2. ARTICLE IN PRESS
2 P. Heraud et al.
analyses to each aliquot of cells destined for pharmaceutical
screening or transplantation is impractical.
Fourier transform infrared (FTIR) spectroscopy has been
used to study biological molecules and biological systems in
vitro since the early 1980s (Mantsch et al., 1986). Subse-
quently, infrared (IR) biospectroscopy was utilised as a tool
to classify bacteria (Naumann et al., 1991) and to show that
different human cells possessed unique infrared spectral
signatures (Wood et al., 1998; Diem et al., 1998). These
biospectroscopic studies demonstrated that different cell
types possessed distinct IR spectral phenotypes, which
reflected the contribution of all the macromolecules
present, including proteins, lipids, carbohydrates, and
nucleic acids.
One of the earliest lineage commitment decisions made
during in vitro differentiation of hESCs is the choice to form
mesoderm and endoderm rather than to commit toward a
“default” ectodermal outcome (Di-Gregorio et al., 2007). In
this study we have evaluated the ability of FTIR spectroscopy
to distinguish between undifferentiated hESCs and cells that
had committed to mesendoderm (precursor cells of the
mesoderm and endoderm germ layers), or were differentiating
to ectoderm (precursor cells of the nervous system and skin).
We demonstrate that their distinct FTIR spectroscopic
“signatures” can be used to characterise and discriminate
hESCs from differentiated progeny, even at very early stages of
development. Moreover, the information that is garnered from
these spectra provides insights into the biochemical and
molecular composition of cells that model very early stages of
human embryonic development. These experiments represent
an important “proof of principle” that indicates the value of
infrared spectroscopy as a rapid and accessible analytical tool
for the characterisation of hESCs and their derivatives.
Results
Human ESCs were differentiated toward mesendoderm in Figure 1 Flow diagram demonstrating the generation of
serum-free medium supplemented with bone morphogenetic cells analysed by FTIR spectroscopy. Undifferentiated
protein (BMP) 4 and activin A (BMP4/Act A medium) or MIXL1GFP/w hESCs were dissociated and cells analysed by
toward ectodermal lineages in medium supplemented with flow cytometry for surface E-CADHERIN and GFP expression
fibroblast growth factor (FGF) 2 (Ng et al., 2008b) (Fig. 1). from the MIXL1 locus, cytospun onto MirrIR-reflective slides
These experiments utilized a genetically modified hESC line for FTIR spectroscopy, or used to set up spin aggregation EBs
expressing green fluorescent protein (GFP) from the MIXL1 in medium supplemented with either BMP4/Act A or FGF2.
locus (MIXL1GFP/w) (Davis et al., 2008), because it provided a After 4–5 days, spin EBs, shown as bright-field (BF) and
means to validate and quantify the differentiation of the fluorescence (GFP) images, were dissociated and cells were
samples subsequently used for the FTIR analysis. The analysed by flow cytometry for GFP and surface E-CADHERIN,
MIXL1GFP/w hESC line enabled objective, quantifiable dis- analysed for hematopoietic colony-forming activity in meth-
crimination between GFP+mesendoderm cells and GFP−ec- ylcellulose, and cytospun onto MirrIR-reflective slides for FTIR
toderm cells and enabled us to confirm that the emergence spectroscopy. Magnification: undifferentiated hESCs × 200;
of MIXL1+cells was dependent on BMP/Act A stimulation. Day 4 spin EBs × 100; hematopoietic colonies × 100.
Undifferentiated hESCs expressed pluripotency-related
genes (including OCT4, REX1, and NANOG), displayed a suite
of stem cell surface antigens (including E-CADHERIN, SSEA-4, similar manner confirmed down-regulation of stem cell
TRA-1-60, and CD9), and did not express GFP from the MIXL1 genes and induction of genes associated with mesendoderm
locus (Fig. 1, Supplementary Fig. 1, and Supplementary including MIXL1, BRACHYURY, and GOOSECOID (Supplemen-
Table 1). The majority of MIXL1GFP/w hESCs differentiated in tary Table 1). Disaggregation and culture of these EBs in
BMP4/Act A medium down-regulated cell surface E-CAD- methylcellulose cultures revealed a high frequency of
HERIN and expressed the mesendoderm marker MIXL1 (74 ± hematopoietic progenitors (means of 174 and 521 colony
9% GFP+cells; mean ± SEM, for 3 experiments analysed by forming cells per 104 cells for triplicate cultures for 2
FTIR spectroscopy) (Fig. 1). Gene expression analysis of experiments analysed by FTIR spectroscopy), consistent
MIXL1GFP/w spin embryoid bodies (EBs) differentiated in a with the induction of mesoderm.
Please cite this article as: Heraud, P., et al., Fourier transform infrared microspectroscopy identifies early lineage commitment in
differentiating human embryonic stem cells, Stem Cell Res. (2010), doi:10.1016/j.scr.2009.11.002
3. ARTICLE IN PRESS
Spectroscopic signature of differentiating hESCs 3
MIXL1GFP/w spin EBs differentiated in FGF2 medium showing the greatest variation and hence most responsible
retained surface E-CADHERIN and did not express MIXL1 for the clustering observed in the scores plots and contrib-
(1.5 ± 0.8% GFP+cells; mean ± SEM, for 3 experiments uting most to the classification results based on the PLS
analysed by FTIR spectroscopy) (Fig. 1). However, there discriminant analysis (Figs. 3B and C and Supplementary Fig.
was down-regulation of other stem cell surface markers in 3). PC1 and PC3 coefficient loadings were very similar
FGF2-stimulated spin EBs, especially CD9 (Supplementary between all three experiments. These showed that differ-
Fig. 1), a reduction in expression of stem cell genes and the ences in bands associated with lipids, proteins, nucleic acids,
induction of early neurectodermal genes, notably neuronal and carbohydrates were responsible for the separation of the
pentraxin 1 (NP1) (Supplementary Table 1). Unlike the BMP4 clusters of the differentiation groups seen in Fig. 3A. Lipid
and activin A-induced MIXL1GFP/w spin EBs, no hematopoietic bands (at 2920, 2850, 1730, 1450, and 1390 cm−1, respec-
colony forming cells were observed from FGF2-stimulated tively) were predominant, corroborating the large difference
spin EBs. in lipid absorbance observed in the average spectra (Fig. 2,
Undifferentiated MIXL1GFP/w hESCs and FGF2-and BMP4/ Figs. 3B and C, and Supplementary Figs. 2 and 3). Regression
Act A-stimulated spin EBs were dissociated into single cells coefficient loadings for protein bands (1700–1500 cm−1)
and cytospun onto MirrIR reflective glass slides. A finely suggested differences in protein secondary structure be-
distributed silver coating, overcoated with tin oxide, on tween the three experimental groups, whereas complex
these slides allowed IR spectra to be acquired in transflection changes in bands associated with nucleic acids and carbohy-
mode (Chalmers and Dent, 2006), effectively doubling the drates (1500–900 cm−1) appeared to be involved in differ-
signal to the detector while still allowing visible microscopic ential clustering of the spectra from the three experimental
examination. An FTIR microscope with a focal plane array groups along both PC1 and PC3 (Figs. 3B and C and
detector (Wood and McNaughton, 2005) simultaneously Supplementary Fig. 3).
acquired 1024 spectra from the cell monolayer. Imaging Considering the distinct clustering observed in the PLS
software automatically selected for high-quality spectra, scores plots, we tested the hypothesis that these early stages
rejecting those acquired from areas of the slide devoid of of hESC differentiation to mesendoderm or ectoderm could
cells or where cells overlapped (Figs. 2A–C). An overlay of be reliably discriminated from undifferentiated hESC and
the quality tested, unprocessed hESC spectra from one from each other using solely FTIR spectra. Two different
experiment is shown in Fig. 2D. types of classification were compared: partial least-squares
Following quality testing, spectra were converted to discriminant analysis (PLS-DA) (Geladi, 1988) and artificial
second derivatives, which eliminated broad baseline slopes neural network (ANN) analysis (Hornik et al., 1989). The first
in the spectra and enabled individual bands that may overlap is based on observation of linear correlations within the data
in the unprocessed spectra to be visualised. Spectra were set, analogous to simple linear regression between two
normalised to correct for differences in sample thickness factors Y and X, except in PLS-DA, X is multivariate and in
(Heraud et al., 2006). the present case represents the absorbance for each
Highly reproducible spectral differences were evident wavenumber value in the spectrum. The PLS2 algorithm
between the three cell populations, with changes indicating was used (Abdi, 2003) with three Y variables coding for class
relatively higher lipid and glycogen levels in the undifferen- membership using nominal values of either+1 or 0, with each
tiated hESCs compared with the differentiated cell popula- Y variable indicating hESC, FGF2, or BMP4/Act A spectra,
tions (Fig. 2E and Supplementary Fig. 2). This is shown in a respectively. Spectral data were randomly divided into
histogram displaying mean integrated areas for the promi- calibration (training) and validation sets comprising two-
nent lipid and glycogen bands at 2920 and 1155 cm−1, thirds and one-third of spectra, respectively. A regression
respectively (Fig. 2F). Bands attributed to proteins and plot of the calibration data set for one experiment yielded a
nucleic acids showed more subtle variation between the correlation coefficient 0.98, indicating that the PLS closely
three experimental groups in band profile shape as well as in modeled the data (Fig. 3D). Y values were then predicted for
total absorbance (Fig. 2E). each spectrum using the separate validation set of spectra
Principal component analysis (PCA) and partial least- (Fig. 3E and Supplementary Fig. 4). Better than 97% of
squares (PLS) (Geladi, 1988) modeling was used to examine spectra from independent validation sets were correctly
the variability of all the FTIR spectra data sets acquired from classified into each of the assigned classes in all experiments
the different cell populations sampled in each experiment. (Table 1).
The first three principal components (PCs) explained more An ANN classification was also examined for its predictive
than 90% of the variance in the 3 experimental datasets in value, because it is based on an entirely different, nonlinear
PCA models. Distinct clustering of spectra from the three cell mode of operation to PLS-DA, thus providing an independent
populations was most clearly visualised in PC1 versus PC2 way of testing the validity of the spectroscopic classification
versus PC3 scores plots from the PLS modeling (Fig. 3 and approach. For the ANN modeling, feed-forward, backpropa-
Supplementary Fig. 3). In all experiments hESC spectra were gation networks were developed. Three neuronal layers
separated from BMP4/Act A and FGF2 spectra along PC1, were found to be optimal with neurons in the input layer
whereas BMP4/Act A and FGF2 spectra were clustered accepting absorbance values from the FTIR spectra, one for
separately along PC3 (Fig. 3A and Supplementary Fig. 3). each spectral absorbance value used (within the ranges
No separation of cluster groups was observed along PC2. 3000–2760 and 1860–900 cm−1), 20 neurons in the hidden
Spectra from replicate samples co-located within each layer, and 3 output neurons for class outputs. The same
respective cluster grouping in all three experiments, which calibration and validation sets used in the PLS-DA analysis
further demonstrated the reproducibility of the results. were employed to allow direct comparison. ANNs correctly
Regression coefficient plots identified the spectral bands classified better than 98% of spectra with independent
Please cite this article as: Heraud, P., et al., Fourier transform infrared microspectroscopy identifies early lineage commitment in
differentiating human embryonic stem cells, Stem Cell Res. (2010), doi:10.1016/j.scr.2009.11.002
4. ARTICLE IN PRESS
4 P. Heraud et al.
validation tests in all experiments. A comparison between more, this study demonstrated that these profiles could be
the accuracy of the PLS-DA and ANN classifications for all used in bioinformatic models to classify independent
three experiments is shown in Table 1. validation sets of spectra correctly in over 97% of cases.
These observations were validated in three independent
experiments that showed consistent differences between
Discussion the spectral patterns of hESCs and their differentiated
progeny. We are expanding our database of spectroscopic
We have shown that undifferentiated hESCs and cells profiles for undifferentiated and differentiated hESCs in
committed to mesendoderm or ectoderm can be readily order to develop a classification system that prospectively
discriminated by their IR spectroscopic profiles. Further- assigns lineage on the basis of FTIR spectra. Thus far, using
Please cite this article as: Heraud, P., et al., Fourier transform infrared microspectroscopy identifies early lineage commitment in
differentiating human embryonic stem cells, Stem Cell Res. (2010), doi:10.1016/j.scr.2009.11.002
5. ARTICLE IN PRESS
Spectroscopic signature of differentiating hESCs 5
our current limited data sets (spectra pooled from experi- and efficiently direct complementary proteomic or lipidomic
ments 1 and 2), undifferentiated hESCs could be prospec- investigations, for example, in order to determine the
tively distinguished from differentiated samples, with 82% of specific protein or lipid components responsible for spectro-
spectra from undifferentiated hESCs in experiment 3 scopic differences.
correctly classified. In conclusion, this study demonstrates that FTIR micro-
Whilst these data were acquired using defined culture spectroscopy, combined with multivariate and artificial
conditions to efficiently direct differentiation toward these neural network modeling, has the potential to provide low-
lineage choices, it is clear from our analyses that there was cost, potentially automatable measurements for the quality
heterogeneity in our populations of undifferentiated and control of stem cell-derived populations to be used in
differentiated hESCs. Future studies, in which the spectral regenerative medicine. This is analogous to the US Food
profiles of more homogeneous flow cytometrically sorted and Drug Administration's recent Process Analytical Tech-
subsets of undifferentiated and differentiated cells will be nologies (PAT) initiative (see http://www.fda.gov/
examined, will enable more accurate identification of cell AboutFDA/CentersOffices/CDER/ucm088828.htm), which
and developmental stage-specific FTIR spectra which we specifies spectroscopic automation of quality control for
anticipate will also improve the accuracy of prospective pharmaceuticals. We envisage that the spectroscopic-based
lineage classification using FTIR spectra. classification system would provide a high-throughput
There are only two previous studies in which FTIR methodology for the verification of cell differentiation
spectroscopy has been applied to stem cells and no studies status, complementing existing approaches provided by
to date have examined hESCs. In one study, the authors expression of genes or cell surface markers.
demonstrated that IR spectroscopy could distinguish be-
tween subpopulations of undifferentiated human mesenchy-
mal stem cells and cells differentiated toward osteogenic Experimental procedures
lineages (Krafft et al., 2007). In a second study, mouse ESCs
were examined at intervals during spontaneous differentia- Embryonic stem cell culture and differentiation
tion in a serum-containing medium (Ami et al., 2008). While
changes in the IR spectra were observed in the heteroge- Undifferentiated hESC adapted to enzymatic passaging
neous cultures in this latter study, they could not be (Costa et al., 2008) were differentiated toward ectodermal
correlated to either a specific inducing stimulus or a or mesendodermal lineages by culture for 4–5 days in serum-
differentiated cell lineage. In contrast, our study examined free APEL medium supplemented with 100 ng/ml FGF2 or
human ES cells differentiated under defined serum-free with 30–40 ng/ml BMP4 and 20 ng/ml Activin A, respectively
conditions in response to specific exogenously added growth (Davis et al., 2008; Ng et al., 2008a). Single cell suspensions
factors. The commitment to specific lineages was validated of undifferentiated or differentiated hESCs were deposited
by evaluation of cell surface markers, gene expression, and by a cyto-centrifuge (Cytospin III, Thermo Fischer Scientific,
the use of a hESC line carrying a mesendoderm-specific Waltham, MA) to produce monolayers of cells on IR reflective
reporter gene. We have shown that commitment to glass slides (MirrIR slides; Tienta Sciences, OH) and then
mesendoderm or ectoderm was associated with specific dried under light vacuum (1 atm less than ambient) in a
infrared spectroscopic “signatures” and that these differed desiccator. Flow cytometry for stem cell surface markers
from the profiles displayed by undifferentiated hESCs. using antibodies directed against human E-CADHERIN
Infrared spectroscopy complements existing approaches (Zymed, San Francisco, CA), TRA-1-60 (Millipore Corpora-
to the classification of stem cells and their differentiated tion, MA), SSEA-4 (Millipore Corporation), and CD9 (BD
progeny because it uses information from a broad range of Biosciences, CA) was performed as previously described
cellular macromolecules, rather than relying on information (Davis et al., 2008; Ng et al., 2008b). Hematopoietic colony
obtained from the expression of single molecules. Further- forming cells were assayed by seeding 104 cells from
more, examination of the spectral loadings that discriminate dissociated d4 EBs into 0.5 ml serum-free methylcellulose
between samples can potentially be directly correlated with cultures (Methocult SF H4236, Stem Cell Technologies,
changes in cellular biochemical components. This informa- Canada) supplemented with 20 ng/ml BMP4 (R & D systems,
tion could be usefully employed in future studies to inform MN), 50 ng/ml vascular endothelial growth factor (VEGF),
Figure 2 Analysis of undifferentiated and differentiated hESCs using FTIR spectroscopy. (A) hESCs cytospun onto a MirrIR-reflective
slide, overlayed with a colored grid showing the area of the slide imaged by the FPA. Each colored pixel represents the area of the slide
where a single FTIR spectrum was acquired. The color scale indicates the absorbance of the amide I protein band, used as an estimation of
total spectral absorbance by the sample. Arrows indicate areas of low absorbance (blue) where there were no cells, or where cells
overlapped, and where absorbance was high (red). Scale bar, 100 μm. (B) The same FPA image grid as shown in (A) at full optical opacity.
(C) Spectral quality testing rejected spectra that were too high or low in absorbance. Black areas indicate where spectra have been
rejected from the dataset, including those areas indicated by the arrows in panel A. (D) The spectral data set for hESCs in one experiment
(n = 192), following quality testing, but prior to spectral preprocessing. The two prominent amide I and II protein bands at ∼1650 and
1540 cm−1, respectively, are indicated. (E) Average spectra from the experiment in panel D, for hESCs (n = 192), cells treated with BMP4/
Act A for 4 days (n = 137), and those with FGF for 4 days (n = 132). Prominent bands in the spectra have been assigned to functional group
vibrations and corresponding macromolecular classes. (F) Histograms showing mean integrated areas for prominent lipid and glycogen
bands (asterisked at 2920 and 1155 cm−1 in panel E) in normalised second-derivative spectra from the experiment in panel D. The
differences between the means of areas for both bands were significantly different between hESCs and differentiated progeny in this
experiment (P b 0.001, by ANOVA). Error bars indicate standard errors of the means (hESC, n = 192; BMP4/Act A, n = 137; FGF2, n = 132).
Please cite this article as: Heraud, P., et al., Fourier transform infrared microspectroscopy identifies early lineage commitment in
differentiating human embryonic stem cells, Stem Cell Res. (2010), doi:10.1016/j.scr.2009.11.002
6. ARTICLE IN PRESS
6 P. Heraud et al.
Please cite this article as: Heraud, P., et al., Fourier transform infrared microspectroscopy identifies early lineage commitment in
differentiating human embryonic stem cells, Stem Cell Res. (2010), doi:10.1016/j.scr.2009.11.002
7. ARTICLE IN PRESS
Spectroscopic signature of differentiating hESCs 7
Table 1 The percentage correct classification for spectra each spectrum in the data array sampled ∼11 × 11 μm of the
in independent validation sets drawn from each of the three cell monolayer. The acquisition of each FTIR spectral image
treatments (BMP4 and Activin A, FGF2, or hESCs), using data array was completed in about 2 min. The Varian system
either partial least-squares discriminant analysis (PLS-DA) was controlled by an IBM-compatible computer running Win
or artificial neural network analysis (ANN), for three IR Pro 3.0 software (Varian Inc.). The absorbance spectra
independent experiments (1–3) were acquired in reflectance mode at a spectral resolution of
8 cm−1 with 64 scans coadded. Apodization was performed
Treatment Experiment PLS-DA ANN
using a Happ-Genzel function.
BMP4 and Activin A 1 100 100
2 100 98 Spectral imaging selection and preprocessing of
3 100 100
spectral data
FGF2 1 98 98
2 98 98
Images based on total protein absorbance were interrogated
3 97 100
using Cytospec 1.03 IR imaging software (Cytospec Inc., NY,
hESCs 1 98 98
USA) to select spectra from sample regions with absorbances
2 98 99
within the linear range of the detector (absorbance 0.2 –
3 100 100
0.8), with spectra extracted from the data array and
exported to Unscrambler 9.2 (Camo, Oslo, Norway) for
statistical analysis and classification. Spectra were prepro-
cessed by calculating second derivatives (Savitsky-Golay
100 ng/ml stem cell factor (SCF), 30 ng/ml interleukin (IL)-3, with 9 point smoothing), hence minimizing the effects of
30 ng/ml thrombopoietin (TPO), and 3 units/ml erythropoi- variable baselines and allowing individual bands that were
etin (EPO) (all from Peprotech Inc. Haifa) (Ng et al., 2008b). overlapping in the raw spectra to be readily visualised as
Colonies were counted after 10–12 days. cDNA generated separate minima bands. Spectra were then normalized to
from undifferentiated hESC and d4 FGF2-and BMP4/Act A- account for possible differences in sample thickness using
stimulated spin EBs was hybridised to Illumina whole genome Extended Multiplicative Signal Correction (EMSC) (Martens
expression bead chips (Illumina Inc., CA) and results analysed et al., 2003).
using Bead Studio (Illumina Inc.) and GeneSpring GX (Agilent
Technologies, USA) software.
Statistical analysis
Infrared microspectroscopy Principal component analysis and partial least-squares
discriminant analysis
Analyses were performed on replicate cultures of undiffer- PCA was performed with The Unscrambler 9.2 using mean
entiated hESCs and compared with analyses performed centered data and employing full cross-validation, with any
simultaneously on hESCs differentiated for 4–5 days toward data outliers identified and removed from the dataset by the
ectodermal lineages or toward mesendodermal precursors. examination of residual influence plots. Scores plots were
Infrared spectral images were acquired from an area used to visualise any clustering of data, and loadings plots
352 × 352 μm on the sample plane. Each image consisted of used to determine which spectral region most contributed to
a 32 × 32 array of 1024 spectra, resulting from binning of the the variance in the dataset. PLS-DA employed PCA models,
signal from each square of 4 detectors on a liquid N2 cooled with the optimal number of PCs chosen to achieve the lowest
64 × 64 element MCT Stingray (Varian Inc., Paolo Alto, CA, residual variance. Spectral data were randomly divided into
USA) focal plane array detector, which was coupled to an calibration and validation sets comprising two-thirds and
infrared microscope (Model 600 UMA, Varian Inc.) and a FTIR one-third of spectra, respectively. The PLS2 training data
spectrometer (Model FTS 7000, Varian Inc.). Accordingly, matrix comprised the spectra (multivariate X) and three Y
Figure 3 (A) Scores plots for PLS analysis of the spectral data for the three treatment groups from one experiment showing replicate
samples for each cell type. Each spectrum is represented as a point in PC1 vs PC2 vs PC3 space. (B) Regression coefficient plots used to
explain the clustering observed in the PLS scores plot shown in panel A. PC1 regression coefficient loadings indicated spectral
differences explaining clustering along PC1, which separates hESCS from the cells differentiated in FGF2 or BMP4/Act A. PC3
regression coefficient loadings explained the separation of the spectra from FGF2-and BMP4/Act A-treated sample groups along PC3.
Limited spread in the PC2 direction did not contribute significantly to the clustering of samples. (C) Table of band assignments for the
mean spectra shown in Fig. 2E compared with the associated PC1 and PC3 regression coefficient loadings from Fig. 3B. Arrows for PC1
loadings indicate up-or down-regulation of the macromolecular component in the change from the hESCs to the differentiated progeny
in the PC1 direction. Arrows for PC3 regression coefficient loadings indicate up-or down-regulation of the macromolecular component
in the change from the FGF2 state to the BMP4/Act A state in the PC3 direction. The references for band assignments are described in
full in the Supplementary References. (D) Partial least-squares (PLS) regression plot for the calibration (training) set in one
experiment, which plotted measured Y against predicted Y (hESCs = +1; FGF2 = 0; BMP4/Act A = 0). The correlation coefficient (0.98)
indicated that the training set was well modeled using PLS analysis. (E) Predicted Y values for hESCs using PLS discriminant analysis
(PLS-DA) for the validation set for the experiment in panel D, shown as box plots (narrow blue rectangles) that indicate the 95%
confidence interval around each predicted Y (center of rectangle). All spectra were classified correctly as hESC using PLS-DA. Spectra
from the FGF2 or BMP4 and Activin A classes were also accurately classified (see Supplementary Fig. 4).
Please cite this article as: Heraud, P., et al., Fourier transform infrared microspectroscopy identifies early lineage commitment in
differentiating human embryonic stem cells, Stem Cell Res. (2010), doi:10.1016/j.scr.2009.11.002
8. ARTICLE IN PRESS
8 P. Heraud et al.
variables with integer values of 0 or+1 coding for the three Costa, M., Sourris, K., Hatzistavrou, T., Elefanty, A.G., Stanley, E.G.,
modeled spectral classes (hESC, FGF2, BMP4/ActA, respec- 2008. Expansion of human embryonic stem cells in vitro. Curr.
tively) (Abdi, 2003). PLS discriminant analysis used the Protoc. Stem Cell Biol. Chap. 1 1C.1.1–1C.1.7.
Davis, R.P., Ng, E.S., Costa, M., Mossman, A.K., Sourris, K., Elefanty,
spectral ranges containing bands of biological origin between
A.G., Stanley, E.G., 2008. Targeting a GFP reporter gene to the
3000–2760 and 1860–900 cm−1.
MIXL1 locus of human embryonic stem cells identifies human
primitive streak-like cells and enables isolation of primitive
Artificial neural network classification hematopoietic precursors. Blood 111, 1876–1884.
ANNs (Neurodeveloper 3.4, Synthon GmbH, Heidelberg, Diem, M., Boydston-White, S., Chiriboga, L., 1998. Infrared
Germany ) used to classify individual experiments employed spectroscopy of living cells: shining light onto a novel subject.
the same training and validation sets as those used with PLS- Appl. Spectrosc. 53, 148A–161A.
DA to allow direct comparison of classification outcomes. A Di-Gregorio, A., Sancho, M., Stuckey, D.W., Crompton, L.A.,
three-layer feed-forward backpropagation design was used, Godwin, J., Mishina, Y., Rodriguez, T.A., 2007. BMP signalling
where the number of input neurons corresponded to the inhibits premature neural differentiation in the mouse embryo.
Development 134, 3359–3369.
number of absorbance measurements in the same wavenum-
Geladi, P., 1988. Notes on the history and nature of partial least
ber range used in the PLS discriminant analysis (within the squares (PLS) modelling. J. Chemometrics 2, 231–246.
ranges 3000–2760 and 1860–900 cm-1), and the output Heraud, P., Beardall, J., Wood, B.R., McNaughton, D., 2006. Effects
neurons to the groups being classified. The number of of pre-processing of Raman spectra on in vivo classification of
neurons in the hidden layer was varied by trial and error nutrient status of microalgal cells. J. Chemometrics 20, 193–197.
based on the root mean square errors of calibration and Hornik, K., Stinchcombe, M., White, H., 1989. Multilayer feedfor-
validation (RMSEC and RMSEV) achieved and the efficacy of ward networks are universal approximators. Neural Netw. 2,
the classification accuracy of the resulting ANN. The 359–366.
calibration process employed an internal validation set and Krafft, C., Salzer, R., Seitz, S., Ern, C., Schieker, M., 2007.
the number of iterations used to establish a model was Differentiation of individual human mesenchymal stem cells
probed by FT-IR microscopic imaging. Analyst 132, 647–653.
chosen by minimizing the RSMEC.
Mantsch, H.H., Yang, P.W., Casal, H.L., 1986. Infrared spectrometry
of living systems: current trends and perspectives. J. Mol. Struct.
Acknowledgments 141, 237–242.
Martens, H., Nielsen, J.P., Engelsen, S.B., 2003. Light scattering and
light absorbance separated by extended multiplicative signal
This work was supported by the Australian Stem Cell Centre,
correction: application to near infrared transmission analysis of
the Juvenile Diabetes Research Foundation, and the National powder mixtures. Anal. Chem. 75, 394–404.
Health and Medical Research Council (NHMRC) of Australia. Murry, C.E., Keller, G., 2008. Differentiation of embryonic stem cells
A.G.E. is a Senior Research Fellow of the NHMRC. to clinically relevant populations: lessons from embryonic
development. Cell 132, 661–680.
Naumann, D., Helman, D., Labischinski, H., 1991. Microbiological
Appendix A. Supplementary data characterizations by FT-IR spectroscopy. Nature 351, 81–82.
Ng, E.S., Davis, R., Stanley, E.G., Elefanty, A.G., 2008a. A protocol
Supplementary data associated with this article can be found, describing the use of a recombinant protein-based, animal
in the online version, at doi:10.1016/j.scr.2009.11.002. product free medium (APEL) for human embryonic stem cell
differentiation as spin embryoid bodies. Nat. Protoc. 3,
768–776.
References Ng, E.S., Davis, R.P., Hatzistavrou, T., Stanley, E.G., Elefanty, A.G.,
2008b. Directed differentiation of human embryonic stem cells as
Abdi, H., 2003. PLS-regression: multivariate analysis. In: Lewis-Beck, M., spin embryoid bodies and a description of the hematopoietic
Bryman, A., Futing, T. (Eds.), Encyclopedia for Research Methods blast colony forming assay. Curr. Protoc. Stem Cell Biol. Chap. 1
for the Social Sciences. Sage, Thousand Oaks, CA. 1D.3.1–1D.3.23.
Ami, I., Neri, T., Natalello, A., Mereghetti, P., Doglia, S.M., Zanoni, Wood, B.R., McNaughton, D., 2005. FPA imaging and spectroscopy
M., Zuccotti, M., Garagna, S., Redi, C.A., 2008. Embryonic stem for monitoring chemical changes in tissue. In: Bhargave, R.,
cell differentiation studied by FT-IR spectroscopy. Biochim. Levin, I. (Eds.), Spectrochemical Analyses Using Multichannel
Biophys. Acta Mol. Cell Res. 1783, 98–106. Infrared Detectors. Blackwell, Oxford.
Chalmers, J.M., Dent, G., 2006. Vibrational spectroscopic methods Wood, B.R., Quinn, M.A., Tait, B., Ashdown, M., Hislop, T., Romeo,
in pharmaceutical solid-state characterization. In: Hilfiker, R. M., McNaughton, D., 1998. FT-IR microspectroscopic study of cell
(Ed.), Polymorphism in the Pharmaceutical Industry. Wiley-VCH, types and potential confounding variables in screening for
Weinheim, Germany. cervical malignancies. Biospectroscopy 4, 75–91.
Please cite this article as: Heraud, P., et al., Fourier transform infrared microspectroscopy identifies early lineage commitment in
differentiating human embryonic stem cells, Stem Cell Res. (2010), doi:10.1016/j.scr.2009.11.002