Patent Application
20160153032
The invention pertains in some aspects to a method for predicting a manifestation of an outcome measure of a cancer patient based on a tumor DNA-containing tissue sample from the cancer patient. The invention further relates to a method for determining a function that allows for the prediction of the manifestation of an outcome measure (such as the development of a metastasis vs. no development of a metastasis or response to therapy vs. no response to therapy) of a cancer patient.
Cancer, in particular solid tumor cancer, is a group of diseases that can occur in every organ of the human body and affects a great number of people. Colorectal cancer, for example, affects 73,000 patients in Germany and approximately 145,000 patients in the United States. It is the second most frequent solid tumor after breast and prostate cancer. Treatment of patients with colorectal cancer differs dependent on the location of the tumor, the stage of the disease, various additional risk factors and routine practice in various countries. Standard treatment for patients with colon cancer that is locally defined (stage I and stage II) or has spread only to lymph nodes (stage III) always involves surgery to remove the primary tumor. Standard treatment for patients with rectum cancer may differ from country to country and from hospital to hospital as a significant part of these patients will receive neo-adjuvant radio/chemotherapy followed by surgery to remove the tumor tissue.
The five-year survival rates of patients with colorectal cancer depend on the clinical stage of the individual patient, the histopathological diagnosis, stage-specific treatment options as well as on routine medical practice that differs from country to country, and often also from hospital to hospital. There are also significant differences in the routine treatment of patient with colorectal cancer in the western world.
In most countries, patients with UICC stage I disease will not receive any additional chemotherapy after surgery as their five-year survival is approximately 95%.
Treatment options for patients with UICC stage II colon cancer differ in many Western countries. The five-year survival of patients with UICC II disease is approximately 80% to 82%, meaning that 18% to 20% will experience a progression of disease—often liver or lung metastasis. Once the disease will have spread to distant organs the outcome of the patients is much worse, and the majority of these patient will die relatively quickly despite heavy treatment of these patients. Therefore guidelines in some Western countries recommend offering adjuvant chemotherapies to patients with UICC stage II disease including 5-flourouracil and leucovorine or in combination with oxaliplatin. In other European countries including Germany, the guidelines do not recommend to offer patients with UICC stage II disease adjuvant chemotherapy. There is a controversy if adjuvant chemotherapy should be offered to UICC II patients or not. Randomized clinical data that show a benefit of adjuvant chemotherapy is still missing for these patient cohorts.
Patients with locally advanced colorectal cancer—loco-regional lymph nodes are infiltrated with cancer cells—have a five-year survival rate of 49%. The treatment guidelines therefore recommend that after surgery all patients should receive adjuvant chemotherapy, either a triple combination of 5-FU, leucovorin and oxaliplatin (FOLFOX4 or FOLFOX6 regimes) or dual combination of capecitabine (an orally available 5-FU derivative) and oxaliplatin (CAPOX). For elderly patients with low ECOG performance scores or known toxicities, the dual 5-FU/leucovorin scheme should be used. In the routine practice only 60 to 80% of patients with UICC stage III disease will however receive adjuvant chemotherapy. In Germany, only 60% of UICC stage III patients will be treated with FOLFOX or 5-FU/leucovorin. There is also a difference in treatment between low density areas and city populations. In general, approximately 50% of patients with UICC stage III disease will experience progression of disease within 1 to 2 years after surgery. Once distant metatastasis is diagnosed, these patients will be offered additional therapies including treatment with targeted antibody drugs that inhibit the EGFR receptor including cetuximab or panitumumab, or antibodies directed against the VGFA ligand (bevacizimab). Several lines of therapies are offered, but most of these patients with disease progression will die within a five-year interval.
The five-year survival rate for patients with advanced, metastatic disease is dramatically low. Only 8% will survive the first five years after surgery. It is these patients for which most of the treatment options with targeted therapies were developed over the last ten years, however, with limited success. The first targeted antibody therapy involved an anti-EGFR antibody (cetuximab) that was approved in 2004 by the FDA as monotherapy or in combination with Irinotecan, for patients with metastatic CRC (mCRC) that failed prior chemotherapy with irinotecan. In the original BOND study the response rate of the patients for the cetuximab was approximately 11%. In 2007, a second anti-EGFR antibody, panitumumab, was approved for the treatment of mCRC patients. However, the FDA approved panitumumab only in combination with a KRAS wildtype (wt), as it was shown in 2007 that only patients with wt KRAS gene would benefit from panitumumab. However, the data also showed that many patients with mCRC and wt KRAS did not benefit from panitumumab. Also, there were some mCRC patients with mutations in the KRAS gene that showed response to panitumumab. Similar data was also published in 2008 to 2009 for cetuximab that led to a label change for the approval of cetuximab. At the moment, both cetuximab and panitumumab are only approved for patients with mCRC and wildtype KRAS status.
Accurate prediction of response/nonresponse to therapy is a prerequisite for individualized approaches to treatment. Current clinical practice in the treatment of patients with solid tumors does not offer effective and accurate prediction of response/nonresponse to chemotherapy and hormone therapy.
In prostate cancer no predictive biomarkers are known or established that predict response to radiation, hormone therapy or chemotherapy with taxanes. The same is true for advanced non-small cell lung cancer (NSCLC). Approximately 70% to 80% of all NSCLC patients have stage IIIB or stage IV disease at the time of first diagnosis. For the majority of these patients no predictive markers exist that allow prediction of response to small molecule drugs like erlotinib or iressa that inhibit the kinase function of the EGF receptor. Response to erlotinib was observed only in a small cohort of NSCLC patients with EGFR mutations in the kinase domain. Still 90% of the NSCLC patients of stage IIIB and IV have a five-year survival of less than 8% despite treatment.
The situation in breast cancer is more complex. For example, most patients with early breast cancer (lymphnode negative, estrogene (ER+) and/or progesterone receptor positive (PR+)) will receive radiation, chemotherapy and hormone therapy with tamoxifen after surgical removal of the tumor. The five-year survival of these patient cohorts is between 90 to 95%. However, only 4% of the patients will benefit from the addition of chemotherapy. Current treatment guidelines still recommend the overtreatment of 100% of these patients with chemotherapy in order to reach the 4% patients that may benefit. Similarly, a significant portion of the patients do not benefit from tamoxifen although they are ER positive. Effective methods to predict response to the chemotherapy or hormone therapy are not available.
There is one FDA approved companion diagnostic (CDx) in breast cancer. Determination of the HERII status is predictive of response to trastuzumab, an anti HERII antibody. Thus patients with HERII positive breast cancer will receive trastuzumab at some point in their treatment. However, only 25% of all breast cancer patients are HERII positive and of those only 20-25% of the patients benefit from trastuzumab, meaning that 75-80% of HERII positive breast cancer patients are over treated and have no benefit from this expensive treatment.
In colorectal cancer, no predictive biomarkers are established in the adjuvant treatment of UICC II or UICC III patients.
At time of first diagnosis, 70% of the CRC patients are in UICC stage II and UICC stage III. 20% of the UICC stage II and 49% of the UICC stage III patients will suffer from progression of disease within 1 to 2 years after surgery. The majority of the patients are diagnosed with metastasis in the liver, about 20% are diagnosed with metastatic disease in the lung. Hence, anti-EGFR antibody drugs like cetuximab and panitumumab would be ideal drugs to treat these patients before metastasis will occur if responders to these drugs could be identified and separated from non-responders. Recently, two randomized phase III trials, one in the US and one in Europe, evaluating cetuximab vs. cetuximab plus FOLFOX in UICC stage III patients did not meet their endpoints. Secondary endpoint analysis showed that patients with wild type KRAS did not benefit in the Cetuximab/FOLFOX arm in comparison to patients in the FOLFOX arm (ASCO, 2010).
Therefore, there is a large clinical need in the art to predict whether a patient with cancer of a certain type and/or of a certain stage will respond to a particular treatment. In addition, there is a large clinical need in the art to predict how the cancer of a certain type and/or of a certain stage of a patient will develop over time.
The present invention provides methods for predicting a manifestation of an outcome measure of a cancer patient based on a tumor DNA-containing tissue sample from the cancer patient as well as methods for determining a function that allows for the prediction of the manifestation of an outcome measure, for example development of a metastasis vs. no development of a metastasis or response to therapy vs. no response to therapy, of a cancer patient based on a tumor DNA-containing tissue sample from the patient.
In one aspect, the invention provides a method for determining a function that predicts the manifestation of an outcome measure (for example the development of a metastasis vs. no development of a metastasis, or response to therapy vs. no response to therapy) of a cancer patient.
The method is based on a tumor DNA-containing tissue sample obtained from the patient. In certain embodiments of the method, the tumor DNA-containing tissue sample is tumor tissue, sputum, stool, urine, bronchial lavage, cerebro-spinal fluid, blood, plasma, or serum.
The tumor DNA-containing tissue sample can, in some embodiments, be a fresh-frozen sample, or a formalin-fixed paraffin-embedded sample.
The cancer is preferably a solid-tumor cancer, such as a cancer of the colon, breast, prostate, lung, pancreas, stomach, ovary or melanoma. The cancer can be of various clinical stages.
The method comprises determining the DNA sequence of segments of at least two genes in a group of cancer patients, which is comprised of patients with at least two disjunctive manifestations (sequence variation) of the outcome measure. For this purpose, the at least two genes are each divided in segments of a size that allows for the reliable determination of the DNA sequence. Segments can be, for example, between 20 and 500 base pairs. Segments of 100 to 250 base pairs are preferred in some embodiments.
The determination of the DNA sequence can be performed using any appropriate method known in the the art. Preferred is DNA sequencing of the segments (amplicons) of at least two cancer genes using oligonucleotides as sequencing primers. Also preferred is the use of next-generation sequencing methods (NGS), e.g., pyrosequencing or other sequencing-by-synthesis method, which are also known as “deep sequencing” methods.
In some embodiments, the method comprises the step of determining the sequence variation of the at least two genes of the tumor DNA as either “present” (i.e. containing a sequence variation), if at least one significant sequence variation can be identified, or as “absent” (i.e. not containing a sequence variation), if no significant sequence variation can be identified. In some embodiments, a significant sequence variation is a variation that changes the amino acid sequence of the encoded protein.
In some embodiments, the method comprises the step of combining the sequence variation statuses of the at least two genes using a logical operator, thereby generating a prediction function, such that patients with one specific manifestation of the outcome measure are distinguishable from patients with another disjunctive manifestation of the same outcome measure.
By combining sequence variation statuses using at least one logical operator, the biological information contained in each sequence variation status is aggregated and thereby maximized. In other words, using logical operators, the biological information contained in each sequence variation status is aggregated and thereby the overall information is maximized. Thus, the prediction function is a maximization function. For example, in one embodiment of the invention, the existence of a sequence variation within segments of a first gene of the tumor DNA and of a second gene of the tumor DNA is determined as present or absent, respectively. Subsequently, the existence of sequence variations of the first and the second gene are combined using a logical operation (prediction function). It is then possible to determine the existence of a sequence variation within segments of a third gene of the tumor DNA as present or absent and combine the existence of sequence variations of the third gene using a logical operation with the sequence variation of the first and of the second gene such that the prediction function is maximized, i.e. that the prediction value is maximized (e.g. based on AROC).
In various embodiments, predicting the outcome measure of the cancer patient comprises predicting disease progression, such as the local recurrence of the cancer, the occurrence of secondary malignancy, or the occurrence of metastasis (vs. no progression of disease). In other embodiments of the method, predicting the outcome measure of the cancer patient comprises predicting response vs. nonresponse of the patient to a cancer treatment with a drug, such as adjuvant chemotherapy, neo-adjuvant chemotherapy, palliative chemotherapy, or the use of targeted drugs in combination with a chemotherapy or radio-chemotherapy. In certain embodiments, the drug is one or more of Bevacizumab, Cetuximab, Panitumumab, IMC-11F8, FOLFOX, FOLFIRI and Oxaliplatin.
Bevacizumab (Avastin®) is a drug that blocks angiogenesis. It is used to treat various cancers, including colorectal cancer. Bevacizumab is a humanized monoclonal antibody that binds to vascular endothelial growth factor A (VEGF-A), which stimulates angiogenesis.
Oxaliplatin (Eloxatin®, Oxaliplatin Medac®) is [(1R,2R)-cyclohexane-1,2-diamine](ethanedioato-O,O′)platinum(II) and is known in the art as a cancer chemotherapy drug.
Cetuximab (IMC-C225, Erbitux®) is a chimeric (mouse/human) monoclonal antibody, an epidermal growth factor receptor (EGFR) inhibitor, usually given by intravenous infusion. Cetuximab is administered for the treatment of cancer, in particular for treatment of metastatic colorectal cancer and head and neck cancer. Cetuximab binds specifically to the extracellular domain of the human epidermal growth factor receptor. It is composed of the Fv regions of a murine anti-EGFR antibody with human IgG1 heavy and kappa light chain constant regions and has an approximate molecular weight of 152 kDa. Cetuximab is produced in mammalian (murine myeloma) cell culture.
Panitumumab, also known as ABX-EGF, is a fully human monoclonal antibody specific to the epidermal growth factor receptor (EGFR). Panitumumab is manufactured by Amgen and sold as VECTIBIX.
IMC-11F8 is a potent, fully human monoclonal antibody that targets the epidermal growth factor receptor (EGFR). It is currently in Phase II studies for metastatic colorectal cancer with one or more Phase III trials planned in 2009. IMC-11F8 is in development by Eli Ully.
In some embodiments, the method comprises analyzing (e.g., identifying) sequence variations that alter the protein sequence and/or analyzing sequence variations that do not alter the protein sequence (silent or synonymous variations) of the encoded protein. For example, sequence variations that alter the amino add sequence include missense variations, nonsense variations (sequence variations introducing a premature STOP codon), splicing variations, deletion variations, Insertion variations, or frame shift variations. Sequence variations that do not alter the protein sequence comprise silent sequence variations (silent amino acid replacements) and synonymous variations.
The logical operation is part of a prediction function. The prediction function comprises the existence of sequence variations or its negation as variables and at least one logical operator. The logical operator is preferably conjunction (And), negation of conjunction (Nand), disjunction (OR), negation of disjunction (Nor), equivalence (Eqv), negation of equivalence (exclusive disjunction, Xor) material implication (Imp), or negation of material implication (Nimp) combining the variables. Within a prediction function, the same or different logical operators may be used, if the prediction function comprises more than one logical operator.
In one embodiment, the use of the conjunction (And) is excluded. In another embodiment, the use of the disjunction (OR) is excluded. In yet another embodiment, the use of the conjunction (And) together with the disjunction (OR) is excluded. In one embodiment of the invention, the prediction function comprises at least three logical operators, for example, three, four, five, six, seven or more logical operators.
With respect to the logical operators, all standard logic rules of Boolean algebra apply, namely the law of the excluded middle, double negative elimination, law of noncontradiction, principle of explosion, monotonicity of entailment, idempotency of entailment, commutativity of conjunction, and De Morgan duality. Therefore, it is often possible to replace a given prediction function comprising the existence of sequence variations or its negation as variables and at least one logical operator with another prediction function comprising the existence of sequence variations or its negation as variables and at least one logical operator without obtaining a different result.
The prediction function is preferably optimized (i.e. maximized or minimized) for at least one of the following: sensitivity, specificity, positive predictive value, negative predictive value, correct classification rate, miss-classification rate, area under the receiver operating characteristic curve (AROC), odds-ratio, kappa, negative Jaccard ratio, positive Jaccard ratio, combined Jaccard ratio or cost.
In some embodiments of the invention, the step of constructing a prediction function combining the sequence variation statuses comprises the construction of a prediction function on a subset of patient data (sequence variation status and manifestation of the outcome measure) and prospective evaluation of the performance on patient data not used for construction of the prediction function. For this purpose, a classification method is preferably used.
In certain embodiments of the invention, the relative frequency of sequence variations within segments of the at least two genes is at least 2% in a given patient population, preferably 5%.
The at least two genes used in the method are so-called cancer genes, i.e. they are associated with the outcome measure of the patient. In one embodiment, the two genes (e.g., 2, 3, 4, 5, 6, 7, or 8) are chosen from genes listed in Tables 1 to 8.
In some embodiments, the logical operation predicts that the patient is in a high risk group, and the patient is subsequently treated, for example, with adjuvant or neoadjuvant chemotherapy, or a targeted therapy. Exemplary therapies are described herein. In some embodiments, the logical operation predicts that the patient is in a low risk group, and the patient is not given said therapy.
In another aspect, the invention provides a method for predicting a manifestation of an outcome measure of a cancer patient based on a tumor DNA-containing tissue sample from the cancer patient. Use is made in this method of a function that allows for the prediction of the manifestation of an outcome measure, of a cancer patient based on a tumor DNA-containing tissue sample from the patient as described above and herein.
Specifically, the method for predicting a manifestation of an outcome measure of a cancer patient based on a tumor DNA-containing tissue sample from a cancer patient comprises determining an existence of a significant sequence variation within segments of at least two genes of the tumor DNA. The existence of a significant sequence variation is determined to be “present” (containing a sequence variation) if at least one significant sequence variation can be determined, or as “absent” (not containing a sequence variation) if no significant sequence variation can be determined.
As stated above, the at least two genes of the tumor DNA are associated with the outcome measure of the patient. In other words, the at least two genes used in the method are so-called cancer genes, i.e. they are associated with the outcome measure of the patient. In one embodiment, the two genes are chosen from genes listed in Tables 1 to 8.
The method further comprises the step of combining the existence of significant sequence variations of the at least two genes using a logical operation (i.e., a prediction function, as described above and herein), and predicting based on the results of the logical operation the manifestation of an outcome measure of the patient.
Exemplary prediction functions are listed together with clinical performance for different outcome measures in Tables 9 to 20.
The method is based on a tumor DNA-containing tissue sample obtained from the patient. In certain embodiments of the method, the tumor DNA containing tissue sample is tumor tissue, sputum, stool, urine, bronchial lavage, cerebro-spinal fluid, blood, plasma, or serum.
The tumor DNA-containing tissue sample can, in some embodiments, be a fresh-frozen sample, or a formalin-fixed paraffin-embedded sample.
The cancer is preferably a solid-tumor cancer, such as a cancer of the colon, breast, prostate, lung, pancreas, stomach, or melanoma. The cancer can be of various clinical stages.
In a certain embodiments of the method, predicting the manifestation of an outcome measure of the cancer patient comprises the prediction of progression of disease of a cancer of the patient, such as the local recurrence of the cancer, the occurrence of secondary malignancy, or the occurrence of metastasis (vs. no progression of disease). In other embodiments of the method, predicting the manifestation of an outcome measure of the cancer patient comprises the prediction of the response vs. nonresponse of the patient to a cancer treatment with a drug, such as adjuvant chemotherapy, neo-adjuvant chemotherapy, palliative chemotherapy or the use of targeted drugs in combination with a chemotherapy or radio-chemotherapy.
In preferred embodiments of the invention, the step of the prediction of the sequence variation comprises analyzing sequence variations that alter the protein sequence and/or analyzing sequence variations that do not alter the protein sequence (silent or synonymous variations) of the encoded protein.
The sequence variation that alters the protein sequence comprises missense variations, nonsense variations (sequence variations introducing a premature STOP codon), splicing variations, deletion variations, insertion variations, or frame shift variations. The sequence variations that do not alter the protein sequence comprise silent sequence variations (silent amino acid replacements) and synonymous variations.
The logical operator is part of a prediction function. The prediction function comprises the existence of sequence variations or its negation as variables and at least one logical operator. The logical operator is preferably conjunction (And), negation of conjunction (Nand), disjunction (OR), negation of disjunction (Nor), equivalence (Eqv), negation of equivalence (exclusive disjunction, Xor) material implication (Imp), or negation of material implication (Nimp) combining the variables. Within a prediction function, the same or different logical operators may be used, if the prediction function comprises more than one logical operator.
With respect to the logical operators, all standard logic rules of Boolean algebra apply, namely the law of the excluded middle, double negative elimination, law of noncontradiction, principle of explosion, monotonicity of entailment, Idempotency of entailment, commutativity of conjunction, and De Morgan duality. Therefore, it is often possible to replace a given prediction function comprising the existence of sequence variations or its negation as variables and at least one logical operator with another prediction function comprising the existence of sequence variations or its negation as variables and at least one logical operator without obtaining a different result.
The prediction function is preferably optimized (i.e. maximized or minimized) for at least one of the following: sensitivity, specificity, positive predictive value, negative predictive value, correct classification rate, miss-classification rate, area under the receiver operating characteristic curve (AROC), odds-ratio, kappa, negative Jaccard ratio, positive Jaccard ratio, combined Jaccard ratio or cost.
The sequence variations are in certain embodiments of the method filtered by the type of variation, preferably by missense, nonsense, silent, synonymous, frame shift, deletion, insertion, splicing, noncoding, or combinations thereof.
In some embodiments of the methods described above, the invention provides a method for predicting a manifestation of an outcome measure of a cancer patient based on a tumor DNA-containing tissue sample from the cancer patient. The method comprises determining an existence of an encoded amino acid sequence variation (e.g., by DNA sequencing) within segments of at least two genes of the tumor DNA, with at least two genes (but in some embodiments 3, 4, 5, or 6 genes) being selected from Tables 1 to 8. The sequence information is then analyzed, e.g., computationally, to determine whether it satisfies the logical operator that is predictive of an outcome. The logical operator is constructed or trained with historical cancer specimens having a known outcome. Patients that are determined to be in a high risk group, may then be subjected to more aggressive treatment (e.g., adjuvant or neoadjuvant treatment or targeted therapy) as described herein. Patients determined to be in a low risk group may not receive such treatment.
In another aspect, the invention provides a computer program that is adapted to perform the methods described above and herein.
In certain embodiments, the computer program computer program that is adapted to perform the steps of determining an existence of a significant sequence variation within segments of at least two genes of the tumor DNA as “present” (containing a sequence variation), if at least one significant sequence variation can be determined, or as “absent” (not containing a sequence variation), if no significant sequence variation can be determined, wherein the at least two genes of the tumor DNA are associated with the outcome measure of the patient; and/or combining the existence of significant sequence variations of the at least two genes using a logical operation (prediction function), and/or predicting based on the results of the logical operation the manifestation of the outcome measure of the patient.
In another aspect, the invention provides a storage device comprising the computer program as described above and herein.
In another aspect, the invention provides a kit, comprising oligonucleotides for sequencing the segments (amplicons) of at least two cancer associated genes, and the computer program described above and herein.
The present invention provides methods for predicting a manifestation of an outcome measure of a cancer patient based on a tumor DNA-containing tissue sample from the cancer patient as well as methods for determining a function that allows for the prediction of the manifestation of an outcome measure, for example development of a metastasis vs. no development of a metastasis or response to therapy vs. no response to therapy, of a cancer patient based on a tumor DNA containing tissue sample from the patient.
The methods in various embodiments comprise filtering of significant sequence variations, functional filtering of the sequence variations, and construction of a prediction function to link sequence variations to the manifestation of an outcome measure.
The invention in various embodiments comprises sequencing of two or more target nucleotide sequences (e.g., genomic or cDNA sequences) of the patient sample. For example, the invention can involve deep sequencing (also known as NGS), which is sequencing with high coverage, of the DNA of at least two segments of at least two genes. Several technologies exist that perform this task. In some embodiments, the method can employ the Illumina technology platform for deep sequencing (Illumina, Inc., San Diego, Calif. 92122 USA); or a similar platform. Common to all deep sequencing methods are the results, namely sequence alignment maps (SAM/BAM-files) of the sequenced bases which makes up the DNA and an analysis of sequence variation data (VCF-files). The sequence alignment uses the human reference genome provided by the Genome Reference Consortium. It is publicly available from the National Institute of Biotechnology Information of The National Institute of Health of the United States of America.
Table A displays a small part of deep sequencing results of an analysis of a gene segment, namely KRAS. For each unique chromosome position it needs to be decided whether a significant variation is present or not. This invention exploits the fact that oncologists are dealing with a mixture of normal and tumor DNA. Given a solid tumor sample, the fraction of tumor cells' is always significantly lower than 100 percent, because there is always some fraction of normal tissue, muscle cells, and stromal cells present. The preparation of the tumor tissue can ensure that the tumor fraction is at least 10%. In cell-free DNA extracted from blood plasma the vast majority stems from normal tissue, and it cannot be ascertained how big the fraction of tumor DNA is. Thus, the decision whether a significant variation is present must be made without the knowledge of the human reference genome.
The overall hypothesis, whether a significant variation is present or not, can be split into four null hypotheses:
1.) The fraction of the overall most frequent nucleotide is not significantly smaller than 99% of the overall coverage.
2.) The fraction of the most frequent nucleotide on allele I is not significantly smaller than 99% of the coverage of allele I.
3.) The fraction of the most frequent nucleotide on allele II is not significantly smaller than 99% of the coverage of allele II.
4.) The fraction of the overall second most frequent nucleotide is not significantly higher than 1% of the overall coverage.
If hypothesis 1 and either hypothesis 2 or hypothesis 3 and hypothesis 4 is rejected by an appropriate statistical test, then there is a statistically significant variation present. Appropriate statistical tests are among others the Poisson test or the binomial exact test. Depending on the number of unique chromosome positions sampled it is good statistical practice to adjust the overall error of first kind, which is called alpha, to account for multiple testing. In the presented examples of deep sequencing the number of unique chromosome positions is 26711 as several segments of 48 cancer genes were simultaneously sequenced for each patient. Hence the statistical tests are made at the alpha=0.05/26711 level, and the upper and lower confidence limits are computed accordingly. In case that another panel with a different number of unique chromosome positions is used, the correction for multiple testing must be adjusted accordingly.
In biological terms, hypothesis 1 and hypothesis 4 ensure that the observed sequence variation is not measurement noise, whereas hypothesis 2 and hypothesis 3 ensure that the variation is not a heterozygous sequence variation.
The manufacturer of the panel ensures that the average measurement noise at each unique position is 1%, which has been confirmed by scientific publications. However, using 315 own samples the inventors used the observed noise levels for each position across all samples to ascertain valid variations above the noise level.
As shown in Table A, the analysis of DNA segments results in counts of the four bases, namely Arginine (A), Cytosine (C), Guanine (G), and Tyrosine (T), which make up the genetic code. To demonstrate the statistical tests, the code for the publicly available R-statistical software package is given for chromosome 12 position 25380290:
Hypothesis 1: poisson.test(x=(361−9), T=361, r=0.99, alternative=“less”, conf.level 20=1−0.05/26711) results in a p-value of 0.4011
Hypothesis 2: poisson.test(x=(172−4), T=172, r=0.99, alternative=“less”, conf.level=1−0.05/26711) results in a p-value of 0.4507
Hypothesis 3: poisson.test(x=(180−5), T=180, r=0.99, alternative=“less”, conf.level=1−0.05/26711) results in a p-value of 0.4246
Hypothesis 4: poisson.test(x=9, T=361, r=0.01, alternative=“greater”, conf.level=1−0.05/26711) results in a p-value of 0.01186
Since all p-values are greater than 0.05/26711=0.0000181 none of the null-hypotheses can be rejected, thus there is no statistically significant variation.
This is a little different for chromosome 12 position 25380293, again the R-code is given so that any knowledgeable person can repeat the following hypothesis tests:
Hypothesis 1: poisson.test(x=(136+145), T=361, r=0.99, alternative=“less”, conf.level=1−0.05/26711) results in a p-value of 1.580681e-05
Hypothesis 2: poisson.test(x=(176−40), T=176, r=0.99, alternative=“less”, conf.level=1−0.05/26711) results in a p-value 0.001539
Hypothesis 3: poisson.test(x=(185−40), T=185, r=0.99, alternative=“less”, conf.level=1−0.05/26711) results in a p-value of 0.002028
Hypothesis 4: poisson.test(x=80, T=361, r=0.01, alternative=“greater”, conf.level=1−0.05/26711) results in a p-value of 2.2e-16
In this instance, hypothesis 4 needs to be rejected, but not hypotheses 1, 2, and 3. Thus, even if a variation of 80 out of 361 appears to be significant, this does not hold if strict bio-statistical principles are employed. This also exemplifies that a high overall coverage is required to detect statistically significant variations. This filtering of significant variation does not require knowledge about a reference.
Next, the functional filtering is described.
Some genetic variations lead to a change in the sequence of the coded proteins, while others do not. Table B lists some properties of the most frequent types of functions of variations. Unfortunately the functional changes are not clearly disjunctive.
It is important for biologists and oncologists if a sequence variation in a known cancer gene changes the protein structure of the cancer gene. Only if the protein encoded by the cancer gene is significantly altered can the linkage of sequence variations to clinical outcome measures in the cancer patient be explained.
It is has become apparent from scientific publications that just the frequency of somatic sequence variations of a tumor is clearly related to outcome measures. Cancer patients with many, in fact hundreds of somatic sequence variations of their tumor can have a significantly better outcome than patients with few genetic variations in their tumor DNA.
Logical Operation with One or Two Operands
First, it is determined whether a predefined segment of a gene, here indicated with A, contains a particular type of genetic variation or not. A=TRUE is assigned if and only if at least one particular genetic variation (or a combination of types of genetic variations) is present on segment A, otherwise A=FALSE is assigned. In mathematical terms, the inventors conjoin the presence of a particular genetic variation (or combinations of types of genetic variations) over all positions of a gene segment and assign the results of this conjunction to a variable, here A. If advantageous for the prediction, the inventors can use the negation of the result of such a conjunction, here denoted with an exclamation mark in front of the symbol assigned to this segment, here A. Table C shows the truth table of the negation.
Such variables, denoting the existence of a particular type of genetic variation on disjunctive gene segments, here denoted with A and B, can be combined using one of the logical operators given in Tables B and C. It is known to skilled persons that such functions are ambiguous and are easily transformed using the rules of Boolean algebra. For example, A And B is the same as B And A, the law of commutability applies to all operators but the material implication and their negation. In digital electronics the Nand gate is used to represent other logical operations, as one can show using the truth tables that IA is equivalent to A Nand A, A And B is equivalent to (A Nand B) Nand (A Nand B), and A Or B is equivalent to (A Nand B) Nand (B Nand B).
Such transformations would defeat one of the purposes of the intervention, namely to produce prediction functions that are interpretable by biologists and/or oncologists. Likewise, the inventors could transform all logical operations in conjunctive or disjunctive normal form to make them unambiguous again with the loss of biological interpretability.
The reason for using logical operators to combine information on sequence variations is as follows. Typically, sequence variations in tumors are sparse. There are a few so-called hot-spots, which harbor up to 16% of all known variations in a tumor entity. Most importantly, the vast majority of sequence variations in tumors occur in a random fashion. Therefore, the information needs to be aggregated to be useful for
Next, the results of the aggregates, of better results of logical functions needs to be related to a particular manifestation of an outcome measure. This is facilitated by the cross classification of the result of one or more logical operations on two or more results of sequence variation analysis, see table F.
When aggregated over some observations that are patients with analyzed DNA, typical performance measures can be derived as shown in Table G. These measures can be used to evaluate and optimize the relation between the aggregation of sequence variations using logical operations and manifestations of clinical outcome measures. Optimization means minimization of miss-classification rate or costs, or maximization of one of the other measures. Keep in mind that any function with an area under the receiver operating characteristic curve (AROC) of 0.5 or higher has potential clinical utility.
The inventors implanted two strategies to construct predictive functions, a retrospective approach and a prospective approach. While the retrospective approach uses all available data, the prospective approach uses a double nested bootstrap procedure.
Briefly, in the double nested bootstrap procedures data of all available case/observation are split in three groups:
The inner loop procedure: After construction of the prediction function, and assessments of its performance, the prediction function is applied to the internal validation set. If the performance within the internal validation set is within the 95% confidence limits of the performance of the learning set, the prediction function is a candidate for prospective validation. The discovery set is randomly re-split in a set for construction of a prediction function, and an internal validation set. Again, the performance is evaluated on both sets. The inner loop is repeated many times, typically 100 times or more. The means of the measures of the performance of the repetitions is used to decide which prediction function shall be evaluated in a strict prospective fashion on the prospective validation set.
The outer loop procedure: In the outer loop the “best” prediction functions of the inner loops are assessed. Then the total set is again split randomly into the two sets of a prospective validation set and learning/internal validation set.
The outer loop procedure is also repeated many times, typically 100 or more times. Thus, the final result is a representation of 10000 or more repetitions.
The advantage of this approach is two-fold. First, the outer loop generates second order unbiased estimates for a future clinical validation. Second, the results are not prone to over fitting. The results are generalizable.
The disadvantage of this approach is also clear, only about 40% of the data are utilized for construction of prediction function and assessment of the performance.
The function may perform better if more data are used. Hence the retrospective approach might perform better, in particular in small datasets. Of course, using all data is prone to over fitting the prediction function to the actual data and loss of generalizability.
In some sense one could argue that the bootstrap gives a pessimistic estimate of the performance while the retrospective approach results in optimistic estimates.
The construction of the prediction function can be likened to regression trees. The nodes are the values of the distinct segments of the genes, TRUE if a particular sequence variation is detected, false otherwise. Additionally, the negations are used as nodes. However, those and only those gene segments can be used which are two-valued with respect to the filtered function(s) in the dataset.
For example, the inventors observed 3 segments of 3 genes, namely KRAS, BRAF, and APC. The nodes would be KRAS, IKRAS, BRAF, IBRAF, APC, and IAPC. Next, the inventors note the performance of each node using the measure of the outcome, either using the bootstrap or the retrospective approach.
Next, the inventors used the logical functions given in tables D and E, to generate logical combinations, or prediction functions. Just to give the first using the node KRAS from the KRAS-BRAF-APC example, the next layer of nodes within the tree would represent: KRAS And BRAF, KRAS And IBRAF, KRAS And APC, KRAS And IAPC, KRAS Nand BRAF, KRAS Nand IBRAF, KRAS Nand APC, KRAS Nand IAPC, KRAS Or BRAF, KRAS Or IBRAF, KRAS Or APC, KRAS Or IAPC, KRAS Nor BRAF, KRAS Nor IBRAF, KRAS Nor APC, KRAS Nor IAPC, KRAS Eqv BRAF, KRAS Eqv IBRAF, KRAS Eqv APC, KRAS Eqv IAPC, KRAS Xor BRAF, KRAS Xor IBRAF, KRAS Xor APC, KRAS Xor IAPC, KRAS Imp BRAF, KRAS Imp IBRAF, KRAS Imp APC, KRAS Imp IAPC, KRAS Nimp BRAF, KRAS Nimp IBRAF, KRAS Nimp APC, KRAS Nimp IAPC.
Once the information on one gene segment is part of the prediction function, is not used again; this restricts the number of layers in the tree to the number of different segments plus 1. However, the number of nodes within each layer is enormous. The foremost reason not to reuse a segment again is biological interpretability. [Recursive partitioning in contrast may resume the same variable over and over again.]
Attempts to just add segment information that increase the performance measure showed that it is possible to and a local maximum in the solution space, but that is not necessarily the overall maximum. Then, the inventors decided to compute the permutations of all possible combinations.
Taken together, the invention in some embodiments provides a method to identify and aggregate somatic sequence variation information contained in tumors of cancer patients in functions that have clinical use for prediction of manifestations of clinical outcome measures on those cancer patients, which allow for biological interpretation.
Exemplary Embodiments with Solid Tumors
In the following, the invention is described in relation to several types and stages of solid tumors, namely breast cancer, lung cancer, skin cancer (melanoma), ovarian cancer, pancreas cancer, prostate cancer, stomach cancer, and colorectal cancer. It will be understood by a person skilled in the art that the invention can also be practiced in relation to other types of solid tumor cancer based on the general knowledge of the skilled person together with the description provided herein.
In the method predicting a manifestation of an outcome measure of a cancer patient, at least two genes are analyzed for sequence variations. For this purpose, the genes are partitioned into segments of appropriate length. The length of the segments may vary from 20 base pairs to 500 base pairs, preferably from 50 base pairs to 250 base pairs. Such segments allow for a convenient and accurate determination of the sequence in order to find sequence variations in the DNA sample form the cancer patient.
The at least two genes that are analyzed are associated with the outcome measure of the patient, i.e. they are associated with the solid tumor cancer disease of the patient. In some embodiments of the invention, the at least two genes that are analyzed are chosen from a list of genes of Tables 1 to 8. Specifically, the genes associated with breast cancer are listed in Table 1; the genes associated with lung cancer are listed in Table 2; the genes associated with skin cancer (melanoma) are listed in Table 3; the genes associated with ovarian cancer are listed in Table 4; the genes associated with pancreas cancer are listed in Table 5; the genes associated with prostate cancer are listed in Table 6; the genes associated with stomach cancer are listed in Table 7; and the genes associated with colorectal cancer are listed in Table 8. For each gene listed with regard to a certain type of cancer, the number of sequence variations (“mutations”) Is given together with the number of samples that were analyzed and the mutation frequency resulting therefrom.
In the following, the invention will be described in relation to several types and stages of solid tumors, namely in respect to colorectal cancer of stage II (predicting outcome), colorectal cancer of stage IV (predicting response to treatment), and in patient derived xenografts (PDXs) of colorectal tumors.
In the following, the invention will be described in relation to several types and stages of solid tumors, namely in respect to colorectal cancer of stage II (predicting outcome), colorectal cancer of stage IV (predicting response to treatment), and in patient derived xenografts (PDXs) of colorectal tumors.
173 patients with colorectal cancer of UICC stage II for which follow-up data of 3 years was available were selected from the prospective MSKK study. Macro-dissection of FFPE samples of 173 Patients with Stage II Colorectal Cancer were used, for which a 3 year follow-up was available. 40/173 patients were diagnosed with metastases in liver, lung, or peritoneum. 27/173 patients were diagnosed with secondary malignancies. 12/173 patients were diagnosed with local recurrences. 94/173 patients had no progression of disease event. Tumor tissues of all 173 patients were deep sequenced using a cancer panel of known cancer genes. 96 tumor tissues were also subjected to exome sequencing using the Illumina HISeq. Raw sequence data was collected and analyzed.
Following DNA isolation, deep sequencing of selected cancer genes (oncogenes and tumor suppressor genes) with approximately 200 amplicons (˜30 kb). 2 gigabases raw sequence per run was performed. Multiplexing was between 12 fold, 24 fold, 48 fold and 96 fold. At 96 plex, coverage within the 200 amplicons is 200 to 2,000 fold. At 24 plex, coverage is 1,000 to 8,000 fold.
The number of screened patients from prospective multicenter MSKK study was 1481; 173 patients were selected from this group.
Progression of disease events are defined as: No progression within 3 years after resection of primary tumors, diagnosis of metastasis (liver, lung, peritoneal), diagnosis of local recurrence, and diagnosis of secondary malignancy. The following selection criteria were applied:
Below, examples of predictions functions that were found in retrospective analyses are described with respect to the tables. The prediction functions are based on missense sequence variations only (A) or on missense and nonsense sequence variations only (B) or on missense and nonsense and silent and synonymous mutations only (C).
Table 9 shows prediction functions and their performance based on sequence variations of one gene only.
Mutations: N1=396, N2=296
Minimum 2 Patients mutated in any given cancer gene
N=134, 40 Patients with Metastases, 94 Patients with no Recurrence
As can be seen in Table 9, !TP53 is the strongest single marker followed by KRAS and !APC, if optimization is performed for AROC (area under the curve). !TP53 is the strongest single marker followed by KRAS and PIK3CA if optimization is performed for combined Jaccard ratio. Preferred are prediction functions that comprise !TP53 or its equivalent TP53.
Table 10 shows the performance of prediction functions for 1 to 6 genes, based on missense mutations only.
As can be seen in Table 9, !TP53 has the largest single impact. The second best marker is XOR BRAF or its logic equivalence XOR !BRAF. The third best marker is OR SMO or ist logic equivalent. The fourth, fifth and six marker IAPC AND IPTEN AND IRET contribute only to the specificity of the function and increases specificity by 6% or 32 false positives versus 37 false positives in the function of 3.
If !TP53 is omitted completely in a function, the sensitivity decreases. Example: BRAF OR SMO AND !APC AND IPTEN AND IRET S+0.15, S−0.936, PPV 0.500, NPV 0.721, AROC 0.540, CJR 0.409. With a function length of six, the maximum of performance is reached. Longer functions do not perform better. After N=7, the performance decreases.
Functions optimized for AROC have a better performance with respect to sensitivity than strings optimized for combined Jaccard ratio. The position of a given marker in the string is not critical. !TP53 can be at the first, second or third position in a function of 3 or even at the sixth position in a function of 6.
The position of XOR BRAF or of OR SMO as well as the position of IAPC or !PTEN or !RET can be changed without change of performance.
Table 11 shows further preferred prediction functions.
Mutations N1=354; N2=465
Table 12 shows preferred prediction functions based on missense and nonsense sequence variations only and their clinical performance (sequence variations N1=354; N2=465), Performance of Best One to Six Genes.
As can be seen in Table 12, adding further genes up to 8 does not change performance of a function. Adding more than 8 sequence variation statuses leads to a decrease of performance.
Table 13 shows further preferred prediction functions for determining progression of disease in Stage II Colorectal Cancer as an outcome measure. The addition of nonsense sequence variations does not change the structure of the signatures, as there are only 42 additional sequence variations and preferentially only in TP53 and APC.
Mutations N1=1044; N2=800
Table 14 shows preferred prediction functions based on missense and nonsense and silent and synonymous sequence variations Only (sequence variations N1=1044; N2=800) and their performance.
Table 15 shows further preferred prediction functions based on missense and nonsense and silent and synonymous sequence variations Only (sequence variations N1=1044; N2=800) and their performance.
As can be seen, the use of missense sequence variations for predicting progression of disease is preferred in this example. Nonsense mutations add a little in performance, especially regarding specificity. Silent and synonymous sequence variations in functions do not add performance to functions of missense mutations alone. A function length of between 1 and 6 sequence variation statuses is preferred.
Table 16 shows best performing functions with missense and nonsense sequence variations and with a sensitivity >70%.
Table 17 shows best performing functions with missense mutations only and with a sensitivity >70%.
Table 18: Results of prediction functions were compiled based on missense and nonsense sequence variations in a prospective study. Data not adjusted.
Tables 19-26
33 Patients with Stage IV Colorectal Cancer for which Follow-up according to RECIST criteria was available. Patients were treated with Bevacizumab in combination with different chemotherapy schemes (Irinotecan, FOLFIRI or FOLFOX). 11 of 33 patients experienced response to treatment according to RECIST (total remission, partial remission). 22 of 33 patients experienced no response to treatment according to RECIST (stable disease, progression of disease).
Primary tumor tissue samples (FFPE, frozen samples) were macro-dissected, followed by DNA isolation. Deep sequencing of 212 amplicons in a panel of 40 selected cancer genes were performed in each of the 33 patients allowing high coverage for each base pair (ca. 34 kilobases of sequence for each patient). The coverage per base was 300-4,000 fold. This high coverage allows mutations to be identified with great confidence.
Table 19 shows prediction functions and performance data for the Prediction of Response to Treatment to Bevacizumab plus Chemotherapy in Patients with Advanced, Metastatic Colorectal Cancer of UICC Stage IV (Mutations N1=256 N2=96; Minimum of 1 Patient mutated in any given cancer gene; N=33: 11 Patients with Response; 33 Patients with no Response); the Performance of Single Genes is shown.
!TP53 is the strongest single marker followed by KRAS and IAPC if AROC (area under the curve) is optimized. !TP53 is the strongest single marker followed by KRAS and PIK3CA if AROC (Combined Jaccard Ratio) is optimized. For this application, a function of two genes is preferred comprising at least !TP53 or ist equivalent TP53.
Mutations Count 1: Gene must be mutated at least in 1/33 Patients
Table 20 shows the performance of 1 to 6 Genes wherein a gene must be mutated at least in 1/33 patients.
Mutations Count 2: Gene must be mutated at least in 2/33 Patients (>5% frequency)
Table 21 shows the performance of 2 to 6 Genes wherein a gene must be mutated at least in 2/33 patients.
Mutations Count 5: Gene must be mutated at least in 5/33 Patients (5% to 30% frequency)
Table 22 shows the performance of 2 to 6 Genes wherein a gene must be mutated at least in 2/33 patients.
The data presented above show that TP53, PIK3CA, !SMAD4 and !CTNNB1 have the largest single impact on performance of the prediction function. The second best marker after !TP53 is OR Kit or AND PIK3CA. The second best marker after PIK3CA is AND KRAS. The second best marker after ISMAD is OR ATM, and the second best marker after !CTNNB1 is AND !TP53.
With a function length of four genes, the maximum performance for AROC and CJR is reached for !CTNNB! AND !TP53 OR KIT AND MET and its equivalent string !TP53 OR KIT AND !CTNNB1 AND MET.
All gene markers can be moved freely from position 1 to 4 within the function without loosing performance.
With string length of five genes, the maximum performance for AROC is !TP53 OR KIT AND CTNNB1 AND !MET OR SMAD4, and for the combined Jaccard ration (CJR) the maximum performance is !CTNNB1 AND !TP53 AND !KDR AND !MET OR PIK3CA.
The difference between the performance of the seven best performance signatures is marginal and within the 95% confidence limits. Most signatures reach maximum performance with a function length of 5 genes, only one signature with a function length at 4 or 6 genes. Longer functions with more than 5 or 6 genes do not have an increased performance. Functions optimized by AROC have a better performance with respect to sensitivity than functions optimized by combined Jaccard ratio. The position of a given marker in the string is not critical.
Table 23 shows the performance of functions containing 3, 4 and 5 sequence variation statuses, based on missense sequence variations only.
The table shows that a function obtained with missense mutations alone has a slightly lower performance than function with missense and nonsense mutations. This might be due to the slightly increased number of mutations.
Table 24 shows the performance of functions containing 5, 6 and 7 sequence variation statuses, based on missense and synonymous sequence variations only.
Table 25 shows the performance of functions containing 4, 5, and 3 (the latter with mutation count 5) sequence variation statuses, based on missense and nonsense and synonymous sequence variations only.
Table 25B shows performance of exemplary functions
Transplantation of 239 human, primary colorectal tumors of patients with colorectal cancer of all four UICC stages was performed onto nude mice. 149 xenograft models were successfully engrafted. 133 xenograft models were quality checked versus matched primary human tumors. 75 tumors/xenograft models were selected for large therapy treatment experiments with three approved drugs in mCRC patients: Oxaliplatin, Cetuximab, and Bevacizumab. For each drug and each of the 67 xenograft models, five mice were treated in addition to five control animals (335 animals plus 335 controls per drug). At the end of the therapy experiment, the median diameter of the tumors (C) of the 5 control animals is devided by the median diameter of the five treated animals (T).
Table 26 shows the performance of functions containing 1, 2, 3, 4, 5, 6, 7, and 8 sequence variation statuses, based on missense and nonsense and synonymous sequence variations only. N1=131, N2=131.
Table 27: shows the performance of a preferred function (T/C<25. Mutation Count 5; R=11; NR=56; Tumor growth of PDXs must be inhibited by at least 75%).
Table 28: shows the performance of preferred functions (T/C<35. Mutation Count 5; R=19; NR=48).
Table 29 shows the best performing signatures with missense and nonsense, a mutation count of 2 (5% frequency) and with a sensitivity >70%.
Table 30 shows the best performing signatures with missense and nonsense, a mutations count of 5 (5-30% frequency) and with a sensitivity >70%.
Table 31: shows performance of preferred functions (T/C</=30. 13 Responder PDXs, 54 Nonresponder PDXs, Tumor growth of PDXs are inhibited by at least 70%.
Table 32: shows performance of preferred functions (T/C</=35. 19 Responder PDXs, 48 Nonresponder PDXs, Tumor growth of PDXs are inhibited by at least 65%.)
Table 33: shows performance of preferred functions (T/C</=25. 11 Responder PDXs, 56 Nonresponder PDxs, Tumor growth of PDXs are inhibited by at least 75%.—
From the above, the following can be concluded. The most useful information for predicting response to treatment with bevacizumab and chemotherapy are missense and nonsense mutations of cancer genes. Nonsense mutations add a little bit in performance, especially with regard to specificity. Silent and synonymous mutations in functions add performance to functions base on missense and nonsense mutations alone. Function length is best between 2 and 6 genes.
350 patients with colorectal cancer of UICC stage III for which follow-up data of at least two years was available were selected from the prospective MSKK study. The following selection criteria were applied:
Patients had received standard adjuvant chemotherapy including 5-fluorouracil, leucovorin, and oxaliplatin (FOLFOX scheme), or 5-fluorouracil and leucovrin. Some patients received oral capecitabine instead of infusional 5-fluorouracil. Progression of disease events are defined as: (i) no progression within 3 years, four years or five years after resection of primary tumors, (ii) diagnosis of metastasis (liver, lung, peritoneal), (iii) diagnosis of local recurrence, and diagnosis of secondary malignancy.
Of the 350 patients with a two year follow up 24 patients had distant metastasis (mainly liver metastasis), 4 patients had a local recurrence or a secondary malignancy, and 13 patients had death as progression event. 309/350 patients had no progression of disease event. Of the 289 patients with a three year follow up, 42 patients distant metastasis (mainly liver metastasis), 6 had a local recurrence or a secondary malignancy, and 14 patients had death as progression event. 227/289 patients had no progression of disease event. Of the 242 patients with a four year follow up, 57 patients had distant metastasis (mainly liver metastasis), 8 had a local recurrence or a secondary malignancy, and 16 patients had death as progression event. 161/242 patients had no progression of disease event. Of the 186 patients with a five year follow up, 66 patients had distant metastasis (mainly liver metastasis), 6 patients had a local recurrence or a secondary malignancy, and 20 patients had death as progression event. 94/186 patients had no progression of disease event.
Macro-dissection of cryo tumor and FFPE tumor samples of 350 Patients with stage III colorectal cancer were used. Tumor DNA was isolated using an automated method on the Qiacube robot (Qiagen, Germany). Tumor DNA was quantified, and at least 250ng of tumor DNA of all 350 patients were deep sequenced using the illumine MiSeq sequencer and a cancer panel of 37 known cancer genes organized in 120 distinct amplicons. Up to 96 sequenced samples were multiplexed per MiSeq run. Raw sequence data was collected and analyzed.
Below, examples of predictions functions that were found in retrospective analyses are described with respect to the tables. The prediction functions are based on missense and nonsense sequence variations only which alter the function of the encoded protein.
Table 34 shows various prediction functions of the best performing genes for predicting metastasis in distant organs as progression of disease in patients with colorectal cancer of stage III who underwent RO resection and were treated using adjuvant chemotherapy. Overall survival is the event time.
In the group of patients with a three year follow up (N=233), 42 patients had a metastasis event while 191 patients remained without any progression of disease event. SMAD4mi (nonsense mutations in the SMAD4 gene) was the strongest single marker of 11 cancer genes which showed missense and nonsense mutations in at least five patients. SMAD4mi showed a sensitivity S+ of 0.262 and a specificity S− of 0.937, and an area under the receiver operating characteristic curve (AROC) of 0,600. Adding the next marker OR KITmi improved S+ to 0.500, reduced S− to 0.817 and improved AROC to 0.658. The prediction function of two markers reads as follows: missense mutations in the SMAD4 gene, or missense mutations in the KIT gene, or missense mutations in both the SMAD4 gene and the KIT gene predict patients with colorectal cancer of stage III with higher risk of metastasis as progression of diseases who have a three year follow up time. Adding a third marker OR FBXW7mi improves the AROC to 0.684. The prediction function of three markers reads as follows: missense mutations in the SMAD4 gene, or missense mutations in the KIT gene, or missense mutations in the FBXW7 gene, or missense mutations in any two of the three genes, or missense mutations in all three genes predict patients with colorectal cancer of staOR SMADge III with higher risk of metastasis as progression of disease who have a three year follow up time. The prediction function can be further improved by adding two markers XOR ATMmi and XOR METmi. The prediction function with these five markers has an AROC of 0.716. Any further marker does not increase the accuracy of the prediction function.
In the group of patients with a four year follow up (N=192), or a five year follow up (N=142), we observed the same prediction function of three markers: IAPCns OR SMAD4mi OR FBXW7mi. IAPCns (no nonsense mutations in the APC genes) turned out to be the strongest single marker of the 11 cancer genes which showed missense and nonsense mutations in at least five patients. IAPCns showed a sensitivity S+ of 0.509, a specificity S− of 0.696, and a area under the operating receiver characteristics curve AROC of 0.603 (four year follow up). In the patient group with five year follow up IAPCns had a S+ of 0.485, S− of 0.763, and an AROC of 0.624. The next strongest marker was OR SMAD4mi improving the AROC to 0.642 and 0.658 in the patients with four or five year follow up, respectively. Finally the maximum of the prediction curve was reached by adding as third marker OR FBXW7mi. This signature showed an AROC of 0.660 and 0.678 in the patients with four years or five years observation time, respectively.
Table 35 shows various prediction functions in the same patient groups with colorectal cancer of stage III if progression free survival (PFS) is the event time and not overall survival and using distant metastasis as the event. Prediction functions are very similar to those shown in Tab. 34. The best performing signature for patients with a follow up time of 5 years is IAPCns OR FBXW7 OR SMAD4mi with a S+ of 0.629, a S− of 0.678 and an AROC of 0.653. This prediction function differs only from Table 34 in that OR FBXW7 is at the second position and OR SMAD4mi is at the third position.
Table 1: Genes associated with breast cancer.
Table 2: Genes associated with lung cancer.
Table 3: Genes associated with skin cancer (melanoma).
Table 4: Genes associated with ovarian cancer.
Table 5: Genes associated with pancreas cancer.
Table 6: Genes associated with prostate cancer.
Table 7: Genes associated with stomach cancer.
Table 8: Genes associated with colorectal cancer.
Table 9: Prediction of Progression of Disease in Stage II Colorectal Cancer, Missense Mutations Only (Sequence variations: N1=396, N2=296, Minimum 2 Patients mutated in any given cancer gene; N=134, 40 Patients with Metastases, 94 Patients with no Recurrence).
Table 10: Prediction of progression of disease in Stage II Colorectal Cancer, Missense Mutations Only, Performance of One to Six Genes.
Table 11: Prediction of Progression of Disease in Stage II Colorectal Cancer, Missense sequence variations only, Other preferred prediction functions.
Table 12: Prediction of Progression of Disease in Stage II Colorectal Cancer, Missense and Nonsense sequence variations Only (sequence variations N1=354; N2=465), Performance of Best One to Six Genes.
Table 13: Prediction of Progression of Disease in Stage II Colorectal Cancer, Preferred prediction functions.
Table 14: Prediction of Progression of Disease in Stage II Colorectal Cancer, Missense and Nonsense and Silent and Synonymous sequence variations only (sequence variations N1=1044; N2=800); Performance of Best One to Six Genes.
Table 15: Prediction of Progression of Disease in Stage II Colorectal Cancer, Missense and Nonsense and Silent and Synonomous Mutations only (sequence variations N1=1044; N2=800); preferred prediction functions.
Table 16: Prediction of Progression of Disease in Stage II Colorectal Cancer, Best performing prediction function with missense and nonsense mutations and with a sensitivity >70%.
Table 17: Prediction of Progression of Disease in Stage II Colorectal Cancer, best performing prediction function with missense mutations only and with a sensitivity >70%.
Table 18: Results of prediction functions were compiled based on missense and nonsense sequence variations in a prospective study. Data not adjusted.
Tables 19 to 33: Prediction of Response to Treatment to Bevacizumab plus Chemotherapy in Patients with Advanced, Metastatic Colorectal Cancer of UICC Stage IV.
Table 19: Prediction of Response to Treatment to Bevacizumab plus Chemotherapy in Patients with Advanced, Metastatic Colorectal Cancer of UICC Stage IV. Shows prediction functions and performance data (Sequence variations N1=256, N2=96; Minimum of 1 Patient mutated in any given cancer gene; N=33: 11 Patients with Response; 33 Patients with no Response); Performance of Single Genes is shown.
Table 20: Prediction of Response to Treatment to Bevacizumab plus Chemotherapy in Patients with Advanced, Metastatic Colorectal Cancer of UICC Stage IV. Performance of 1 to 6 Genes wherein a gene must be mutated at least in 1/33 patients.
Table 21: Prediction of Response to Treatment to Bevacizumab plus Chemotherapy in Patients with Advanced, Metastatic Colorectal Cancer of UICC Stage IV. Shows the performance of 2 to 6 Genes wherein a gene must be mutated at least in 2/33 patients.
Table 22: Prediction of Response to Treatment to Bevacizumab plus Chemotherapy in Patients with Advanced, Metastatic Colorectal Cancer of UICC Stage IV. Shows the performance of 2 to 6 Genes wherein a gene must be mutated at least in 5/33 patients.
Table 23: Prediction of Response to Treatment to Bevacizumab plus Chemotherapy in Patients with Advanced, Metastatic Colorectal Cancer of UICC Stage IV. Shows the performance of functions containing 3, 4 and 5 sequence variation statuses, based on missense sequence variations only.
Table 24 shows the performance of functions containing 5, 6 and 7 sequence variation statuses, based on missense and synonymous sequence variations only.
Table 25 shows the performance of functions containing 4, 5, and 3 (the latter with mutation count 5) sequence variation statuses, based on missense and nonsense and synonymous sequence variations only.
Table 26 shows the performance of functions containing 1, 2, 3, 4, 5, 6, 7, and 8 sequence variation statuses, based on missense and nonsense and synonymous sequence variations only. N1=131, N2=131.
Table 27: (T/C<25. Mutation Count 5; R=11; NR=56; Tumor growth of PDXs must be inhibited by at least 75%)
Table 28: (T/C<35. Mutation Count 5; R=19; NR-48) Table 29 shows the best performing signatures with missense and nonsense, a mutation count of 2 (5% frequency) and with a sensitivity >70%.
Table 30 shows the best performing signatures with missense and nonsense, a mutations count of 5 (5-30% frequency) and with a sensitivity >70%.
Tables 31 to 33: Response to bevacizumab monotherapy in patient derived xenografts (PDXs)
Table 34: Prediction functions and performance data for the prediction of progression of disease in patients with colorectal cancer of stage III who underwent surgical RO resection followed by standard adjuvant chemotherapy. Prediction functions were based on deep sequencing data of 37 key cancer genes organized in 120 amplicons and analysis of missense and nonsense mutations if they occurred in at least five patients using Boolean operators. Patients had different follow up times: 365 days (1 year), 731 days (2 years), 1.096 days (3 years), 1.461 days (4 years), and 1.826 days (5 years). Metastasis to distant organs was the measured event compared to patients who did not show any event (metastasis, local recurrence, secondary malignancy, death) in the same follow up period. Event time is overall survival (OS).
Tab 35: Prediction functions and performance data for the prediction of progression of disease in patients with colorectal cancer of stage III who underwent surgical RO resection followed by standard adjuvant chemotherapy. Prediction functions were based on deep sequencing data of 37 key cancer genes organized in 120 amplicons and analysis of missense and nonsense mutations if they occurred in at least five patients using Boolean operators. Patients had different follow up times: 365 days (1 year), 731 days (2 years), 1.096 days (3 years), 1.461 days (4 years), and 1.826 days (5 years). Metastasis to distant organs was the measured event compared to patients who did not show any event (metastasis, local recurrence, secondary malignancy, death) in the same follow up period. Event time is progression-free survival (PFS).
From the 13 genes displaying statistically significant missense or nonsense mutations (also found in the COSMIC database), TPS3 has the largest single gene impact on performance of the signature with respect to predicting metastasis. The element !TP53 which reads “No missense and nonsense mutations in TP53” has a sensitivity (S+) of 0.59, a specificity (S−) of 0.63, a positive predictive value (PPV) of 0.41 and negative predictive value (NPV) of 0.78.
The first element !TP53 is now connected with the second element IBRAF using the Boolean operator Eqv. The meaning of the first two elements of the signature !TP53 Eqv IBRAF is as follows: “Patients who have neither missense nor nonsense mutations in TP53 and BRAF, or patients who have missense or nonsense mutations in both genes, have the highest likelihood of developing metastatic disease”. !TP53 Eqv IBRAF has the following performance: S+ 0.74, S− 0.65, PPV 0.48, NPV 0.86, AROC 0.69.
The addition of Eqv IBRAF increases S+ by 0.15 and S− by 0.02. The addition of OR SMAD4 missense or nonsense mutations shows no improvement. This holds not up in the prospective validation.
Further extension of the signature by OR ATM OR KRAS does not improve overall performance as measured by the AROC. However, a signature with five elements !TP53 Eqv IBRAF Or SMAD4 OR ATM or KRAS leads to increased sensitivity of 0.89, however on the expense of a lower specificity of 0.39. Such a signature with high sensitivity might be of use for selection of patients at high risk of metastasis for a chemotherapy study. The signature would predict 36 True Positives of the 40 patients with the risk of metastasis correctly. Only 4 patients with high risk of metastasis would not be identified and would be False Negatives. However, of the 94 patients with no risk of progression the signature would only identify 37 correctly as True Negatives, thus leading to 57 False Positive patients.
The results of the prospective discovery can be complemented by the retrospective analysis shown in
In the retrospective analysis the addition of OR PIK3CA to the function of four elements !TP53 XOR BRAF AND !FLT3 OR ATM leads to an increased sensitivity of 0.775 and a decreased specificity of 0.543. Thus 32 of the 40 high risk patients and 51 of the 94 patients with no risk of progression of disease were identified correctly. Addition of OR KRAS instead of OR PIK3CA leads to a further increase of sensitivity similar to the prospective analysis.
The signature !TP53 XOR BRAF AND !PIK3CA has a sensitivity of 55% and a specificity of 71%. By exchanging AND ! PIK3CA through OR PIK3CA one achieves a sensitivity of 77.5% and a specificity of 54.3%, hence one has swapped sensitivity for specificity without change to positive, or negative predictive value, or AROC.
Extraction of nucleic acids from the tissue samples was performed using the AIIPrep DNA/RNA Mini Kit (Qiagen, Hilden). The preparation was done on a Qiacube robot from Qiagen. Starting material was approximately 10-20 mg of cryo preserved tumor tissue cut in 4 μm slices on a cryotom.
Before starting the protocol the following things need to be prepared:
Add 350 μl of Buffer RLT Plus and vortex well until tissue gets dissolved. Centrifuge 3 minutes at maximum speed (14000 g). Transfer the supernatant directly into a 2 ml Safe-Lock tube.
Prepare the Qiacube robot:
Prepare the Qiacube robot:
After finishing the program the RNA tubes are stored at −80° C. The used rotor adapter are discarded and the robot is cleaned up.
During this step, a custom pool containing upstream and downstream oligos specific to the targeted regions of interest is hybridized to your genomic DNA samples.
This process removes unbound oligos from genomic DNA using a filter capable of size selection.
This process connects the hybridized upstream and downstream oligos. A DNA polymerase extends from the upstream oligo through the targeted region, followed by ligation to the 5′ end of the downstream oligo using a DNA ligase. This results in the formation of products containing your targeted regions of interest flanked by sequences required for amplification.
In this step, your extension-ligation products are amplified using primers that add index sequences for sample multiplexing (i5 and i7) as well as common adapters required for cluster generation (P5 and P7).
Xenograft models provide sufficient tissue material for molecular studies of biomarkers that are predictive for response/nonresponse to therapy and can be used as companion diagnostics (CDx).
Shortly after surgery, original colorectal cancer tumor pieces were shipped in gentamicin containing RPMI-1640 medium to the mouse facility. After arrival at the mouse facilities they were transplanted onto immunodeficient mice and were further passaged until a stably grown tumor xenografts has developed.
Surgical colorectal tumor samples were cut into pieces of 3 to 4 mm and transplanted within 30 min s.c. to 3 to 6 immunodeficient NOD/SCID mice (Taconic); the gender of the mice was chosen according to the donor patient. Additional tissue samples were immediately snap-frozen and stored at −80° C. for genetic, genomic, and protein analyses. All animal experiments were done in accordance with the United Kingdom Co-ordinating Committee on Cancer Research regulations for the Welfare of Animals and of the German Animal Protection Law and approved by the local responsible authorities. Mice were observed daily for tumor growth. At a size of about 1 cm3, tumors were removed and passaged to naive NMRI: nu/nu mice (Charles River) for chemosensitivity testing. Tumors were passaged no more than 10 times. Numerous samples from early passages were stored in the tissue bank in liquid nitrogen and used for further experiments. Several rethawings led to successful engraftment in nude mice. All xenografts as well as the corresponding primary tumors were subjected to histological evaluation using snap-frozen, haematoxylin-eosin-stained tissue sections.
75 xenograft models were used in therapy experiments testing responsiveness towards drugs approved in the treatment of patients with colorectal cancer including cetuximab as an anti-EGRF antibody, bevacizumab, and oxaliplatin. Each of the 75 tumors was transplanted onto 20 mice (5 controls and 5 for each drug). Models with treated-to-control ratios of relative median tumor volumes of 20% or lower were defined as responders.
The chemotherapeutic response of the passagable tumors was determined in male NMRI: nu/nu mice. For that purpose, one tumor fragment each was transplanted s.c. to a number of mice. At palpable tumor size (50-100 mm3), 6 to 8 mice each were randomized to treatment and control groups and treatment was initiated. If not otherwise mentioned, the following drugs and treatment modalities were used: Bevacizumab (Avastin®; Genentech Inc., South San Francisco, Calif., USA) 50 mg/kg/d, qd 7×2, i.p., Cetuximab (Erbitux; Merck) 50 mg/kg/d, qd 7×2, i.p.; Oxaliplatin (Eloxatin, Sanofi-Avensis), 50 mg/kg/d, qd1-5, I.p. Doses and schedules were chosen according to previous experience in animal experiments and represent the maximum tolerated or efficient doses. The injection volume was 0.2 ml/20 g body weight.
Tumor size was measured in two dimensions twice weekly with a caliper-like instrument. Individual tumor volumes (V) were calculated by the formula: V=(length+[width]2)/2 and related to the values at the first day of treatment (relative tumor volume). Median treated to control (T/C) values of relative tumor volume were used for the evaluation of each treatment modality and categorized according to scores (− to ++++;). The mean tumor doubling time of each xenograft model was calculated by comparing the size between 2- and 4-fold relative tumor volumes. Statistical analyses were done with the U test (Mann and Whitney) with P<0.05. The body weight of mice was determined every 3 to 4 days and the change in body weight was taken as variable for tolerability.
Genomic DNA and total RNA were simultaneously extracted with AllPrep DNA/RNA Mini Kit (automated protocol using the QIACube) according to the manufacturer's instructions. DNA and RNA concentrations (ng/μl) were measured using UV spectrophotometer (Nanovue, GE Healthcare).
| Number | Date | Country | Kind |
|---|---|---|---|
| 13152610.5 | Jan 2013 | EP | regional |
| 13152797.0 | Jan 2013 | EP | regional |
This application claims the benefit of U.S. Provisional Application No. 61/756,801 filed Jan. 25, 2013, which is hereby incorporated by reference in its entirety. This application further claims priority to EP 13152610.5 filed Jan. 25, 2013 and to 13152797.0 filed Jan. 25, 2013, both of which are hereby incorporated by reference in their entireties.
| Number | Date | Country | |
|---|---|---|---|
| 20140342925 A1 | Nov 2014 | US |
| Number | Date | Country | |
|---|---|---|---|
| 61756801 | Jan 2013 | US |
