Trials that accrue participants on the basis of genetic biomarkers are a powerful means of testing targeted drugs, but they are often complicated by the rarity of the biomarker-positive population. Umbrella trials circumvent this by testing multiple hypotheses to maximize accrual. However, bigger trials have higher chances of conflicting treatment allocations because of the coexistence of multiple actionable alterations; allocation strategies greatly affect the efficiency of enrollment and should be carefully planned on the basis of relative mutation frequencies, leveraging information from large sequencing projects.

We developed software named Precision Trial Drawer (PTD) to estimate parameters that are useful for designing precision trials, most importantly, the number of patients needed to molecularly screen (NNMS) and the allocation rule that maximizes patient accrual on the basis of mutation frequency, systematically assigning patients with conflicting allocations to the drug associated with the rarer mutation. We used data from The Cancer Genome Atlas to show their potential in a 10-arm imaginary trial of multiple cancers on the basis of genetic alterations suggested by the past Molecular Analysis for Personalised Therapy (MAP) conference. We validated PTD predictions versus real data from the SHIVA (A Randomized Phase II Trial Comparing Therapy Based on Tumor Molecular Profiling Versus Conventional Therapy in Patients With Refractory Cancer) trial.

In the MAP imaginary trial, PTD-optimized allocation reduces number of patients needed to molecularly screen by up to 71.8% (3.5 times) compared with nonoptimal trial designs. In the SHIVA trial, PTD correctly predicted the fraction of patients with actionable alterations (33.51% [95% CI, 29.4% to 37.6%] in imaginary v 32.92% [95% CI, 28.2% to 37.6%] expected) and allocation to specific treatment groups (RAS/MEK, PI3K/mTOR, or both).

PTD correctly predicts crucial parameters for the design of multiarm genetic biomarker-driven trials. PTD is available as a package in the R programming language and as an open-access Web-based app. It represents a useful resource for the community of precision oncology trialists. The Web-based app is available at https://gmelloni.github.io/ptd/shinyapp.html.

The availability of sequenced cancer genomes and the generation of new targeted drugs offer unprecedented opportunities for personalized treatment for patients with cancer. Indeed, treatment within genomic-based clinical trials significantly improves clinical outcome.1 However, identification of the best biomarker-drug combination for patient stratification and treatment remains a critical research and ethical challenge.

Large multi-institutional projects such as The Cancer Genome Atlas (TCGA) revealed extreme heterogeneity of cancer mutational landscapes, with a prevalence of low-frequency mutations, including those critical for tumor growth (driver mutations) and those that might be targetable by specific drugs (actionable mutations).2,3 This scenario renders the conduction of adequately powered genome-based clinical trials prohibitive in many cases: the biomarker-positive population is often rare, and identifying it requires screening large numbers of patients.

Strategies to circumvent this issue are proposed through novel trial designs in which multiple drugs and/or their matched biomarkers are tested simultaneously. These novel designs, usually defined as basket (one biomarker-drug pair, multiple histologies), umbrella (multiple biomarker-drug pairs, one histology), or platform (multiple biomarker-drug pairs, multiple histologies),4-6 may allow reducing the number of patients needed to molecularly screen (NNMS) to achieve the desired statistical power for multiple treatment subgroups. However, these approaches have potential disadvantages: first is the occurrence of a large fraction of patients with potential allocation conflicts resulting from the concomitance of genetic alterations associated with different drugs,2 and second is the high costs and turnaround time of the molecular screening, which must include sequencing enough genomic space to accommodate the study’s need. The latter will become more critical in the near future as more complex, polygenic genomic signatures are used as stratification biomarkers (eg, multiple homologous recombination genes to establish BRCAness as a predictor of response to poly [ADP-ribose] polymerase [PARP] inhibitors7). For these reasons, patient allocation to individual trial arms significantly deviates from expectations in multigenic umbrella trials, which leads to inadequate statistical power8,9 and the inability to reach conclusions.

Here, we present Precision Trial Drawer (PTD), a comprehensive bioinformatics tool that takes advantage of patient-level genomic information from TCGA or alternative sources to simulate genomically characterized cohorts and provide critical parameters for designing genetic biomarker-driven trials. We analyzed the usefulness of PTD in an imaginary trial that was based on recommendations from the Molecular Analysis for Personalised Therapy (MAP)10 consensus conference for metastatic cancer and demonstrated its predictive power by retrospective evaluation of the SHIVA (A Randomized Phase II Trial Comparing Therapy Based on Tumor Molecular Profiling Versus Conventional Therapy in Patients With Refractory Cancer) trial, the largest genome-based umbrella trial published to date.8

PTD was developed in the R programming language. A full compendium of all the package features and methodology can be found at https://gmelloni.github.io/ptd/PrecisionTrialDrawer.html (vignette). To reproduce the analyses proposed in this work, refer to https://gmelloni.github.io/ptd/PTD_panel_showcase.html (MAP simulation), https://gmelloni.github.io/ptd/SHIVA_analysis.html (SHIVA simulation).

More precise discussion on the statistical considerations for trial design can be found in the following chapters of the vignette: https://gmelloni.github.io/ptd/PrecisionTrialDrawer.html#priority_trial and https://gmelloni.github.io/ptd/PrecisionTrialDrawer.html#basket_trial_design:_tumorfreqs_and_tumorweights_parameters.

The Web-based interface was developed by using R shiny version and can be accessed at https://gmelloni.github.io/ptd/shinyapp.html. The original R code is available upon request.

Definitions

PTD analyses can be applied to the design of clinical trials in which patient recruitment is based on genetic alterations as assessed by next-generation sequencing of specific genomic regions obtained by probe hybridization or polymerase chain reaction amplification (target enrichment). The ensemble of target genomic coordinates for the molecular screening is called the “sequencing panel.” Detected fraction (DF) is the fraction of patients who carry at least one of the target molecular alterations. DF is used to calculate the NNMS (ie, the minimum number of patients that should be tested to obtain the desired sample size as calculated on the expected treatment effect by using conventional power analyses). NNMS is usually a fraction of the number needed to screen (NNS), which incorporates nongenomic reasons for screening failure (such as poor clinical conditions, disease progression during screening, or withdrawal of consent) and is not considered by PTD. “Actionable mutations” refers to those genetic alterations that dictate a therapeutic choice within the trial.

Description of the Package

PTD includes a series of functions in the R programming language to (1) translate genomic information into genomic coordinates and optimize coordinates for the design of a targeted sequencing panel, (2) retrieve single or grouped genetic alteration frequencies from public databases (by default TCGA, but other sources can be used) that maintain patient-level associations, (3) simulate genomically defined patient cohorts by extracting frequencies of co-occurrence of single or grouped mutations, and (4) calculate NNMS for defined power and/or sample size (Fig 1A). A Web-based interface with slightly reduced capabilities is available for quick analyses at https://gmelloni.github.io/ptd/shinyapp.html.

To describe a typical PTD workflow and illustrate its potential, we simulated an umbrella trial on the basis of recommendations issued by the recent MAP consensus conference.10 MAP has proposed panels of genomic biomarkers for three tumor types (lung, stomach, and breast) to be assessed in daily practice through molecular screening programs. We also added colorectal cancer to our analysis, which is largely covered by the panel of genes proposed by MAP for stomach cancer. In our imaginary trial, patients are allocated to 10 possible drugs, including inhibitors of PARP, NOTCH, MET, HER2, FGFR, EGFR, BRAF, ALK, AKT, and immune checkpoints, and covering all the allocations linked to the MAP biomarkers. Allocation to checkpoint inhibitors was based on the detection of inactivating mutations in MLH or MSH genes, regardless of tumor type, which are highly correlated with microsatellite instability, a now-established predictor of immune checkpoint response.11,12 PTD was then used to design a dedicated custom targeted sequencing panel, identify the optimal allocation rule, and estimate sample size and statistical power.

The list of genomic features associated with each drug is provided in Table 1. Illustrations of the workflow, the codes used for generating data, and detailed PTD functionalities are available at https://gmelloni.github.io/ptd/shinyapp.html.

Table

Table 1. Genetic Biomarkers and Drug Allocation Scheme for the MAP Consensus Analysis

Input Definition

The first step involves designing the sequencing panel, which is imported as a standard Excel table with official gene symbols, type and exact alteration to be considered, and two possible grouping variables (gene-linked drug and, if desired, a generic group of any sort). Table 1 presents the panel used in the MAP analysis. The panel can also be expanded to include genes or genomic positions without therapeutic information, which may nevertheless be relevant for prognostic stratification, prediction of drug toxicity, or a generic biologic interest (grouped as driver genes in the example).

Four classes of genetic alterations can be explored: single nucleotide variants (SNVs), copy number alterations (CNAs), translocation-associated fusion transcripts, and gene expression. Alterations can be further specified, for example, to include amplifications or deletions among CNAs or specific nucleotide substitutions among SNVs. Variants can be encoded using any of several accepted formats. For instance, the BRAF V600E mutation can be coded using amino acid notation (V600E), the Single Nucleotide Polymorphism Database (dbSNP) identifier (rs113488022), a set of genomic coordinates (eg, chr7: 140753336-140753337), or as genomic variants (7:140753336 A>C). Once the input file is generated and uploaded into the working environment, the exact list of variants can be further refined in several ways. Mutations of interest can be batch filtered with any set of genomic coordinates to eliminate variants likely to be passengers.3,13 In this example, we selected only variants listed as probably pathogenic in the Catalogue of Somatic Mutations in Cancer (COSMIC) database,14 thus eliminating 33.5% of the initial variants. Alternatively, individual genes can be visually inspected and edited by using an interactive interface built on the LowMACA (Low Frequency Mutations Analysis via Consensus Alignment) algorithm15 (manual filtering; Fig 1B). In this example, we manually excluded EGFR T790M and PIK3CA exon 20 mutations associated with resistance to first-generation EGFR inhibitors.16,17 Specific mutations and/or drugs can be associated with specific tumors and excluded from others. For example, one can decide to associate BRAF inhibitors with BRAF mutations in lung cancer but not in colorectal cancer, or to exclude T790M carriers (even when EGFR mutated elsewhere) from being associated with first-generation tyrosine kinase inhibitors.

Simulation

Once the coordinates and types of genetic alterations are defined, PTD will sample the patient population database iteratively to obtain for each alteration a frequency of occurrence with empirically calculated CIs (by default, 100 random subsamples for each estimate). Default databases are cBioPortal and Tumor Fusion Gene Data Portal,18,19 but any appropriately formatted mutation data set can be used. For panels that contain multiple alteration types (eg, SNVs and CNAs), PTD will construct a reference data set with the intersection of the relative data sets in the source database (not all patients in TCGA have data from all sequencing platforms).

In projects that allow inclusion of multiple tumor histologies, their frequency of occurrence in the real world is likely to differ significantly from the relative representation in the reference data set, and this can be corrected by PTD by weighting the probability of sample extraction. In this example, we used relative frequencies obtained from the National Cancer Institute SEER Program (Fig 2A shows the comparison between the TCGA and the example populations).

Output

The percentage of patients with at least one actionable alteration (detected fraction; DF) is probably the most crucial estimate that can be obtained through PTD, which is 65.4% in this example (Fig 2B). Notably, 20.1% of patients showed more than one actionable mutation (up to nine; Fig 2B). PTD allows breakdown of the DF by drug or by disease histology (Fig 2C-D) or any other prespecified group.

To optimize panel design, one can evaluate the trade-off between the feature of interest (eg, available drugs or the inclusion of genes with uncertain predictive power) and the size of the genomic space to be sequenced. The decision to include additional mutations may depend on gain in terms of a biomarker-positive population against an increase in the target genomic space to be sequenced. This trade-off can be visualized with a saturationPlot that expresses the mean number of alterations per patient or the number of patients with at least x alterations as a cumulative function of the genomic space, with alterations ranked in descending frequency order (absolute or gene length–normalized; Fig 2E). The same plot can be obtained by using drugs or other grouping variables (Fig 2F).

The coocMutexPlot indicates the degree of co-occurrence of grouped genetic alterations, thus providing an estimate of the frequency of allocation conflicts (Fig 2G). This is particularly useful in designing umbrella trials, in which allocation conflicts may increase together with the number of drugs being simultaneously investigated. This plot can be used to decide which drugs should be combined in the same trial because prioritizing those associated with highly mutually exclusive mutations will maximize the biomarker-positive population.

Power Analysis and Calculation of the NNMS

Multidrug trials can be powered to test the experimental arm as a whole, ignoring the number of patients allocated to each individual drug, in which case the NNMS is a simple linear function of the target sample size divided by the detection power (in our example, if the given target sample size = n, the NNMS will be n/0.654; no allocation rule design). However, this approach may well have insufficient power to detect the efficacy of single drugs. If this is a desired end point, the trial can be divided into a series of drug-specific subtrials, each having its own target sample size.9 Probability of allocation can vary significantly if the associated biomarkers have very different prevalence. Furthermore, increasing the number of competing arms also increases the likelihood of conflicting allocations; rules to follow in such cases should be specified in advance to avoid allocation biases. Such rules can be based on general a priori considerations that do not take into account mutation frequencies (eg, allocation to endocrine therapy only if no mutation is present8; manual allocation design). Alternatively, rules may be set to maximize the fraction of allocated patients in an unbiased way. If information regarding the co-occurrence of mutations is not considered, proper NNMS-based sample size calculation is not possible because considering each subtrial independently leads to gross overestimation. With PTD, arms can be ranked in order of expected frequency of allocation, so that if a patient bears mutations associated with two drugs, he or she will be systematically assigned to the one with the lowest frequency. PTD can provide such prediction, identify the rule, calculate NNMS based on the rule, and simulate patient allocation so that each drug-specific arm reaches at least the desired level of power (optimal allocation design; Fig 3A). We can thus precisely quantify the advantage of an optimized umbrella design: for the given example, if one were to run 10 independent one-arm studies with a target sample size of 13 to show a response rate of 40% against historical controls of 10% (for α = .05 and power = 0.8), a total of 4,276 patients would need to be sequenced to obtain 130 (ie, 10 × 13) biomarker-positive patients. Testing the same 10 hypotheses in a single umbrella study using an optimized allocation rule requires sequencing of only 1,235 patients to obtain the same 130 biomarker-positive patients, only 28.2% for a 3.5-fold reduction (Fig 3B). PTD can estimate the NNMS across a range of trial designs, allowing the researcher to specify different types of end points (time to event, progression-free survival, overall survival, time to progression, or proportions such as response rate and survival rate), number of arms, allocation ratios (symmetric or asymmetric), and other conventional trial parameters. Simulation results for one- or two-arm studies with time-to-event or proportion end points and with no, manual, or optimal allocation rules are provided in Fig 3B. Simulation results across a range of powers and target proportions are provided in Fig 3C.

Retrospective Evaluation of the SHIVA Trial

To test PTD in a real-world scenario, we retrospectively evaluated its ability to correctly predict population composition and treatment allocation in the recent SHIVA trial, a phase II randomized trial comparing conventional therapy to multiple targeted agents on the basis of tumor molecular profiling. In SHIVA, 741 patients were subjected to a triad of molecular tests, including sequencing with a targeted enrichment panel (AmpliSeq Cancer Panel, covering 45 tumor-associated genes or mutational hotspots), gene copy number analysis by Cytoscan HD, and immunohistochemistry for three endocrine receptors. Although PTD can include information on gene expression in the simulation, we removed endocrine receptors from the analyzed parameters because the correlation between RNA-based and immunohistochemistry-based quantification of endocrine receptors is poor (Appendix Fig A1). In addition, we removed patients with rare tumors that have not been sequenced in TCGA (see Appendix Table A1 for disease conversion). After data cleanup, 417 of 741 patients were left for analysis (Fig 4A).

We constructed a PTD panel by using the actionable alterations in SHIVA and ran 100 trial simulations, extracting tumor histologies from TCGA with probabilities equal to their frequencies in SHIVA and with allocation rules reproducing those adopted in the actual trial. Our results were remarkably similar: detection power was 33.51% (95% CI, 29.4% to 37.6%) in the simulation versus 32.92% (95% CI, 28.2% to 37.6%) in the SHIVA trial (Fig 4B). Alteration frequencies for individual genes were within 95% CIs in most cases (Appendix Fig A1A-B), except for low-frequency genes, for which 95% CIs can be expected to be wide. Grouping genes by pathway yielded more consistent results, all within 95% CIs (Fig 4C and Appendix Fig A1C). Consequently, observed versus predicted NNMS was also quite similar (Fig 4D; Appendix Fig A1D), demonstrating that PTD can correctly predict the number of patients needed to be subjected to molecular screening.

We report a versatile bioinformatics platform that is based on mutational databases to simulate genetic biomarker-driven clinical trials. In a retrospective analysis of the SHIVA trial, we demonstrated that our tool is able to predict the actual number of patients to be sequenced and has the power to detect overall and arm-specific efficacy signals. We believe that our tool will be instrumental in overcoming emerging issues in the design of precision oncology trials.

The conceptual framework of precision medicine implies an increased fragmentation of disease entities into distinct subgroups identified by the presence of biomarkers that predict response to targeted agents. As previously highlighted by several authors,6,20 for trials aimed at demonstrating efficacy of targeted drugs,20 more biomarkers means rarer biomarkers, which in turn implies higher and more variable numbers of patients needed to be screened. This simple consideration makes classic randomized clinical trials hardly feasible, an issue that could be circumvented by novel trial designs in which multiple targeted drugs, each associated with multiple biomarkers, are tested simultaneously, possibly in multiple tumor histologies.4,6 However, implementation of these so-called umbrella trials poses additional problems, such as designing the most comprehensive analytical platform, identifying the optimal drug combination and allocation strategy to optimize accrual, and appropriately calculating trial power because the false-positive rate increases with the number of independent hypotheses being tested.5 The difficulty in carefully evaluating these parameters results in high rates of screening failures, as evidenced by recent studies.9,21 In a basket trial on thoracic malignancies, accrual could be completed for only two of 15 arms because of the extreme rarity of the selected genetic alterations.9 In the National Cancer Institute’s MATCH Screening Trial (Targeted Therapy Directed by Genetic Testing in Treating Patients With Advanced Refractory Solid Tumors, Lymphomas, or Multiple Myeloma), the largest trial of this type to date, an interim analysis after 795 patients were screened showed that only 2.5% of patients could be allocated to seven of the 10 targeted agents on trial.21 The possibility of better predicting accrual rate afforded by PTD would allow trialists to drop inefficient arms and/or modify trial design to maximize patient allocation.

An unavoidable limitation of our methodology is that simulations are affected by the prevalence of mutations in available databases, which may not be representative of what is found in study populations. For instance, TCGA predominantly contains data from surgical specimens of patients’ primary tumors. Instead, most trials with targeted drugs are initially aimed at the metastatic setting, and the mutational profile of metastatic tumors can be significantly different from that of primary tumors. The recent release of the MSK-IMPACT (Memorial Sloan Kettering-Integrated Mutation Profiling of Actionable Cancer Targets) data,22 available through the same interface as that currently used by PTD,19 will bring patient prediction in the metastatic setting even closer to reality. An ideal setting for PTD-based simulations with current TCGA data may be the neoadjuvant/preoperative setting, which is increasingly used to evaluate novel drugs.23 The availability of clinical metadata, such as tumor size or nodal involvement, can be used to strengthen the representativeness of database samples.

Finally, some disease tumor types are under-represented in mutational databases, which results in high variability and low confidence of corresponding simulations. Thus, PTD performance is likely to be better for highly represented diseases such as breast cancer.

Although PTD was conceived for studies that use small targeted sequencing panels, its usefulness will extend to projects involving whole-genome or whole-exome sequencing because, even in these cases, only a small fraction of the sequenced genome will be used to dictate drug allocation, which can be handled by PTD in the same fashion as for targeted sequencing panels.

The implementation of PTD as a publicly available online tool may be useful to the growing community of physicians and statisticians who conduct clinical research in precision medicine to deal with issues of feasibility and efficiency in the context of a rapidly evolving environment.6

© 2018 by American Society of Clinical Oncology

Supported by Grants No. AIRC-15988 (A.G., G.C., and L.M.), RF-2013-02357231 from the Italian Ministry of Health (A.G., G.C., and L.M.), ANR-10-EQPX-03 (M.K. and C.L.T.), and by the Ludwig Center at Harvard University (G.E.M.M.).

Conception and design: Giorgio E.M. Melloni, Alessandro Guida, Giuseppe Curigliano, Edoardo Botteri, Ruggero de Maria, Piergiuseppe Pelicci, Luca Mazzarella

Financial support: Giuseppe Curigliano, Piergiuseppe Pelicci

Administrative support: Giuseppe Curigliano

Provision of study materials or patients: Giuseppe Curigliano, Angela Esposito, Maud Kamal, Christoph Le Tourneau

Collection and assembly of data: Giorgio E.M. Melloni, Alessandro Guida, Giuseppe Curigliano, Maud Kamal, Christoph Le Tourneau

Data analysis and interpretation: Giorgio E.M. Melloni, Alessandro Guida, Giuseppe Curigliano, Edoardo Botteri, Angela Esposito, Laura Riva, Alberto Magi

Manuscript writing: All authors

Final approval of manuscript: All authors

Accountable for all aspects of the work: All authors

The following represents disclosure information provided by authors of this manuscript. All relationships are considered compensated. Relationships are self-held unless noted. I = Immediate Family Member, Inst = My Institution. Relationships may not relate to the subject matter of this manuscript. For more information about ASCO's conflict of interest policy, please refer to www.asco.org/rwc or ascopubs.org/po/author-center.

Giorgio E.M. Melloni

No relationship to disclose

Alessandro Guida

No relationship to disclose

Giuseppe Curigliano

Consulting or Advisory Role: Genentech, Pfizer

Speakers’ Bureau: Genentech, Novartis, Pfizer

Travel, Accommodations, Expenses: Genentech

Edoardo Botteri

No relationship to disclose

Angela Esposito

No relationship to disclose

Maud Kamal

No relationship to disclose

Christophe Le Tourneau

Honoraria: Novartis, Bristol-Myers Squibb

Consulting or Advisory Role: Amgen, MSD, Bristol-Myers Squibb, Merck Serono, AstraZeneca

Travel, Accommodations, Expenses: MSD, Bristol-Myers Squibb, AstraZeneca

Laura Riva

Employment: Philips Healthcare (I)

Alberto Magi

No relationship to disclose

Ruggero De Maria

No relationship to disclose

Piergiuseppe Pelicci

Stock and Other Ownership: Genextra

Luca Mazzarella

No relationship to disclose

Table

Table A1. Conversion of SHIVA Tumor Types Into TCGA Nomenclature

1. Schwaederle M, Zhao M, Lee JJ, et al: Association of biomarker-based treatment strategies with response rates and progression-free survival in refractory malignant neoplasms: A meta-analysis. JAMA Oncol 2:1452-1459, 2016 Crossref, MedlineGoogle Scholar
2. Meric-Bernstam F, Brusco L, Shaw K, et al: Feasibility of large-scale genomic testing to facilitate enrollment onto genomically matched clinical trials. J Clin Oncol 33:2753-2762, 2015 LinkGoogle Scholar
3. Vogelstein B, Papadopoulos N, Velculescu VE, et al: Cancer genome landscapes. Science 339:1546-1548, 2013 Crossref, MedlineGoogle Scholar
4. Woodcock J, LaVange LM: Master protocols to study multiple therapies, multiple diseases, or both. N Engl J Med 377:62-70, 2017 Crossref, MedlineGoogle Scholar
5. Cunanan KM, Gonen M, Shen R, et al: Basket trials in oncology: A trade-off between complexity and efficiency. J Clin Oncol 35:271-273, 2017 LinkGoogle Scholar
6. Biankin AV, Piantadosi S, Hollingsworth SJ: Patient-centric trials for therapeutic development in precision oncology. Nature 526:361-370, 2015 Crossref, MedlineGoogle Scholar
7. Mirza MR, Monk BJ, Herrstedt J, et al: Niraparib maintenance therapy in platinum-sensitive, recurrent ovarian cancer. N Engl J Med 375:2154-2164, 2016 Crossref, MedlineGoogle Scholar
8. Le Tourneau C, Delord JP, Gonçalves A et al: Molecularly targeted therapy based on tumour molecular profiling versus conventional therapy for advanced cancer (SHIVA): A multicentre, open-label, proof-of-concept, randomised, controlled phase 2 trial. Lancet Oncol 16:1324-1334, 2015 Crossref, MedlineGoogle Scholar
9. Lopez-Chavez A, Thomas A, Rajan A, et al: Molecular profiling and targeted therapy for advanced thoracic malignancies: A biomarker-derived, multiarm, multihistology phase II basket trial. J Clin Oncol 33:1000-1007, 2015 LinkGoogle Scholar
10. Swanton C, Soria JC, Bardelli A, et al: Consensus on precision medicine for metastatic cancers: A report from the MAP conference. Ann Oncol 27:1443-1448, 2016 Crossref, MedlineGoogle Scholar
11. Le DT, Durham JN, Smith KN, et al: Mismatch repair deficiency predicts response of solid tumors to PD-1 blockade. Science 357:409-413, 2017 Crossref, MedlineGoogle Scholar
12. Hause RJ, Pritchard CC, Shendure J, et al: Classification and characterization of microsatellite instability across 18 cancer types. Nat Med 22:1342-1350, 2016 Crossref, MedlineGoogle Scholar
13. Kandoth C, McLellan MD, Vandin F, et al: Mutational landscape and significance across 12 major cancer types. Nature 502:333-339, 2013 Crossref, MedlineGoogle Scholar
14. Forbes SA, Beare D, Boutselakis H, et al: COSMIC: Somatic cancer genetics at high-resolution. Nucleic Acids Res 45:D777-D783, 2017 Crossref, MedlineGoogle Scholar
15. Melloni GE, de Pretis S, Riva L, et al: LowMACA: Exploiting protein family analysis for the identification of rare driver mutations in cancer. BMC Bioinformatics 17:80, 2016 Crossref, MedlineGoogle Scholar
16. Pao W, Miller VA, Politi KA, et al: Acquired resistance of lung adenocarcinomas to gefitinib or erlotinib is associated with a second mutation in the EGFR kinase domain. PLoS Med 2:e73, 2005 Crossref, MedlineGoogle Scholar
17. Mao C, Yang ZY, Hu XF, et al: PIK3CA exon 20 mutations as a potential biomarker for resistance to anti-EGFR monoclonal antibodies in KRAS wild-type metastatic colorectal cancer: A systematic review and meta-analysis. Ann Oncol 23:1518-1525, 2012 Crossref, MedlineGoogle Scholar
18. Yoshihara K, Wang Q, Torres-Garcia W, et al: The landscape and therapeutic relevance of cancer-associated transcript fusions. Oncogene 34:4845-4854, 2015 Crossref, MedlineGoogle Scholar
19. Cerami E, Gao J, Dogrusoz U, et al: The cBio cancer genomics portal: An open platform for exploring multidimensional cancer genomics data. Cancer Discov 2:401-404, 2012 Crossref, MedlineGoogle Scholar
20. Sleijfer S, Bogaerts J, Siu LL: Designing transformative clinical trials in the cancer genome era. J Clin Oncol 31:1834-1841, 2013 LinkGoogle Scholar
21. Conley BA, Gray R, Chen A, et al: NCI-molecular analysis for therapy choice (NCI-MATCH) clinical trial: Interim analysis. Cancer Res 76, 2016 (suppl; abstr CT101) CrossrefGoogle Scholar
22. Zehir A, Benayed R, Shah RH, et al: Mutational landscape of metastatic cancer revealed from prospective clinical sequencing of 10,000 patients. Nat Med 23:703-713, 2017 Crossref, MedlineGoogle Scholar
23. Barker AD, Sigman CC, Kelloff GJ, et al: I-SPY 2: An adaptive breast cancer trial design in the setting of neoadjuvant chemotherapy. Clin Pharmacol Ther 86:97-100, 2009 Crossref, MedlineGoogle Scholar
Downloaded 617 times

COMPANION ARTICLES

No companion articles

ARTICLE CITATION

DOI: 10.1200/PO.18.00015 JCO Precision Oncology - published online August 3, 2018

ASCO Career Center