Subscriber access provided by UNIV OF DURHAM
Article
Epi-detected Hyperspectral Stimulated Raman Scattering Microscopy for Label-free Molecular Subtyping of Glioblastomas Bae Kideog, Wei Zheng, Kan Lin, See Wee Lim, Yuk Kien Chong, Carol Tang, Nicolas King, Christopher Beng Ti Ang, and Zhiwei Huang Anal. Chem., Just Accepted Manuscript • DOI: 10.1021/acs.analchem.8b01677 • Publication Date (Web): 02 Aug 2018 Downloaded from http://pubs.acs.org on August 2, 2018
Just Accepted “Just Accepted” manuscripts have been peer-reviewed and accepted for publication. They are posted online prior to technical editing, formatting for publication and author proofing. The American Chemical Society provides “Just Accepted” as a service to the research community to expedite the dissemination of scientific material as soon as possible after acceptance. “Just Accepted” manuscripts appear in full in PDF format accompanied by an HTML abstract. “Just Accepted” manuscripts have been fully peer reviewed, but should not be considered the official version of record. They are citable by the Digital Object Identifier (DOI®). “Just Accepted” is an optional service offered to authors. Therefore, the “Just Accepted” Web site may not include all articles that will be published in the journal. After a manuscript is technically edited and formatted, it will be removed from the “Just Accepted” Web site and published as an ASAP article. Note that technical editing may introduce minor changes to the manuscript text and/or graphics which could affect content, and all legal disclaimers and ethical guidelines that apply to the journal pertain. ACS cannot be held responsible for errors or consequences arising from the use of information contained in these “Just Accepted” manuscripts.
is published by the American Chemical Society. 1155 Sixteenth Street N.W., Washington, DC 20036 Published by American Chemical Society. Copyright © American Chemical Society. However, no copyright claim is made to original U.S. Government works, or works produced by employees of any Commonwealth realm Crown government in the course of their duties.
Page 1 of 33 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Analytical Chemistry
1
Epi-detected Hyperspectral Stimulated Raman Scattering Microscopy for Label-
2
free Molecular Subtyping of Glioblastomas
3 4
Bae Kideog1, Wei Zheng1,2, Kan Lin1,2, See Wee Lim3, Yuk Kien Chong3, Carol
5
Tang3,4,5, Nicolas K King3,4, Christopher Beng Ti Ang3,4,6,7, and Zhiwei Huang1*
6 7
1
8
Engineering, National University of Singapore, Singapore 117576
9
2
Optical Bioimaging Laboratory, Department of Biomedical Engineering, Faculty of
Department of Medicine, Yong Loo Lin School of Medicine, National University of
10
Singapore, Singapore 119260
11
3
National Neuroscience Institute, Singapore 308433
12
4
Duke-National University of Singapore Medical School, Singapore 169857
13
5
Division of Cellular and Molecular Research, Humphrey Oei Institute of Cancer
14
Research, National Cancer Centre, Singapore 169610
15
6
16
Singapore, Singapore 117597
17
7
18
Research (A*STAR), Singapore 117609
Department of Physiology, Yong Loo Lin School of Medicine, National University of
Singapore Institute for Clinical Sciences, Agency for Science, Technology and
19 20
*Correspondence to:
21
Dr. Zhiwei Huang, Optical Bioimaging Laboratory, Department of Biomedical
22
Engineering, Faculty of Engineering, National University of Singapore, 9 Engineering
23
Drive 1, Singapore 117576
24
Tel: +65- 6516-8856
25
Fax: +65- 6872-3069
26
E-mail:
[email protected] 27 28
Running title: SRS imaging for label-free molecular subtyping of GBMs
29 30
Key words: glioblastoma; stimulated Raman scattering microscopy; multivariate
31
curve resolution, principal component analysis, linear discriminant analysis
32
1
ACS Paragon Plus Environment
Analytical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
1
ABSTRACT
2
We report the development and implementation of an epi-detected spectral-focusing
3
hyperspectral stimulated Raman scattering (SRS) imaging technique for label-free
4
biomolecular subtyping of glioblastomas (GBMs). The hyperspectral SRS imaging
5
technique developed generates SRS image stacks (from 2800 to 3020 cm-1 at 7 cm-1
6
intervals) within 30 s through controlling the time delay between the chirped pump
7
and Stokes beams. SRS images at representative Raman shifts (e.g., 2845, 2885 and
8
2935 cm-1) delineate the biochemical variations and morphological differences
9
between proneural and mesenchymal subtypes of GBMs. Multivariate curve
10
resolution (MCR) analysis on hyperspectral SRS images enables the quantification of
11
major biomolecule distributions in mesenchymal and proneural GBMs. Further
12
principal component analysis (PCA) and linear discriminant analysis (LDA) together
13
with leave-one SRS spectrum-out, cross-validation (LOOCV) yields a diagnostic
14
sensitivity of 96.7% (29/30) and specificity of 88.9% (28/36) for differentiation
15
between mesenchymal and proneural subtypes of GBMs. This study shows great
16
potential of applying hyperspectral SRS imaging technique developed for rapid, label-
17
free molecular subtyping of GBMs in neurosurgery.
18 19
2
ACS Paragon Plus Environment
Page 2 of 33
Page 3 of 33 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Analytical Chemistry
1
1. Introduction
2
Glioblastoma (GBM) is the most aggressive brain tumor with fatality rate of
3
13,000 deaths each year in the United States1. Despite the implementation of surgical
4
resection and the postsurgical therapies (radiotherapy, chemotherapy) together, the
5
overall median survival remains at ~12 to 15 months2–5. Inter-patient variability to
6
treatment response is also commonly observed. In recent years, with the development
7
of The Cancer Genome Atlas (TCGA) efforts to deep-profile GBM tumors, evidence
8
has suggested that the poor prognosis could be attributed to the molecular complexity
9
and heterogeneity of GBM, accounting for significant differences in response to the
10
treatments6,7. These findings underscore the limitations of traditional pathologic
11
analyses to diagnose and subsequently influence treatment decisions. The current
12
histological system for brain tissue classification is unable to distinguish the
13
molecular subtypes of GBMs due to the lack of their cellular differences by
14
histopathology8. Recent research on the genomic and proteomic profiling can be used
15
to differentiate the subtypes with features that are described as proneural, proliferative
16
and mesenchymal9. The genetic approaches involve multiple steps for sample
17
preparation (e.g., homogenisation of tissue specimens, extraction and amplification of
18
DNA strains, etc) which make the overall process very labor-intensive and time-
19
consuming, unsuited for intraoperative diagnostic uses. It is highly desirable to
20
develop a rapid screening imaging technique with molecular specificity for GBM
21
molecular subtyping in neurosugery.
22
Stimulated Raman scattering (SRS) microscopy is a nonresonant background-free
23
vibrational spectroscopic imaging technique with high biochemical specificity10–12. In
24
SRS imaging, a pump wave (ωp) and a Stokes wave (ωs < ωp) are tightly focused onto
25
the sample; if the frequency difference (ωp − ωs) matches the vibrational frequency of
3
ACS Paragon Plus Environment
Analytical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
1
the specific chemical bond, the SRS signal (translated as a loss of the pump beam
2
intensity or an increase of the Stokes beam intensity) can be resonantly enhanced up
3
to 104-fold with phase-sensitive detection. While the fixed wavelengths of the two
4
beams only allow resonance with a single Raman band, a broader spectral information
5
can be obtained by tuning the frequency of either the pump or Stokes beams for
6
acquisition of hyperspectral SRS images. With the high biochemical selectivity and
7
sensitivity, SRS microscopy has emerged as an appealing tool for label-free,
8
quantitative imaging in biological and biomedical systems13–19. SRS technique has
9
been explored for diagnosis and tumor margin delineation of GBMs in brain tissue
10
samples20–23. But to date, SRS microscopy has not been investigated for
11
differentiation between different molecular subtypes of GBMs in brain tissue. In this
12
study, we report the development and implementation of an epi-detected hyperspectral
13
stimulated Raman scattering (SRS) imaging technique based on spectral-focusing24
14
for label-free biomolecular subtyping of GBMs without pretreatments of tissue
15
samples. The hyperspectral SRS images of GBMs acquired delineate the
16
morphological and spectroscopic differences between the mesenchymal and proneural
17
GBMs for histopathological and biomolecular assessments without labeling.
18
Multivariate curve resolution (MCR) analysis on hyperspectral SRS images and
19
multivariate statistical analysis (e.g., principal components analysis (PCA), linear
20
discriminant analysis (LDA)) on SRS spectra are conducted to establish diagnostic
21
model for differentiation between the molecular subtypes of GBMs.
22 23
2. Materials and Methods
24
2.1 Epi-detected spectral-focusing hyperspectral SRS microscopy
25
Figure 1 shows the schematic of an epi-detected spectral-focusing hyperspectral
4
ACS Paragon Plus Environment
Page 4 of 33
Page 5 of 33 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Analytical Chemistry
1
SRS microscopy system developed for label-free tissue imaging. The excitation light
2
source consists of a femtosecond (fs) laser (Insight DS dual, Spectra-Physics Inc.)
3
with two laser outputs: The fixed laser output generates 1041 nm fs laser pulses
4
(pulsewidth of 120 fs) with a repetition rate of 80 MHz, serving as the Stokes beam
5
for SRS imaging. The tunable output ranging from 680 to 1300 nm (pulsewidth of
6
100 fs) is used as the pump beam for SRS imaging. For SRS imaging, the intensity of
7
1041 nm Stokes beam is modulated at 20 MHz by an electro-optic modulator (EOM)
8
(APE-Berlin). The 1041 nm Stokes beam are collinearly combined with the pump
9
beam through a dichroic mirror and delivered into a multiphoton scanning microscope
10
(MPM-4R, Thorlabs Inc.). For epi-SRS detection, a home-built polarization control
11
module consisting of a quarter-wave plate (AQWP10M-980, Thorlabs) and a
12
polarizing beam-splitter (PBS252, Thorlabs), whose optical axis are set at 450 with
13
respect to each other, is placed before the objective to circularly polarize the Stokes
14
and pump beams. The circularly polarized excitation laser beams are scanned with a
15
pair of galvo mirrors and focused onto the sample by the objective (XLUMPLFLN
16
20X, NA=1.0, Olympus Inc.) for tissue imaging. A portion of the scattered beams is
17
epi-collected by the objective, and directed towards the same polarization control
18
module. Another round of transmission through the quarter-wave plate in the module
19
vertically polarizes the collected beam, which then allows them to be reflected
20
towards a large area photodiode (FDS1010, Thorlabs Inc.) by the subsequent
21
polarizing beam-splitter. Before the photodiode, a bandpass filter is used to spectrally
22
isolate the Stokes beam so that only the pump beam alone will be detected. The
23
modulation of the pump beam intensity due to the stimulated Raman loss (SRL)
24
process is demodulated by a lock-in amplifier to obtain the SRL signal. The time
25
constant of the lock-in amplifier is set to be 2 µs to match to the dwell time set for
5
ACS Paragon Plus Environment
Analytical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
1
SRS imaging. For hyperspectral SRS imaging, the fs pump and Stokes laser beams
2
are chirped by using SF-57 glass rods with the lengths of 48 and 50 cm to generate the
3
1.7 ps pump beam and 2 ps Stokes beam, respectively, before combined onto a
4
dichroic mirror. The Raman shift differences between the pump and Stokes is scanned
5
by controlling the time delay between the pump and Stokes pulses. The movement of
6
the motorized delay stage is synchronized to the frame trigger of the microscope to
7
achieve automatic spectral scanning of SRS images at different Raman shifts. The
8
SRS intensity changes due to differences in the pulse overlap of the pump and Stokes
9
beam during the scanning are calibrated by measuring the intensity changes of
10
Rhodamine 6G over the same scanning range24. We have evaluated the relationship of
11
epi-detected SRS signal versus the thickness of fresh pork brain tissue (Figure S1). It
12
shows that epi-detected SRS signal increases with the increasing of brain tissue
13
thickness, and reaches a maximum signal level when tissue thickness goes up to 1
14
mm, probably due to the strong back-scattering effect in brain tissue25. The above
15
results indicate that the epi-detected spectral-focusing hyperspectral SRS microscopy
16
system developed in this work has an advantage of rapid characterization of bulky
17
brain tissues without tissue pretreatment in clinical settings.
18 19
2.2 Tissue Samples
20
Graded brain tumor specimens were obtained with written informed consent, as part
21
of a National Neuroscience Institute study protocol approved by the SingHealth
22
Centralized Institutional Review Board. The brain tissue samples are obtained from
23
National Neuroscience Institute (NNI), Singapore. The tissue samples are snap-frozen
24
in liquid nitrogen and stored at -80°C before sending for SRS imaging. The tissue
25
specimens (~3x2x1 mm3 in size) are thawed first, and then sandwiched between a
6
ACS Paragon Plus Environment
Page 6 of 33
Page 7 of 33 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Analytical Chemistry
1
microscope slide and coverslip for epi-detected hyperspectral SRS imaging. In this
2
preliminary study, 9 brain tissue samples are used for SRS imaging, including 5
3
mesenchymal and 6 proneural GBM subtypes based on The Cancer Genome Atlas
4
(TCGA) Research Network.
5 6
2.3 Multivariate statistical analysis
7
In this study, multivariate curve resolution (MCR) analysis is used to decompose
8
the hyperspectral SRS image stack into the MCR spectrum of each major component
9
as well as the chemical distributions in the tissue. The underlying principle of MCR is
10
based on the bilinear model of spectrum of an unknown mixture where the measured
11
data matrix (D) may be decomposed into pure concentration maps (C) and pure
12
spectra (ST) of the k species of the sample as follows:
13
= ∑ +
(1)
14
where E is the error matrix containing the residual variation of the data. The number
15
of components (k) contributing to the data matrix can be determined by principal
16
component analysis (PCA). With an initial estimation for spectra matrix S, the C and
17
ST are calculated and then optimized iteratively using alternative least squares (ALS)
18
algorithm. In each iteration, the fitting result with raw data is compared with the
19
percentage of variance explained. The process continues until the fitting error reaches
20
the convergence set (1%). In the optimization process, non-negative concentrations
21
and spectra are imposed as constraint. All MCR analysis is performed using the
22
MCR-ALS toolbox in MATLAB (MathWorks, Natick, MA).
23
For tissue classification, the multivariate statistical analysis technique (e.g.,
24
principal components analysis and linear discriminant analysis (PCA-LDA)) is used
25
as a statistical model. First, PCA reduces the dimensionality of the SRS spectrum
7
ACS Paragon Plus Environment
Analytical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
1
ranging from 2800 to 3020 cm−1 with a subset of 30 data points. The SRS spectra are
2
first standardized to ensure that the mean of the spectra is zero and the standard
3
deviation (SD) is one, accounting for the influence of inter- and/or intra-subject
4
spectral variability on PCA. Mean centering ensure that the principal components
5
(PCs) form an orthogonal basis. Thus, PCA extracts a set of orthogonal PCs
6
comprising loadings and scores that account for most of the total variance in original
7
spectra. Each loading vector is associated with the original spectrum via PC score,
8
which represents the weight of that particular component against the basis spectrum.
9
The most diagnostically significant PCs are determined with unpaired two-sided
10
Student’s t-test (p < 0.001) and then fed into LDA modeling for tissue classification.
11
The model determines the discriminant function that maximizes the variances in the
12
dataset between different groups while minimizing the variances between members of
13
the same group. The performance of the diagnostic algorithms rendered by the PCA–
14
LDA models for correctly predicting the subtypes of GBMs is estimated in an
15
unbiased manner using the leave one spectrum-out, cross-validation method on all the
16
SRS spectra. In this approach, the spectra from each same tissue are withdrawn from
17
the data set and the PCA–LDA modeling is redeveloped based on the remaining SRS
18
spectra. The redeveloped PCA–LDA diagnostic algorithm is then used to classify the
19
withdrawn spectra. This process is repeated until all the spectra are tested. Receiver
20
operating characteristic (ROC) curves are generated by successively changing the
21
thresholds to determine the correct and incorrect classifications for all tissue
22
examined. Multivariate statistical analysis is performed using the in-house written
23
scripts in the Matlab programming environment.
24
8
ACS Paragon Plus Environment
Page 8 of 33
Page 9 of 33 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Analytical Chemistry
1
3. Results
2
The epi-detected spectral-focusing hyperspectral SRS imaging technique enables
3
rapid acquisition of hyperspectral SRS image stacks (ranging from 2800 to 3020 cm-1
4
at 7 cm-1 intervals) of GBMs without tissue pretreatments. The entire hyperspectral
5
SRS stack can be obtained within 30 s with 2.4 µs of pixel dwell time for 352 x 352
6
pixels (160 x 160 µm). Figures 2(a-h) show the representative epi-detected
7
hyperspectral SRS images of mesenchymal and proneural GBMs at the Raman shifts
8
of 2845, 2885 and 2935 cm-1, respectively. Each image provides distinct biomolecular
9
information on the two subtypes of GBMs. At the Raman shift of 2845 cm−1, lipid
10
distributions in GBMs are visualized due to the symmetric stretching of CH2 bonds of
11
the lipid molecules (Figures 2a and 2e). The hyperspectral SRS image at 2885 cm−1
12
shows a large overlap with the SRS image at 2845 cm−1, as the SRS signals are
13
generated due to the asymmetric stretching of the CH2 bonds in the same lipid
14
molecules (Figures 2b and 2f). The relatively higher intensity at 2885 cm−1 over 2845
15
cm−1 suggests that the lipids tend to be in a solid phase instead of liquid, and thus,
16
they are possibly in the form of spherical micelles26. At the Raman shift of 2935 cm−1,
17
CH3 stretch band of both lipids and proteins are strongly resonated, giving rise to the
18
uniform brightness of the tissues (Figures 2c and 2g). To increase the specific contrast
19
of the proteins in the hyperspectral SRS images, we subtract the SRS image at 2935
20
cm−1 from the SRS image at 2845 cm−1, providing a better visualization of protein
21
distributions in GBMs (Figures 2d and 2h).
22
To assess the cellular morphology of GBM tissues, the SRS at 2845 cm−1 (Figures
23
2a and 2e) and the subtracted images (Figures 2d and 2h) are overlaid together (Figure
24
3: SRS at 2845 cm−1 in green channel and the subtracted image in red channel). In
25
Figure 3a, the mesenchymal GBM is characterised by lipid droplets (white arrow) and
9
ACS Paragon Plus Environment
Analytical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
1
amorphous tissue background (orange arrow). On the other hand, the proneural GBM
2
shows greater extent of biological features which include myelins and cells (white and
3
orange arrows in Figure 3b). Figure 3c shows the average SRS spectra of the two
4
subtypes of GBM covering the high wavenumber region from 2800 to 3020 cm-1. The
5
corresponding SRS image stacks are available as movies in the supporting document
6
(Videos S1 and S2). The unique spectroscopic patterns suggest their remarkable
7
differences in terms of biomolecular makeups. The mesenchymal GBM spectrum
8
shows higher intensities at 2845 and 2885 cm-1, possibly due to the dense lipid
9
droplets in tissue; whereas a more intense peak (2935 cm-1) in proneural SRS
10
spectrum could be contributed by the abundance of cellular proteins in proneural
11
GBMs.
12
We further fed the hyperspectral SRS image stack into the MCR algorithm for
13
decomposition of SRS spectra obtained from the hyperspectral SRS images. The
14
entire hyperspectral SRS image stack is divided into 4 to 6 segments (~ square ROI of
15
120 x 120 pixels) from which average SRS spectra are obtained. First, we determine
16
the number of major chemical components in the tissues based on the eigenvalues
17
computed using principal component analysis (PCA). The first two eigenvalues make
18
up of >88% of the total variance and thus, the number of components in our MCR
19
analysis is set as two. Figures 4a and 4b show the two MCR spectra obtained for each
20
subtype. Based on the reference to the Raman spectra of the relevant chemicals27, the
21
first MCR spectra in Figure 4a are most likely to represent cellular proteins. However,
22
there are no observable spectral differences between the two for further classification.
23
In contrast, the second MCR spectra of the mesenchymal and proneural GBM in
24
Figure 4b belong to a family of fatty acids, oleic acid and arachidonic acid,
25
respectively, based on the relevant Raman spectra measured (Figure S2). In addition
10
ACS Paragon Plus Environment
Page 10 of 33
Page 11 of 33 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Analytical Chemistry
1
to the spectral analysis, the concentration maps retrieved by MCR analysis on Figures
2
4c and 4d give a clearer separation between the lipid (green color) and protein (red
3
color) distributions with no cross-talks between the two channels, thereby enabling
4
better visualization of the key cellular features such as myelins, cells and lipid
5
droplets in GBMs. In the concentration maps retrieved by MCR analysis, the pixel
6
intensities are contributed by specific biomolecules. Thus, we can quantify the relative
7
lipid content in the two subtypes of GBM tissues by comparing the pixel intensities of
8
the concentration maps (Figures 4c and 4d) with those of pure oleic acid (100% v/v).
9
Figure 5 shows that the average lipid content in the mesenchymal GBM is ~3-fold
10
higher compared to the proneural GBM (unpaired two-sided Student’s t-test, p
