(419a) Metabolomic Modularity Analysis (MMA) to Quantify Human Liver Perfusion Dynamics

Biological cells are complex dynamic systems, comprising
thousands of interacting nucleic acids, proteins, and metabolites acting in
concert to perform coordinated and well regulated tasks. The ability to effectively
quantify system-wide cellular
responses to various stimuli is paramount to our understanding of complex disease-states
and how pharmacological agents elicit desired outcomes. Investigators are therefore
increasingly relying on large scale ?omics data (transcriptomics,
proteomics, metabolomics) to experimentally capture changes
in signaling and metabolic pathway components. While the burden of data
acquisition has lightened with improved instrumentation, data interpretation in
the proper biological context remains challenging. Metabolomics data, in
particular, are difficult to interpret in the context of pathways because a
single metabolite may be utilized by several functional modules.  Both univariate and multivariate statistical tools
can determine which metabolites explain a differential response among
experimental treatment groups. However, these data-driven methods do not
utilize the vast domain-specific biochemical knowledge of known stoichiometric
and regulatory interactions that can be used to systematically uncover which
signaling or metabolic pathways are activated by a perturbation. To this end,
mapping metabolomics data onto well curated metabolic graph networks can help
identify interactions that may not be intuitive from simply observing the
metabolic pathway maps offered by the KEGG database.  

In this study, we introduce Metabolomic
Modularity Analysis (MMA) as a graph-based algorithm to systematically identify
modules of reactions enriched with metabolites flagged to be statistically significant
based on univariate analysis. Briefly, a metabolic network is abstracted as a
reaction-centric graph network where a non-directional edge is drawn between
two reactions if a metabolite produced by the first reaction is consumed by the
second. The length of the edge between reactions is weighted such that if the
two reactions are involved in metabolites that are statistically significant, their
edge distance is shorter. When applying Newman's hierarchical partition
algorithm based on relative reaction-pair shortest paths, the resulting modules
feature reactions from the surrounding local topological neighborhood, which
favors adjacent reaction pairs that involve statistically significant
metabolites. A defining feature of determining reaction-centric modularity is
that interactions between reactions mediated by the production and consumption
of cofactors and other hub metabolites are also accounted for.

We apply MMA on time-course metabolomics data collected from
biopsies during subnormothermic machine perfusion (SNMP)
of nine discarded human livers that were rejected for transplantation. These
livers had endured various degrees of warm ischemia time (WIT) prior to organ
procurement and three of them exhibited over 30% macrovesicular
steatosis. Of the 155 primary metabolites measured,
the pre-perfusion levels (t=0 hours) of 33 of them were found to be
significantly correlated with WIT, suggesting that they could be putative
biomarkers for ischemic injury. MMA was performed on the human ReconX metabolic network (7439 reactions and 2626 metabolites)
after flagging these 33 metabolites in the cytosolic compartment as being
significant. For a network this large, computing the adjacency distance matrix
requires ₇₄₃₉C₂, or 2.77 ·10⁷ computations for
the initial network, which causes total run times on the order of days with a
laptop, rendering the method impractical for repeated analysis. Fortunately, Mathworks® now offers Matlab's
Distributed Computing Server (MDCS) on Amazon's Elastic Compute Cloud (EC2). For
this study, one cluster node (c3.8xlarge) with 16 workers was rented at a rate
of $1.68/hour for the cluster and $1.12/hour for MDCS. Several key parts of the
MMA algorithm are parallelizable and was therefore able to complete a full run
in 2.7 hours, with a total cost of $8.40 for the 3 hours of server time. The MMA
partitioning resulted in 4755 hierarchical modules, of which 223 contain at
least one significant metabolite. To highlight an example, one such module contained
four metabolites significantly correlated with WIT; arachidic
acid (arach[c]), cholesterol (chsterol[c]),
stearic acid (ocdca[c]), and palmitic
acid (hdca[c]). The stoichiometry of reactions in the
module suggests that ischemia may have an impact on coenzyme A (coA[c])
pools as a cause for why these four metabolite levels are affected (Figure 1).
Ongoing work involves applying MMA to identify modules activated by perfusion
itself by flagging metabolites significantly different between various time
points and determine how ischemia or fat content affects the modules
uncovered.    

In this study, we demonstrate that graph network-based
analytics to quantify metabolomics data is made practical and feasible using
parallel computing on Amazon's cloud. Prospectively, similar graph-based
algorithms can be employed to analyze proteomics and transcriptomics
data as well and ultimately devise ways to incorporate all information to best
characterize cellular dynamics. 

Figure 1 Example
of module identified using MMA. Metabolite nodes are represented as ellipses,
while reaction nodes are represented as rectangles. Metabolite nodes colored in
red are those whose pre-perfusion levels are correlated to WIT (p<0.10).