tasks, and exemplary classifiers include extra tree classifiers, KNN classifiers, RBF or linear support vector classifiers, and decision tree classifiers. , Naive Bayes Classifier, Quad Discretionary Classifier, Ridge Classifier, Gaussian Process Classifier, Random Forest Classifier, Random Forest Classifier or Decision Tree Based Estimate. There are various univariate classification algorithms for predictive tasks known in the art, and the optimal classification threshold has been found using Uden analysis as an example. Such an algorithm will provide various accuracy metrics, as shown below. In addition, it is generally preferable to use a classifier having the highest accuracy (or accuracy gain) for creating a reaction prediction model. Especially with relatively high accuracy, the intended method, when tested in patients with unknown cancer (eg, mCRC patients), depends on the type of classifier used, at least 70%, at least 80%, at least. It enabled a prediction accuracy of 85%. Most preferably, about 86% accuracy was achieved when using the K-nearest neighbor classifier.

Archived FFPE tissue sections were obtained from 41 patients with metastatic colorectal cancer who received TMZ in one of the three Phase II clinical trials of the FELDPAR cohort. Tumor samples were available from 41 TMZ treated patients for analysis, as shown in Table 1. These patients had a median age of 69 years and had received three chemotherapy regimens with the median prior to TMZ. The majority of patients had 0 or 1 ECOG status (85%) and at least 2 metastatic sites (56%), with the most frequent site being the liver. All patients eventually progressed to TMZ, as expected by mCRC. The ORR was as follows. That is, 26 patients (63%) had progressive disease, 9 (22%) had partial reaction, and 6 (15%) had stable disease. Of these 41 samples, 39 successfully passed the quality control standard for RNAseq sequencing, and 35 successfully passed the quality control standard for methyl BEAMing (digital MB), as shown in FIG. The following is a brief analysis of this selected sample.

All 41 archive samples were evaluable by LC-MS, 35 were analyzed by digital MB, and 39 were of sufficient quality for MGMT assessment by RNA-seq (FIGS. 1, Table 1). Of the patients evaluated by LC-MS-based proteomics, 18 (44%) were negative in the MGMT protein test (<200 amol / μg of tumor protein) and were therefore considered likely to respond to TMZ. It was. The rest (n = 23) were MGMT protein positive and were considered unlikely to respond to TMZ therapy. In this molecularly enriched population of clinical trial participants, the MGMT rate was high at 44%. By comparison, an MGMT-negative incidence of 16% was found in all CRC patient samples (n = 114) delivered to our laboratory for a one-year proteomic study. MGMT promoter methylation above the 63% cutoff by MB was observed in 12 (34%) patients, the rest having unmethylated MGMT status (Table 2). For 35 tumors analyzed by both MB and LC-MS, the concordance rate between methods was 77%, p = 0.004.

Ability of MGMT assay to predict response and survival: Quantitative proteomics retrospectively identified 9 of 9 RECIST-regulated responders to TMZ, all 9 responders having negative MGMT protein expression by LC-MS. It was. Nine additional patients with MGMT-negative proteins did not have a RECIST-defined response to TMZ (ORR, MGMT-negative patients: 50%). None of the patients with positive MGMT protein expression responded to TMZ (ORR, MGMT positive patients: 0%, p = 0.0001, Table 2, FIG. 2A).

Of the patients analyzed by MB (n = 35), MGMT hypermethylation status retrospectively identified 6 of 8 responders to TMZ, and an additional 6 patients with MGMT hypermethylation by MB. , Non-responder (ORR, MGMT hypermethylated patients: 50%). Two patients with negative methylation status responded to TMZ (ORR: 9%, p = 0.011, Table 2, FIG. 2B). FIG. 3 shows a graph combining data on methylation status and MGMT protein expression levels (MGMT protein <200 amol / μg (n = 18, dark blue) and MGMT ≧ 200 amol / μg (n = 23, light blue). Percentage change (from baseline) of tumor volume per patient with. MGMT methylation status by MB (red bar, n = 35) with positive status defined as> 60% (red line).

In survival analysis, patients with negative MGMT protein expression by LC-MS had longer median PFS (mPFS) than proteomic-positive patients (3.7 months vs. 1.8 months, HR: 0). .504 [95% CI 0.27 to 0.94], p = 0.014) (Fig. 4A). MGMT levels were 12 latent confounders (BRAF and KRAS mutation status, gender, ECOG, number of pretreatments, LDH baseline level, number of metastatic sites, neutrophil to lymphocyte ratio, peritoneal disease in a bivariate Cox proportional hazards model. , Primary tumor location, archived tissue site, and age) maintained a statistically significant predictor of PFS.

Although some of these clinical variables were outcome-related (eg LDH), MGMT protein expression was the most statistically significant predictor of PFS (Table 3). The difference in OS due to MGMT protein expression was similar to the difference in PFS, but did not reach statistical significance (8.7 vs. 7.4 months, HR: 0.593 [95% CI: 0.32-]. 1.12], p = 0.078) (Fig. 4B). There were no statistically significant differences in PFS or OS between patients stratified by MGMT MB (FIGS. 5A and 5B).

MGMT by RNA-Seqing: Low MGMT mRNA expression was observed in the majority of samples using experimental cutpoints for data-fitted mRNA expression (≤3.5log2 [TPM + 1]). %) (Table 2). Tumors of patients with low MGMT RNA expression due to RNA-seq had a non-significantly higher ORR than higher mRNA expressors (35% vs. 6%, p = 0.115, Table 2). There is no statistically significant survival difference between patients stratified by MGMT mRNA expression (FIGS. 6A and 6B). 7A and 7B show that among TMZ treated patients (n = 41), those with MGMT protein level <200 amol / μg (n = 18) are longer median PFS (mPFS) than patients with higher MGMT protein levels. ) Is shown in another graph showing that it had. All patients eventually progress with TMZ. (Fig. 7A) Progression is redefined by RECIST criteria that reflect clinical response, and patients with partial response or stable disease for ≧ 6 months are defined as responders (n = 18) or non-responders (n = 23). did. Overall survival results were consistent and nearly statistically significant (8.7 vs. 7.4 months, HR = 0.6, p = 0.077) (Fig. 7B).

Correlation between RNAseq cutoffs and MS-proteomic cutoffs: In an attempt to identify the corresponding cutoff values for RNAseq relative to observed / practical MS-proteomics cutoff values, we have identified all COADs and READs. MGMT expression levels were obtained from TCGA samples and natural breaks that could match the 200 amol cutoff proposed by prior proteomics were searched for in expression patterns. As can be seen from FIGS. 8A and 8B, the distribution of MGMT TPM appears in two modes with a natural break near 3.5log2 (TPM + 1) (indicated by the vertical line). At the 3.5 log2 (TPM + 1) cutoff, good agreement is observed between the 200 amol / mL MS proteomics cutoff and the RNAseq class, as is also apparent from Table 4 below. Fisher's exact p-value for this association level between the protein threshold and the RNA threshold is p <0.00082.

In follow-up analysis, we found that the selected cutoff (3.5log2 [TPM + 1]) was optimal for matching between RNAseq and proteomic values in this cohort by the Uden analysis obtained from FIGS. 9A and 9B. I found that. At this 3.5 cut point, a TPR of 0.71 and an FPR of 0.11 are obtained in this cohort when compared to the proteomic MGMT class. FIG. 9C shows a bar graph showing non-respondent and respondent RNAseq values for TMZ (statistic = −1.04, p-value = 0.305). FIG. 9D shows a bar graph suggesting that the concordance rate between the MGMT proteomic assay and MB was 80%, p = 0.0011 Fisher's exact test (35 tumors analyzed by MB).

Definition of MGMT subgroups related to outcomes: In the first analysis, we reproduced the previous study to demonstrate PFS response in patients with <200 amol / mL MGMT protein. Typical results are shown in FIG. 10A. The log rank test between proteomic classes has p ≦ 0.0186, and the Cox proportional hazard results are shown in Table 5.

It should be noted that the log-rank test gives p <0.0186 in these two subgroups (shown in the previous study), but the Cox proportional hazard ratio gives p <0.021. The Cox model is slightly conservative and takes into account imbalances in the arm. Both statistics are considered in the results below. It should also be noted that this split loses the significance of p <0.05 when using OS as the survival metric and death as the endpoint, as can be seen in FIG. 10B. In this case, the log rank test between proteomic classes was p ≦ 0.0879, and the Cox proportional hazards were as shown in Table 6.

The inventor then investigated the associations observed using the RNAseq subgroup. In this case, the 3.5 log2 (TPM + 1) RNAseq cutoffs established in the TGCA COAD / READ data were used to define subgroups as shown in FIG. This provides a log-rank test between RNAseq classes with p ≦ 0.1731, and the Cox proportional hazards are shown in Table 7. As can be easily seen, the RNAseq class was less predictive than the proteomic subgroup and did not achieve significance in this cohort size. Similarly, using OS as a survival metric did not achieve significance.

The inventor then investigated the observed associations using combinations of RNAseq subgroups and proteomic classes. If the sample had a high MGMT in either RNAseq or protein, it was considered high in MGMT. Typical results of such an analysis are shown in FIG. In this case, the log rank test between the combination RNA + protein classes is p ≦ 0.0350, and the Cox proportional hazard results are shown in Table 6. It should be noted that in this case, differential viability was significant and improved RNAseq alone, but not as significantly as proteomic 200amol split alone.

In further analysis, we investigated the associations observed using the above combination + MGMT promoter CpG methylation subgroup. More specifically, samples with> 60% methylation were expected to have inhibition of MGMT expression, and the optimal combinations were: That is, high MGMT: either low methylation and high RNA or high protein, and low MGMT: high methylation or low RNA or low protein.

FIG. 13A shows exemplary results of such an analysis. In this case, the log rank test between the ternary combination classes is p ≦ 0.0378, and the Cox proportional hazards are shown in Table 9.

This separation was not as distinguishable as RNA + protein or protein alone, but remained significant in both the Logrank test and the Cox PH test when OS was a survival metric, as shown in FIG. 13B. In that case, the log rank test between the ternary combination classes is p ≦ 0.0419, and the Cox proportional hazards are shown in Table 10.

Temozolomide Reaction Prediction Example I: The present inventor evaluated a plurality of methods for constructing a prediction model of temozolomide reaction based on MGMT-omics values. More specifically, the inventor used each of the MGMT assays, ie RNAseq-expressing TPM, protein amol / mL, and percent methylation, and combinations and partial combinations thereof to predict a model of temozolomide reaction. It was constructed. Further models were constructed using the raw continuous values of each of these features as well as their subgrouping values (3.5log2 (TPM + 1), 200amol / mL, and 60% CpG methylation, respectively). .. In combination, this resulted in 10 different "datasets":
1. 1. Expression alone 2. Protein alone 3. Methylation alone 4. Expression + protein 5. Expression + protein + methylation 6. Expression (subgrouping) + protein 7. Expression (subgrouping) + protein (subgrouping)
8. Expression (subgrouping) + protein (subgrouping) + methylation (subgrouping)
9. Expression + protein + methylation (subgrouping)
10. Expression + protein (subgrouping) + methylation (subgrouping)

To evaluate predictive performance, we used leave pair-out cross-validation (LPOCV). This verification method is necessary to build a prediction model with 37/39 samples and then test the prediction performance with one unknown positive sample and one unknown negative sample. This is repeated with all possible combinations of positive and negative samples, resulting in a 308 performance rating in this cohort. The average performance across these 308 unknown test sets is the accuracy reported for a given prediction algorithm.

For all multi-feature datasets (4-10 above), the inventor evaluated 13 different classification algorithms for this prediction task. With single feature datasets (1, 2, and 3 above), the Uden J statistic was used to establish new optimal cutoffs with training samples and the performance of the new cutoffs was tested with unknown sample pairs.

These 10 datasets and 14 classification algorithms were combined into 140 different modeling strategies. In order to evaluate the prediction performance of these 140 strategies with unknown samples using LPOCV, it was necessary to build an additional 2772 unique prediction submodel. FIG. 14 shows the average accuracy of unknown samples for each of these modeling strategies.

As can be seen from the calculations and FIG. 14, the best overall modeling strategy was 87% predictive accuracy of the temozolomide reaction in the unknown sample. This performance (87% accuracy) is significantly better than the majority classification strategy (ie, all samples are estimated to be resistant: 71%) and is improved over protein value alone (80%). You should be aware. We selected this modeling strategy and proposed a final prediction model.

The best-performing modeling strategy uses the K-nearest neighbor approach, utilizing all three features (RNA, protein, and methylation) for their subgrouping transformations. This approach predicts the temozolomide reaction for new samples as follows. 1. 1. The pre-defined cutoffs described above are used to define MGMT mRNA expression status, protein levels, and promoter methylation status, and predicted training instances and novels using all 2.3 MGMT-related features. 2. Calculate the pairwise Minkowski distance between each of the samples (ie the brute tree). 4. Identify 5 closest matches in each new sample. The new sample is attributed to the majority response class of the closest training sample.

With a strong belief that the predictive performance of the new sample is similar to that of the cross-validation configuration, we propose a final model in which all available samples are used for training. Since trained with three binary features, the final model describes the probabilities of temozolomide susceptibility in eight distinguishable states (Table 11). New samples can be subgrouped using the same cutoffs as described above and assigned to one of these eight states. A susceptibility prediction probability of> 0.5 suggests that the condition is sensitive to temozolomide with a certainty of about 87%. Conversely, a temozolomide reaction probability of <0.5 is associated with resistance to temozolomide.

Temozolomide Reaction Prediction Example II: We trained a robust prediction model of TMZ reaction based on three individual quantitative MGMT assays (promoter quantitative methylation, RNA expression, and protein abundance) to patients with unknown mCRC. I tried to verify its accuracy. Rather than identifying a single type of predictor, from a variety of perspectives, we have incorporated multiple variables to derive a predictive model with high sensitivity and accuracy in multiple machine learning settings. Attempts were made to identify predictors.

In one example, the model was trained with 41 archive tumor samples from three TMZ safety studies (INT study n.20 / 13, INT study 20/13, and EudraCT2012-002766-13). Responses to TMZ are described in RECIST v. Specified by 1.1 standards. MGMT status was evaluated by three methods: digital PCR / methyl-BEAMing (MB), RNAseq, and liquid chromatography mass spectra. Several multivariate modeling strategies (kNN, SVM, decision tree, etc.) were evaluated using cross-validation (CV) within the training set. Due to the lack of clinical grade methylation trials, we also explored a model that first predicted MGMT methylation (based on total RNAseq) and then used the predicted methylation to classify TMZ reactions. The most accurate model of CV was validated with 14 unknown tumor samples from a similarly assayed follow-up study. Predefined thresholds for each MGMT assay were used as criteria for comparison.

As can be seen from the exemplary results in Table 1 below, when multiple variables were used for training and validation, they were compared to a single variable (ie, individually used methylation or protein or expression) for temozolomide. Response prediction was significantly improved. In fact, the TMZ response in refractory mCRC is nearly predictable, based on the incorporation of multiple variables. The combination of predicted methylation, transcriptional levels, and protein abundance gives the most accurate and robust method of predicting response (82% -87% certainty).

In one set of experiments, we determined by MGMT protein (measured by LC-MS), MGMT expression (measured by TPM), and MGMT promoter methylation (digital PCR / methyl-BEAMing (MB)). The training cohort prediction performance was investigated for (measured). More specifically, to assess the ability of pre-defined cutoffs to predict response to TMZ, we used a leave pair-out cross-validation strategy. Predefined search cutoffs were evaluated on unknown samples 330, 308, and 250 times with LC-MS, RNAseq, and MB data, respectively. Predefined cutoffs in LC-MS and RNA-seq show better mean predictive performance (82.1% and 72.2%, respectively) than the MB model (68.0%), which is typical. The results are shown in FIG.

To further investigate the effects of various classification algorithms and training data (ie, single-variable vs. multi-variable), we set protein, RNA, and methylation data alone or in combination and with predetermined thresholds. Several learning approaches were performed with or without use. As can be seen from FIG. 16, the protein-based model had a relatively high predictive accuracy, and the model using all three variables was even better. In a further attempt to improve accuracy and simplify clinical or sample requirements, we constructed 10 candidate models (+3 pre-defined cutoffs) using previous TMZ trials as training data. , The measured methylation was replaced with "predicted methylation" using a regression model based on total RNAseq. Performance was then tested in an unknown test cohort (TEMIRI), as shown exemplary in FIG. In this case, the training dataset was a TMZ cohort and included 41 mCRC patients treated with TMZ from three Phase II trials. Continuous MGMT protein levels by mass spectrum as well as RNA expression data and continuous MGMT methylation percent data by RNAseq were available in all patients. The drug reaction is described as binary drug reaction data. The study dataset included 32 mCRC patients treated with TMZ + irinotecan. Dual drug response data are missing in 3 patients. Gene expression values were available in 14 patients and MGMT protein expression data were available in 21 patients. See Table 13.

We further contemplate that it is possible to construct a regression model for machine learning with higher accuracy that describes methylation and / or uses asserted methylation values. FIG. 18 shows a regression and classification pipeline for building a regression model. As shown in Table 14, when both MGMT protein expression levels and MGMT RNAseq were used as datasets, the regressor model's RMSE (the square root of the variance of the residuals, which is the absolute fit of the model to the data, ie. , Represents how close the observed data points are to the predicted values of the model) is smaller (better fit).

FIG. 19 shows the average accuracy values of various regressor models when the total expression (expression level of all RNA) and the MGMT protein expression level are used as a data set, which are also summarized in Table 15.

FIG. 20 shows the mean squared RMSE values of various regressor models when the levels of MGMT gene expression and MGMT protein expression are used as a data set, which are also summarized in Table 16.

We then constructed a regressor model using the asserted methylation values. 21 and 22 show the average accuracy values of the various regressor models when the asserted methylation values were used as the dataset, which are also summarized in Tables 17 and 18, respectively.

FIG. 23 shows a heatmap with exemplary results of response prediction for most of 1,000 variable genes over 44 samples using preset thresholds as described, and Table 14 shows the selected dataset and A list of exemplary classification algorithms used in the dataset combination. As you will see again, the use of MGMT RNAseq, MGMT protein, and MGMT promoter methylation provided excellent training and test accuracy for predicting response to temozolomide. Similarly, sensitivity, specificity, and F1 scores were all substantially increased over other classifiers and individual datasets. The sample level predictions for the best models in Table 19 are listed in Table 20, suggesting that models that simultaneously consider protein, MGMT methylation, and mRNA work better when compared to other models. Will be done.

In this retrospective analysis of patients with refractory mCRC treated with TMZ, MS-based trials for MGMT protein had 50% sensitivity and 100% specificity. In this small patient cohort, proteomic trials were superior to both digital MB and RNA-seq in predicting response to TMZ. In addition, MGMT protein expression below a predetermined threshold (200 amol / μg) was associated with a 2-fold increase in mPFS, which association was independent of the 12 predictors. Interestingly, patients with positive MGMT protein expression by LC-MS had a PFS similar to that reported in patients with mCRC who participated in TMZ clinical trials. The disappointing results of such tests may reflect the limited capabilities of standard MGMT assessment methods such as MSP to select the best candidate for TMZ. This study supports the assessment of alternative MGMT platforms such as LC-MS in prospective trials.

The accuracy and robustness of the MGMT assay as a predictor of TMZ reaction was rigorously tested under cross-validation settings. At the time of this run, the proteomic MGMT test was superior to the other test platform with an average accuracy of 82.1%.

It should be noted that the importance of identifying potential responders to TMZ is recognized when the DNA mismatch repair of TMZ-treated mCRC is impaired. In this case, the TMZ responder is eligible for immune checkpoint inhibitor therapy by switching from stable microsatellite to unstable microsatellite (MSI). In other words, patients who relapse with TMZ may initiate immunotherapy.

The importance of identifying potential responders to TMZ is highlighted by recently published findings regarding impaired DNA mismatch repair in TMZ-treated mCRCs, in which case TMZ responders change from stable microsatellite to unstable microsatellite (MSI). The switch qualifies for therapy with immune checkpoint inhibitors. This suggests that patients who relapse with TMZ may start immunotherapy. Therefore, we further contemplate that it is possible to recommend or update cancer treatments for patients based on the predicted results. For example, if a response prediction model predicts that a patient will respond to TMZ, TMZ can be administered to the patient at a dose and schedule that is effective in treating the tumor. In another example, the patient no longer responds to TMZ or has substantially reduced responsiveness to TMZ (eg, with other individuals having a similar prognosis for cancer compared to before TMZ treatment). Patients can be administered immunotherapeutic agents (eg, checkpoint inhibitors, cancer vaccines, etc.) if predicted by response prediction models to be at least 30%, at least 50%, and at least 70% reduced in comparison. is there.

As used herein, the term "administering" a drug or cancer therapeutic agent means both direct and indirect administration of the drug or cancer therapeutic agent. Direct administration of the drug or cancer therapeutic agent is typically performed by a healthcare professional (eg, doctor, nurse, etc.), and indirect administration is direct administration (eg, injection, oral ingestion, topical application, etc.). Includes the step of providing a drug or cancer therapeutic agent to a healthcare professional or making it available to a healthcare professional for (via).

As used herein and throughout the claims below, the meanings of "a", "an", and "the" include plural references unless expressly specified in the context. Also, as used herein, the meaning of "in" includes "in" and "on" unless explicitly defined in the context. Also, as used herein, unless otherwise specified in the context, the term "bonded to" refers to direct binding (two elements coupled to each other contacting each other) and indirect coupling (at least one). It is intended to include both (an additional element is located between the two elements). Therefore, the terms "combined with" and "combined with" are used synonymously. Finally, unless contextually reversed, all ranges set forth herein should be construed to include their endpoints, and open-ended ranges include commercially practical values. Should be interpreted. Similarly, all lists of values should be considered to contain intermediate values unless the context suggests the opposite.

It should be apparent to those skilled in the art that more modifications than those already described can be made without departing from the concepts of the invention described herein. Therefore, the subject matter of the present invention should not be limited unless it is within the scope of the appended claims. Moreover, in interpreting both the present specification and the claims, all terms should be interpreted in the broadest possible sense consistent with the context. Specifically, the terms "comprises" and "comprising" should be construed and referenced as non-exclusive references to elements, components, or processes. It suggests that elements, components, or processes may exist, or that they may be utilized, or that they may be combined with other elements, components, or processes that are not explicitly referenced. The claims of the present specification are A, B, C.I. .. .. .. When referring to at least one of something selected from the group consisting of and N, the wording should be interpreted as requiring only one element of the group, not A + N, B + N, etc.

Claims (18)

A method of predicting therapeutic response to temozolomide in patients.
To provide RNAseq information, protein quantification information, and methylation information from the patient's tumor.
To calculate the reaction prediction for temozolomide by the reaction prediction model, that the reaction prediction model uses the RNAseq information, the protein quantification information, and the methylation information.
How to include. The method of claim 1, wherein the reaction prediction model uses a K-nearest neighbor approach. The method according to claim 1 or 2, wherein the RNAseq information, protein quantification information, and methylation information are subgrouped. The method of claim 3, wherein the RNAseq information is subgrouped using a log2 (TPM + 1) cutoff value of 3.5. The method of claim 3, wherein the protein quantification information is subgrouped using a cutoff value of 200 amol / mL. The method of claim 3, wherein the methylation information is subgrouped using the cutoff value of 60% promoter CpG methylation. The method according to any one of claims 1 to 6, wherein the RNAseq information, protein quantification information, and methylation information are provided from an FFPE sample or a fresh tumor sample. The method according to any one of claims 1 to 7, wherein the tumor is a solid tumor. The method according to any one of claims 1 to 8, wherein the tumor is metastatic colon cancer, glioblastoma, or melanoma. The method according to any one of claims 1 to 9, wherein the reaction prediction model has a prediction accuracy of at least 85%. The method of claim 1, wherein the RNAseq information, protein quantification information, and methylation information are subgrouped. The method of claim 11, wherein the RNAseq information is subgrouped using a log2 (TPM + 1) cutoff value of 3.5. The method of claim 11, wherein the protein quantification information is subgrouped using a cutoff value of 200 amol / mL. The method of claim 11, wherein the methylation information is subgrouped using the cutoff value of 60% promoter CpG methylation. The method of claim 1, wherein the RNAseq information, protein quantification information, and methylation information are provided from an FFPE sample or a fresh tumor sample. The method of claim 1, wherein the tumor is a solid tumor. The method of claim 1, wherein the tumor is metastatic colon cancer, glioblastoma, or melanoma. The method of claim 1, wherein the reaction prediction model has a prediction accuracy of at least 85%.

Images