838 Articles | JNCI Vol. 99, Issue 11 | June 6, 2007
Mass Spectrometry to Classify Non – Small-Cell
Lung Cancer Patients for Clinical Outcome After
Treatment With Epidermal Growth Factor Receptor
Tyrosine Kinase Inhibitors: A Multicohort
Fumiko Taguchi , Benjamin Solomon , Vanesa Gregorc , Heinrich Roder , Robert Gray , Kazuo Kasahara ,
Makoto Nishio , Julie Brahmer , Anna Spreafico , Vienna Ludovini , Pierre P. Massion , Rafal Dziadziuszko ,
Joan Schiller , Julia Grigorieva , Maxim Tsypin , Stephen W . Hunsucker , Richard Caprioli ,
Mark W . Duncan , Fred R . Hirsch , Paul A . Bunn Jr , David P . Carbone
Background Some but not all patients with non – small-cell lung cancer (NSCLC) respond to treatment with epidermal
growth factor receptor (EGFR) tyrosine kinase inhibitors (TKIs). We developed and tested the ability of a pre-
dictive algorithm based on matrix-assisted laser desorption ionization (MALDI) mass spectrometry (MS)
analysis of pretreatment serum to identify patients who are likely to benefit from treatment with EGFR TKIs.
Methods Serum collected from NSCLC patients before treatment with gefitinib or erlotinib were analyzed by MALDI
MS. Spectra were acquired independently at two institutions. An algorithm to predict outcomes after treat-
ment with EGFR TKIs was developed from a training set of 139 patients from three cohorts. The algorithm
was then tested in two independent validation cohorts of 67 and 96 patients who were treated with gefi-
tinib and erlotinib, respectively, and in three control cohorts of patients who were not treated with EGFR
TKIs. The clinical outcomes of survival and time to progression were analyzed.
Results An algorithm based on eight distinct m/z features was developed based on outcomes after EGFR TKI ther-
apy in training set patients. Classifications based on spectra acquired at the two institutions had a concor-
dance of 97.1%. For both validation cohorts, the classifier identified patients who showed improved
outcomes after EGFR TKI treatment. In one cohort, median survival of patients in the predicted “good”
and “poor” groups was 207 and 92 days, respectively (hazard ratio [HR] of death in the good versus poor
groups = 0.50, 95% confidence interval [CI] = 0.24 to 0.78). In the other cohort, median survivals were 306
versus 107 days (HR = 0.41, 95% CI = 0.17 to 0.63). The classifier did not predict outcomes in patients who
did not receive EGFR TKI treatment.
Conclusion This MALDI MS algorithm was not merely prognostic but could classify NSCLC patients for good or poor
outcomes after treatment with EGFR TKIs. This algorithm may thus assist in the pretreatment selection of
appropriate subgroups of NSCLC patients for treatment with EGFR TKIs.
J Natl Cancer Inst 2007;99: 838 – 46
The epidermal growth factor receptor (EGFR) was identifi ed as
a potential target for lung cancer therapy because of its frequent
expression in lung cancer tissue and the importance of EGF and
transforming growth factor ? in supporting the growth of lung
Affiliations of authors: Departments of Medicine (FT, DPC), Medicine,
Pulmonary Division (PPM), Biochemistry (RC), and Cancer Biology (DPC),
Vanderbilt-Ingram Cancer Center, Vanderbilt University Medical Center,
Nashville, TN; Departments of Medical Oncology (BS, FRH, PAB), and
Pediatrics (SWH, MWD), University of Colorado at Denver and Health
Sciences Center, Aurora, CO; Department of Oncology, Scientific Institute
University Hospital San Raffaele, Milan, Italy (VG, AS); Biodesix, Steamboat
Springs, CO (HR, JG, MT); Department of Biostatistics and Computational
Biology, ECOG Biostatistical Office, Boston, MA (RG); Department of
Respiratory Medicine, Kanazawa University, Kanazawa, Japan (KK); Thoracic
Oncology Center, Japanese Foundation for Cancer Research, Tokyo, Japan
(MN); Department of Oncology, Johns Hopkins University School of
Medicine, Baltimore, MD (JB); Department of Medical Oncology, Azienda
Ospedaliera di Perugia, Perugia, Italy (VL); Department of Medical Oncology,
Medical University of Gdansk, Poland (RD); Department of Medicine,
University of Texas Southwestern Medical Center, Dallas, TX (JS) .
Correspondence to: David P. Carbone, MD, PhD, Vanderbilt-Ingram Cancer
Center, Vanderbilt University Medical Center, Nashville, TN 37232-6838
(e-mail: firstname.lastname@example.org ).
See “Notes” following “References.”
© 2007 The Author(s).
This is an Open Access article distributed under the terms of the Creative Com -
mons Attribution Non-Commercial License (http://creativecommons.org/licenses/
by-nc/2.0/uk/), which permits unrestricted non-commercial use, distribution, and
reproduction in any medium, provided the original work is properly cited.
by guest on June 6, 2013
JNCI | Articles 839
cancer cells ( 1 ). The tyrosine kinase domain of EGFR is neces-
sary for its activity, and two highly selective EGFR tyrosine
kinase inhibitors (TKIs), gefi tinib and erlotinib, produce ob -
jective tu mor responses and clinically important stable disease
in some pa tients with advanced non – small-cell lung cancer
(NSCLC) ( 2 , 3 ). However, many patients do not respond to
EGFR TKIs, indicating that growth of only a subset of tumors is
dependent on this signaling pathway. Nevertheless, randomized
trials in unselected NSCLC patients comparing erlotinib or gefi -
tinib with placebo ( 4 , 5 ) show a small survival advantage for erlo-
tinib-treated patients and a trend toward improved survival for
those treated with gefi tinib. Thus, although a benefi t for these
drugs may exist for unselected patients, it is clearly desirable to
identify — before initiation of therapy — patients who will benefi t
and those who will not.
Specifi c mutations in the tyrosine kinase domain of the EGFR
gene ( 6 – 8 ), increased EGFR gene copy number as assessed by
fl uorescence in situ hybridization (FISH) ( 3 , 9 – 11 ), and high
EGFR protein levels as assessed by immunohistochemistry ( 3 , 9 , 11 )
are associated with tumor sensitivity to gefi tinib and erlotinib.
Mutations of the KRAS gene have also been associated with re -
sistance to EGFR TKIs ( 12 ). To date, EGFR copy number and
EGFR protein levels are the only molecular features of tumors
that have been shown to predict survival benefi t in patients treated
with gefi tinib or erlotinib in randomized placebo-controlled
trials ( 3 , 11 ). Some clinical parameters, especially a history of never
smoking, are also associated with responsiveness to EGFR TKIs,
but, importantly, a statistically signifi cant survival benefi t for erlo-
tinib was observed in all of the clinical subsets evaluated ( 13 ). This
fi nding indicates that clinical parameters alone are not suffi cient
to identify subsets of patients who will benefi t from therapy and
that this benefi t cannot be explained by receptor mutation or gene
amplifi cation ( 13 ). Thus, better tools are needed to predict which
patients with NSCLC will benefi t from EGFR TKIs.
Matrix-assisted laser desorption/ionization (MALDI) time-of-
fl ight mass spectrometry (MS) is a rapid, inexpensive, and simple
technique for analyzing complex biologic samples, such as serum,
urine, and tissue ( 14 , 15 ). Peaks in the mass spectrum correspond to
ions formed from relatively abundant species in the sample, pre-
dominantly peptides and proteins. In this study, we tested whether
mass spectrometric analysis of pretreatment peripheral blood
could assist in the identifi cation of patients who will benefi t from
treatment with gefi tinib and erlotinib. To do so, we developed a
prediction algorithm on a training set that comprised three patient
cohorts and tested it on two independent validation patient cohorts
and three independent control patient cohorts. We also examined
the concordance of mass spectra independently acquired at two
institutions to assess the reproducibility of the approach.
Patients and Samples
The training set included 139 patients with NSCLC who were
treated systemically with gefitinib and from whom sera had been
collected before treatment. These patients primarily had advanced
stage disease, but a few were medically inoperable or refused sur-
gery, as shown in Table 1 . There were three training cohorts: from
Scientific Institute Hospital San Raffaele, Milan, Italy (n = 70,
“Italian A”); from Kanazawa University, Kanazawa, Japan (n = 26,
Japan A); and from the Japanese Foundation for Cancer Research,
Tokyo, Japan (n = 43, Japan B). There were two validation cohorts.
One was an independent sequential cohort of patients with late-
stage or recurrent NSCLC from the Scientific Institute Hospital
San Raffaele (n = 67, “Italian B”) from whom sera was obtained
before treatment with single-agent gefitinib. The second validation
cohort included patients with NSCLC who were treated with first-
line erlotinib on Eastern Cooperative Oncology Group (ECOG)
protocol E3503, a single-arm phase II study (n = 96). Pretreatment
samples of both serum and plasma were available from 73 of the
patients in the ECOG study; only pretreatment serum was available
for 13 patients, and only pretreatment plasma samples were avail-
able for the remaining 10 patients. Sera were also collected from
three additional cohorts of NSCLC patients who did not receive
treatment with EGFR TKIs. Two of these control cohorts con-
sisted of patients with unresectable disease: Azienda Ospedaliera di
Perugia, Perugia, Italy (n = 32, “Italian C”), and Vanderbilt-Ingram
Cancer Center, Nashville, TN (n = 61, “VU”). The third control
cohort included patients with resectable disease from the Medical
CONTEXT AND CAVEATS
Some patients with non – small-cell lung cancer respond to treat-
ment with the epidermal growth factor receptor (EGFR) tyrosine
kinase inhibitors (TKIs) gefitinib or erlotinib, but others do not.
Clinical parameters alone are not sufficient to identify which
patients are likely to benefit.
Matrix-assisted laser desorption/ionization (MALDI) time-of-flight
mass spectrometry (MS) analysis of a training set of patients was
used to develop an algorithm to classify patients as having “good”
or “poor” outcomes after EGFR TKI treatment. The algorithm was
then tested in several independent validation and control cohorts.
The algorithm was able to classify patients in the validation cohorts
in terms of their outcomes after treatment with gefitinib or erlo-
tinib. In one validation set, the patients classified as “good” sur-
vived for a median of 306 days, whereas those classified as “poor”
survived for a median of 107 days. The algorithm did not predict
outcomes in control cohorts of patients who were not treated
with EGFR TKIs.
The algorithm was able to classify patients according to their out-
comes after EGFR TKI treatment. This classification algorithm, if
confirmed in other cohorts, may help to identify appropriate sub-
groups of non – small-cell lung cancer patients for treatment with
Some studies have shown poor reproducibility of MALDI MS profil-
ing, although this study reported that the profile was reproducible
in different institutions. The identity of the proteins that make
up the MALDI MS features in the classification algorithm is not
by guest on June 6, 2013
840 Articles | JNCI Vol. 99, Issue 11 | June 6, 2007
University of Gdansk, Poland (n = 65). The clinical characteristics
of the patients in the study are shown in Table 1 . Samples were
obtained after patients provided written informed consent, and all
analyses were performed under protocols approved by the local in -
stitutional review boards. Sera were separated by centrifugation at
1000 g (Japan B) or 2000 g (all other cohorts) for 10 minutes at 4 °C,
separated into aliquots, and frozen at – 80 °C. Duplicate samples
were shipped on dry ice to Vanderbilt University (VU) and to the
University of Colorado at Denver and Health Sciences Center
(UCDHSC), where they were stored at − 80 °C until analysis.
Mass spectra for all training samples, the Italian and Vanderbilt
control samples, and the Italian B test samples were generated
independently at both VU and UCDHSC on a Voyager
DE-STR MALDI-TOF mass spectrometer and a Voyager
DE-PRO MALDI-TOF mass spectrometer, respectively
(Applied Biosystems, Framingham, MA). The ECOG and Polish
samples were analyzed only at UCDHSC. Serum or plasma
samples were thawed on ice and diluted 1 : 10 in deionized water.
One microliter of each diluted sample was spotted at a unique
location on the MALDI target (in triplicate), and 1 µ L of matrix
solution (35 mg/mL sinapinic acid; Sigma, St Louis, MO) in
50% acetonitrile (Burdick & Jackson, Muskegon, MI) and 0.1%
trifluoroacetic acid (Sigma) was then added. The solutions were
mixed by drawing the mixture into the pipette tip and then
expelling it five times. Plates were allowed to dry at room tem-
perature. Positive ion mass spectra were then acquired in linear
Table 1 . Characteristics of the NSCLC patient sets used in this analysis *
Training set Validation setsControl sets
A and B (n = 139)Italian B (n = 67) ECOG (n = 96) Italian C (n = 32)VU (n = 61)
stage (n = 65)
Sex, No. (%)
Stage, No. (%)
Histology, No. (%)
ECOG PS, No. (%)
Smoking history, No. (%)
Current or former
Previous chemotherapy † ,
RECIST, No. (%)
37 – 77
36 – 9038 – 9141 – 9337 – 7440 – 84
0 1 (1.0)
* For some clinical attributes, we did not have complete data; only the available data are shown, with the percentage of the total available calculated. NSCLC =
non – small-cell lung caner; ECOG = Eastern Cooperative Oncology Group; VU = Vanderbilt University; NOS = not otherwise specified; PS = performance status;
N/A = not available; RECIST = Response Evaluation Criteria for Solid Tumors.
† Number of prior chemotherapy regimens.
by guest on June 6, 2013
JNCI | Articles 841
mode in an automated manner. Results from 500 to 525 inde-
pendent spectrum acquisitions for each sample were averaged to
generate each spectrum. All mass spectra were output as two-
column text files of intensity versus m/z . Spectra were calibrated
externally with mixtures of pure, well-characterized protein
standards. At UCDHSC, a mixture of insulin (bovine), thiore-
doxin ( Escherichia coli ), and apomyoglobin (equine) (Applied
Biosystems) was used; at VU, calibration was based on a mixture
of insulin, cytochrome C, myoglobin, and ubiquitin (Bruker
Daltonics, Bremen, Germany).
Raw spectra were sent electronically to Biodesix (Steamboat
Springs, CO) from each institution for analysis. Mass spectra gen-
erated from the same sample but by different personnel, institu-
tions, and instruments can exhibit variations. To enable analysis of
these spectra, we applied a suite of preprocessing procedures ( 16 –
20 ) and developed some additional procedures. In brief, the back-
ground was estimated and then subtracted from each spectrum
based on local noise estimators ( 16 – 20 ), and peaks were detected
using a signal-to-noise ratio cutoff of 3.0, which was found to be
a good compromise between overdetection and sensitivity. To
account for day-to-day and interinstrument variations in the m/z
axis scale, spectra were aligned by using a set of common peaks ( m/z
6434.5, 6632.1, 11686.9, 12864.8, 15131.1, 15871.5, and 28102.5).
Normalization of spectra was complicated by the observed large
variability of some intense peaks between individual samples, which
can suppress other signals if total ion current is used for normali-
zation. To overcome this problem, we used partial ion current
normalization techniques ( 21 ), which are based on the union of the
m/z ranges (6100 – 7500), (8500 – 10 700), and (13 300 – 16 400). The
entire preprocessing procedure was optimized by using the train-
ing set data and was held fixed for the classification of testing sets.
Preprocessing gave comparable spectra for mass spectra generated
at the two source institutions.
Training and Classifier Optimization
Each spectrum was characterized by a set of features. Features
were defined as integrated, background-subtracted, and normalized
intensities over a chosen m/z range containing a peak. A classifier
was then constructed to map a (sub)set of these features to “good”
or “poor” outcome (survival or time to progression) defined from
the clinical data. The classification algorithm we used is a straight-
forward implementation of a k -nearest neighbor (KNN) algorithm
( 22 ). The KNN algorithm requires as parameters a set of represen-
tative and labeled “instances” (i.e., a list of selected feature values).
In brief, to classify a new spectrum, the KNN algorithm first calcu-
lates the Euclidean distance of the feature values of the new spec-
trum to those of its representative spectra. This calculation yields a
list of distances from the test spectrum to each representative spec-
trum. For the k nearest neighbors (those with the k smallest dis-
tances) the labels are compared. The calculated label is a simple
majority vote over the KNN labels.
To optimize parameters for the KNN algorithm, representa-
tive spectra need to be selected from training set patients that are
themselves representative of the clinical groups. For this analysis
we initially examined fi ve clinical groups: progressive disease – early
(i.e., disease progression in <1 month), progressive disease
(i.e., disease progression in 1 – 3 months), partial response, stable
disease – short (i.e., stable disease for ≤ 6 months), and stable
disease – long (i.e., stable disease for >6 months). Visual inspection
of all available training spectra showed that the most spectrally
distinct clinical groups were those from patients with disease pro-
gression in less than 1 month and those with stable disease for
more than 6 months (Supplementary Fig. 1, A and B; available
online). We chose 13 total spectra from these groups. The remain-
ing spectra in the training set were used to optimize parameters
(the value of k , preprocessing parameters, and the integration
range for feature values) and to select the features that are the most
discriminating. As an optimization criterion, we used the leave-
one-out cross-validation (LOOCV) error. Candidate features for
the classifi cation algorithm were identifi ed as differentially
expressed m/z values from spectra from patients with rapid pro-
gressive disease (training label “poor”) and from spectra from
patients with long-term stable disease (training label “good”) by
using univariate testing (Mann – Whitney U test). The list of fea-
tures used in the classifi cation is shown in the Supplementary
Table 1 (available online).
After training and optimization, all parameters were frozen. No
changes in the classification algorithm were allowed during the
analysis of the validation and control sets. Classification labels for
two groups, “good” (i.e., closest to the stable disease – long pattern)
and “poor” (i.e., closest to the progressive disease – early pattern),
were determined by KNN analysis using the following procedure:
mass spectra were generated in triplicate; the spectra were prepro-
cessed using the fixed, predetermined parameters for alignment
and normalization; for each spectrum, the required feature values
for the eight features defined in Supplementary Table 1 (available
online) were determined; and for each replicate spectrum, these
feature values were presented to the fixed KNN classifier ( k = 7),
which then returned a label, either “good” or “poor,” or gave a
message that the values were unclassifiable. If the labels for all
replicate samples were the same, that label was assigned to the
patient sample; if the replicate labels disagreed or if no designa-
tion could be made, the patient sample was labeled “undefined.”
Time to progression and overall survival were calculated using the
Kaplan – Meier method, and graphs were generated using Graph-
Pad Prism software (GraphPad Software Inc, San Diego, CA).
Association of clinical variables, stage, sex, age, performance status,
smoking status (not available for the ECOG validation cohort), and
histology with survival was evaluated in univariate analyses and in
multivariable analyses using Cox proportional hazards regression
modeling. Proportionality was checked visually by examining plots
of log – log survival curves and of Schoenfeld residuals. The data for
all cohorts except ECOG are available as supplemental data online,
and access to the clinical data for the ECOG set is available by
contacting Robert Gray at the ECOG biostatistical office. SAS/
JMP software ( http://www.jmp.com , Cary, NC) and R ( http://
www.r-project.org , Boston, MA) were used. All P values are
by guest on June 6, 2013
842 Articles | JNCI Vol. 99, Issue 11 | June 6, 2007
Development and Assessment of the Prediction Algorithm
The set of eight discriminating features indicated in Supplementary
Table 1 gave the least LOOCV error in the classification of patients
in the training set (data not shown). Although the feature at m/z
5843.2 is the doubly charged form of the peak at m/z 11 685 and
is, therefore, not a completely independent feature, its inclusion
improved the performance of the classification algorithm (data not
The interlaboratory reproducibility of these eight MALDI
MS features was determined, and Supplementary Fig. 2 (available
online) shows a graph of the feature values from the spectra
obtained at VU against those obtained at UCDHSC for two of
these features. Good agreement in the feature values was observed
across the two institutions. The concordance of the classifi cation
results of the MALDI mass spectral data obtained at the two insti-
tutions using these eight features is shown in Table 2 . The overall
concordance with which the 206 samples constituting the training
set and the Italian B validation set were labeled as “good,” “poor,”
or “undefi ned” was 97.1%. Thus, the spectral preprocessing tech-
niques that we adopted enabled the generation of similar MALDI
mass spectra (i.e., with consistent m/z values and amplitudes)
across different institutions and nearly identical patient
Validation of the Classification Algorithm
The classification algorithm was then validated in an indepen-
dent cohort of 67 sequential NSCLC patients from Italy treated
with second- or greater line gefitinib (validation set Italian B, see
Table 1 ). This validation was performed in a blinded manner in
that MALDI MS analysis was performed and classifications
generated before the clinical outcome data were made available
to the investigators. One of the 67 samples did not yield inter-
pretable spectra and was excluded from the analysis. In Kaplan –
Meier analysis ( Fig. 1, A ), patients classified as being in the
“good” group had a statistically significantly longer time to pro-
gression than the patients in the predicted “poor” group (medi-
ans of 84 versus 61 days, respectively), with a univariate hazard
ratio (HR) of progression of 0.56 (95% confidence interval
[CI] = 0.28 to 0.89, log-rank P = .02). A statistically significant dif-
ference in overall survival was also observed between the “good”
and “poor” groups (medians of 207 versus 92 days; HR of death =
0.50, 95% CI = 0.24 to 0.78, log-rank P = .0054) ( Fig. 1, B ).
Multivariable Analysis of the Italian B Validation Set
A Cox multivariable analysis of overall survival was performed
using data from the Italian B validation cohort to compare the
Fig. 1 . Kaplan – Meier analysis of outcomes in the Italian B validation
cohort. These patients with non – small-cell lung cancer received second-
or later-line treatment with gefi tinib alone. A ) Time to progression.
B ) Overall survival (n = 67, one sample was undefi ned, one patient had
no survival data). C ) Overall survival among smokers in the Italian B
validation cohort (n = 54). Solid lines = event-free fraction; dashed
lines = 95% confi dence intervals; tick marks = censored patients.
Table 2 . Concordance between classification labels from two
different institutions *
* Samples from the combined training set and Italian B validation set (n = 206 )
were analyzed independently at Vanderbilt University (VU) and the University
of Colorado at Denver and Health Sciences Center (UCDHSC).
by guest on June 6, 2013
JNCI | Articles 843
classification obtained from the MALDI MS – based test with clas-
sification according to clinical parameters that have been previously
associated with responsiveness to EGFR TKIs, i.e., being a never
smoker and adenocarcinoma histology ( 3 ). The analysis ( Table 3 )
showed that only performance status and the MALDI MS result
were independently associated with survival benefit. Patients who
were classified in the “good” group had a statistically significantly
lower risk of death than patients classified in the “poor” group
(HR = 0.74, 95% CI = 0.55 to 0.99, log-rank P = .048).
We also asked whether it was possible to use the classifi er to
identify subgroups of patients with improved outcomes in clinical
groups known to have low response rates to EGFR TKIs. Indeed,
even in current or former smokers in the Italian B validation set
(54 of the 67 patients), the group identifi ed as “good” by the clas-
sifi cation algorithm had statistically signifi cantly better median
survival (178 days) than the group identifi ed as “poor” (88 days,
HR = 0.52, 95% CI = 0.25 to 0.87, log-rank P = .017) ( Fig. 1, C ).
Eastern Cooperative Oncology Group Validation Cohort
We also applied the prediction algorithm to a second validation
cohort. This cohort consisted of 96 previously untreated patients
participating in ECOG protocol E3503, a phase II trial of erlo-
tinib, for whom blinded samples of pretreatment serum, plasma, or
both were available. Plasma and serum samples gave similar values
for all eight MALDI MS features used in our classification algo-
rithm (Supple mentary Fig. 3, available online). Each sample was
then classified using the eight-feature classification algorithm, and
the results were sent to the ECOG biostatistical office for correla-
tion with the clinical data (data not shown). Using just the 73
patients for whom both serum and plasma were available, classifica-
tion of patients into good and poor groups in terms of overall sur-
vival was equally powerful whether based on analysis of serum or
plasma (data not shown). Consequently, we classified all 96 ECOG
patients, using spectra from serum if available (n = 86) or plasma if
not (n = 10). The patients classified in the “good” outcome group
indeed had better survival than those classified in the “poor” out-
come group ( Fig. 2 ; median survivals of 306 versus 107 days, HR =
0.41, 95% CI = 0.24 to 0.70, log-rank P <.001). With the available
Table 3 . Outcomes in the patient sets included in this analysis *
Training setValidation sets Control sets
A and B (n = 139)Italian B (n = 67) ECOG (n = 96) Italian C (n = 32)VU (n = 61)
stage (n = 65)
Classification from MALDI
MS algorithm, No. (%)
HR (95% CI)
Median time to death,
Time to progression
HR (95% CI)
Median time to
Multivariable analysis of
overall survival †
HR (95% CI)
0.9 (0.4 to 1.9)
0.45 (0.19 to 0.63)
0.5 (0.24 to 0.78)
0.4 (0.24 to 0.70)
0.74 (0.3 to 1.6)
0.81 (0.4 to 1.6)
0.5 (0.23 to 0.74)
0.56 (0.28 to 0.9)
0.53 (0.33 to 0.85)
0.74 (0.55 to 0.99) 0.53 (0.30 to 0.94)
* ECOG = Eastern Cooperative Oncology Group; VU = Vanderbilt University; MALDI = matrix-assisted laser desorption ionization; MS = mass spectrometry;
HR = hazard ratio; CI = confidence interval; N/A = not available; ND = not done.
† In the multivariable analysis, the cofactors included were performance status (0 – 5), sex (male/female), histology (adenocarcinoma, squamous cell carcinoma,
large cell carcinoma, or not otherwise specified), smoking history (no versus current or former), and MALDI MS classification (good versus poor) in the Italian B
set and performance status (0 – 5), number of involved sites (1 – 5), prior weight loss (≥5% or <5%), histology, and MALDI MS classification (good versus poor) in
the ECOG validation set.
Fig. 2 . Kaplan – Meier analysis of overall survival in the Eastern
Cooperative Oncology Group validation cohort (n = 96). These patients
had advanced non – small-cell lung cancer and had been treated fi rst line
with erlotinib alone. Solid lines = event-free fraction; dashed lines =
95% confi dence intervals; tick marks = censored patients.
by guest on June 6, 2013
844 Articles | JNCI Vol. 99, Issue 11 | June 6, 2007
follow-up (the median follow-up for time to progression was 6.7
months for those eight of the 96 patients for whom no progression
was observed), patients classified in the “good” group had statisti-
cally significantly longer time to progression than those classified
in the “poor” group (median time to progression was 3.2 and 1.9
months, respectively; HR = 0.53, 95% CI = 0.33 to 0.85, log-rank
P = .007; Table 3). In a Cox multivariable analysis that included
the parameters that were most statistically significant in the uni-
variate model — performance status (0, 1, or 2), number of involved
sites ( ≤ 3 or >3), and prior weight loss (<5% or ≥ 5%) — the MALDI
MS classification algorithm was independently statistically signifi-
cant (HR = 0.53, 95% CI = 0.30 to 0.94, Wald P = .03; Table 3 ).
Smoking status was not available in these patients, and, thus, this
important cofactor could not be analyzed.
Predictive or Prognostic?
Because the preceding analyses were all based on outcomes in
patients treated with gefitinib or erlotinib, it was important to show
that the classifications of survival outcomes from the MALDI MS
algorithm were not simply prognostic but rather identified patients
who would benefit from therapy with EGFR TKIs. For this analy-
sis, we examined outcomes in three separate cohorts of patients
with advanced NSCLC, none of whom received treatment with
EGFR TKIs. In the first cohort, a group of 32 NSCLC patients
from Perugia (Italian C) from whom serum was collected immedi-
ately before second-line chemotherapy, no statistically significant
differences were seen in the overall survival curves of patients clas-
sified in the “good” and “poor” groups ( Fig. 3, A ; HR = 0.74, 95%
CI = 0.33 to 1.6, log-rank P = .42). Because this set was so small, we
performed a permutation analysis to investigate the possibility that
such a result could have been obtained by chance, using the Italian
B set as a reference population (see Supplementary Fig. 4, available
online). There was only a 6.6% chance that these results could have
been obtained by chance.
Similarly, in the second control cohort, of 61 patients with
advanced NSCLC from VU, no difference in survival was observed
between patients classifi ed in the “good” and “poor” groups ( Fig. 3, B ;
HR = 0.81, 95% CI = 0.4 to 1.6, log-rank P = .54). Finally, in the
third control cohort, of 65 patients with resected early-stage (i.e.,
pathologic stage IA – IIB) NSCLC from Gdansk, Poland, again
survival was the same in the “good” and “poor” groups ( Fig. 3, C ;
HR = 0.90, 95% CI = 0.43 to 1.89, log-rank P = .79). Thus, the
classifi cation algorithm did not accurately classify patient out-
comes among patients not treated with EGFR TKIs.
In this study, we developed a classification algorithm based on
MALDI MS analysis of pretreatment serum and plasma that could
identify subgroups of NSCLC patients with improved time to pro-
gression and overall survival after treatment with the EGFR TKIs
gefitinib and erlotinib. On multivariable testing in two independent
validation cohorts, this algorithm retained its predictive value
independent of clinical factors associated with sensitivity to EGFR
TKIs. The classifier thus performed well for both gefitinib and
erlotinib. However, it did not perform well for traditional chemo-
therapy or surgery, based on its inability to identify patients with
Fig. 3 . Kaplan – Meier analysis of overall survival in control cohorts of
patients with non –small-cell lung cancer . A ) Overall survival in the
Italian C control cohort. These patients received chemotherapy alone
and no epidermal growth factor receptor (EGFR) tyrosine kinase inhibi-
tors (TKIs) (n = 32). B ) Overall survival in the Vanderbilt control cohort.
These patients received chemotherapy alone and no EGFR TKIs (n = 61,
one sample undefi ned). C ) Overall survival in the Polish control cohort.
These patients received surgery alone and no EGFR TKIs (n = 65). Solid
lines = event-free fraction; dashed lines = 95% confi dence intervals; tick
marks = censored patients.
by guest on June 6, 2013
JNCI | Articles 845
poor outcomes in the control cohorts. In addition, the MALDI
MS algorithm also performed well for the same classifier peaks
in both plasma and serum.
The best studied predictive tumor markers for benefi t from
treatment with EGFR TKIs in NSCLC are EGFR protein expres-
sion, specifi c EGFR mutations, and EGFR gene copy number.
Specifi cally, EGFR protein expression, as assessed by immunohis-
tochemical analysis, has been shown to be modestly associated with
overall survival benefi t after EGFR TKI treatment in some studies
( 3 , 9 ). Activating mutations in the EGFR tyrosine kinase domain
( 6 – 8 ) have been shown to be associated with dramatic responses to
EGFR TKIs, and in some retrospective single-arm studies muta-
tions have been associated with improved survival of patients
treated with EGFR TKIs compared with those without mutations
( 23 – 27 ). However, in other studies ( 3 , 9 ), including a prospective
randomized study of erlotinib compared with placebo ( 3 ), EGFR
mutations were not associated with survival benefi t from EGF TKI
treatment. In addition, an observed survival benefi t with erlotinib
was also found in groups less likely to have EGFR mutations, such
as males, smokers, and patients with squamous cell carcinoma ( 13 ).
A predictive classifi cation algorithm would have added value in
the identifi cation of patients who would benefi t from relatively
nontoxic treatment with EGFR TKIs. One molecular feature of
tumors, amplifi cation or high polysomy of the EGFR gene (i.e.,
FISH positivity), was associated with improved survival of lung
cancer patients treated with EGFR TKIs in a multivariable analysis
( 9 ) (HR = 0.44), a fi nding that has been confi rmed in a similar study
of adenocarcinoma patients (HR = 0.50) ( 10 ) and in randomized
studies of erlotinib and gefi tinib compared with placebo ( 3 , 11 ).
The classifi cation ability of the MALDI MS test described in
this study appears to be similar to that of this tumor tissue – based
assay. The univariate hazard ratio for death from any cause for
predicted “good” compared with predicted “poor” groups in the
Italian B validation set (HR = 0.50, 95% CI = 0.24 to 0.78,
P = .008) was similar to that for EGFR FISH-positive compared
with FISH-negative patients (HR = 0.44, 95% CI = 0.23 to 0.82)
( 9 , 10 ). The multivariable analysis showed that the MALDI
MS algorithm provides information over and above the clinical
parameters proposed to be predictive of response to EGFR TKIs,
spe cifi cally sex, smoking history, and histology. The algorithm
even identifi ed subgroups of smokers with statistically signifi cantly
improved survival after gefi tinib treatment. Therefore, even in
patients with clinical features associated with low response rates
to EGFR TKIs as a group, it was possible to use the algorithm to
identify a subset with a substantial predicted survival benefi t.
Both FISH and mutation analysis are tumor-based assays that
require well-preserved biopsy material, are technically diffi cult,
have a substantial cost, and have a slow turnaround time. By con-
trast, the MALDI MS method that we have described can be per-
formed on less than 1 µ L of pretreatment serum, at low cost, and
rapidly, and the method can easily be fully automated. It is thus
much more readily applied in a clinical setting than the other
An important and appropriate criticism of many previous stud-
ies using MALDI MS profi ling of serum is lack of reproducibility
( 28 , 29 ). However, here we have demonstrated that processed mass
spectra independently obtained from two institutions on two
different instruments can yield highly reproducible classifi cation
by using appropriate preprocessing methods. The observed con -
cordance of 97.1% between the two institutions that generated
the MALDI MS data for our study compares favorably with inter-
laboratory variability of well-established tests, such as immuno-
histochemistry or FISH testing for HER2 ( 30 ).
In the clinical development of biomarkers for the individual-
ization of therapy, it is important to distinguish between biomark-
ers that can accurately classify patients according to whether they
will benefi t from an intervention and those that simply portend a
favorable or unfavorable prognosis, independent of the planned
intervention. Biomarkers predictive for survival benefi t from an
intervention are much more useful for guiding management. The
discriminating features that we have identifi ed in the mass spectra
of serum and plasma are unlikely to represent markers of poor
prognosis, given the lack of prognostic signifi cance of the classifi -
cation algorithm when it was used to analyze three independent
cohorts of patients with NSCLC who did not receive EGFR
TKIs. Moreover, in multivariable analysis, the MALDI MS test
was predictive of survival independent of performance status,
which also suggests that it was not merely prognostic.
One limitation of our analysis is the lack of smoking data in
the ECOG cohort, because smoking is a clear predictive factor
for response to EGFR TKIs. However, our classifi er predicted
outcomes independent of smoking status in the Italian B valida-
tion cohort. Moreover, in a US-based trial of fi rst-line treatment
for advanced disease such as the ECOG study, the number of
never smokers is likely to be too low to account for a substantial
portion of the discriminatory power of our signature. Another
limitation is the unknown biology underlying the ability of these
features to predict benefi t. The identifi cation and analysis of
the informative peaks might lead to important insights into the
mechanism of the association, and these studies are under way.
This study represents the fi rst comprehensive and rigorously
validated attempt to use MALDI MS methods to classify patients
for their clinical benefi t from a molecularly targeted anticancer
agent. MALDI MS analysis of pretreatment serum performed in
parallel at two institutions and based on samples from three con-
tinents offers a robust and reproducible method. In two blinded
validation studies, in patients receiving both fi rst- and second-
line treatment and using both gefi tinib and erlotinib, the test had
a classifi cation ability similar to that of tumor-based assays and
independent of other clinical parameters associated with response
to EGFR inhibitors. It will be important to confi rm the clinical
value of this strategy in randomized trials with larger cohorts of
patients treated with EGFR TKIs.
(1) Arteaga CL . ErbB-targeted therapeutic approaches in human cancer . Exp
Cell Res 2003 ; 284 : 122 – 30 .
(2) Fukuoka M , Yano S , Giaccone G , Tamura T , Nakagawa K , Douillard JY ,
et al . Multi-institutional randomized phase II trial of gefi tinib for previ-
ously treated patients with advanced non-small-cell lung cancer (The
IDEAL 1 Trial) [corrected] . J Clin Oncol 2003 ; 21 : 2237 – 46 .
(3) Tsao MS , Sakurada A , Cutz JC , Zhu CQ , Kamel-Reid S , Squire J , et al .
Erlotinib in lung cancer—molecular and clinical predictors of outcome .
N Engl J Med 2005 ; 353 : 133 – 44 .
by guest on June 6, 2013
846 Articles | JNCI Vol. 99, Issue 11 | June 6, 2007 Download full-text
(4) Shepherd FA , Rodrigues Pereira J , Ciuleanu T , Tan EH , Hirsh V ,
Thongprasert S , et al . Erlotinib in previously treated non-small-cell lung
cancer . N Engl J Med 2005 ; 353 : 123 – 32 .
(5) Thatcher N , Chang A , Parikh P , Rodrigues Pereira J , Ciuleanu T , von
Pawel J , et al . Gefi tinib plus best supportive care in previously treated
patients with refractory advanced non-small-cell lung cancer: results from
a randomised, placebo-controlled, multicentre study (Iressa Survival
Evaluation in Lung Cancer) . Lancet 2005 ; 366 : 1527 – 37 .
(6) Lynch TJ , Bell DW , Sordella R , Gurubhagavatula S , Okimoto RA ,
Brannigan BW , et al . Activating mutations in the epidermal growth factor
receptor underlying responsiveness of non-small-cell lung cancer to gefi -
tinib . N Engl J Med 2004 ; 350 : 2129 – 39 .
(7) Paez JG , Janne PA , Lee JC , Tracy S , Greulich H , Gabriel S , et al . EGFR
mutations in lung cancer: correlation with clinical response to gefi tinib
therapy . Science 2004 ; 304 : 1497 – 500 .
(8) Pao W , Miller V , Zakowski M , Doherty J , Politi K , Sarkaria I , et al . EGF
receptor gene mutations are common in lung cancers from “never smok-
ers” and are associated with sensitivity of tumors to gefi tinib and erlotinib .
Proc Natl Acad Sci U S A 2004 ; 101 : 13306 – 11 .
(9) Cappuzzo F , Hirsch FR , Rossi E , Bartolini S , Ceresoli GL , Bemis L ,
et al . Epidermal growth factor receptor gene and protein and gefi tinib
sensitivity in non-small-cell lung cancer . J Natl Cancer Inst 2005 ; 97 :
643 – 55 .
(10) Hirsch FR , Varella-Garcia M , McCoy J , West H , Xavier AC , Gumerlock
P , et al . Increased epidermal growth factor receptor gene copy number
detected by fl uorescence in situ hybridization associates with increased
sensitivity to gefi tinib in patients with bronchioloalveolar carcinoma
subtypes: a Southwest Oncology Group Study . J Clin Oncol 2005 ; 23 :
6838 – 45 .
(11) Hirsch FR , Varella-Garcia M , Bunn PA Jr . Molecular predictors of out-
come with gefi tinib in a phase III placebo-controlled study in advanced
non-small cell lung cancer . J Clin Oncol 2006 ; 24 : 5034 – 42 .
(12) Pao W , Wang TY , Riely GJ , Miller VA , Pan Q , Ladanyi M , et al . KRAS
mutations and primary resistance of lung adenocarcinomas to gefi tinib or
erlotinib . PLoS Med 2005 ; 2 : e17 .
(13) Clark GM , Zborowski DM , Santabarbara P , Ding K , Whitehead M ,
Seymour L , et al . Smoking history and epidermal growth factor receptor
expression as predictors of survival benefi t from erlotinib for patients with
non-small-cell lung cancer in the National Cancer Institute of Canada
Clinical Trials Group Study BR.21 . Clin Lung Cancer 2006 ; 7 : 389 – 94 .
(14) Schwartz SA , Weil RJ , Thompson RC , Shyr Y , Moore JH , Toms SA ,
et al . Proteomic-based prognosis of brain tumor patients using direct-
tissue matrix-assisted laser desorption ionization mass spectrometry .
Cancer Res 2005 ; 65 : 7674 – 81 .
(15) Yanagisawa K , Shyr Y , Xu BJ , Massion PP , Larsen PH , White BC , et al .
Proteomic patterns of tumour subsets in non-small-cell lung cancer .
Lancet 2003 ; 362 : 433 – 9 .
(16) Amann JM , Chaurand P , Gonzalez A , Mobley JA , Massion PP , Carbone
DP , et al . Selective profi ling of proteins in lung cancer cells from fi ne-
needle aspirates by matrix-assisted laser desorption ionization time-of-
fl ight mass spectrometry . Clin Cancer Res 2006 ; 12 : 5142 – 50 .
(17) Chaurand P , Norris JL , Cornett DS , Mobley JA , Caprioli RM . New
developments in profi ling and imaging of proteins from tissue sections by
MALDI mass spectrometry . J Proteome Res 2006 ; 5 : 2889 – 900 .
(18) Cornett DS , Mobley JA , Dias EC , Andersson M , Arteaga CL , Sanders
ME , et al . A novel histology-directed strategy for MALDI-MS tissue pro-
fi ling that improves throughput and cellular specifi city in human breast
cancer . Mol Cell Proteomics 2006 ; 5 : 1975 – 83 .
(19) Meistermann H , Norris JL , Aerni HR , Cornett DS , Friedlein A , Erskine AR ,
et al . Biomarker discovery by imaging mass spectrometry: transthyretin
is a biomarker for gentamicin-induced nephrotoxicity in rat . Mol Cell
Proteomics 2006 ; 5 : 1876 – 86 .
(20) Roder H , Grigorieva J , Tsypin M . The use of mass spectra for cancer bio-
marker detection . Biodesix ; 2005. Available at: http://www.biodesix.com/
(21) Tsypin M , Grigorieva J , Rotthoff E , Roder H . A061107. Partial ion cur-
rent nor malization of mass spectra and biomarker reproducibility .
Proceedings of the 54th ASMS Conference on Mass Spectrometry and
Allied Topics, Seattle, WA, May 28–June 1, 2006; p. 121S .
(22) Webb A . Statistical pattern recognition . Chichester (U.K.) : Wiley ; 2002 .
(23) Cortes-Funes H , Gomez C , Rosell R , Valero P , Garcia-Giron C , Velasco A ,
et al . Epidermal growth factor receptor activating mutations in Spanish
gefi tinib-treated non-small-cell lung cancer patients . Ann Oncol 2005 ;
16 : 1081 – 6 .
(24) Han SW , Kim TY , Hwang PG , Jeong S , Kim J , Choi IS , et al . Predictive
and prognostic impact of epidermal growth factor receptor mutation in
non-small-cell lung cancer patients treated with gefi tinib . J Clin Oncol
2005 ; 23 : 2493 – 501 .
(25) Mitsudomi T , Kosaka T , Endoh H , Horio Y , Hida T , Mori S , et al . Mu -
tations of the epidermal growth factor receptor gene predict prolonged
survival after gefi tinib treatment in patients with non-small-cell lung can-
cer with postoperative recurrence . J Clin Oncol 2005 ; 23 : 2513 – 20 .
(26) Takano T , Ohe Y , Sakamoto H , Tsuta K , Matsuno Y , Tateishi U , et al .
Epidermal growth factor receptor gene mutations and increased copy
numbers predict gefi tinib sensitivity in patients with recurrent non-small-
cell lung cancer . J Clin Oncol 2005 ; 23 : 6829 – 37 .
(27) Taron M , Ichinose Y , Rosell R , Mok T , Massuti B , Zamora L , et al .
Activating mutations in the tyrosine kinase domain of the epidermal
growth factor receptor are associated with improved survival in gefi tinib-
treated chemorefractory lung adenocarcinomas . Clin Cancer Res 2005 ;
11 : 5878 – 85 .
(28) Diamandis EP . Mass spectrometry as a diagnostic and a cancer biomarker
discovery tool . Mol Cell Proteomics 2004 ; 3 : 367 .
(29) Diamandis EP . Proteomic patterns in biological fl uids: do they represent
the future of cancer diagnostics? Clin Chem 2003 ; 49 : 1272 .
(30) Perez EA , Suman VJ , Davidson NE , Martino S , Kaufman PA , Lingle WL ,
et al . HER2 testing by local, central, and reference laboratories in speci-
mens from the North Central Cancer Treatment Group N9831 inter-
group adjuvant trial . J Clin Oncol 2006 ; 24 : 3032 – 8 .
F. Taguchi and B. Solomon contributed equally to this work.
Supported by grants from National Cancer Institute SPOREs in Lung
Cancer CA90949 (to D. P. Carbone) and CA58187 (to P. A. Bunn); cancer
center core grants to Vanderbilt and Colorado CA68485 and CA046934,
respectively; International Association for the Study of Lung Cancer fellow-
ship to R. Dziadziuszko; ECOG support for J. Schiller, J. Brahmer, and
R. Gray; and unrestricted funds from Genentech to David Johnson, supporting
F. Taguchi. OSI Pharmaceuticals provided some funding for sample collection
and processing for the E3503 ECOG study. H. Roder is the Chief Technical
Offi cer of Biodesix, which is developing tools for MS-based clinical diagnostics.
J. Brahmer is conducting research sponsored by AstraZeneca (manufacturer of
gefi tinib). F. R. Hirsch is doing research in collaboration with AstraZeneca,
OSI Pharmaceuticals, and Genentech under research agreements with the
University of Colorado. F. R. Hirsch has served on advisory boards, and
P. A. Bunn has consulted for OSI/Genentech/Roche (manufacturers of erlo-
tinib) and AstraZeneca. J. Schiller has served on advisory boards for Genentech.
Genentech and the other study sponsors played no role in the design of the
study; the collection, analysis, or interpretation of the data; and the preparation
of the manuscript or the decision to submit the manuscript for publication.
Funding to pay the Open Access publication charges for this article was
provided by Biodesix, Inc, Steamboat Springs, CO.
Manuscript received November 20 , 2006 ; revised March 26 , 2007 ; accepted
April 13 , 2007.
by guest on June 6, 2013