MS Metabolomics Core -BCM 2014
Metabolomics Tier 1 Data Analysis
Targeted Mass Spectrometry: For MRM data acquired using Agilent 6430 or 6490
QQQ, data handling will be performed using both Qualitative and Quantitative Analysis
software from Agilent Technologies, that will ascertain retention time, verify MRM
transitions and calculate absolute levels of metabolites using calibration curves.
Low-level analysis. The initial low level analysis of metabolomic data will use a series
of steps including preliminary review of data, visualization for spotting patterns, data
cleaning,
and
Data
imputation
normalization.
cleaning
basic
and
statistical
analyses will include
identifying
potential
outliers, checking for
normality,
examining
the proportion and/or
variance
for
each
variable.
As samples will be obtained from multiple data sources (Bruker vs. Varian NMRs) or will
be acquired in batches, this descriptive analysis will first be conducted stratifying by
data source and batch to determine if there are any important site differences that will
need to be adjusted for, during the data normalization steps.
For mass spec data, the mode of acquisition (targeted vs. unbiased) will impact the
imputation procedure employed. For targeted acquisition, it is not expected the data to
have any missing values. For unbiased acquisition, missing values are a common
feature. In that case, only metabolites with fewer than 50% missing values for studies
involving a relatively large number of tissues/biofluids will be considered, and with fewer
than 20% missing values in the case of cell lines. Missing values will be imputed either
at the minimum detection level or through a k=5 nearest neighbor procedure (pamr
MS Metabolomics Core -BCM 2014
package in the R programming language). Depending on the study design several
different approaches will be available, ranging from simple median centering, to
centering and scaling based on the values of internally spiked standards, to employing
more advanced fixed effects analysis of variance procedure that use factors, data
platform, batch information, ionization mode, etc.
For samples run on days that are fairly apart, batch effects occur that need to be
corrected. To correct such batch effects, we either use analysis of variance techniques,
or when the focus is on unsupervised techniques such as clustering and principal
components analysis, the function “removeBatchEffect” available in the R package
“limma” is employed.
Standardization of Isotopic Enrichment data and Determination of Steady State:
For every substrate or metabolite, a standard curve of known isotopic enrichments
versus measured enrichments is first performed. A Regression analysis is performed
and raw isotopic data first corrected based on the slope and intercept of the curve when
R2=0.99. To calculate flux from the corrected data, steady state isotopic enrichment
(plateau) has to be determined by performing a regression analysis of enrichment
versus time to ascertain whether slope is not different from zero. If the slope is greater
than zero, a non-linear regression is performed to estimate the asymptote (plateau).
Normalization of Biolog metabolic phenotyping data: First the data from these metabolic
phenotyping assays will be corrected for background signal, obtained from the median
reading of the three empty wells on the plate. Following this, the data will be scaled by
dividing each measurement on the plate by the inter-quartile range of the data. In our
experience, this simple normalization has worked well even for samples that have been
run on different days or belong to different conditions (control vs. treatment).
Mid-level analysis (Figure 1): These types of analysis involve identification of
differentially expressed metabolites, model building for classificatory or survival analysis
purposes, dimension reduction for extracting broad patterns from the data and
identification of groups of samples and/or metabolites.
MS Metabolomics Core -BCM 2014
Specifically, differentially expressed compounds across two classes will be identified
using parametric (t-tests) and non-parametric (rank-sum) tests, while for multiple
classes we will employ analysis of variance models. The latter models in addition to the
key treatment factors being tested also allow for the incorporation of key covariate
information, such as clinical (stage of the disease, indices of physiological impairments),
as well as demographic and health habits of the samples (e.g. age, race, gender,
education, smoking and drinking status in the case of humans and strain, gender and
housing conditions in the case of mice).
Given the large number of markers that are likely to be identified as significantly
different between groups, as well as the number of conditions and differences in any
given experiment, we recognize that these multiple comparisons increase the possibility
of Type I error (false positives). Hence, family-wise error rate (FWER) methods and
false discovery rate (FDR) methods would be employed to reduce or eliminate false
positives.
In many studies, classificatory and/or prognostic models will be built. Such models are
important for delineating metabolomics signatures associated with clinical outcomes,
including disease/normal status, recurrence times, stage progression, etc. For
categorical outcomes (e.g. disease/normal status, etc.) there are several standard
models in the machine learning literature that will be employed, including logistic
regression, random forests and support vector machines, whereas for outcomes
capturing event times (e.g. disease recurrence, survival, etc) Cox-proportional hazards
models will be used. An important aspect of this modeling will be to enforce sparsity
through penalization (e.g. lasso or group lasso penalties) that lead to more
parsimonious models that exhibit good theoretical properties in terms of inference and
predictive ability. Especially in the case of structured penalties (e.g. group lasso that is
implemented in the R package grplasso), one can impose a priori biological information,
such as pathway structure. The performance of these classificatory and prognostic
models will be assessed through K-fold cross-validated error rates, and through
receiver-operator characteristic (ROC) curve, while the area under the curve (AUC) will
be used as an overall measure of model fit. The significance of the AUC metric for each
MS Metabolomics Core -BCM 2014
fitted model will be assessed through the Mann-Whitney U-statistic and it will also be
used to select between competing models.
Finally, depending on project needs the core will also provide several other standard
analyses to users to understand and gain insight into their data including dimension
reduction techniques, such as principal components analysis and penalized (for
sparsity) variants for obtaining more robust low-dimensional representations of the
samples and/or the metabolites, clustering of samples and/or metabolites into groups
using a wide range of algorithms (hierarchical, model-based, partition such as k-means
and robust variants and graph based ones, such as normalized cuts). In addition, it will
provide enhanced visualization capabilities by mapping results into pathways (see also
Section on pathway mapping).