Fast reconsonstruction of compact context-specific network models

Recent progress in high-throughput data acquisition has shifted the focus from data generation
to the processing and understanding of now easily collected patient-specific information.
Metabolic models, which have already proven to be very powerful for the integration and analysis
of such data sets, might be successfully applied in precision medicine in the near future.
Context-specific reconstructions extracted from generic genome-scale models like Reconstruction
X (ReconX) (Duarte et al., 2007; Thiele et al., 2013) or Human Metabolic Reconstruction
(HMR) (Agren et al., 2012; Mardinoglu et al., 2014a) thereby have the potential to become a
diagnostic and treatment tool tailored to the analysis of specific groups of individuals. The use
of computational algorithms as a tool for the routinely diagnosis and analysis of metabolic diseases
requires a high level of predictive power, robustness and sensitivity. Although multiple
context-specific reconstruction algorithms were published in the last ten years, only a fraction
of them is suitable for model building based on human high-throughput data. Beside other
reasons, this might be due to problems arising from the limitation to only one metabolic target
function or arbitrary thresholding.
The aim of this thesis was to create a family of robust and fast algorithms for the building
of context-specific models that could be used for the integration of different types of omics
data and which should be sensitive enough to be used in the framework of precision medicine.
FASTCORE (Vlassis et al., 2014), which was developed in the frame of this thesis
is among the first context-specific building algorithms that do not optimize for a biological function
and that has a computational time around seconds. Furthermore, FASTCORE is devoid of
heuristic parameter settings. FASTCORE requires as input a set of reactions that are known to
be active in the context of interest (core reactions) and a genome-scale reconstruction. FASTCORE
uses an approximation of the cardinality function to force the core set of reactions to
carry a flux above a threshold. Then an L1-minimization is applied to penalize the activation of reactions with low confidence level while still constraining the set of core reactions to carry a
flux. The rationale behind FASTCORE is to reconstruct a compact consistent (all the reactions
of the model have the potential to carry non zero-flux) output model that contains all the core
reactions and a small number of non-core reactions.
Then, in order to cope with the non-negligible amount of noise that impede direct comparison
within genes, FASTCORE was extended to the FASTCORMICS workflow (Pires Pacheco
and Sauter, 2014; Pires Pacheco et al., 2015a) for the building of models via the integration
of microarray data . FASTCORMICS was applied to reveal control points regulated
by genes under high regulatory load in the metabolic network of monocyte derived macrophages
(Pires Pacheco et al., 2015a) and to investigate the effect of the TRIM32
mutation on the metabolism of brain cells of mice (Hillje et al., 2013).
The use of metabolic modelling in the frame of personalized medicine, high-throughput data
analysis and integration of omics data calls for a significant improvement in quality of existing
algorithms and generic metabolic reconstructions used as input for the former. To this aim
and to initiate a discussion in the community on how to improve the quality of context-specific
reconstruction, benchmarking procedures were proposed and applied to seven recent contextspecific
algorithms including FASTCORE and FASTCORMICS (Pires Pacheco et al., 2015a). Further, the problems arising from a lack of standardization of building and annotation
pipelines and the use of non-specific identifiers was discussed in the frame of a review.
In this review, we also advocated for a switch from gene-centred protein rules (GPR rules) to
transcript-centred protein rules (Pfau et al., 2015).