Extensible and Executable Stochastic Models

of Fatty Acid and Lipid Metabolism

Argyris Zardilis1,2,5 , Joao Dias3,4 , Animesh Acharjee5 , and James Smith1,4,5
1

Cambridge Computational Biology Institute, Department of Applied Mathematics

and Theoretical Physics, University of Cambridge, U.K.
2
SynthSys, C.H. Waddington Building, University of Edinburgh, U.K.
3
Cambridge Systems Biology Centre, University of Cambridge, U.K.
4
Department of Biochemistry, University of Cambridge, U.K.
MRC Human Nutrition Research, Elsie Widdowson Laboratory, Cambridge, U.K.
A.Zardilis@sms.ed.ac.uk, James.Smith@mrc-hnr.cam.ac.uk

Abstract. Stochastic reaction-centric views are suitable for exploring

hybrid minimal mechanism-statistical models of fatty acid and lipid
metabolism, the basis of de novo lipogenesis. In this work, we demonstrate a reduced model for the core fatty acid synthesis and elongation
process with a regulatory mechanism. This allows us to explore fatty acid
profiles from lipid metabolomics data. This is part of a current study to
assess the programming languages for capturing inherent probabilistic
behaviour of the hierarchical chemical transformations of complex lipid
species.

Introduction

Understanding lipid metabolism and its regulation is essential in metabolic disorders and diabetes-related diseases. Many of the metabolites and transformation
reactions involved are still poorly characterised. Because of the low numbers of
carbons owing through the lipid pathways, the eect of the inherent probabilistic nature of chemical events is amplied. These two factors make lipid pathways
dicult to analyse with current modelling approaches used for metabolic processes, such as constraint-based analyses (including FBA) [1], which is focused
on deterministic average-case behaviour (of population growth). Here, we use
an alternative stochastic and reaction-centric view to capture these pathways.
In particular, we focused on Fatty Acid (FA) synthesis and elongation which is
central in lipid metabolism. FAs are the core building blocks, modied for more
complex lipids within the cell and tissues. FA synthesis and elongation pathways
in particular display some characteristics that make models benet from this alternative view: local iterative processes, probabilistic decisions at dierent levels
and between pathway control mechanisms that aect decision making. We hope
the eventual models will be useful for examining metabolomics data from model
organisms, clinical trials or large-scale epidemiological cohort studies.
P. Mendes et al. (Eds.): CMSB 2014, LNBI 8859, pp. 244247, 2014.
c Springer International Publishing Switzerland 2014


Extensible and Executable Stochastic Models of FA and Lipid Metabolism

245

Methods

An important aspect of this work was to assess possible languages for capturing this stochastic behaviour in a reaction-centric projection. Recently, there
has been a general trend towards constructing executable models, from the distributed systems world [2]. Here, we use Petri Nets [3] mainly and pi-calculus
(SPiM [4] variant in particular) which is an example of a Process Algebra [5].
Petri Nets provide a vivid and intuitive graphical notation with a natural correspondence to chemical reactions and they have been used before for other
metabolic pathways [6]. The main unit of denition for Petri Nets is the transition. The main unit of denition for Process Algebra is the species but the
operational semantics are in terms of interactions, which again make them suitable for our reaction-centric view. Since these languages have been designed to
handle distributed systems and therefore concurrency, non-determinism representing decisions is inherent in the structure of Petri Nets and in the syntax of
pi-calculus. FA biosynthesis and elongation was taken from a reference annotation (KEGG hsa00062) with source and sink metabolites. Sinks were dened
as the Cn :0, even-chain FAs. Elongation was combined with multiple steps in
synthesis ignoring transport processes, under the assumption of non-reversable
(net-forward) eects and constant reaction rates.

Results

We modelled the iterative FA elongation process as the combined eect of synthesis and elongation pathways, reduced down to a series of binary decisions or
Bernoulli trials. An FA under elongation at each point makes a binary decision
of whether to stay at its current length or continue to form longer FAs. This
decision can be captured in the Petri Net language as a race condition between
two enabled transitions. According to the operational semantics of the system
the probabilities of the decision outcomes are controlled by the rates of the reactions corresponding to the transitions. The entire process can be seen as a
series of binary decisions. An FA starts its journey in the net and moves along
making the decisions along the way before getting trapped in one of sinks of the
net that represents an FA reaching its nal length (Figure 1).
We also modelled the Acetyl-CoA ow decision between the Krebs cycle and
FA synthesis, controlled by the immediate energy requirements of the cell. Again
this decision is captured very naturally as a race condition between two enabled
transitions, the transition taking an Acetyl-CoA molecule to the rst step of
FA synthesis and the transition taking it towards the TCA cycle. The other
pathways involved in this process were not modelled explicitly. Instead all the
involved reactions were grouped into a single Petri Net transition respecting the
stoichiometric constraints of the pathway. In this case, the transition rates are
functions of ATP to display this control mechanism. The strength of the change
in the likelihoods of the two outcomes can be captured by an exponent parameter
on the two corresponding rates (Figure 2).

p5
1

p4
1

p2
1

t1

t2

p1

t2

p3

A. Zardilis et al.

246

t1
p1
Cn

C12

Cn+2

p2

p3

C14

p4

C16

C18

p5
C20

1
C22

Fig. 1. On the left, the decision taken at a specific point by a metabolite during the
elongation process. On the right, the entire process can then be seen as a series of
binary decisions that the FA takes whether to stay at the current length or continue
for further elongation.

k
ATPb
TCA flow

Acetyl-CoA

ATP k
FA flow

Fig. 2. Depending on the energy requirements of the cell, for which ATP levels act as a
proxy, Acetyl-CoA can either go towards the TCA cycle to produce energy or towards
FA synthesis. This decision is captured by a race condition. The likelihoods of the two
outcomes are functions of ATP naturally. As ATP from the TCA feeds to the FA flow,
this is described as a feed-forward motif (FFM).

The models were tuned with real experimental data, donated from a GC-FID
mouse adipose metabolomics study. Relative metabolite abundances was used
to parameterise the output states and therefore sample the relevant posteriors.
Since the entire FA synthesis and elongation process can be seen as a series of
Bernoulli trials, the only parameters of the model are the success probabilities
for the decisions. A Maximum Likelihood estimation of these can be done from
the data by recognising that the number of successes is the number of the corresponding species (for example C4 for the rst decision) and the number of
trials the number of the corresponding species and the sum of the numbers of
all longer species.

Extensible and Executable Stochastic Models of FA and Lipid Metabolism

247

Non-parametric Dirichlet process mixture model clustering was used to partition the mouse adipose metabolite data, giving four clusters representing four
distinct metabolic states and drug-dose treatment conditions. Ratios of evenchain FAs in each cluster were the posteriors per metabolic state for the models.
Clustering of metabolic data adds an interesting dimension to the study as we can
observe changes in the model parameters across dierent conditions and identify
mechanistic detail or topological structures that show dynamical changes under
perturbation.

Discussion and Conclusion

Reaction-centric network projections have advantages in model building. Pathways can be reduced to single composite transitions leaving interface metabolites (e.g. in the TCA cycle). To remain biochemically valid, stochastic statistical
models are also hybrid, retaining mechanistic detail that can be parameterised to
discriminate between metabolic states. Posteriors for model extensions are simply the ratios of additional metabolites taken from each metabolic state. In the
next extensions to the current models, the additions could be parameterised from
odd-chain, unsaturated or even combinations of these in the complex lipid species
such as DAGs and phospholipids, sphingolipids and TAGs. Metabolic processes
such as degradation of FAs and the anabolism and catabolism of more complex
lipids can be added as extension modules to our core models. It is important that
phenomenological-statistical models for metabolic processes retain some essential mechanistic detail that allow them to discriminate between metabolic states
in health and disease. The impact to the lipid metabolomics community is that
these models allow a re-examination of lipid metabolite proles in large scale
epidemiological data. Lipid metabolites proled using extraction and separation
of complex lipid species can be compared to model predictions of endogenous
metabolic intermediates such as FAs in dierent metabolic states.

References
1. Orth, J.D., Thiele, I., Palsson, B.: What is flux balance analysis? Nature Biotechnology 28(3), 245248 (2010)
2. Fisher, J., Henzinger, T.A.: Executable cell biology. Nature Biotechnology 25(11),
12391249 (2007)
3. Murata, T.: Petri nets: Properties, analysis and applications. Proceedings of the
IEEE 77(4), 541580 (1989)
4. Phillips, A., Cardelli, L., Castagna, G.: A graphical representation for biological
processes in the stochastic pi-calculus. In: Priami, C., Ing
olfsd
ottir, A., Mishra, B.,
Riis Nielson, H. (eds.) Transactions on Computational Systems Biology VII. LNCS
(LNBI), vol. 4230, pp. 123152. Springer, Heidelberg (2006)
5. Fokkink, W.: Introduction to Process Algebra. Computer Science Monograph
(English), 2nd edn. Springer, Berlin (2007)
6. Baldan, P., Cocco, N., Marin, A., Simeoni, M.: Petri Nets for Modelling Metabolic
Pathways: a Survey. J. Natural Computing (9), 955989 (2010)