Biology Reference
In-Depth Information
unwanted intersample variance and/or address
the large dynamic range of metabolite concentra-
tions, were recently evaluated by Kohl et al.
169
Eleven normalization methods were assessed:
normalization to creatinine, cubic spline,
contrast, cyclic loess, linear baseline, Li-Wong,
probabilistic quotient, quantile, variance stabili-
zation, auto scaling, and pareto scaling. A quick
overview of the logic behind each technique
was also presented. Two sample sets were used
for testing: (1) eight pooled human urine samples
spiked with eight endogenous metabolites
following a Latin square design and (2) human
urine from 54 autosomal polycystic kidney
disease patients and 46 healthy volunteers. The
NMR spectra were binned using a constant bin
width of 0.01 ppm. The authors concluded that
quantile normalization was the only method
that performed well for all tests and was recom-
mended for datasets with more than 50 samples.
For smaller datasets, cubic spline normalization
was suggested.
Natural metabolite concentration does not
necessarily directly translate to biological impor-
tance; therefore, scaling is usually performed in
addition to data normalization to speci
(PLSDA), and orthogonal projections to latent
structures (OPLS). The reader is referred to three
excellent reviews on multivariate statistical anal-
ysis of NMR data by Ebbels and Cavill, Trygg
et al., and Lindon et al.
172
e
174
PCA is particularly
useful for detecting outliers. Group separation is
not always observable in PCA models because
changes due to various pathological states may
be small compared to other intra- and intersample
variations. Supervised methods such as PLSDA
and OPLS are employed in such cases. Statistical
interpretation is simpli
ed in OPLS models,
where contributions to group separation are
located in the very
rst component.
175
NMR METABOLITE
IDENTIFICATION
s biggest advantage compared to other
metabolomics platforms is the ability to identify
and con
NMR
'
rm metabolites in a mixture. Different
functional groups have very characteristic
NMR frequencies, which can be used to identify
components of an unknown metabolite. NMR
frequency ranges for numerous functional
groups have been tabulated in texts such
as Pretsch et al.
176
and Silverstein et al.
177
and the web page of the chemistry depart-
ment of the University of Wisconsin
e
Madison
index-chem.htm
. Complete structure determina-
tion by NMR is based on characteristic NMR
chemical shifts, scalar couplings from peak
multiplicity patterns, molecular connectivity,
and spatial information obtainable from a suite
of 1D and 2D experiments. For the interested
reader, Berger and Braun contains more than
200 NMR experiments for small molecule NMR
spectroscopy.
178
A quick introduction with prac-
tical details on how NMR experiments can be
used to determine skeletal connectivity, relative
stereochemistry, and structure veri
cally
address this issue and to prevent bias towards
changes in abundant metabolites in subsequent
statistical modeling. Scaling techniques include
centering, autoscaling, pareto, range scaling,
vast scaling, level scaling, log transformation,
and power transformation. These scaling tech-
niques are explained in the work of van den
Berg et al.
170
The free, web-based program
MetaboAnalyst 2.0 provides
comprehensive
choices for data
filtering, normalization, scaling,
statistical modeling, and time-series and
pathway analysis capabilities.
171
Local installa-
tion of the program is also possible.
Multivariate Statistical Analysis
The commonly used multivariate statistical
models are principal component analysis
(PCA), partial least square discriminate analysis
cation can
be found in the mini-review by Kwan and
Huang.
179
For a more detailed explanation, the
Search WWH ::
Custom Search