Pharmacogenetics of minimal residual disease response in children with B-precursor acute lymphoblastic leukemia: a report from the Children's Oncology Group

Abstract

Minimal residual disease (MRD) as a marker of antileukemic drug efficacy is being used to assess risk status and, in some cases, to adjust the intensity of therapy. Within known prognostic categories, the determinants of MRD are not known. We measured MRD by flow cytometry at day 8 (in blood) and at day 28 (in bone marrow) of induction therapy in more than 1000 children enrolled in Pediatric Oncology Group therapy protocols 9904, 9905, and 9906. We classified patients as “best risk” if they had cleared MRD by day 8 of therapy and as “worst risk” if they had MRD remaining in bone marrow at day 28, and tested whether MRD was related to polymorphisms in 16 loci in genes hypothesized to influence response to therapy in acute lymphoblastic leukemia (ALL). After adjusting for known prognostic features such as presence of the TEL-AML1 rearrangement, National Cancer Institute (NCI) risk status, ploidy, and race, the G allele of a common polymorphism in chemokine receptor 5 (CCR5) was associated with more favorable MRD status than the A allele (P = .009, logistic regression), when comparing “best” and “worst” risk groups. These data are consistent with growing evidence that both acquired and host genetics influence response to cancer therapy.

Introduction

Variable response to drugs has been shown in many clinical settings to be due to polymorphisms in key genes involved in drug transport, uptake, metabolism, or targeting. A number of studies have shown that pharmacogenetic variants can influence response to therapy and toxicity of therapy in children with acute lymphoblastic leukemia (ALL). The best-studied example is the gene thiopurine methyltransferase (TPMT) that modifies metabolism of 6-mercaptopurine (6-MP).^1–3 A series of studies over 20 years have shown that TPMT genotype can modify both disease control and toxicity in children with ALL.^4–6 However, multiple drugs are used to treat children with ALL, and more recent studies have identified additional loci that may modify response to therapy, including genes participating in folate metabolism, steroid response and drug transport, metabolism, and detoxification.^7–9 In addition, analysis of the numerous “nongenetic” characteristics, such as age, white cell count, and race that influence treatment response in addition to polymorphic genotypes is necessary to adequately assess the importance of germ line variation on antileukemic response.

The presence of minimal residual disease (MRD) as a marker of antileukemic drug efficacy is being used to assess risk status in children with ALL and, in some cases, to adjust the intensity of therapy.^10–14 Multiple studies have shown that the presence of MRD at the end of induction therapy is associated with inferior survival. A number of cooperative groups, including the United States Children's Oncology Group, are using the presence or absence of MRD at the end of induction as a key element of risk-group assignment to determine the intensity of subsequent therapy. Although MRD correlates to some extent with known prognostic factors, it is not know what governs whether patients within known prognostic categories have MRD or not. In this study we investigated whether specific host germ line pharmacogenetic polymorphisms had any relationship to level of MRD after remission induction, after accounting for other prognostic factors. A candidate gene approach was used to select polymorphisms likely to have functional importance and in pathways likely to be involved in the response to chemotherapy treatment for childhood ALL. Specifically, we determined whether the presence of MRD in bone marrow, measured by flow cytometry at day 8 and day 28 of remission induction therapy in more than 1000 children enrolled on Pediatric Oncology Group (POG) protocol 9900, was related to the following germ line polymorphisms: MTHFR 1298A>G and 677C>T, NQO1, GSTP1, TYMS, ADRB2, RFC 80A/G, VDR intron 8 and Fok I sites, MDR1 3435C>T and 2677G>T/A, CCR5, CONNEXIN, MBPC1, P22, and TPMT. The data indicate that in addition to TEL rearrangement, National Cancer Institute (NCI) risk status, ploidy, and race, CCR5 genotype modifies clearance of MRD.

Methods

Patients

The study included 1197 patients enrolled in the study POG 9900. Informed consent was obtained from patients or their parents for both the treatment and biology studies in accordance with the Declaration of Helsinki. The clinical study was approved by the local institutional review boards at all participating institutions (see Document S1, available on the Blood website; see the Supplemental Materials link at the top of the online article), and the biological study was approved by the institutional review board at Cincinnati Children's Hospital and St Jude Children's Research Hospital. The 9900 protocol included 2 parts: (1) laboratory classification of all patients with ALL, and (2) remission induction therapy for patients over 1 year of age classified as having B-precursor ALL. To be eligible for enrollment on subsequent treatment studies, every newly diagnosed patient with ALL was required to be registered on the classification study. Molecular classification studies, performed at the reference laboratory at the University of New Mexico, included flow cytometry for DNA index, molecular testing for TEL-AML1, E2A-PBX1, BCR-ABL, and t(^4,11) and fluorescence in situ hybridization (FISH) testing for trisomies 4 and 10, and for MLL rearrangements other than t(^4,11). Immunophenotyping was performed on all cases at the reference laboratory at Johns Hopkins University, and the original blast cell immunophenotypic signature was subsequently used to assess MRD.

Induction chemotherapy was assigned according to NCI risk group, with NCI standard risk patients (age < 10 years and white cell count at presentation < 50×10⁹/L [50 000/mm³]) eligible for a 3-drug induction (dexamethasone, vincristine, and PEG-asparaginase with intrathecal Ara-C and methotrexate, given at separate times) and NCI high-risk patients (age > 10 years or white cell count at presentation > 50×10⁹/L [50 000/mm³]) eligible for a 4-drug induction regimen (prednisone, vincristine, daunomycin, and L-asparaginase with intrathecal methotrexate).

Minimal residual disease

Four-color multiparameter flow cytometry was used to detect minimal residual disease as previously described.¹⁵ Briefly, aliquots of blood or marrow were stained with 2 different 4-color combinations of antibodies: CD19-APC/CD45-perCP/CD20-PE/CD10-FITC and CD19-APC/CD45-perCP/CD9-PE/CD34-FITC in order to detect cells with aberrant leukemic phenotypes different from those seen in normal or regenerating marrows. In day-8 blood specimens, cell numbers were generally very low, and a single tube containing the most informative combination based on the pretreatment phenotype was used in most cases. Samples were run on a FACSCAlibur flow cytometry (BD Biosciences, San Jose, CA) and data analyzed using Paint-A-Gate software. For day-29 marrow, 750 000 events were collected while only 100 000 events were collected from blood specimens, although the latter had no contaminating normal B-cell precursors, making it possible to achieve high sensitivity in the absence of a larger number of cells. The term “events” refers to the number of cells collected by the flow cytometer. The different number of cells at the 2 time points largely reflects the cellularity of the 2 samples provided. Sensitivity of MRD detection is dependent in part upon the number of events (cells) difference between the abnormal phenotype and that of any normal cells.

In practice blood is a much “cleaner” sample than marrow so that fewer abnormal cells can be confidently recognized as MRD in blood. Thus, more events must be collected to achieve equivalent sensitivity for the marrow. Overall, in POG 9904/5/6 studies satisfactory MRD studies at a sensitivity of .01% were obtained on 92% of day-29 samples and 90% of day-8 blood samples. Sensitivity was 1/10 000 cells in the great majority of cases.

Genotyping

Genotyping was performed essentially as described previously.⁷ Genotypes were performed for the following loci (Table 2): ADRB21 −468 C>G rs11168070, CCR5 + 246 A>G rs1799987, CONNEXIN 1019 C>T rs1764391, GSTP1 313 A>G rs947894, MBP codon 54 rs1800450, MDR1 exon 21 2677 rs 2032581, MDR1 exon 26 3435 C>T rs1045642, MTHFR 677 C>T rs1801133, MTHFR 1298 A>C rs1801131, NQO1 609 C>T rs 1800566, CYBA P22 rs4673, RFC 80 A>G rs1051266, TPMT 460 G>A rs 1142345, TYMS - 827 promoter enhancer repeat, VDR Fok1 Start site C>T rs1073581, VDR intron 8 G>A rs1544410.¹⁶

Table 2.

Number of patients by demographic, prognostic feature, and genotype categories who were in best-versus-worst MRD groups

	Best*	Worst
Sex
Male	153	153
Female	124	99
Race
White	173	160
Hispanic	54	56
African American	20	9
Other	30	27
NCI
Standard risk	211	136
High risk	66	116
Hyperdiploidy
No	228	183
Yes	47	69
Missing	2	0
MLL rearrangement
No	275	241
Yes	2	11
Missing	0	0
TEL-AML1
No	167	223
Yes	109	27
Missing	1	2
Trisomy 4 and 10
No	230	193
Yes	47	58
Missing	0	1
ADRB21 -468 C>G, rs11168070
CC	118	113
CG and GG	134	127
Missing	25	12
CCR5 +246 A>G, rs1799987
AA	67	73
AG and GG	189	145
Missing	21	34
CONNEXIN 1019 C>T, rs1764391
CC	120	108
CT and TT	134	114
Missing	23	30
GSTP1 313 A>G, rs947894
AA and AG	221	193
GG	33	27
Missing	23	32
MBP codon 54, rs1800450
AA	2	4
AG and GG	258	215
Missing	17	33
MDR1 exon 21 2677, rs 2032581
GG	96	85
Others	158	150
Missing	23	17
MDR1 exon 26 3435 C>T, rs1045642
CC	76	57
CT and TT	189	190
Missing	12	5
MTHFR 677 C>T, rs1801133
CC	114	100
CT and TT	157	145
Missing	6	7
MTHFR 1298 A>C, rs1801131
AA	129	125
AC and CC	142	120
Missing	6	7
NQO1 609 C>T, rs 1800566
CC	154	112
CT and TT	102	112
Missing	21	28
CYBA P22rs4673
CC	121	92
CT and TT	138	124
Missing	18	36
RFC 80 A>G, rs1051266
AA and AG	182	157
GG	73	79
Missing	22	16
TPMT 460 G>A, rs 1142345
AA and AG	16	16
GG	246	208
Missing	15	28
TYMS - 827 promoter enhancer repeat
3AND3	79	77
Others	176	161
Missing	22	14
VDR Fok1 Start site C>T, rs1073581
CC	105	92
CC and CT	156	133
Missing	16	27
VDR intron 8 G>A, rs1544410
AA and AG	71	61
GG	51	48
Missing	155	143

^*“Best” refers to children who are day-8 blood and day-29 bone marrow MRD negative (<0.01%); “worst” refers to children who are bone marrow day-29 MRD positive (> 0.01%).

Statistical methods

Our data set consisted of all patients whom we could classify by MRD status. We have previously shown that patients who are MRD positive (defined as > 0.01%) in the day-29 bone marrow have a poor outcome, while those who are negative (defined as < 0.01%) in the blood at day 8 have a good outcome.¹⁷ Our goal was to identify features that best differentiated patients who would go on to have the “best” MRD results (those whose MRD is undetectable, defined as < .01%) at day 8 in blood and at day 29 in bone marrow from those with the “worst” MRD status, that is, those whose bone marrow was positive at day 29 at more than .01% level. Table 2 lists all the prognostic features tested, along with the 16 germ line polymorphism genotypic groupings. We sought to find predictors of best or worst MRD status using classification and regression tree methods (CART) as implemented in the R statistical environment in the function named rpart. We considered various demographic characteristics, tumor cytogenetics, and polymorphisms when building the decision rules predicting MRD status.

A CART analysis was performed to take into account the multiple possible important factors influencing MRD clearance, and possible interactions between them. CART analysis is a form of binary recursive partitioning. The term “binary” implies that each group of patients, represented by a “node” in a decision tree, can only be split into 2 groups. Thus, each node can be split into 2 daughter nodes, with each as homogeneous as possible in terms of MRD. The term “recursive” refers to the fact that the binary partitioning process can be applied over and over again. The term “partitioning” refers to the fact that the dataset is split into sections or partitioned.

Tree building begins at the root node, which includes all patients in the dataset. Beginning with this node, the CART software finds the best possible variable (in this case, presence or absence of TEL-AML1; Figure 1) to split the node into 2 daughter nodes. In order to find the best variable, the software checks all possible splitting variables (splitters), as well as all possible values of the variable to be used to split the node. In choosing the best splitter, the program seeks to maximize the average “homogeneity” of the 2 daughter nodes. The process of node splitting, followed by the assignment of a predicted class (MRD positive or negative) to each node, is repeated for each daughter node (in this case the second best splitter is NCI risk group) and continued recursively until it is impossible to continue due to small numbers of cases in each daughter node.

Figure 1.

Classification and regression tree (CART) for predictors influencing those patients destined to be in the “worst” versus “best” minimal residual disease (MRD).

When building our decision trees, we included patients if they were missing genotype information via so-called surrogate splits. This is a form of imputation that uses covariate information. The results were very similar when we included patients only as far as their nonmissing data allowed.

We determined the best splits by minimizing the misclassification error (Gini index).¹⁸ We followed the recommendation in Breiman et al, setting the minimum number of observations in any terminal node to 5.¹⁸ The program required at least 15 observations in a node before the program would try splitting it. We estimated the misclassification error via 10-fold cross-validation. The misclassification rate is a measure of heterogeneity in the subsets, estimated from the proportion of patients who are in the minority MRD group within each subset. Permutation testing (1000 permutations) allowed us to determine the significance of the final model, relative to a model without any splits. We chose our final models according to the “1 standard error” (1-SE) rule. This rule calls for choosing the smallest tree for which the cross-validation prediction error is within one standard error of the tree with the smallest cross-validation prediction error.

We fit a logistic regression model to compare with the CART results and to allow us to determine P values for each covariate in the full model. We also fit a stepwise logistic regression for covariates and polymorphisms associated with MRD status, using Akaike's Information Criterion to determine the best model.

Results

Of 1197 potentially eligible patients, 252 had “worst” MRD status (detectable bone marrow disease at day 28) and 277 had the “best” MRD status (no detectable leukemia at day 8 or day 28; Table 1). Subsequent analyses were performed on this subset of 529 patients.

Table 1.

Distribution of patients based on MRD status in blood at day 8 and in bone marrow at day 28

MRD present on day 8	MRD present on day 28
	Yes	No	Missing	Total
Yes	213	490	51	754
No	16	277*	28	321
Missing	23	74	25	122
Total	252*	841	104	1197

^*252 worst-risk (positive) and 277 best-risk (negative) patients from 9900 studies.

In Table 2, we present the numbers of patients in the 2 MRD categories cross-classified by various categorical predictors. Using 7 characteristics previously associated with treatment response and the 16 germ line genotypes (Table 2), the CART analysis indicated that a tree incorporating 5 splits (allowing missing values and surrogate splits) satisfied the 1-SE rule.

The CART analysis identified 5 characteristics that distinguished patients who were destined for the best versus worst MRD status (Figure 1, permutation P value for the CART = 0.0006). Logistic regression using these 5 characteristics (Tel/AML1 status of blasts, NCI risk group, race, ploidy, and the CCR5 genotype) allows assigning a P value for each characteristic. The characteristic that best distinguished MRD status was an acquired genetic feature of their ALL blasts, with the presence of the Tel/AML1 translocation in their leukemic blasts conferring a lower probability of worst MRD status (20%) than those whose blasts were negative for Tel/AML1 (53%; P < .001). Among the 136 patients positive for Tel/AML1, no other features were prognostic. Among the 393 patients negative for Tel/AML1, the most important feature to distinguish MRD status was NCI risk group (P = .001). Among the NCI low-risk patients, white patients whose blasts lacked hyperdiploidy had their ultimate MRD status differentiated by a germ line polymorphism in chemokine receptor 5 (CCR5). Patients with a homozygous AA genotype were more likely to be in the group with the worst MRD status (69%) than those with at least one G allele (41%; P = .009).

The misclassification rate for this tree is 0.31 (164 of 529), and the correct classification rate is 0.69. The proportion of observations correctly classified as “best” is 0.65, and the proportion of observations correctly classified as “worst” is 0.73. This is significantly better than random assignment of patients to best and worst group would be, where (based on the starting distribution of this population) 52.4% of patients would be considered best and 47.6 considered worst, with permutation results being significant for this CART (P = .006). Various other ways of building the tree, such as removing the variables with many missing values or restricting attention to those variables that associated with MRD when patients are first divided by their NCI risk group, made little difference to factors that classified MRD status. We also examined other techniques for building the CART, such as excluding patients if they had any data missing at a node in the tree, but the characteristics of the resulting CART were quite similar to those of the tree in Figure 1. In addition, a logistic regression using the 5 covariates identified by the CART indicated that P values for each of the 5 factors were less than .05, except for race (P = .12).

We tested whether the genotype/phenotype associations might be confounded by associations between germ line genotypes and other prognostic features (Table 3). We found that genotypes at 8 loci differed by race, and genotypes at 3 loci (including CCR5) differed by NCI risk group. As the CART approach adjusted for race and NCI risk group, it can be inferred that CCR5 is associated with MRD after accounting for confounding of genotypes with these other prognostic factors. Moreover, multiple logistic regression confirmed a significant association of CCR5 genotype (P = .009) with MRD status in a model including race (P = .12) and NCI risk group (P = .001), in addition to Tel/AML1 status (P = .001) and hyperdiploidy (P = .043).

Table 3.

Genotype by race and by NCI risk group

	Race			NCI risk group
	White	Black	Other	Standard	High
ADRB21 -468 C>G, rs11168070
CC	113*	67	5	150	81
CG and GG	190	39	32	167	94
CCR5 +246 A>G, rs1799987
AA	90	33	17	106*	34
AG and GG	205	70	59	219	115
CONNEXIN 1019 C>T, rs1764391
CC	142	51	35	153	75
CT and TT	155	51	42	170	78
GSTP1 313 A>G, rs947894
AA and AG	261*	84	69	279	135
GG	34	19	7	42	18
MBP codon 54, rs1800450
AA	4	2	0	3	3
AG and GG	293	103	77	324	149
MDR1 exon 21 2677, rs 2032581
GG	103*	39	39	121	60
Others	201	65	42	203	105
MDR1 exon 26 3435, rs1045642
CC	72*	32	29	78*	55
CT and TT	252	74	53	259	120
MTHFR 677 C>T, rs1801133
CC	141*	29	44	133	81
CT and TT	185	78	39	205	97
MTHFR 1298 A>C, rs1801131
AA	139*	75	40	163	91
AC and CC	187	32	43	175	87
NQO1 609 C>T, rs 1800566
CC	182*	45	39	190	76
CT and TT	117	58	39	137	77
CYBA P22, rs4673
CC	124	55	34	135*	78
CT and TT	174	46	42	189	73
RFC 80 A>G, rs1051266
AA and AG	208	70	61	225	114
GG	96	35	21	97	55
TPMT 460 G>A, rs 1142345
AA and AG	19	10	3	23	9
GG	284	95	75	308	146
TYMS - 827 promoter enhancer repeat
3AND3	85*	36	35	100	56
Others	224	69	44	222	115
VDR Fok1 Start site, rs1073581
CC	130	39	28	127	70
CC and CT	175	65	49	205	84
VDR intron 8 G>A, rs1544410
AA and AG	93	24	15	94	38
GG	51	25	23	67	32

^*Significantly different frequency of genotypic groups by race or by NCI risk group.

Discussion

Pharmacogenetic variation in genes such as TPMT, glutathione transferases, and thymidylate synthase has been shown to explain at least part of the variable clinical outcomes seen in children with ALL treated with uniform therapy.^{4–8,19–22} The potential involvement of multiple genes in modifying response to chemotherapy must be interpreted in the context of clinical and blast characteristics known to affect treatment response. CART analysis has a number of advantages over other classification methods, including multivariate logistic regression. First, it is inherently nonparametric. In other words, no assumptions are made regarding the underlying distribution of values of the predictor variables. Thus, CART can handle numerical data that are highly skewed or multimodal, as well as categorical predictors with either ordinal or nonordinal structure. CART identifies “splitting” variables based on an exhaustive search of all possibilities. Since efficient algorithms are used, CART is able to search all possible variables as splitters, even in problems with many hundreds of possible predictors.

The presence of MRD after remission induction therapy is known to predict poor long-term treatment outcome and is now often used to adjust postremission induction therapy.^10–14 MRD clearance is influenced by the biology of the leukemic blasts¹⁷ and has been shown to be influenced by the TPMT genotype of the patient.²³ In this study, we assessed the extent to which germ line genetic polymorphisms predict MRD status, accounting for multiple known or suspected other predictors of MRD. We focused on MRD at end induction and even earlier, rather than long-term event-free survival (EFS), because this allowed us to isolate genetic effects relevant to the small number of drugs used in induction therapy.

Previous work investigating clinical factors that influence clearance of MRD, from our group and others, have shown slow clearance of MRD in NCI high-risk children, or with leukemic blasts expressing BCR-ABL.^10,12,13 Clearance of MRD is particularly rapid in children with blasts expressing TEL-AML1.¹² The mechanism underlying poor treatment response in NCI high-risk and BCR-ABL cases remains largely unknown. Incorporation of these “non–germ line” clinical risk factors along with putative pharmacogenetic risk factors confirmed the powerful favorable effect of the TEL-AML1 translocation in our study. Among those positive for TEL-AML1, no other factors were predictive of MRD status. Among those without the favorable TEL-AML1 translocation, NCI risk group, ploidy, and race are important. The importance of race as a predictor of MRD despite inclusions of multiple pharmacogenetic variants in models is of interest, suggesting perhaps that differences in outcome driven by race are due to pharmacogenetic variants other than those tested in this study or due to non–genetically-based differences, such as subtle differences in underlying health status or nutrition.

Given the powerful effect of many of the known prognostic factors, it is not surprising that many of the genotypic factors we investigated did not show additional predictive ability. However, our data indicate that chemokine receptor 5 (CCR5) genotype (with the SNP in the CCR5 promoter) predicts clearance of MRD, after accounting for variables such as presence or absence of TEL rearrangement, NCI risk group, ploidy, and race. Children with at least one CCR5 G allele (A/G or G/G genotypes) were more likely than children with an AA genotype to be MRD negative. CCR5 has been associated in other models with apoptosis and has been related to the aggressiveness of other leukemias, and of solid tumors such as breast cancer,^24,25 and thus it is plausible that it might play a role in early response to chemotherapy. Moreover, the G allele for this CCR5 polymorphism has been associated with 45% lower promoter activity than the A allele, and HIV-1–positive individuals with the GG genotype progressed to AIDS much more slowly than those with the AA genotype,²⁶ suggesting that the G allele has salutary effects in other systems as well. Similar to other pharmacogenetic studies, the observation of a significant role for CCR5 genotype needs to be replicated in an independent dataset, and investigated in biological studies to determine its importance and any potential clinical use of genotyping.

The data in our study differ in some respects from the findings of Stanulla et al.²³ These authors genotyped thiopurine methyltransferase (TPMT) in children enrolled in the clinical trial ALL-BFM 2000 and measured MRD load at days 33 and 78 of treatment. The children received a 4-week cycle of 6-mercaptopurine, a TPMT substrate, between the 2 measurements. Patients heterozygous for allelic variants of TPMT conferring lower activity had a significantly lower rate of MRD compared with patients with homozygous wild-type alleles. In our study TPMT genotype did not influence clearance of MRD, a not-unexpected finding because we measured MRD load at early time points (days 8 and 28) prior to use of 6-mercaptopurine.

In summary, our study confirmed the importance of known predictors of MRD status and found that in a subset of patients at higher risk for MRD at the end of induction, a germ line polymorphism in the chemokine receptor further predicted MRD status. While these data are of interest, they are not sufficiently robust to use for treatment assignment at this time. Future work should address the extent to which genomic variation is the basis for some of mechanism by which patient and blast characteristics affect each child's probability of a good MRD response and thus of a good long-term prognosis.

Authorship

Contribution: S.M.D. designed the study, participated in genotyping and data interpretation, and wrote the manuscript; M.B. designed the study, performed the minimal residual disease measurements, and edited the manuscript; G.L.R. designed and supervised the statistical analysis; K.R. performed the statistical analysis; M.D. participated in the preparation of the dataset and the statistical analysis; N.W. lead a clinical study and edited the manuscript; P.L. lead a clinical study and edited the manuscript; P.B. lead a clinical study and edited the manuscript; J.S.E. participated in genotyping; C.L.W. participated in the minimal residual disease studies and edited the manuscript; S.D. participated in genotyping; E.H.C. participated in genotyping; and M.V.R. designed the study, participated in genotyping and data interpretation, and wrote the manuscript.

A list of participating institutions in the Children's Oncology Group can be found in Document S1.

Conflict-of-interest disclosure: The authors declare no competing financial interests.

Correspondence: Stella M. Davies, Cincinnati Children's Hospital and Medical Center, 3333 Burnet Ave, Cincinnati, OH 45230; e-mail: stella.davies@cchmc.org.