Kinase Pathway Dependence in Primary Human Leukemias Determined by Rapid Inhibitor Screening

Gene-targeted cancer therapies have achieved remarkable clinical outcomes in recent years (1–4). In particular, cell-permeable small molecules that exhibit inhibitory activity against tyrosine kinases have generated great interest. Tyrosine kinases represent a gene family widely implicated in cancer pathogenesis (5, 6), and dysregulation of specific tyrosine kinases has been observed in most hematologic malignancies, including chronic myeloid leukemia (CML; ref. 7), chronic myelomonocytic leukemia (CMML; refs. 8–11), other myeloproliferative neoplasms (MPN; refs. 12–15), acute myeloid leukemia (AML; refs. 9, 16–20), acute lymphoblastic leukemia (ALL; refs. 21–26), and chronic lymphocytic leukemia (CLL; refs. 27–30). Although a minority of patients with hematologic malignancies are successfully treated with kinase inhibitors, most patients remain ineligible for this form of targeted therapy due to lack of knowledge of the specific kinase pathways involved.

Many strategies exist to better understand kinase dysregulation in cancer including the recent development of deep sequencing techniques, which are accelerating our understanding of cancer genetics. Thus far, however, many studies of malignancies with predicted kinase pathway dependence have not found frequent mutations in kinase genes (31–34). These findings suggest that kinase pathway dependence in malignant cells often occurs due to complex genetic mechanisms. Hence, although deep sequencing represents an immensely powerful technique, it may not independently allow for prediction of kinase targets and kinase inhibitor therapies. Instead, understanding of the best kinase inhibitor therapies for patients will likely require the combination of deep sequencing with complementary studies that can define kinase targets regardless of mutational status. These functionally important kinase pathways can then be correlated with genetic profiles that have been revealed by deep sequencing.

To better define the use of kinase inhibitor therapies in hematologic malignancies, we have developed a small-molecule kinase inhibitor panel designed to identify kinase pathway dependence in primary leukemia samples. To analyze kinase pathway dependence based on this functional data, we have developed an accompanying bioinformatics approach to predict the kinase targets underlying inhibitor sensitivity profiles. This algorithm takes advantage of our knowledge of the gene products that are targeted by each drug as well as the fact that these target profiles are partially overlapping. Using the overlap of effective drugs and eliminating targets of ineffective drugs, we are able to predict critical kinase targets and signaling pathways for individual patient samples. These kinase target predictions represent a manner by which functional data from drug screening could be integrated with genomics data such as deep sequencing to aid in prioritization of sequence variants and, thus, accelerate our understanding of molecular etiologies of cancer as well as application of individualized therapeutic approaches for patients.

Kinase inhibitors were purchased from or were generously provided by the sources outlined in Supplementary Table S7.

All clinical samples were obtained with informed consent with approval by the Institutional Review Boards of Stanford University (Stanford, CA), Oregon Health & Science University (Portland, OR), the Children's Oncology Group, and Erasmus Medical Center/Sophia Children's Hospital (Rotterdom, the Netherlands). Blood or bone marrow from patients was separated on a Ficoll gradient and mononuclear cells were treated with ammonium-chloride-potassium (ACK) lysis buffer. The only exceptions to this procedure were cases of atypical CML or chronic neutrophilic leukemia, where samples were only processed with ACK lysis buffer to preserve the neoplastic granulocytes that would otherwise be lost on the Ficoll gradient. Cells from myeloid leukemia samples were cultured in R10 [RPMI-1640 medium supplemented with 10% FBS (Atlanta Biologicals), l-glutamine, penicillin/streptomycin (Invitrogen), and fungizone (Invitrogen)] supplemented with 10⁻⁴ M 2-mercaptoethanol (Sigma). Cells from lymphoid leukemia samples were cultured in R20 [RPMI-1640 medium supplemented with 20% FBS (Atlanta Biologicals), l-glutamine, penicillin/streptomycin (Invitrogen), and fungizone (Invitrogen)] supplemented with 10⁻⁴ M 2-mercaptoethanol (Sigma) insulin-transferrin-sodium selenite (Invitrogen).

Kinase inhibitors were stored at 10 to 100 mmol/L in dimethyl sulfoxide (DMSO; stock concentration was 1,000 times the final concentration of the highest tested dose). Drugs were used at final concentrations shown in Supplementary Table S4. For creation of replicate plates of the library, each drug concentration was diluted to twice the final concentration and 50 μL were plated into 96-well plates using a Hydra 96-channel automated pipettor (Matrix Technologies). Plates were sealed with adhesive lids (Bio-Rad; microplate seal B), wrapped in aluminum foil, and stored at −20°C until use. Upon receipt of a patient sample, plates were thawed at 37°C, 5% CO₂ for 1 hour and centrifuged at 800 × g before removal of adhesive lids. Subsequently, patient samples were suspended into culture media at a concentration of 1,000,000 cells per mL, such that addition of 50 μL to each well would deliver 50,000 cells to that well (this also dilutes the drugs to their final, desired concentration). Cells were incubated for 3 days at 37°C, 5% CO₂ and subjected to a CellTiter 96 AQ_ueous One solution cell proliferation assay (Promega). Each plate contained 7 wells without any drug. The average absorbance value of these 7 wells was used for data normalization and the kill curve of each drug gradient was assessed relative to this average no-drug point.

An algorithm was designed and implemented using Excel and Visual Basic to provide automated IC₅₀ calculation and therapeutic target identification. IC₅₀ values were calculated using second-degree polynomial regression curves fit through 5 data points (average of no drug wells and 4 serial dilution points). All curves were manually inspected and a small number of IC₅₀s were corrected in 2 circumstances: (1) The curve fit intersected the IC₅₀ at 2 distinct points — the lower concentration intersect was used in these instances; (2) the polynomial curve fit yielded an artificial IC₅₀ not reflected in the data points (generally due to increasing cell viability over the course of the drug titration). For a given sample, drug IC₅₀ values were considered effective if they were less than or equal to 5-fold below the median IC₅₀ for all samples tested. Where an IC₅₀ was not achieved for a given drug, an IC₅₀ value equal to the highest drug concentration used was arbitrarily assigned. After effective and ineffective drugs were determined for each sample, a drug target score was assigned by the program for each potential therapeutic target.

The drug target score is based on the IC₅₀ measured effectiveness of panel drugs against a given therapeutic target for a given patient. Each drug is associated with a “tiered” ranking of target kinases for which the drug has been shown to biochemically associate (35, 37–43). Weights are used to give a stronger quantitative emphasis to targets with a higher ranking (Figs. 4A and 4C). Scores are determined empirically for a given sample by assigning positive weighted scores to targets of effective drugs and negative weighted scores to targets of ineffective drugs. Drug effectiveness threshold and tiered ranking weights were determined empirically using patient samples and cell lines with known kinase signaling abnormalities. The algorithm generates a cumulative drug target score for each target according to the following equations:

Where Weight_Tier (i) is a given drug target's tier ranking for drug i and n is the number of effective drugs.

And

Where Weight_Tier (j) is a given drug target's tier ranking for drug j and m is the number of ineffective drugs.

n + m = 66.

Final drug target score = effective drug target score + ineffective drug target score

Hierarchical clustering was conducted using GenePattern software (Broad Institute). Sample clustering and 2-way clustering by row (drug) and column (patient sample) were conducted using Pearson correlation distance shown in Fig. 6 and Supplementary Figs. S2 and S3.

The ubiquitous role of tyrosine kinases in regulating critical cellular processes leading to malignancy suggests that a large percentage of leukemia (and other malignancy) patient samples would exhibit sensitivity to inhibition of one or more kinase pathways. To test this hypothesis, we compiled a library of 66 small-molecule kinase inhibitors with collective activity against two-thirds of the tyrosine kinome (Supplementary Table S1). Because many nontyrosine kinases are also critical regulators of cellular growth/survival, we also included drugs with activity against select families of nontyrosine kinases including phosphoinositide-3 kinase (PI3K)/AKT, PKC, PKA, IκK, RAF/MEK/ERK, c-jun-NH2-kinase (JNK), p38, AMPK, aurora kinases, and cyclin-dependent kinases (Supplementary Table S1). Each inhibitor was plated at 4 graded concentrations that bracket the predicted on-target IC₅₀ value. Primary patient samples were incubated with this panel of drugs for 3 days at which point a tetrazolium-based cell viability assay (MTS) was conducted for assessment of cell viability. All values were normalized to cells incubated in the absence of drug (Supplementary Fig. S1).

Over a 2-year period, we accrued and tested 151 fresh, primary leukemia patient samples against this panel of kinase inhibitors. The cohort was comprised of 34 AML, 42 ALL, 31 MPN, and 44 CLL patients. Detailed clinical and demographic information about this patient cohort can be found in Supplementary Table S2. Assessment of kinase inhibitor hypersensitivity profiles of these 151 leukemia patient samples revealed a wide diversity of responses to kinase inhibitors, even when patients were grouped according to diagnostic subsets and kinase inhibitors grouped according to predicted kinase target spectra (Fig. 1). Despite this heterogeneity of responses, certain trends emerged, such as more frequent sensitivity to PI3K/AKT inhibitors in lymphoid samples. In addition, select cases could be identified with universal sensitivity to whole families of kinase inhibitors. For example, AML case 07335 exhibited universal sensitivity to all ERBB family inhibitors on the panel, suggesting involvement of an ERBB family member in maintenance of the viability of malignant cells from that patient (Fig. 1). Overall, 70% of patients exhibited hypersensitivity to one or more kinase inhibitors (IC₅₀ data for each drug for each patient specimen listed in Supplementary Table S3; raw data points used to generate these IC₅₀s listed in Supplementary Table S4). Nine of the drugs on our panel are currently approved by the U.S. Food and Drug Administration (FDA; imatinib, nilotinib, dasatinib, sunitinib, erlotinib, gefitinib, lapatinib, sorafenib, and pazopanib), and approximately 40% of samples exhibit hypersensitivity to one or more of these 9 drugs. Hypersensitivity to a drug was determined by comparison of the response of each individual sample with the response of all other samples (patient IC₅₀ compared with whole cohort median IC₅₀). In this way, we could define outlier samples that were truly hypersensitive to a given drug versus responses at higher concentrations that might occur due to off-target toxicity of the compound. Rank ordering of patient IC₅₀s for each drug helps illustrate this point (Supplementary Fig. S2). To better distinguish inhibitor sensitivity profiles that were similar from patient to patient, we applied one-way Pearson correlation for hierarchical clustering of the data (Supplementary Fig. S3). Notably, although drug responses are clearly not uniform among diagnostic subsets, there were large groups of patients diagnosed with the same type of leukemia [notably ALL (shaded red) and CLL (shaded yellow); Supplementary Fig. S3] with similar responses to these kinase inhibitors. However, the segregation of sample responses by cell type was far from complete with a number of isolated lymphoid samples clustering amidst myeloid cases and vice versa. In addition, 2-way hierarchical clustering revealed drugs with similar activity profiles across patients (Supplementary Fig. S4). Many of these drugs clustered in groups that would be predicted on the basis of known target profiles of the compounds such as BIRB-796 and VX-745 (both p38 inhibitors), flavopiridol and BMS-387032 (both CDK inhibitors), and EKB-569 and CI-1033 (both ERBB-family inhibitors).

Clinical and research interest in the application of kinase inhibitors has led to a concerted effort to develop techniques that characterize the targets to which each compound can effectively bind (35–38). Because a majority of the kinase inhibitors on our panel have been characterized in this manner, we realized that this information could be used to predict the critical kinase targets and signaling pathways that underlie the observed kinase inhibitor sensitivity patterns. Development of this bioinformatics approach relies on the fact that all kinase inhibitors on our panel bind multiple targets and the target spectra for these drugs are partially overlapping. Thus, if a sample exhibits sensitivity to 2 different drugs, the kinase targets that are commonly inhibited by both drugs are implicated as most likely to explain this sensitivity pattern (unless there are 2 different pathways operational in that sample). A second step can further narrow the candidate kinase list by elimination of targets of drugs to which a sample is not sensitive.

The results from AML patient 08024 illustrate our target identification strategy. Malignant cells from this patient were hypersensitive to 3 drugs: AST-487, sunitinib, and SU-14813. As noted above, drug hypersensitivity is defined by comparison of the IC₅₀ values for this individual sample with the median IC₅₀ values that these drugs achieved across the entire cohort. Cells from patient 08024 exhibited IC₅₀ values that were at least 5-fold lower than the cohort median IC₅₀ values for each of these 3 drugs (Fig. 2A). Analysis of the known targets of these 3 compounds revealed 4 kinases, KIT, PDGFR, CSF1R, and FLT3, that are common targets between all 3 drugs (Fig. 2B). Because this sample was also unaffected by the 2 drugs, dasatinib and nilotinib (no difference between patient 08024 IC₅₀ and median IC₅₀; Fig. 2A), the kinase targets of these 2 drugs can be eliminated from consideration as kinases that could mechanistically explain cell viability in this patient sample. The kinase targets that are effectively inhibited by dasatinib and/or nilotinib are shaded blue, whereas the kinases not targeted by either dasatinib or nilotinib are shaded red (Fig. 2B). This reveals that the only kinase target in common between AST-487, sunitinib, and SU-14813 that is also not a target of dasatinib or nilotinib is FLT3. Further analysis of AML patient 08024 revealed presence of the FLT3-ITD and no wild-type FLT3 alleles (Fig. 2C).

Application of the logic illustrated in Fig. 2B can be conducted manually for a small number of kinase inhibitors; however, expansion of this process to evaluate data from the entire panel of 66 drugs requires computational support. We have developed an algorithm to carry out the same logical steps outlined in Fig. 2, for all 66 kinase inhibitors in 4 steps (Sweave document with instructions for generating application to perform this algorithm can be found in Supplementary Material). First, the kinase target information for each drug was curated from published sources (35, 37–43) into a database (Fig. 3A and Supplementary Table 1). Second, the K_d or IC₅₀ values (depending on the format of assay used to define drug targets) for each drug were subdivided into 5 tiers on a Log₁₀ scale. The first tier is defined as the kinase with the lowest K_d or IC₅₀ value as well as all other kinases with K_d or IC₅₀ values less than or equal to 10-fold than lowest value. Kinases with K_d or IC₅₀ values ranging from 10-fold to 100-fold, the lowest values are considered tier 2 targets. Each subsequent Log₁₀ increase in K_d or IC₅₀ comprises a new tier of kinase targets (Fig. 3A; bottom). Next, the IC₅₀ results from an individual patient sample are subdivided into drugs to which the sample was hypersensitive (defined by a patient-specific IC₅₀ that is 5-fold less than the cohort median IC₅₀) as well as the drugs to which the sample is not sensitive (Fig. 3B). We next devised a scoring system that assigns points to each kinase based on whether inhibitors defined to block that kinase target were effective (positive points awarded to the target's score) or ineffective (negative points deducted from the target's score). The addition or subtraction of points is conducted in a graded manner with tier 1 kinases having the most points added or subtracted from their score and the score for kinases from each subsequent tier being modified to lesser degrees (Fig. 3C). Finally, the cumulative scores for each kinase are tabulated and ranked such that the highest scoring kinases for any given patient are those kinases predicted to be most probable in explaining the observed drug response and, therefore, the most probable in playing a pathogenic role for that patient's malignancy (Fig. 3D).

To test this algorithm, we chose 3 specimens from patients with leukemia with known, dysregulated tyrosine kinase pathways and one specimen from a patient without a known kinase mutation. The first example, AML patient 08024, was described in Fig. 3 whereby the FLT3-ITD gene target was predicted on the basis of analysis of the response pattern of the cells to 5 small-molecule kinase inhibitors. To determine whether our algorithm would also successfully identify FLT3 as a high probability target when applying data from all 66 drugs, we conducted the algorithm on kinase inhibitor panel screening outcomes from this patient sample. This exercise revealed that FLT3 was the second highest scoring kinase on the target list with a score of 89 points (Fig. 4A). The highest scoring kinase (EGFR) scored 90 points due to near complete hypersensitivity of this specimen to ERBB family inhibitors. Analysis of target profiles of these ERBB family inhibitors reveals that they do not exhibit off-target effects against FLT3 (37), indicating that there may be cross-talk between the FLT3-ITD oncogene and ERBB family members. We also applied this screen to cells from a patient with CML in blast crisis. In this case, the algorithm correctly identified ABL1 as the top-scoring kinase (Fig. 4B). Finally, we applied this technique to cells from a patient with MPN positive for the oncogene, MPL-W515L. In this case, 3 of the top 5 targets are JAK family kinases, which represent therapeutic targets downstream of MPL (MPL is not a kinase, but signals through JAK kinases; refs. 44, 45; Fig. 4C). Thus, application of this algorithm to 3 proof-of-principle examples shows that the approach correctly identifies known oncogenic signaling pathways in each case. In addition, a patient with AML (08102) with no known mutations in tyrosine kinases exhibited hypersensitivity to 12 of the kinase inhibitors on our panel (Fig. 4D). Among the kinases predicted by our algorithm to be likely involved in pathogenesis of this specimen was FLT3, which is commonly mutated in AML (Fig. 4E). However, as noted above, this patient did not exhibit any of the common FLT3 abnormalities. This prompted us to conduct more extensive sequencing of FLT3, and we identified a point mutation in the extracellular region of FLT3 (S451F; Fig. 4F), this rare mutation has been previously shown to exhibit transformative capacity (46).

To better understand the kinase targets that might underlie drug sensitivity patterns in our entire cohort of leukemia patient samples, we applied our target prediction algorithm to all 151 specimens that were interrogated by the kinase inhibitor panel. The data reveal a heterogeneous list of probable kinase targets that arise in a highly patient-specific manner (Supplementary Table S5). One clear conclusion from these results is the need for a personalized approach to application of kinase-targeted therapies. However, the data can also be used to identify kinases that may be more frequently implicated in any one diagnostic subset of malignancy. As such, we tabulated the average kinase scores for each patient within the broad diagnostic categories of ALL, CLL, AML, CML, CMML, and other MPN. Kinase scores were normalized for comparison between patients by dividing all scores of a given profile by the maximum score for that profile. Negative scores were removed resulting in a normalized profile score between 0 and 1. Next, we computed average normalized scores for each kinase across diagnostic groups. Finally, the kinase scores for each diagnostic group were again normalized to the highest scoring kinase within each group so that each diagnostic category would be represented on the same scale. These values were then expressed on a heatmap to visually represent the kinases that are predicted to be more frequently involved in the pathogenesis of each of these broad diagnostic subsets of hematologic malignancies (Fig. 5; input data found in Supplementary Table S6). As anticipated, ABL1 is the highest cumulative scoring kinase among CML patients tested, serving as a proof-of-principle for this strategy. Interestingly, CaM kinases seem to score highly in lymphoid but not myeloid malignancies. One possible link between that CaM kinases and lymphoid malignancies may be the activation of CBP. Recent studies have identified mutations in CBP (and loss of activity) as a strong correlation with relapse in ALL (47). It has been previously established that phosphorylation of CBP by CaM kinase IV is important for its activity (48). Therefore, inhibition of CaM kinases may provide further insight to the functional importance of CBP in lymphoid viability. PI3K and AKT isoforms score highly in lymphoid malignancies as well as MPN, but are infrequently represented in patients with AML. BTK scores highly in patients with CLL, consistent with the notion that B-cell receptor signaling, which depends partly on BTK, is critical for viability of cells from many patients with CLL (28, 30). Cyclin-dependent kinases are predicted to be involved more frequently in MPN than in AML or lymphoid malignancies. Ephrin receptors score strongly across all diagnostic groups. In addition, p38 appears frequently involved in all malignancy subsets, and especially so in AML and CMML. There are a wide diversity of genes and signaling pathways that are predicted to play a role in the pathogenesis of each type of hematologic malignancy. However, our drug target scoring algorithm indicates particular pathways are more frequently represented in some diagnostic subsets than in others. Follow-up investigation will be required to validate the complete genetic etiology of these observations.

The clinical use of this type of test is predicated on a meaningful correlation between in vitro and an in vivo response to kinase inhibitors. As a proof-of-concept, we tested this correlation in a patient with refractory AML. A 36 year-old patient with a white blood cell (WBC) count of 133,000 was diagnosed with AML with inversion of chromosome 2 and trisomy 8. The FLT3-ITD was noted to be weakly positive with an allelic ratio of 0.02. Following leukapheresis and a standard 7+3 induction therapy, he was found to have refractory AML and was reinduced with HAM chemotherapy. Although he achieved a remission and soon after underwent an unrelated donor transplant, he relapsed 60 days later. The donor was not available for donor leukocyte infusions and the patient was refractory to FLAG-IDA salvage therapy. Our inhibitor panel showed dramatic sensitivity to numerous kinase inhibitors, including several drugs that are already FDA-approved and have been used for treatment of AML (Fig. 6A). Because this patient had no other standard therapeutic options, he elected treatment with one of the inhibitors predicted to be effective by the inhibitor panel assay. Daily treatment with this drug, sorafenib, induced a rapid normalization of WBC counts with decreased blasts in both the peripheral blood and bone marrow that was maintained for more than 2 months (Fig. 6B). At the time of relapse, a repeat kinase inhibitor sensitivity panel showed that the in vitro response to sorafenib was approximately 3 logs less than the pretreatment cells (Fig. 6C). Interestingly, the analysis also showed that this relapsed AML remained highly sensitive to another FDA-approved kinase inhibitor, sunitinib. (Fig. 6D). Treatment with daily sunitinib resulted in a significant initial response with rapid reduction in WBC counts and peripheral leukemic blasts for a period of 4 weeks.

Here, we show that functional screening of primary cells from leukemia patients with a panel of kinase inhibitors can identify effective kinase inhibitors in 70% of patients in just 3 days, including 40% of patient samples that were hypersensitive to drugs already approved for clinical use. We also show an algorithm that uses the partially overlapping kinase target spectra for each drug for prediction of critical kinase targets that underlie inhibitor sensitivity patterns. Finally, using this algorithm, we identify and rank probable kinase targets in 151 patients with a variety of hematologic malignancies.

The predicted kinase targets for these patient samples show a great deal of heterogeneity, even within diagnostic subsets. This finding highlights one of the strengths of this kinase inhibitor screen—kinase targets and patient-specific therapeutic options are detected regardless of the frequency with which these targeted therapies would be applicable within the given disease subset. In addition, although our target prediction algorithm can suggest candidate pathways for follow-up validation, this technique is empirical in that it does not require specific knowledge of any genetic lesion or biomarker to uncover potential therapeutic options for patients.

It is likely that the algorithm shown here exhibits both false-positive and false-negative kinase targets. Because the algorithm is driven by our knowledge of potential target spectra of kinase inhibitors that are present on the panel, there is certainly an opportunity for target bias due to pathways that are over- or under-represented on the panel. As drug development continues and the panel expands to include other drugs that offer more complete and even coverage of the kinome, this pitfall will be diminished. In addition, further profiling of these inhibitors to fully delineate target spectra will also improve the algorithm output as this output is largely driven, and limited, by our knowledge of these target spectra.

It is also possible that the setup of the assay could lead to false-negative results, as the technique relies on a short 3-day window for assessment of drug sensitivity. For this reason, interpretation of negative results from the drug assay (i.e., lack of sensitivity to a particular drug) must be taken with caution. Correlation of results from this screen with clinical cases in which patients are treated with drugs from the panel will help inform the full clinical relevance of both positive and negative assay results.

It will be critical to integrate this kinase inhibitor screen with other cutting edge techniques. We have already begun to apply this assay in conjunction with a related siRNA screen (49, 50), and initial results obtained within 4 days of receiving fresh primary samples indicate a high concordance of siRNA and kinase inhibitor sensitivities. In addition, application of genomics techniques such as gene expression microarray and deep sequencing in parallel with these functional screening tools will undoubtedly accelerate our understanding of the precise molecular events that underlie the observed kinase sensitivity patterns. Matching of these patient genotypes with drug sensitivity patterns will ultimately enable patients to be treated on the basis of tumor genotypes, and this drug assay and kinase target algorithm offer useful modalities for translation of tumor genotypes into therapeutically relevant clinical strategies. Hence, the kinase inhibitor screen illustrated here represents advancement toward individual patient-tailored cancer therapy. Malignant cells from patients with cancer can be economically screened to identify effective targeted therapies within 3 days, thus matching the appropriate drugs with the appropriate patients on a timescale that affords opportunity for informed, mechanism-based intervention.

S.E. Spurgeon has a honoraria from Speakers Bureau from GSK and Millenium. B.J. Druker has ownership interest (including patents) and is a consultant/advisory board member in Blueprint Medicines. No potential conflicts of interest were disclosed by the other authors.

Conception and design: J.W. Tyner, G. Fan, B.H. Chang, M.M. van den Heuvel-Eibrink, B.J. Druker, M.M. Loriaux

Development of methodology: J.W. Tyner, G. Fan, M.M. van den Heuvel-Eibrink, T. O'Hare, M.M. Loriaux

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): J.W. Tyner, W.F. Yang, L.B. Fletcher, J. Bryant, J.M. Glover, B.H. Chang, S.E. Spurgeon, W.H. Fleming, T. Kovacsovics, J.R. Gotlib, S.T. Oh, M.W. Deininger, C.M. Zwaan, M.L. den Boer, M.M. van den Heuvel-Eibrink, M.M. Loriaux

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): J.W. Tyner, W.F. Yang, A. Bankhead, G. Fan, J.M. Glover, B.H. Chang, W.H. Fleming, C.M. Zwaan, M.L. den Boer, M.M. van den Heuvel-Eibrink, B.J. Druker, M.M. Loriaux

Writing, review, and/or revision of the manuscript: J.W. Tyner, G. Fan, B.H. Chang, W.H. Fleming, T. Kovacsovics, C.M. Zwaan, M.L. den Boer, M.M. van den Heuvel-Eibrink, T. O'Hare, B.J. Druker, M.M. Loriaux

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): J.W. Tyner, L.B. Fletcher, J.R. Gotlib, M.M. van den Heuvel-Eibrink, M.M. Loriaux

Study supervision: J.W. Tyner, M.M. van den Heuvel-Eibrink, B.J. Druker, M.M. Loriaux

This work was supported in part by The Leukemia and Lymphoma Society, the TJ Martell Foundation, and the Doris Duke Charitable Foundation. J.W. Tyner is supported by grants from the National Cancer Institute (NCI; grant # 5K99CA151457-02) as well as grants from the William Lawrence and Blanche Hughes Foundation and the Oregon Clinical and Translational Research Institute (OCTRI) grant number UL1 RR024140 from the National Center for Research Resources (NRCC), a component of the NIH, and NIH Roadmap for Medical Research. B.J. Druker is an investigator of the Howard Hughes Medical Institute.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.