Clinical Chemistry 61:1
249–258 (2015)
Cancer Diagnostics
Validation of DNA Methylation to Predict Outcome in
Acute Myeloid Leukemia by Use of xMELP
Gerald B.W. Wertheim,1,2 Catherine Smith,1 Marlise Luskin,3 Alison Rager,3 Maria E. Figueroa,4
Martin Carroll,3,5 and Stephen R. Master2*
BACKGROUND: Epigenetic dysregulation involving alterations in DNA methylation is a hallmark of various types
of cancer, including acute myeloid leukemia (AML). Although specific cancer types and clinical aggressiveness of
tumors can be determined by DNA methylation status,
the assessment of DNA methylation at multiple loci is
not routinely performed in the clinical laboratory.
METHODS: We recently described a novel microspherebased assay for multiplex evaluation of DNA methylation. In the current study, we validated and used
an improved assay [termed expedited microsphere
HpaII small fragment Enrichment by Ligationmediated PCR (xMELP)] that can be performed with
appropriate clinical turnaround time.
RESULTS: Using the xMELP assay in conjunction with a
new 17-locus random forest classifier that has been
trained using 344 AML samples, we were able to segregate an independent cohort of 70 primary AML patients
into methylation-determined subgroups with significantly distinct mortality risk (P ⫽ 0.009). We also evaluated precision, QC parameters, and preanalytic variables of the xMELP assay and determined the sensitivity
of the random forest classifier score to failure at 1 or more
loci.
CONCLUSIONS: Our results demonstrate that xMELP performance is suitable for implementation in the clinical
laboratory and predicts AML outcome in an independent
patient cohort.
© 2014 American Association for Clinical Chemistry
Cancer is thought of as a disease caused by multiple mutations that confer proliferative advantages to neoplastic
cells (1, 2 ). Extensive investigations have explored the
1
Department of Pathology, Children’s Hospital of Philadelphia; 2 Department of Pathology and Laboratory Medicine, 3 Division of Hematology and Oncology, and 5 Abramson
Research Center, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA; 4 Department of Pathology, University of Michigan Medical School, Ann Arbor, MI.
* Address correspondence to this author at: Department of Pathology and Laboratory Medicine, University of Pennsylvania, 613A Stellar-Chance Labs, 422 Curie Blvd., Philadelphia, PA 19104-6100. Fax 215-746-4650; e-mail [email protected].
Disclaimer: The content is solely the responsibility of the authors and does not necessarily
represent the official views of the NIH.
role of sequence alterations in oncogenesis, and mutational assessment of pathologic tissue aids in determining
diagnosis, prognosis, and therapy of multiple tumor
types (3 ).
Although the mutational profile of tumor cells is
central to tumor biology and the clinical assessment of
patients, it does not encompass the entire biologic dysregulation of tumor cells. Much recent work has demonstrated that cancer is not only driven by mutations but
also by epigenetic events or disrupted chromatin structure (4 ). These epigenetic changes occur at multiple levels, including DNA methylation and histone modifications. Not surprisingly, large-scale analyses of epigenetic
phenomena have shown clear correlations between epigenetic patterns and patient outcome. Correlations between DNA methylation and clinical prognosis have
been observed for many cancers, including glioblastoma,
acute myeloid leukemia (AML),6 T-cell and B-cell lymphoblastic leukemia, lung carcinoma, ovarian carcinoma,
and melanoma (5–16 ).
Despite the relationship between DNA methylation
and prognosis, assays measuring patterns of methylation
are not commonly used in clinical practice. The reasons for this likely involve both techniques and instrumentation required for DNA methylation analysis.
Methods for analyzing DNA methylation typically
utilize methylation-sensitive restriction enzyme digestion, bisulfite treatment of DNA, or precipitation using
proteins specific for methylated DNA, and the choice of
technique depends on a number of factors, including
cost, resolution required, number of loci interrogated,
turnaround time, and instrumentation and technical
skills required. Multilocus methylation analysis often involves platforms such as custom-made arrays or highthroughput sequencing, which substantially raise the cost
of clinical implementation. Thus, assays utilizing tech-
Received July 10, 2014; accepted September 12, 2014.
Previously published online at DOI: 10.1373/clinchem.2014.229781
© 2014 American Association for Clinical Chemistry
6
Nonstandard abbreviations: AML, acute myeloid leukemia; HELP, HpaII small fragment
Enrichment by Ligation-mediated PCR; MELP, Microsphere HELP; xMELP, expedited
MELP; A, absorbance; NEB, New England Biolabs; HOVON, Dutch-Belgian HematoOncology Cooperative Group; SuperPC, supervised principal components; UPenn, University of Pennsylvania; M-score, methylation score; MFI, median fluorescent intensity.
249
niques and equipment that are commonplace in pathology laboratories would be ideal.
We recently described a novel assay that simultaneously assesses the DNA methylation status of 18 prognostically important loci in patients with AML (17 ).
This methodology, based on the HpaII small fragment
Enrichment by Ligation-mediated PCR (HELP) assay,
depends on molecular techniques—restriction digestion,
oligonucleotide ligation, and PCR—that are commonplace in a clinical laboratory (18 ). Unlike HELP, which
employs custom-made solid-phase arrays for locus identity and methylation assessment, our assay (termed Microsphere HELP or MELP) uses oligonucleotides coupled to fluorescent microspheres and flow cytometric
analysis for multilocus DNA methylation assessment.
Microsphere-based techniques are commonly performed
for mutation assessment of patients with AML in clinical
laboratories (19 ). Our studies demonstrated that MELP
is a quantitative method for locus-specific assessment of
methylation levels and is highly correlated with HELP
(17 ). Additionally, a MELP-based DNA methylation
classifier using prognostic loci previously identified with
the HELP assay segregated tumors into subgroups with
significantly distinct outcomes. These data suggested that
MELP is a robust method of multilocus DNA methylation quantification that may be useful for assessing prognosis in patients with AML.
We now describe novel developments in both assay
methodology and the multivariate classification algorithm used to predict AML prognosis. The optimized
MELP technique [expedited MELP (xMELP)] shortens
the assay time to make it more appropriate for use in the
clinical laboratory (20 ). Our methylation-based prognostic algorithm is now based on a random forest classifier using a refined 17-locus panel trained from a
set of 344 AML samples. With these alterations, we
define QC standards for the assay and describe performance characteristics. We further demonstrate that
the xMELP assay and new classification algorithm
strongly predict overall survival in an independent cohort of 70 primary AML samples. These results indicate that xMELP is suitable for prognostic tumor evaluation in the clinical setting.
HpaII (NEB), 2 U T4 DNA ligase (Life Technologies).
Total reaction volume is 50 ␮L. Reactions are performed
at 25 °C for 12 h.
Primers used (nomenclature as previously used with
XXXX indicating xMELP primers):
JHpaII 12XXXX: CGCCTGTTCAT.
JHpaII 24XXXX: CGACGTCGACTATCCATG
AACAGG.
Nucleotides in bold indicate changes to prevent
redigestion of ligated products. Underlined nucleotides
are involved in annealing to genomic DNA. JHpaII
24XXXX is also used for PCR.
For dilution experiments, genomic DNA was diluted with water or with genomic DNA from peripheral
blood of a healthy donor at indicated ratios. Ficoll preparation of bone marrow samples was done according to
standard protocol (Stemcell Technologies).
Median fluorescent intensity was measured to derive raw abundance values from Luminex beads
as previously described (17 ). Log2(HpaII/MspI) values
were scaled by subtracting the mean log2 ratio for
3 loci (MSPI0406S00318682, MSPI0406S00653944,
MSPI0406S00890278) previously shown to represent an
unmethylated baseline within AML samples.
Materials and Methods
LOCUS SELECTION
xMELP ASSAY
xMELP is performed as previously described (17 ) with
the following alterations: digestion and ligation reactions
are combined using 500 ng of DNA along with 7.5 ␮L of
annealed oligonucleotides [3 A/mL JHpaII 12XXXX
(A is absorbance; XXXX indicates xMELP primers) and 6
A/mL JHpa24XXXX], 0.5 ␮L BSA [10 g/L, New England Biolabs (NEB)], 0.5 ␮L ATP (100 mmol/L, pH 7.0,
NEB), 5 ␮L digestion buffer (NEB), 4 U MspI or 2 U
250
Clinical Chemistry 61:1 (2015)
TUMOR BANK
The 344-sample Dutch-Belgian Hemato-Oncology Cooperative Group (HOVON) cohort used for training has
been previously described (6 ), and details are provided in
the Supplemental Methods in the Data Supplement that
accompanies the online version of this article at http://
www.clinchem.org/content/vol61/issue1. An independent cohort of 207 AML samples was obtained from the
Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA (UPenn). All participants provided
informed consent for collection under a protocol approved by the UPenn Institutional Review Board.
xMELP was performed on all 207 samples. A subset of
this group (n ⫽ 70) was randomly selected from the set of
participants for whom survival data were available and
was used for validation.
By use of HOVON methylation data, 18 loci were previously shown to predict outcome in AML (6 ), and we
have described their use with MELP (17 ). Additionally,
we randomly repartitioned the original HOVON data
set and used supervised principal components (SuperPC)
to identify an additional 9 candidate loci. Methylation at
these 27 loci was simultaneously assayed along with
methylation at 3 control loci for each AML sample (see
online Supplemental Table 1). A subset of the 27 loci was
used in the final classification model (see below).
xMELP Assay for Methylation in AML
DATA ANALYSIS AND MULTIPLEX CLASSIFICATION
Data analysis was performed using the R statistical software package (R version 3.0.2) in conjunction with the
survival (version 2.37–7), MethComp (version 1.22),
and scatterplot3d (version 0.3–35) packages (21–24 ).
An R script for analysis and figure creation is included as
online Supplemental Data along with associated data
files. Random forest survival calculations were performed
using the RandomForestSRC package (version 1.4) (25 )
with 1000 trees. The random forest was trained using the
HOVON (n ⫽ 344) HELP data after converting the
HELP values to equivalent MELP scale values using Deming regression as previously described (17 ).
Informative loci were identified using a permutationbased method as described in the online Supplemental
Methods. After identification of 17 informative loci, we
trained a final random survival forest. The output of this
random forest, designated as methylation score (Mscore), represents estimated risk associated with a given
sample. A low M-score corresponds to a long predicted
survival while a high M-score predicts shorter survival.
PERTURBATION ANALYSIS
To determine the effect of “failed” loci on the overall
score function, we used a Monte Carlo perturbation approach. Briefly, we selected random subsets of 1, 2, 3, 5,
or 10 loci to assess the effect of different numbers of
“failed” loci. We then selected random samples from our
21 repeat samples, and for each subset of perturbed loci
we replaced the values of each locus with a value for the
same locus drawn randomly from the cohort of 207
UPenn samples. In this way, we generated a draw of
random values that were biologically plausible but incorrect. This process was repeated 100 times for each number of perturbed loci.
methylation in the original sample DNA. Notably, this procedure is amenable to multiplexing, and we can successfully
quantify ⬎30 loci in a single Luminex experiment utilizing
only 2 reaction tubes. Despite the utility of the MELP assay
in its original form, our experiments indicated that several alterations could improve its suitability for clinical
implementation.
DEVELOPMENT OF XMELP
To improve the turnaround time for clinical methylation
assessment, we simplified the original MELP technique
using a single base-pair substitution in the oligonucleotides used for linker ligation. In both the HELP and
original MELP assays, ligating oligonucleotides to
genomic DNA fragments recreates an HpaII/MspI site
(17, 18 ). Because MspI is not heat labile, standard HELP
requires a phenol-chloroform extraction to prevent redigestion of ligated products. By changing a single base pair
(see Materials and Methods), we are able to maintain the
5⬘ overhang necessary for genomic DNA annealing while
ensuring that the ligated products no longer contain an
HpaII/MspI restriction site. With this subtle alteration,
the restriction digestion and ligation of MELP can be
combined into a single step, reducing the time required
for performing MELP by a full day and significantly decreasing the amount of sample manipulation (Fig. 1A).
This alteration of the MELP technique (referred to now
as expedited MELP or xMELP) was employed to measure the methylation levels at 28 loci (Fig. 1B). The
log2([HpaII]/[MspI]), normalized to the mean methylation level at 3 control loci, was determined using both
MELP and xMELP, and results were highly correlated at
all loci. Thus, we found xMELP to be a valid surrogate of
standard MELP and used it in all subsequent analyses.
QC AND REPRODUCIBILITY OF xMELP
Results
We have previously described the MELP assay and
shown that it can quantitatively assess levels of DNA
methylation at multiple loci (17 ). Briefly, this assay
uses a methylation-insensitive restriction enzyme
(MspI) along with its methylation-sensitive isoschizomer (HpaII) to create 2 differentially digested DNA
aliquots. Linker-specific Taq-mediated PCR generates
pools of amplicons in which relatively short, easily amplifiable products predominate. Amplicons are fluorescently labeled and hybridized to Luminex microspheres,
which are then subjected to flow cytometry to quantify
amplicon abundance at specific loci. For each sample,
median fluorescent intensities (MFI) for each locus being
interrogated are normalized to the mean MFI of 3 control
unmethylated loci evaluated in the same reaction. Differential normalized signals from the MspI and HpaII
digests at each locus are used to identify the level of
We noted occasional xMELP samples showing low fluorescence signals across all loci, suggestive of either inadequate DNA quality (low MspI and HpaII signals) or failure of subsequent reactions (low MspI or HpaII signals).
To address the need for a QC metric, we determined
mean individual scores (HpaII or MspI) across all loci as a
surrogate for global assay performance (Fig. 2A). We
then examined the individual HpaII or MspI values obtained at 3 unmethylated control loci used to normalize
our results. The low tail of the mean score distribution
(Fig. 2A) was associated with control locus values ⬍100
(Fig. 2, B and C). As a result, we established a QC cutoff
of 3/3 control loci ⬍100 (median intensity for Luminex
analysis) that was used in subsequent analyses.
To assess the precision of individual locus measurements, we performed xMELP on 21 frozen cellular aliquots of a single diagnostic bone marrow sample (Fig. 3).
Loci used were a combination of the original prognostic
group (n ⫽ 18) along with an additional 9 loci later
Clinical Chemistry 61:1 (2015) 251
Fig. 1. Comparison of the standard MELP assay to xMELP.
(A), Schematic of standard MELP and xMELP. Alterations used in xMELP decrease both hands-on labor and turnaround time by a full day. (B),
Comparison of methylation levels as measured by MELP and xMELP. DNA from 10 primary AML samples was subjected to both MELP and
xMELP. Methylation levels at 28 loci, measured as log2([HpaII]/[MspI]) normalized to the mean methylation level at 3 control loci, was
determined. Comparable methylation levels were obtained with the 2 methods at all loci examined. The 28 loci interrogated along with
the 3 normalization loci are listed in online Supplemental Table 1.
identified as potentially prognostic (Materials and Methods) and another locus (␤2-microglobulin) that was not
found to be prognostic and is used as a negative control in
later experiments. To assess intra- and interassay reproducibility, samples were run on 3 separate days (7 independent replicates/day). Because it is not appropriate to
calculate %CV with log2-based scores extending below
zero, we compared the reproducibility of replicates to the
range of M-scores measured across 207 AML samples.
With the exception of 1 locus (MSPI0406S00697563),
fluctuations in the replicate samples were small (median
intraassay SD, 0.29; median interassay SD, 0.12) and
were significantly less than the variation seen across all
samples (median range, 9.4).
EXPANSION OF CANDIDATE LOCI FOR AML PROGNOSIS
Having established QC criteria and reproducibility parameters for individual loci, we next developed an improved classifier for AML prognosis. Similar to our previous work (17 ), we used MELP-correlated HELP values
252
Clinical Chemistry 61:1 (2015)
from the HOVON AML data set to train this classifier. In
the current analysis, however, we used the full HOVON
cohort rather than a subset. Further, in contrast to previous
classifiers utilizing SuperPC analysis, we used a random forest method to exploit the robust properties of ensemble machine learning methods (26 ). We recently obtained robust
classification results using a random forest classifier to segregate myeloid neoplasms from reactive conditions (27 ) and
wished to utilize a similar approach for AML survival
prediction.
To assess which of the 27 potential loci to include in
our final model, we used a permutation-based technique
(28 ) to select an informative subset. For each locus, the
distribution (n ⫽ 5) of the variable importance score in
the random forest model was compared with a control
distribution (n ⫽ 50) derived using permuted data for
the given locus. The same analysis was done for the
␤2-microglobulin locus, whose methylation status was
not anticipated to contribute to outcome prediction.
Results for each of the 27 loci are shown in Fig. 4A,
xMELP Assay for Methylation in AML
Fig. 2. QC of xMELP assay.
xMELP was performed on 207 primary AML samples (UPenn cohort). (A) After microsphere flow cytometry, mean MFI across 31 loci (listed in online
SupplementalTable1)wasdetermined.DistributionofmeanMFIisshown.TherelativelylargenumberofsampleswithlowMFIindicatesassayfailure
forthesesamples.(B),ComparisonofMFIfrom3controllociandmeanMFIforeachsampleisshown.MFIofthe3controllociisplotted.Colorsindicate
mean MFI (red, ≤50; green, >50, dark blue ≤200, >200, ≤400; cyan >400, ≤800; black, >800). (C), Enlargement of the group of samples with low
MFI of control loci. The dotted box indicates those samples for which the assay likely failed, corresponding to an MFI signal of <100 for each of the
control loci.
and 17 loci were identified that have a true importance
significantly greater than the control importance (corrected P value of ⬍0.05, Fig. 4B). The final random
forest classifier uses these 17 loci. Importantly, the
single locus with poor precision characteristics
(MSPI0406S00697563, Fig. 3) is not included in
this model, nor is the ␤2-microglobulin locus
(MSPI0406S00708912).
The predictive score generated by this random forest
survival classifier reflects patient risk and therefore in-
Fig. 3. Variability of xMELP-determined methylation levels.
xMELP was performed on the entire primary AML UPenn cohort (n = 207) and on 21 aliquots of a single sample. Methylation levels, measured
by log2([HpaII]/[MspI]), for each locus were determined and are shown. Red dots indicate 21 replicates of a single sample, and gray dots
represent the ratios for all samples in the UPenn cohort in order to illustrate the range of biological variability. Loci included in the final xMELP
classifier are in blue, loci not in the classifier are in gray, normalization loci are in green.
Clinical Chemistry 61:1 (2015) 253
Fig. 4. Variable selection and random survival forest.
(A), A comparison of variable importance (x axis) is shown for 28 loci (listed in online Supplemental Table 1). Distributions of importance scores
from independently trained random forests on the original data are shown in green; control distributions derived from perturbation analysis
are shown in white. Green bars (original values) are not seen for loci in which near complete overlap of original and permuted values is
observed. Loci with original > permuted scores (P < 0.05) were retained for the final model. (B), 17 retained loci from (A). (C), Error rate (left)
and final variable importance (right) for 17 loci in the final 1000-tree random survival forest classifier.
creases with poorer prognosis (29 ). To assess the precision of this aggregate prognostic indication, we used the
21 replicate samples to generate an overall a risk score
(Fig. 5A). Variation in intrasample score is small compared to the intersample distribution among the 207
UPenn AML samples (overall precision, 14.8% CV; intraassay precision, 13.6% CV; interassay precision, 7.4%
CV). Additionally, 6 samples were independently processed and assayed in duplicate; methylation risk scores
for all were highly reproducible (Fig. 5B).
Because the prognostic results of xMELP are obtained from mononuclear AML blasts that have been
enriched by Ficoll gradient centrifugation and then fro254
Clinical Chemistry 61:1 (2015)
zen, and because most clinical samples are not subjected
to Ficoll enrichment or freezing before processing, we
sought to determine the effects that these 2 procedures
would have on xMELP-derived M-scores. For this analysis, we obtained multiple fresh bone marrow samples
from 5 newly diagnosed AML patients. xMELP was performed on DNA extracted at 3 points: (a) before any
manipulation (“no Ficoll”), (b) after Ficoll centrifugation (“fresh”), and (c) after both Ficoll centrifugation
and freezing in cryopreservation media (“frozen”).
One fresh, Ficoll-purified sample was eliminated owing to QC failure. As shown in Fig. 5C, scores among
the 3 types of cellular manipulations are similar, im-
xMELP Assay for Methylation in AML
Fig. 5. Variability of the overall M-score.
(A), M-scores for the entire primary AML UPenn cohort (gray dots) and for 21 replicates of a single primary AML sample (red dots) are
shown. Variability of M-score in the replicates is minimal compared to variability of M-scores across the entire UPenn cohort. (B),
M-scores of 6 duplicate samples. Line of unity is shown. (C), Effect of Ficoll centrifugation and freezing on M-scores. DNA from 5 primary
AML bone marrow aspirates was isolated prior to Ficoll centrifugation (no Ficoll), after Ficoll centrifugation (fresh), and after both Ficoll
centrifugation and cryopreservation in DMSO-containing media (frozen). Comparison of M-scores for these samples is shown along
with the line of unity. A single fresh sample failed QC (Fig. 2), so only 4 samples are shown for plots that include fresh samples. (D) Effect
of normal DNA contamination on M-scores. DNA from 2 primary AML white blood cell (WBC) samples was diluted with varying amounts
DNA from normal peripheral blood. M-scores for each dilution are indicated (blue and red dots). M-scores of the entire primary AML
UPenn cohort (gray dots) are shown as a comparison of the variability across AML samples. E) Robustness analysis of M-score with
random locus perturbation. For reference, the bottom of the figure shows the range of tertiles of M-score seen in 70 patients from the
UPenn cohort (see Fig. 6).
plying that fresh marrow samples are appropriate for
xMELP analysis.
These results were somewhat surprising, because we
considered that maturing granulocytes in the unmanipulated samples might significantly alter the M-score. We
therefore sought to determine the minimal blast percentage for which xMELP is valid. We combined varying
ratios of genomic DNA from primary AML samples with
DNA from normal peripheral blood and performed
xMELP on these mixtures. A 75:25 mixture of leukemic:
normal DNA retains a similar M-score to that of the
leukemic sample alone, whereas a 50:50 ratio shows a
substantial deviation (Fig. 5D).
Because the amount of DNA that can be obtained
from marrow or peripheral blood samples can be highly
variable, we also determined the total amount of DNA
required for xMELP to yield valid results. We serially
diluted DNA from AML samples and performed xMELP
on the dilutions. All dilutions for 1 sample showed similar M-scores, whereas significant deviations of M-scores
for a second sample were seen with approximately 20 ng
of DNA (see online Supplementary Fig. 1).
Clinical Chemistry 61:1 (2015) 255
One advantage of predictors from multiplex measurements is that the aggregate score may be robust even
if a subset of individual components is perturbed. To
explore this possibility, we compared the effect of a simulated “failure” of 1, 2, 3, 5, or 10 components to the
inherent score variation observed in our 21 replicates. To
assess the effect of values that are plausible but wrong, we
selected a random replicate from the 21 available, selected a random subset of j loci (j ⫽ 1, 2, 3, 5, 10), and
replaced the value at that locus with another value randomly chosen from the cohort of 207 UPenn samples.
This process was repeated 100 times for each value of j.
Given that our replicate sample has a low risk score relative to most other samples, this analysis should provide a
conservative estimate of the effects of perturbing our assay, because it is less likely that multiple perturbations
will “offset” each other. As expected, the score distribution shifts higher toward the population mean when the
number of perturbed loci (j) is increased (Fig. 5E). However, the overall distribution is relatively stable compared
with true replicates if a single locus “fails,” suggesting that
multiplex analysis provides some buffer against changes
in the methylation risk score due to analytical problems
at a single locus.
VALIDATION IN AN INDEPENDENT SAMPLE COHORT
Having demonstrated that our assay is reproducible and
robust in the presence of defined preanalytical variables,
we tested its ability to predict survival in a cohort of
samples from 70 AML patients (subset of the 207 tumor
samples) for which we obtained overall survival data. Because our classifier was trained on the HOVON data, this
UPenn data set represents an entirely independent cohort. xMELP-derived methylation risk scores were determined and samples were sorted into tertiles. Survival
analysis showed a highly significant difference between
methylation risk score– based cohorts (Fig. 6, P ⫽
0.009), demonstrating the clinical validity of this assay.
Taken as a whole, these results strongly suggest that
xMELP can predict outcomes of patients with AML in 2
completely independent sets of AML samples (HOVON
for training, UPenn for testing) and that xMELP may
have clinical utility for AML prognostication.
Discussion
We previously used MELP to assess DNA methylation in
select loci and showed that—at the individual locus
level—the assay is specific for the loci of interest, linear over
a 3-log range of signal intensity, as quantitative as methods
involving real-time PCR, and capable of recapitulating levels of DNA methylation determined by the HELP assay and
MassArray Epityper assay (17 ). These results, coupled with
the relatively standard techniques and instrumentation em256
Clinical Chemistry 61:1 (2015)
Fig. 6. Outcome analysis based on M-scores.
Using the random forest trained on the HOVON data set, M-scores
were determined for 70 primary AML samples from the UPenn
cohort. Samples were ranked by M-score and divided into tertiles.
Overall survival for each tertile is plotted (red, lowest M-score
group, n = 24; blue, middle M-score group, n = 23; green, highest M-score group, n = 23; P = 0.009, log-rank test). Prob,
probability.
ployed, suggested that MELP could be a useful clinical assay
for methylation assessment of AML and other diseases.
We have now expanded on the previous study by
significantly improving the techniques and analysis,
characterizing assay performance (including precision),
establishing QC parameters, and demonstrating the predictive potential of xMELP in an independent set of
AML samples. Our results further the argument that
xMELP can be used in a clinical laboratory setting for
determination of prognosis in patients with AML.
With the development of xMELP, we substantially
improved the assay technique, the loci used, and the
method of analysis. A single base-pair change in primers
significantly reduces the amount of hands-on work required for the assay and decreases the turnaround time by
a full day. The entire xMELP assay can now be performed
in 2 days, well within the optimal temporal window between AML diagnosis and treatment initiation. We
showed that xMELP yields virtually identical results regardless of Ficoll blast enrichment, indicating the sample
typically received in a clinical lab— unmanipulated bone
marrow aspirate—is adequate for xMELP analysis. Dilution experiments with normal DNA indicate that a 25%
dilution with normal DNA does not significantly alter
methylation risk score, so a 75% blast count may be taken
either as a cutoff for assay validity or as an indication that
blasts should be enriched by Ficoll centrifugation. This
criterion, however, may be too stringent, because methylation patterns of nonblast cells may not be identical to
xMELP Assay for Methylation in AML
those found in normal peripheral blood cells. Similarly,
DNA dilution results suggest that 50 ng may be an optimal, albeit stringent, cutoff for the total amount of DNA
required for analysis.
The composite, multilocus analysis of MELP data
uses a random forest classifier to determine the methylation risk score. In our analysis of the MELP data, we
tried several multiple-variable analytic approaches, including the SuperPC algorithm used in the original
HELP/MELP analysis of AML (6, 17, 30 ), and found
that the random forest method yielded robust results.
Further, the multiplex classifier retains its predictive
value even if a single locus yields an erroneous result.
Assay precision both at the individual locus level and
in terms of overall methylation risk is likely sufficient for
clinical use. Compared to intersample variation across
207 samples, the intrasample variability is minimal at
most loci, including those included in the 17-locus classifier. Formal testing of methylation risk score variation
using 21 replicates of the same sample shows a %CV of
approximately 15%, and additional experiments showed
little variation in duplicates of 6 samples. Importantly,
when we divide the UPenn AML cohort into prognostic
tertiles, 5/6 replicates are found within the same tertile.
Overall, our results indicate a high level of confidence
associated with xMELP-based risk score.
Because we obtained an AML cohort (UPenn
samples) that is entirely independent of the original
HOVON cohort, we used the full HOVON set to
train our methylation-based classifier. This scenario is
distinct from previous work in which the HOVON data
set was randomly divided into training and test subsets
(6, 17 ). The addition of samples to the training set, along
with reanalysis of HELP data for additional informative
loci, the elimination of uninformative loci, and the use of
the random forest algorithm, further optimized the
xMELP risk score. Importantly, testing the xMELP AML
risk score on a subset of the UPenn samples for which we
currently have outcome data clearly shows that xMELP
can segregate AML patients with distinct outcomes. Although preliminary, our xMELP results identify a subset
of patients with very aggressive disease for whom standard chemotherapy is likely not beneficial, and for whom
alternative treatments, such as up-front bone marrow transplant or experimental regimens, may be warranted. Additionally, the risks of subjecting patients with favorable
M-scores to aggressive therapies may outweigh the ben-
efits, so these treatment protocols may not be appropriate
for this cohort. Further studies will attempt to develop an
integrated, multivariate prognosis classification scheme
for AML that uses all available factors currently known to
influence prognosis to guide clinical management of
AML patients.
The prognostic power of our current classifier,
as well as assay reproducibility, rapid turnaround time,
and technical simplicity, demonstrate the suitability of
xMELP for analysis of AML samples in a clinical setting. Given its general applicability, xMELP warrants
further exploration for other diseases in which DNA
methylation patterns predict clinical outcomes.
Author Contributions: All authors confirmed they have contributed to
the intellectual content of this paper and have met the following 3 requirements: (a) significant contributions to the conception and design, acquisition of data, or analysis and interpretation of data; (b) drafting or revising
the article for intellectual content; and (c) final approval of the published
article.
Authors’ Disclosures or Potential Conflicts of Interest: Upon manuscript submission, all authors completed the author disclosure form. Disclosures and/or potential conflicts of interest:
Employment or Leadership: None declared.
Consultant or Advisory Role: None declared.
Stock Ownership: None declared.
Honoraria: None declared.
Research Funding: G.B.W. Wertheim, pilot project grant from the
Perelman School of Medicine Hematologic Malignancies Translational Center of Excellence, grant IRG-78-002-35 from the
American Cancer Society, and National Institutes of Health
(1R21CA185365-01); M. Carroll, National Institutes of Health
(1R21CA185365-01 and 1R01CA149566-01A1), Veterans Affairs Administration (1I01BX000918-01), and 5R21CA185365-02 (PDQ5)
Integrated Genetic and Epigenetic Prognostication for Acute
Myeloid Leukemia; S.R. Master, National Institutes of Health
(1R21CA185365-01). This project was supported in part by the
UPenn Institute for Translational Medicine and Therapeutics (ITMAT)
through a grant from the National Center for Research Resources, Grant
UL1RR024134 (now at the National Center for Advancing Translational
Sciences, Grant UL1TR000003).
Expert Testimony: None declared.
Patents: G.B.W. Wertheim, provisional patent application 62/
040,821; M. Carroll, provisional patent application 62/040,821; S.R.
Master, provisional patent application 62/040,821.
Role of Sponsor: The funding organizations played no role in the
design of study, choice of enrolled patients, review and interpretation of
data, or preparation or approval of manuscript.
References
1. Hanahan D, Weinberg RA. Hallmarks of cancer: the next
generation. Cell 2011;144:646 –74.
2. Vogelstein B, Kinzler KW. The multistep nature of cancer. Trends Genet 1993;9:138 – 41.
3. Watson IR, Takahashi K, Futreal PA, Chin L. Emerging
patterns of somatic mutations in cancer. Nat Rev Genet
2013;14:703–18.
4. Baylin SB, Jones PA. A decade of exploring the cancer
epigenome: biological and translational implications.
Nat Rev Cancer 2011;11:726 –34.
5. Noushmehr H, Weisenberger DJ, Diefes K, Phillips HS,
Pujara K, Berman BP, et al. Identification of a CpG island
methylator phenotype that defines a distinct subgroup
of glioma. Cancer Cell 2010;17:510 –22.
6. Figueroa ME, Lugthart S, Li Y, Erpelinck-Verschueren
C, Deng X, Christos PJ, et al. DNA methylation signatures identify biologically distinct subtypes in acute
myeloid leukemia. Cancer Cell 2010;17:13–27.
7. Bullinger L, Ehrich M, Döhner K, Schlenk RF, Döhner H,
Nelson MR, van den Boom D. Quantitative DNA methylation predicts survival in adult acute myeloid leukemia.
Clinical Chemistry 61:1 (2015) 257
Blood 2010;115:636 – 42.
8. Alvarez S, Suela J, Valencia A, Fernández A, Wunderlich
M, Agirre X, et al. DNA methylation profiles and their
relationship with cytogenetic status in adult acute myeloid leukemia. PLoS One 2010;5:e12197.
9. Borssén M, Palmqvist L, Karrman K, Abrahamsson J,
Behrendtz M, Heldrup J, et al. Promoter DNA methylation pattern identifies prognostic subgroups in childhood T-cell acute lymphoblastic leukemia. PLoS One
2013;8:e65373.
10. Nordlund J, Bäcklin CL, Wahlberg P, Busche S, Berglund EC, Eloranta ML, et al. Genome-wide signatures
of differential DNA methylation in pediatric acute lymphoblastic leukemia. Genome Biol 2013;14:r105.
11. Sandoval J, Mendez-Gonzalez J, Nadal E, Chen G,
Carmona FJ, Sayols S, et al. A prognostic DNA methylation signature for stage I non-small-cell lung cancer.
J Clin Oncol 2013;31:4140 –7.
12. Wei SH, Balch C, Paik HH, Kim Y-S, Baldwin RL, Liyanarachchi S, et al. Prognostic DNA methylation biomarkers in ovarian cancer. Clin Cancer Res 2006;12:
2788 –94.
13. Sigalotti L, Covre A, Fratta E, Parisi G, Sonego P, Colizzi
F, et al. Whole genome methylation profiles as independent markers of survival in stage IIIC melanoma patients. J Transl Med 2012;10:185.
14. Jiang D, Hong Q, Shen Y, Xu Y, Zhu H, Li Y, et al. The
diagnostic value of DNA methylation in leukemia: a systematic review and meta-analysis. PLoS One 2014;9:
96822.
15. Esteller M, Garcia-Foncillas J, Andion E, Goodman SN,
Hidalgo OF, Vanaclocha V, et al. Inactivation of the DNA-
258
Clinical Chemistry 61:1 (2015)
16.
17.
18.
19.
20.
21.
22.
repair gene MGMT and the clinical response of gliomas
to alkylating agents. N Engl J Med 2000;343:1350 – 4.
Chatterton Z, Burke D, Emslie KR, Craig JM, Ng J, Ashley DM, et al. Validation of DNA methylation biomarkers for diagnosis of acute lymphoblastic leukemia. Clin
Chem 2014;60:995–1003.
Wertheim GBW, Smith C, Figueroa ME, Kalos M, Bagg
A, Carroll M, Master SR. Microsphere-based multiplex
analysis of DNA methylation in acute myeloid leukemia. J Mol Diagn 2014;16:207–15.
Figueroa ME, Melnick A, Greally JM. Genome-wide determination of DNA methylation by Hpa II tiny fragment
enrichment by ligation-mediated PCR (HELP) for the
study of acute leukemias. Methods Mol Biol 2009;538:
395– 407.
King RL, Naghashpour M, Watt CD, Morrissette JJD,
Bagg A. A comparative analysis of molecular genetic
and conventional cytogenetic detection of diagnostically important translocations in more than 400 cases
of acute leukemia, highlighting the frequency of falsenegative conventional cytogenetics. Am J Clin Pathol
2011;135:921– 8.
Bertoli S, Bérard E, Huguet F, Huynh A, Tavitian S,
Vergez F, et al. Time from diagnosis to intensive chemotherapy initiation does not adversely impact the outcome of patients with acute myeloid leukemia. Blood
2013;121:2618 –26.
R Core Team. R: a language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. 2013. http://www.R-project.org/
(Accessed December 2014).
Therneau T. A package for survival analysis in S. R pack-
23.
24.
25.
26.
27.
28.
29.
30.
age version 2.37-7. 2014. http://CRAN.R-project.org/
package=survival (Accessed December 2014).
Carstensen R, Gurrin L, Ekstrom C, Figurski M.
MethComp: Functions for analysis of agreement in
method comparison studies. R package version 1.22.
2013. http://CRAN.R-project.org/package=
MethComp. (Accessed December 2014).
Ligges U, Machler M. Scatterplot3d–an R package for
visualizing multivariate data. J Stat Soft 2003;8:1–
20.
Ishwaran H, Kolalur UB. Random forests for survival,
regression and classification (RF=SRC). R package version 1.4. 2013. http://CRAN.R-project.org/package=
randomForestSRC (Accessed within the time period of December 2013–May 2014).
Hastie T, Tibshirani R, Friedman J. The elements of statistical learning: data mining, inference, and prediction. New York: Springer; 2009.
Raess PW, van de Geijn GJ, Njo TL, Klop B, Sukhachev D,
Wertheim G, et al. Automated screening for myelodysplastic syndromes through analysis of complete blood
count and cell population data parameters. Am J Hematol 2014;89:369 –74.
Hapfelmeier A, Ulm K. A new variable selection approach using Random Forests. Comp Stat Data Anal
2013;60:50 – 69.
Ishwaran H, Kogalur UB, Blackstone EH. Random survival forests. Ann Appl Stat 2008;2:841– 60.
Hastie T, Paul D, Tibshirani R. Prediction by supervised
principal components. J Am Stat Assoc 2006;101:119 –
37.