Anda di halaman 1dari 7

Experimental Hematology 2009;37:784–790

Proteomics-based prediction of clinical response in acute myeloid leukemia

Maher Albitara, Steven J. Pottsb, Francis J. Gilesc, Susan O’Briend,
Iman Jilania, Amber C. Donahuea, Elihu H. Esteyd, and Hagop Kantarjiand
Quest Diagnostics Nichols Institute, San Juan Capistrano, Calif., USA; bAperio Technologies,
Vista, Calif., USA; cDivision of Hematology, Cancer Therapy and Research Center, University of Texas, Health Science
Center, San Antonio, Tex., USA; dLeukemia Department, M.D. Anderson Cancer Center, University of Texas, Houston, Tex., USA
(Received 28 June 2008; revised 27 January 2009; accepted 4 March 2009)

Objective. Response to chemotherapy is achieved in 60% to 70% of patients with acute

myeloid leukemia. The ability to predict responders may help in stratifying patients and
exploring different therapeutic approaches for nonresponders. Proteomics methods were
used to search for predictive factors or combinations of factors.
Materials and Methods. Peripheral blood plasma samples from 41 patients with confirmed
acute myeloid leukemia with intermediate or poor cytogenetics were obtained prior to induc-
tion therapy for proteomic analysis. For each plasma sample, four fractions eluted from
a strong anion column were applied to 3 different ProteinChip array surfaces and 12
surface-enhanced laser desorption/ionization spectra were generated. Peaks that correlated
with response were identified, and decision trees incorporating these peaks along with various
clinical and laboratory findings were constructed to predict response.
Results. Multiple decision trees were constructed. One peak, when combined with age,
provided strong positive prediction of responders with 83% accuracy. A second tree, which
combined one peak with both cytogenetics and the percent of monocytes in peripheral blood,
detected responders with 95% accuracy. A third peak was adequate to predict responders in
the intermediate cytogenetic group with 86% accuracy.
Conclusions. Proteomic analysis should be further explored to define factors important in
predicting clinical response in patients with acute myeloid leukemia. Ó 2009 ISEH - Society
for Hematology and Stem Cells. Published by Elsevier Inc.

Predicting clinical behavior and response to a given therapy from patients with leukemia is enriched by leukemia-specific
in a specific patient is the basis of personalized medicine. DNA, and we have shown that cellular proteins can be readily
This is especially important in patients with acute myeloid detected in plasma from leukemia patients [4–8]. Proteomics
leukemia (AML), because of the relatively poor response to of peripheral blood plasma is particularly promising for the
treatment (60–70% responders) [1]. In particular, heteroge- analysis and prediction of clinical behavior in patients with
neity observed in response among patients exhibiting inter- hematologic diseases [9–11].
mediate cytogenetics requires new markers for use in Although plasma is easily accessible, it is particularly
stratifying these patients [2]. Numerous studies have attemp- challenging to work with because of the broad range of
ted to find new biomarkers for prediction of clinical behavior proteins that are present [12]. Plasma contains many reac-
in AML. However, the majority of these approaches depend tive proteins at levels that may be sufficiently abundant to
on obtaining bone marrow samples, which can vary in their overshadow important proteins present at much lower
composition and in the relative ratio of leukemic cells to levels. Fractionation of intact proteins is therefore essential
residual normal cells, making highly reproducible results for proper proteomic work when using plasma [13]. Here
very difficult to obtain [3]. In contrast, when plasma is we used surface-enhanced laser desorption/ionization
used, the influence of marrow variability is generally not (SELDI) and the Ciphergen ProteinChip system (Ciphergen
considered a factor. We have previously reported that plasma Biosystem Inc., Fremont, CA, USA) to analyze protein
profiles in plasma fractions from untreated AML patients.
Offprint requests to: Maher Albitar, M.D., Quest Diagnostics, Nichols
The SELDI and the ProteinChip arrays capture proteins ac-
Institute, 33608 Ortega Highway, San Juan Capistrano, CA 92690-6130; cording to their physicochemical properties (i.e., via hydro-
E-mail: phobic, hydrophilic, ion exchange, immobilized metal, or

0301-472X/09 $–see front matter. Copyright Ó 2009 ISEH - Society for Hematology and Stem Cells. Published by Elsevier Inc.
doi: 10.1016/j.exphem.2009.03.011
M. Albitar et al./ Experimental Hematology 2009;37:784–790 785

other chemistries). Proteins captured from the plasma are and are present in multiple spectra. Various settings for noise
then volatilized by a laser. The mass-to-charge ratio is subtraction, peak detection, and cluster completion were evalu-
calculated by time-of-flight mass spectrometry. Peaks iso- ated. The final settings chosen were similar to the vendor’s default
lated by this method were analyzed in tandem with a wealth settings, a first-pass peak detection of a signal-to-noise ratio of 5.0
on both peaks and valleys, and a cluster completion window of 1.0
of clinical information, and the resulting decision trees
times peak width, with a second-pass signal-to-noise setting of 2.0
yielded three promising data combinations, the collective
for both peaks and valleys.
presence of which is predictive for response to therapy To compare automatic peak detection with manual peak detec-
with a high degree of accuracy. These findings highlight tion, spectra from 2 of 12 chip types were analyzed in depth (im-
the value of plasma-based proteomics in the search for mobilized metal affinity capture chip/pH 3 fraction and WCX2
biomarkers in AML, and the great potential for this method chip/pH 9 fraction). Peaks were detected manually and results
in other malignancies [14]. were analyzed with Matlab (MathWorks, Natick, MA, USA) fol-
lowed by decision tree analysis using the classification and regres-
sion tree (CART) software created by Salford Systems (San
Diego, CA, USA) and implemented in the Biomarker Patterns
Materials and methods
software (Ciphergen). As minimal substantive differences were
Patients and sample collection found between manual peak identification and automatic peak
Consecutive newly diagnosed patients with AML were selected. All identification, automatic peak detection was employed throughout
patients were treated at the M.D. Anderson Cancer Center of the the rest of the study. As we have previously reported, rather than
University of Texas. Diagnosis of AML was based on morphology, normalizing peaks based on the total ion current of all peaks, we
cytochemical staining, and immunophenotypic analysis. Conven- normalized to neighboring peaks [9]. Accordingly, ratios of each
tional cytogenetic analysis was also performed, and patients with peak with its six nearest neighbors in each direction were calcu-
good cytogenetics (i.e., inversion 16, t(15:17), and t(8;21)) were lated, and these values were included in the decision tree analysis.
excluded. Institutional review board–approved clinical research proto-
cols were followed throughout the study, and written informed consent Statistical analysis
was obtained from all participants. Blood samples from AML patients For each of the spectra generated from the 12 combinations of
were collected prior to initiation of cytotoxic therapy. Plasma was ob- fractions and ProteinChip surfaces, a correlation matrix was calcu-
tained by centrifuging whole blood in the presence of ethylene lated between all of the observational variables (clinical outcomes,
diamine tetraacetic acid at 1500g for 10 minutes at 4 C. Plasma patient demographics, and cellular analysis) and all peaks. A
samples obtained from healthy volunteers were used as controls for program was created in Matlab to plot the 16 peaks within each
each ProteinChip plate. Plasma samples were stored at –70 C. fraction type with the lowest p value curve fitted to each observa-
tional variable. These plots were then manually inspected for
Fractionation and SELDI analysis correlation between the mass spectrometry peaks and observa-
Fractionation was performed as described previously [9]. Briefly, tional variables.
plasma samples were fractionated using the 96-well filter plate Decision tree algorithms were used to identify peaks that might be
anion exchange kit provided by Ciphergen. Samples were first useful for prediction of responders and nonresponders to treatment in
bound to the BioSepra Q Ceramic HyperD F anion exchange sorbent the AML patient population. Observational variables from cellular
on the plates, then eluted in stepwise pH gradient using the buffers and routine laboratory analysis were included, along with peak values
provided by the manufacturer. Four fractions were elution from and peak ratios of nearby neighbors (within five peaks upstream or
a strong anion based on pH as follows: flow through þ pH 9 þ pH downstream). We utilized the observational variables that are known
7, pH 5, pH 4, and pH 3 þ organic wash. All separation was per- to be clinically relevant for clinical behavior in AML in the CART
formed using the Biomek 2000 robotics system. Two aliquots analysis. The cut-off points in these variables were obtained based
from each of the four strong anion fractions were randomly assigned on the CART analysis. When using decision trees, considerable
to eight-well ProteinChip plates (Ciphergen Biosystems Inc.) with caution must be exercised to prevent overfitting [15]. To minimize
the following surfaces: immobilized metal affinity capture, strong overfitting, only two levels were allowed, meaning that the model
anion exchange, and weak cation exchange (WCX2). All samples could only be comprised of two variables, at most, from the set of
were run within a short period (7 – 10 days) to minimize machine- all peak values and all observational variables.
dependent variations. Dendrograms were used to assure consis-
tency. Only results with excellent concordance between the dupli-
cate samples were considered acceptable. Chips were read using Results
the Ciphergen ProteinChip Reader (series PBS II; Ciphergen). We
obtained 24 spectra per patient (four fractions on three plates, in Peak detection and clinical correlations
All samples were collected from patients with AML diag-
Peak detection
nosis prior to initiation of therapy. All patients were then
Peak detection was performed with CiphergenExpress 3.0 soft- treated using standard therapy (idarubicine þ cytosine arabi-
ware (Ciphergen). Spectra were normalized against total ion noside). Seventeen of 41 (41%) achieved response and 24
current between mass-to-charge (m/z) ratios of 2000 and patients failed to achieve response. The characteristics of
160,000. The CiphergenExpress software calculates clusters by these patients are listed in Table 1. Response was defined ac-
determining peaks that are above a given signal-to-noise ratio, cording to the International Working Group criteria [16].
786 M. Albitar et al./ Experimental Hematology 2009;37:784–790

Response evaluation was performed after the first induction. Table 2. The number of peaks detected in each chip/fraction
In this study, we excluded patients with good cytogenetics combination
[inv16, t(8;21), or t(15;17)]. Surface/fraction combination Peaks
A total of 856 peaks were detected in spectra from the 12
combinations of elution fraction and ProteinChip surface IMAC3/pH 3 14
IMAC3/pH 4 99
(Table 2). Numerous peaks with the same m/z were seen
IMAC3/pH 5 11
on more than one fraction/surface. Representative spectra IMAC3/pH 9 14
from the WCX2/pH 9 surface/fraction combination, which SAX2/pH 3 63
include the significant peak at m/z 6611, are shown in SAX2/pH 4 60
Figure 1. We compared spectra obtained from duplicate SAX2/pH 5 153
SAX2/pH 9 39
aliquots in order to evaluate the reproducibility of peak
WCX2/pH 3 99
normalization methods. Median coefficients of variation WCX2/pH 4 198
calculated for duplicate samples, using the various peak WCX2/pH 5 51
normalization approaches, are presented in Table 3. The WCX2/pH 9 55
normalization of a peak to the six neighboring peaks, three
IMAC 5 immobilized metal affinity capture surface; SAX2 5 strong anion
on each side, gave the best coefficient of variation between exchange; WCX2 5 weak cation exchange surface.
duplicate samples. Therefore, normalization to neighboring
peaks was used for clinical correlations. resulting from turnover of leukemic cells. The high predictive
Table 4 gives the top 20 of 856 peaks that showed signifi- value seen for performance status is of particular interest
cant correlation with response in the AML patients. We also because plasma analysis may reflect the effects not just of
looked for correlation of the peaks with various laboratory the tumor, but of the host as well.
data, such as white blood cell (WBC) count, creatinine levels,
and percentage of lymphocytes in the blood. As a representa-
Predicting response
tive example, Table 5 gives the number of peaks from the
Only 41% of the studied AML patients responded to therapy.
WCX2/pH 9 surface/fraction combination that demonstrated
As expected, numerous peaks showed significant correlation
correlation with the analytes listed (cut-off of p value
with response (Table 4). In addition B2-M, cytogenetic
#0.001). Results of these correlations were interpreted with
grouping, age, and performance status showed significant
caution because of the risk of overfitting [15]. The greatest
correlation with response (p ! 0.05). All predictive peaks,
number of peaks from the WCX2/pH 9 combination was
cytogenetics, B2-M, percentages of blasts, monocytes, and
seen to correlate with the performance status of the patient,
lymphocytes in bone marrow and peripheral blood, hemo-
b2 microglobulin levels (B2-M), and WBC count. There
globin, white cell count, platelet count, blood urea nitrogen,
was no significant correlation with lactate dehydrogenase,
and creatinine were all considered in the construction of
blood urea nitrogen, or creatinine, which suggests that most
of the significant peaks do not reflect acute-phase response
proteins. The correlation with the WBC, platelets, and blasts
might suggest that the relevant peaks may represent proteins

Table 1. Acute myeloid leukemia patient characteristics

AML patient values

Characteristic (n 5 41 patients)

Age (y), median (range) 60 (20 – 80)

White blood cell count, median  109/L (range) 4.6 (0.5 – 97.7)
Hemoglobin (g/dL), median (range) 7.65 (3.8 – 11.9)
Platelets, median  109/L (range) 42 (5 – 635)
Zubrod performance status, n (%)
0–1 34 (83)
2–4 7 (17)
Cytogenetics, n (%)
Favorable (inv16, t(8;21), or t(15;17)) 0 (excluded)
Unfavorable (–5, –7, and complex abnormalities) 12 (29)
Intermediate (diploid and other) 29 (71)
FAB classification, (%) Figure 1. Spectra from two representative samples from the weak cation
M0 – 2 31 (76) exchange pH 9 fraction. The peak at 6,611 is a strong distinguishing
M3 0 (0) biomarker for response in acute myeloid leukemia patients. The actual
M4 – 5 10 (24) peak value used is normalized against the average of the six nearest
neighbor peaks. Red indicates a responder patient spectra, and blue indi-
AML 5 acute myeloid leukemia; FAB 5 French-American-British. cates nonresponder patient spectra.
M. Albitar et al./ Experimental Hematology 2009;37:784–790 787

Table 3. An analysis of various normalization approaches Table 5. Correlation of peaks from the WCX2/pH 9 combination with
various analytes
Normalization approach CV (%) Analyte No. of peaks

No normalization 9.8 Age 4

Dividing by the total ion current in a spectrum 11.2 Performance status 7
Dividing each peak by the mean peak intensity of all peaks in 10.6 French-American-British classification 0
a spectrum Cytogenetic classification 2
Dividing each peak by the median peak intensity of all peaks in 9.8 WBC 7
a spectrum Lymphocytes in peripheral blood (%) 3
Dividing each peak by the average value of its closest six peak 6.6 Monocytes in peripheral blood (%) 1
neighbors (three on each side) Monocytes in bone marrow (%) 1
Blasts in peripheral blood (%) 5
Data from one representative chip/fraction combination: The median coef- Creatinine 0
ficient of variation (CV) across each set of two patient sample replicates Lactate dehydrogenase 0
from the weak cation exchange surface/pH 9 combination. Blood urea nitrogen 0
b2 microglobulin 9
decision trees. Two-thirds of the samples were considered in Platelets 6
CART models and one-third were used for testing and vali- Hemoglobin 0
dating the models. To reduce overfitting, only two levels WBC 5 white blood cell; WCX2 5 weak cation exchange surface.
were accepted. As shown in Figure 2A, CARTanalysis showed
that a decision tree incorporating age (cut-off at 68 years old) the same m/z 6611 peak described above was adequate to
with a peak at m/z 3223 predicted response with 83% correct predict responders with 86% accuracy and nonresponders
predictions of responders and 67% correct prediction of nonre- with 73% accuracy (Fig. 3A). A second tree utilizing B2-
sponders. A second decision tree using a peak at m/z 6611 as M and a peak at m/z 10,002 was able to predict responders
well as cytogenetic grouping and percentage of monocytes in with 81% accuracy and nonresponders with 73% accuracy
peripheral blood predicted responders at 95% accuracy and (Fig. 3B). Clearly, the peak at m/z 6611 appears to be
nonresponders at 85% accuracy (Fig. 2B). a strong predictor of response in AML, and further studies
Because patients with intermediate cytogenetics are in identifying this peak may yield an excellent prognostic
well-known to be a heterogeneous group, and it can be indicator for AML. However, by searching the protein data-
difficult to predict their response to therapy, we constructed base, the apolipoprotein C1 appears to be the best possible
decision trees specifically for this group using a random protein corresponding to m/z 6611. Apolipoprotein C1 is
two-thirds of the patients for building the models and a major protein constituent of triglyceride. It has been re-
one-third for testing. CART analysis showed that using ported based on microarray expression data using cell
lysates that the apolipoprotein C1 is expressed at high
Table 4. The 20 peaks that showed the highest correlation with response levels in leukemic cells from patients with Down syndrome
to therapy in acute myeloid leukemia patients, ranked by p value and acute megakaryocytic leukemia [17]. Confirmation of
Surface/fraction combination m/z Value p Value the identity of this protein will require further studies.

WCX2/pH 5 11679 2.24E-05

WCX2/pH 9 2447 2.64E-05
WCX2/pH 9 11022 2.75E-05 Discussion
WCX2/pH 4 17054 4.11E-05 In this study, we hypothesized that plasma protein profiles,
WCX2/pH 4 10002 4.41E-05 when used with various clinical and laboratory findings, may
WCX2/pH 4 13719 4.91E-05 help in stratifying patients for chemotherapy and to identify
WCX2/pH 9 6817 5.45E-05
SAX2/pH 3 11842 7.08E-05
nonresponders. We therefore analyzed peripheral blood
SAX2/pH 3 11632 7.94E-05 plasma using SELDI and the Ciphergen platform to identify
WCX2/pH 9 11335 8.37E-05 protein peaks that have the potential to be biomarkers for
SAX2/pH 3 11484 8.89E-05 response. We used a new approach in normalizing peak inten-
WCX2/pH 5 11630 9.04E-05 sity that is dependent on the average of the six surrounding
WCX2/pH 4 11651 1.12E-04
SAX2/pH 9 11440 1.18E-04
peaks to determine intensity, demonstrating higher reproduc-
SAX2/pH 9 11600 1.32E-04 ibility than other methods. Large number (n 5 856) of specific
SAX2/pH 9 11849 1.32E-04 peaks were identified. However, it is highly possible that some
WCX2/pH 5 11476 1.48E-04 of the individual peaks may represent a post-translation modi-
WCX2/pH 9 11641 1.65E-04 fication of the same protein detected in a different peak.
WCX2/pH 9 3223 1.78E-04
WCX2/pH 9 6611 1.78E-04
Numerous peaks as determined in this fashion showed signif-
icant correlation with response. However, to take advantage of
SAX2 5 strong anion exchange; WCX2 5 weak cation exchange surface all possible parameters, we constructed our decision tree
788 M. Albitar et al./ Experimental Hematology 2009;37:784–790

Figure 2. Two decision trees for predicting response in 41 acute myeloid leukemia patients. (A) For the test set (chosen as 33% of the samples set aside),
patients were first separated by age with a cut-off of 68 years, and then by intensity of the peak at m/z 3,223 (weak cation exchange [WCX2]/pH 9). This
analysis resulted in an accuracy of predicting patient response to therapy of 83%, and predicting nonresponse correctly in 67% of the cases. (B) This decision
tree first separated patients based on intensity at m/z 6,611 (WCX2/pH 9), with a cut-off of 73.7. Patients showing an intensity value equal to or below this
cut-off value were then further separated based on the percentage of monocytes present in the peripheral blood, while those patients with intensity values
higher than the cut-off were separated based on cytogenetics. This decision tree resulted in 95% accuracy in predicting response to therapy, while prediction
of nonresponse was 85% accurate.

incorporating all clinical and laboratory factors together with if it can be extended to routine use in clinical laboratories.
peaks in our analysis. The purpose of this study was to establish the feasibility of
Despite the small number of cases and the requirement for using proteomics in predicting clinical behavior when per-
further validation using a larger patient pool, the data we formed on plasma from peripheral blood. Issues regarding
present here show good prediction of patient response to reproducibility, standardization, and sensitivity must be ad-
therapy. Positive prediction of response was very strong, with dressed before such an approach becomes clinically useful
accuracy ranging from 83% to 95%, while the prediction of in managing patients. However, in the past few years, more
nonresponders ranged from 67% to 85%. This approach also advanced technology has become available, promising
proved to be valuable in predicting response in patients with better reproducibility and specificity than SELDI. With
intermediate cytogenetic abnormalities, a group for which such technology, proteomic testing could become routine
prediction of response is particularly difficult. Both responders in clinical laboratories, and peripheral blood plasma may
and nonresponders were predicted with a high degree of accu- potentially provide valuable information for managing
racy, and in this group of patients, B2-M emerged as a powerful patients with various diseases. Aivado et al. [18] used
predictor of response when combined with a specific peak in a similar approach along with mass spectrometry and iden-
the WCX2/pH 4 surface/fraction combination. Plasma samples tified CXCL4 and CXCL7 proteins in the serum as specific
from normal control individuals showed no significant increase markers to patients with myelodysplastic syndrome distin-
in intensity in any of the predictive peaks described here. guish them from patients with AML.
Notably, all analyses in our study were performed on The success of chemotherapy in treating patients with AML
peripheral blood plasma without the need for leukemic cells is limited [1]. Success of chemotherapy in patients with unfa-
from bone marrow, which represents an important advance vorable cytogenetics is particularly poor, but outcomes of
M. Albitar et al./ Experimental Hematology 2009;37:784–790 789

Figure 3. Decision trees for predicting response among 29 acute myeloid leukemia patients with intermediate cytogenetics. (A) For the test set (chosen as 33% of the
samples set aside), the patients exhibiting intermediate cytogenetics were analyzed using the intensity of the peak at m/z 6611 (weak cation exchange [WCX2]/pH 9)
with a cut-off of 73.7. This decision tree resulted in prediction of response to therapy with an accuracy of 86%, and prediction of nonresponse with 73% accuracy. (B)
Here we analyzed patients with intermediate cytogenetics for their levels of b2 microglobulin levels (B2-M) with a cut-off of 3.05, and then analyzed those patients
with levels greater than the B2-M cutoff for the intensity of the peak at m/z 10,002 (WCX2/pH 4) with a cut-off of 4.33. This decision tree resulted in the prediction of
responders with an accuracy of 81%, and prediction of nonresponders accurately in 73% of the test cases.

chemotherapy in patients with intermediate cytogenetics are into the MRC AML 10 trial. The Medical Research Council Adult and
difficult to determine, and thus there is a need for prognostic Children’s Leukaemia Working Parties. Blood. 1998;92:2322–2333.
3. Amin HM, Yang Y, Shen Y, et al. Having a higher blast percentage in
markers to distinguish responders from nonresponders among circulation than bone marrow: clinical implications in myelodysplastic
this group of patients [2]. Considering alternative therapy, such syndrome and acute lymphoid and myeloid leukemias. Leukemia.
as stem cell transplantation in patients who are unlikely to 2005;19:1567–1572.
respond to chemotherapy, can improve overall outcome and 4. Albitar M, Do KA, Johnson MM, et al. Free circulating soluble CD52
spare patients unnecessary treatment and discomfort. as a tumor marker in chronic lymphocytic leukemia and its implica-
tion in therapy with anti-CD52 antibodies. Cancer. 2004;101:999–
The amount of protein present in the cell is regulated not 1008.
only by levels of messenger RNA, but also by protein turn- 5. Giles FJ, Vose JM, Do KA, et al. Circulating CD20 and CD52 in
over mechanisms, proteomic methods offer a valuable alter- patients with non-Hodgkin’s lymphoma or Hodgkin’s disease. Br J
native approach to identification of potentially useful Haematol. 2003;123:850–857.
biomarkers and measurement of the levels of these proteins 6. Jilani I, Estey E, Manshuri T, et al. Better detection of FLT3 internal
tandem duplication using peripheral blood plasma DNA. Leukemia.
[19]. Data presented here support the use of peripheral 2003;17:114–119.
blood plasma for proteomic analysis in leukemias and 7. Manshouri T, Do KA, Wang X, et al. Circulating CD20 is detectable in
demonstrate that it is a valid approach for developing clin- the plasma of patients with chronic lymphocytic leukemia and is of
ically useful biomarkers for stratifying patients with AML. prognostic significance. Blood. 2003;101:2507–2513.
8. Rogers A, Joe Y, Manshouri T, et al. Relative increase in leukemia-
specific DNA in peripheral blood plasma from patients with acute
myeloid leukemia and myelodysplasia. Blood. 2004;103:2799–2801.
Acknowledgment 9. Albitar M, Potts SJ, Giles FJ, et al. Proteomic-based prediction of clin-
No financial interest/relationships with financial interest relating ical behavior in adult acute lymphoblastic leukemia. Cancer. 2006;
to the topic of this article have been declared. 106:1587–1594.
10. Czibere A, Grall F, Aivado M. Perspectives of proteomics in acute
myeloid leukemia. Expert Rev Anticancer Ther. 2006;6:1663–1675.
11. Kiehntopf M, Siegmund R, Deufel T. Use of SELDI-TOF mass spec-
References trometry for identification of new biomarkers: potential and limita-
1. Stone RM, O’Donnell MR, Sekeres MA. Acute myeloid leukemia. tions. Clin Chem Lab Med. 2007;45:1435–1449.
Hematology Am Soc Hematol Educ Program. 2004;98–117. 12. Anderson NL, Polanski M, Pieper R, et al. The human plasma pro-
2. Grimwade D, Walker H, Oliver F, et al. The importance of diagnostic teome: a nonredundant list developed by combination of four separate
cytogenetics on outcome in AML: analysis of 1,612 patients entered sources. Mol Cell Proteomics. 2004;3:311–326.
790 M. Albitar et al./ Experimental Hematology 2009;37:784–790

13. Chertov O, Simpson JT, Biragyn A, Conrads TP, Veenstra TD, Standards for Therapeutic Trials in Acute Myeloid Leukemia. J Clin
Fisher RJ. Enrichment of low-molecular-weight proteins from bio- Oncol. 2003;21:4642–4649.
fluids for biomarker discovery. Expert Rev Proteomics. 2005;2: 17. Lightfoot J, Hitzler JK, Zipursky A, Albert M, Macgregor PF. Distinct
139–145. gene signatures of transient and acute megakaryoblastic leukemia in
14. Hanash SM, Pitteri SJ, Faca VM. Mining the plasma proteome for Down syndrome. Leukemia. 2004;18:1617–1623.
cancer biomarkers. Nature. 2008;452:571–579. 18. Aivado M, Spentzos D, Germing U, et al. Serum proteome profiling
15. Wiemer JC, Prokudin A. Bioinformatics in proteomics: applica- detects myelodysplastic syndromes and identifies CXC chemokine
tion, terminology, and pitfalls. Pathol Res Pract. 2004;200: ligands 4 and 7 as markers for advanced disease. Proc Natl Acad
173–178. Sci U S A. 2007;104:1307–1312.
16. Cheson BD, Bennett JM, Kopecky KJ, et al. Revised recommenda- 19. Chen G, Gharib TG, Huang CC, et al. Discordant protein and mRNA
tions of the International Working Group for Diagnosis, Standardiza- expression in lung adenocarcinomas. Mol Cell Proteomics. 2002;1:304–
tion of Response Criteria, Treatment Outcomes, and Reporting 313.