Xo Mics To Model

bioRxiv preprint doi: https://doi.org/10.1101/2021.11.08.467803; this version posted November 10, 2021.
The copyright holder for this

preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.
XomicsToModel: Multiomics data integration and generation of

thermodynamically consistent metabolic models
8th November 2021
German Preciat1∗ , Agnieszka B. Wegrzyn1∗ , Ines Thiele2,3,4,5 ,

Thomas Hankemeier1† and Ronan M.T. Fleming1,2†
Abstract
Constraint-based modelling can mechanistically simulate the behaviour of a biochemical sys-
tem, permitting hypotheses generation, experimental design and interpretation of experimental
data, with numerous applications, including modelling of metabolism. Given a generic model,
several methods have been developed to extract a context-specific, genome-scale metabolic model
by incorporating information used to identify metabolic processes and gene activities in a given
context. However, existing model extraction algorithms are unable to ensure that the context-
specific model is thermodynamically feasible. This protocol introduces XomicsToModel, a semi-
automated pipeline that integrates bibliomic, transcriptomic, proteomic, and metabolomic data
with a generic genome-scale metabolic reconstruction, or model, to extract a context-specific,
genome-scale metabolic model that is stoichiometrically, thermodynamically and flux consistent.
The XomicsToModel pipeline is exemplified for extraction of a specific metabolic model from a
generic metabolic model, but it enables multi-omic data integration and extraction of physico-
chemically consistent mechanistic models from any generic biochemical network. With all input
data fully prepared, algorithmic completion of the pipeline takes ~10 min, however manual re-
view of intermediate results may also be required, e.g., when inconsistent input data lead to an
infeasible model.
Keywords
Constraint-based modelling, multi-omics, data integration, thermodynamic consistency, systems bio-

logy.
1
Division of Systems Biomedicine and Pharmacology, Leiden Academic Centre for Drug Research, Leiden University,
Einsteinweg, Leiden, The Netherlands,
2
School of Medicine, National University of Ireland, University Road, Galway, Ireland,
3
Ryan Institute, National University of Galway, Galway, Ireland,
4
Division of Microbiology, National University of Galway, Galway, Ireland,
5
APC Microbiome Ireland, Ireland.
*
These authors contributed equally to this work.
†
Correspondence should be addressed to Thomas Hankemeier ([email protected]) and Ronan M.T. Flem-
ing ([email protected]).
1
bioRxiv preprint doi: https://doi.org/10.1101/2021.11.08.467803; this version posted November 10, 2021. The copyright holder for this
INTRODUCTION
Introduction to genome-scale metabolic model extraction

The main goal of a genome-scale metabolic model is to represent all known metabolic functions and
predict physiochemically and biochemically realistic metabolic fluxes in living systems [23]. One can
distinguish four complementary approaches to the development of a genome-scale metabolic model for
a single organism or anatomical region of interest, e.g., organ, tissue or cell type. In the first instance,
if no network reconstruction is available, a high-quality genome-scale metabolic reconstruction of a
single organism can be generated from scratch by following established protocols [28]. Subsequently,
various metabolic models can be derived from a single reconstruction, by computational application of
different combinations of mathematical modelling assumptions. In the second instance, if a network
reconstruction, or a model, is available for an organism with orthologous genes, methods to generate
metabolic models for related microbial [20] and mammalian [34] species have been proposed. In
the third instance, given a generic metabolic reconstruction, or model, for a multi-cellular organism
containing metabolic reactions from all anatomical regions, there are several methods for deriving a
model that is specific for a particular anatomical region. In literature, they are also referred to as
methods for tailoring a genome-scale metabolic model [21]. When objectively compared in detail, it
was found that the model derivation method most strongly affected the accuracy of gene-essentiality
predictions [21]. In the fourth instance, given a universal metabolic reaction database, containing
metabolic reactions from multiple organisms, an organism-specific model can be derived using one
of the several existing methods , e.g., CarveMe [17]. In all except the first instance, one is given a
generic reconstruction, or model, and the method extracts a specific model, which is a subset of the
reconstruction.
A specific model may be specific to a particular organism or a particular anatomical region. For
exmple, given a generic human metabolic model, one could extract a hepatocyte model [16], with
the capability to model different contexts, depending on the constraints subsequently applied to it.
A specific model may also be context-specific, which is a particular organism or anatomical region
in a particular environmental and internal context. From the same starting point, with different
input data, one could extract a context-specific model, e.g., a hepatocyte model specific to a fasting
state [32]. The distinction between cell-type and context-specific is more a gradation of specificity,
paralleled by increased constraints and decreased volume of the feasible steady state solution set,
which can be reliably quantified if the set is convex [13].
Established model extraction methods ensure that a specific model is flux-consistent, that is, each
reaction admits a non-zero steady-state flux. Consider Flux Balance Analysis [22], which requires
the solution to the following optimisation problem
max ρ(v) := cT v
v∈Rn
s.t. Sv = 0, (1)
lb ≤ v ≤ ub,
where v ∈ Rn is the net rate of each reaction, ρ(v) := cT v is a biologically motivated linear objective,
specified by the coefficient vector c ∈ Rn . The matrix S ∈ Rm×n is a stoichiometric matrix, where m
is the number of metabolites, and n is the number of reactions. The constraint Sv = 0 represents the
assumption of steady-state, that is production = consumption for metabolites not exchanged across
the boundary of the system and production + uptake = consumption + secretion for metabolites
exchanged across the boundary of the system. The inequalities lb ≤ v ≤ ub denote box constraints
from the lower (lb ∈ Rn ) and upper bounds (ub ∈ Rn ) on reaction rates.
The j th reaction in a network is said to be flux inconsistent if, vj = 0 for all steady state flux
vectors in the feasible set Ω := {Sv = 0, lb ≤ v ≤ ub}. Since Flux Balance Analysis only predicts
steady state flux, it is misleading to include unidentified flux inconsistent reactions in a model for Flux
Balance Analysis because one must be able to distinguish between a reaction that cannot carry steady
state flux (independent of the objective chosen) from a reaction that does not carry steady state
2
flux in a particular Flux Balance Analysis solution (dependent of the objective chosen). Therefore,
flux inconsistent reactions should be identified in a model before interpreting the results of methods
such as Flux Balance Analysis. Note that, in a reconstruction, a flux inconsistent reaction may be
supported by experimental evidence, and serves as a starting point for further refinement, e.g., by
gap filling [29].
It is well established that Flux Balance Analysis requires additional constraints to ensure that a
steady state flux vector is also thermodynamically feasible [3, 15], i.e., consistent with energy con-
servation and the second law of thermodynamics [9]. The critical necessity at genome-scale, is
to incorporate any additional constraints, or terms in the objective, in a mathematical form that,
when computationally implemented, retains the computational efficiency and certificates available
with linear optimisation. As exemplified in Figure 1, a stoichiometric matrix may be split into two
sets of columns S := [N, B], where the matrix N represents internal reactions, which do conserve
mass, and the matrix B represents external reactions, which do not conserve mass, e.g., reaction
equation A → ∅. The latter are modelling constructs to represent the exchange of metabolites
with the environment. By splitting internal net fluxes into unidirectional fluxes, and maximising the
entropy of unidirectional fluxes, one can compute steady state, thermodynamically feasible fluxes
in genome-scale biochemical models using convex optimisation [10], which retains the efficiency as
well as feasibility certificates available with linear optimisation [6]. However, this approach assumes
the model admits a thermodynamically feasible flux. We define a model to be thermodynamically
consistent if each of its reactions admits a nonzero thermodynamically feasible flux. A flux vector
may be thermodynamically feasible but the flux is zero for some reaction, which is not enough to
establish that reaction to be thermodynamically consistent.
Figure 1: Network diagram of upper glycolysis (left) with reactions (arrows, upper case labels) and metabolites
(nodes, lower case labels) with corresponding labels. The corresponding Stoichiometric matrix (right, S ∈
R13×10 , N ∈ R13×8 , B ∈ R13×2 ), with reversible reactions shown in green, non-reversible reactions shown
in blue, and exchange reactions shown in red. The upper bound (ub) of internal reactions is unconstrained,
whereas the lower bound (lb) is limited by reaction directionality or by the maximum uptake rate, which can be
seen as a constraint on uptake of extracellular glucose (glc_D[e]) or secretion of dihydroxyacetone phosphate
(dhap[e]) from the environment. All metabolite and reaction abbreviations are with respect to the namespace
in www.vmh.life[19].
A thermodynamically feasible flux does not exist when at least one combination of bounds on re-
action rates forces net flux around a stoichiometrically balanced cycle [8]. Consider the following
stoichiometrically balanced cycle of reactions A ⇌ B ⇌ C ⇌ A. If the bounds are such that
A → B → C → A is the only feasible net flux, then the network containing this cycle cannot carry a
thermodynamically feasible steady state net flux. The second law of thermodynamics requires that a
chemical potential difference between substrates and products is required for net flux. Let µA denote
the chemical potential of metabolite A. Given net flux A → B the second law of thermodynamics
requires µA > µB and similarly µB > µC and µC > µA . However, the first pair of inequalities imply
3
µA > µC , which is inconsistent with the last, unless one assigns more than one chemical poten-
tial to C at the same instant. However, to do so would be inconsistent with energy conservation,
which requires each row of a stoichiometric matrix to be associated with a single chemical potential,
assuming that each row corresponds to a well mixed compartment [9].
Stoichiometrically balanced cycles are biochemically faithful network topological features that are
omnipresent in genome-scale models. It is not the presence of a stoichiometrically balanced cycle,
per say, that presents a problem. It is the thermodynamically inconsistent specification of bounds,
which force net flux around a stoichiometrically balanced cycle and prevents the prediction of a
thermodynamically feasible net flux. As recognised by Desouki et al. [8], provided that a model does
not contain a combination of thermodynamically inconsistent bounds, minimisation of the one-norm
of net flux, subject to additional constraints maintaining the direction of net flux, is guaranteed to
remove that part of a steady state flux vector that is thermodynamically infeasible. This approach
is attractive, as it is based on linear optimisation, but it has yet to be leveraged for generation of
thermodynamically consistent models. Generation of thermodynamically consistent models should, in
principle, increase the biochemical fidelity of constraint-based models as it opens up the possibility of
leveraging established methods to efficiently compute thermodynamically feasible steady state fluxes
[10, 8], that assume an input model is thermodynamically consistent. Given a generic reconstruction
or model, all established model extraction algorithms, including GIMME [4], iMAT [36], MBA [16],
mCADRE [35], FastCore [33], and CarveMe [17], and also gap filling algorithms [29], extract a
specific model based on either a binary (present/absent) or weighted assignment of reactions desired
to be present in a specific model, where each reaction is (net) flux consistent. However, the models
generated from these algorithms are not guaranteed to admit thermodynamically feasible net flux.
Development of the protocol

Overview In this protocol we present the XomicsToModel pipeline, which, given a generic reconstruc-
tion and multi-omics data, enables extraction of a context-specific, genome-scale metabolic model,
which is stoichiometrically, thermodynamically and flux consistent. The pipeline was developed
to generate a context-specific model of a dopaminergic neuronal metabolism [24], given a generic
genome-scale human metabolic reconstruction, Recon3D [7], and transcriptomic, metabolomic and
bibliomic data. Substantia nigra dopaminergic neurons are the most susceptible to degeneration in
Parkinson’s disease, a progressive, neurodegenerative movement disorder, and its biochemical mech-
anisms remain poorly understood [2]. Finally, although the operation of XomicsToModel pipeline is
exemplified for extraction of a specific metabolic model from a generic metabolic model, its imple-
mentation is envisaged to enable multi-omic data integration and extraction of a physicochemically
consistent mechanistic model from any generic biochemical network.
Input The input to the XomicsToModel pipeline is a biochemical network and a set of omics data.
The biochemical network may be a reconstruction, or a model, and is not required to be stoi-
chiometrically, thermodynamically or flux consistent. The XomicsToModel pipeline allows a flexible
and modular integration of transcriptomic, proteomic, and metabolomic data, as well as bibliomc
data abstracted from literature curation. In each case, the input data may be qualitative (present,
absent, unspecified), semi-quantiative, quantitative, or combinations thereof. The XomicsToModel
pipeline is complemented by functions to automatically import omics data.
The application of the XomicsToModel pipeline to extract a dopaminergic neuronal metabolic model
[24] demonstrates its flexibility with respect to incorporation of a variety of qualitative and quantit-
ative constraints. For example, the presence or absence of metabolites in the culture medium, as well
as quantitative metabolite exchange reaction rates, which were applied using quadratic optimisation
to set exchange reaction bounds, weighted by the inverse of measurement uncertainty. Optionally,
the XomicsToModel pipeline enables extension of an input reconstruction or model, with manually
specified metabolic reactions, in the case where a generic model is missing certain key pathways
relevant for a system of interest.
4
Constraint relaxation The XomicsToModel pipeline incorporates a series of tests for flux and op-
tionally thermodynamic consistency after each step that reduces the size of the feasible set, e.g.,
after removal of generic metabolites or reactions assigned not to be present in a specific model. This
approach ensures the early detection of infeasible constraints resulting from inaccurate or inconsist-
ent experimental data. Detection of inconsistency is followed by algorithmic relaxation of constraints
(e.g. bounds) to render the draft model feasible. Specifically, the XomicsToModel pipeline automat-
ically searches for the minimal number of constraint relaxations required to admit a flux-consistent
model in the case that inconsistent or incorrect omics data renders a draft model infeasible. This
approach can be used to feedback into data processing or input to avoid model infeasibility (Figure
3), which is often a challenge with multi-omic data integration.
[24] presented a particular challenge for constraint-based modelling because the extracted model was
required to be representative of neurons that do not grow.Therefore, the maximisation of biomass
growth could not be used as an objective. Instead, a set of coupling constraints [30] were added
to represent cell maintenance requirements, e.g., for the turnover of key metabolites. However,
if incorrectly scaled with respect to constraints on exchange reactions, the application of coupling
constraints may generate an infeasible draft model. Therefore, an algorithmic proposition of a
minimal number of constraint relaxations accelerates the identification of biochemically inconsistent
constraints and refinement of input data.
Model extraction The XomicsToModel pipeline is compatible with various model extraction al-
gorithms (cf Table 1), with an established interface to FastCore [33]. In addition, the XomicsToModel
pipeline builds upon thermoKernel, a novel algorithm used to extract the dopaminergic neuronal meta-
bolic model [24], but applicable for the extraction of any context-specific model that is required to
be stoichiometrically, thermodynamically and flux consistent. The ability of thermoKernel to enforce
thermodynamic consistency during the model extraction process opens up new possibilities for data
integration. Thermodynamic consistency is implemented by constraining the possible relationships
between reaction flux and metabolite chemical potential (detailed below). Consequently, it is possible
to directly specify a ternary (presence, absence, unspecified) or weighted assignment of metabolites
desired to be present in a specific model. More generally, it is possible to specify a ternary or weighted
assignment of metabolites and reactions desired to be present in a specific model.
This enables the thermoKernel algorithm to extract a thermodynamically consistent model that is a
trade-off between ternary or weighted assignments of rows and columns of a stoichiometric matrix.
[24] demonstrates this flexibility with ternary and weighted specification of metabolites and reactions
desired to be present in the dopaminergic neuronal model. The XomicsToModel enables integration
of, for example, qualitative metabolomic data, where the metabolite presence can now be applied
directly as context-specific data input.
Ensemble modelling To exploit the XomicsToModel pipeline flexibility, a supplementary

XomicsToMultipleModels function is provided to generate an ensemble of context-specific genome-scale
models by varying the specific data and technical parameters used in the model extraction process.
Moreover, supplementary functions are provided to estimate the predictive capacity of an extracted
model, given independent data. These features will facilitate analysis by complementary software with
capabilities for machine learning from model ensembles [18]. The application of the XomicsToModel
pipeline to extract an ensemble of dopaminergic neuronal metabolic models is presented elsewhere
[24].
Applications of the XomicsToModel pipeline

The XomicsToModel pipeline applies to any situation where one has a generic (or universal) recon-
struction (or model) and seeks to extract a specific subset of it, based on transcriptomic, proteomic,
5
or metabolomic data, or combinations thereof. There is no restriction with respect to the identity
of the species of the input reconstruction or model, nor to the number of species represented in
either the input or output models. The most computationally intensive steps are the identification
of the largest thermodynamically consistent subset of an input reconstruction or model and the ex-
traction of a thermodynamically consistent subset of minimal size, both of which are achieved with
the thermoKernel algorithm. As the thermoKernel algorithm is based on a sequence of linear optimisa-
tion problems, its performance scales accordingly. Ternary specification of metabolite presence/ab-
sence/unspecified, reaction activity/inactivity/unspecified, gene activity/inactivity/unspecified, and
combinations thereof are all possible. Similarly, weighted specification, to bias for or against the
inclusion of metabolites, reactions, or both, based on input transcriptomic, proteomic, or meta-
bolomic data is also possible, also in combination with a ternary specification. The XomicsToModel
pipeline is interfaced via a function, but multiple options can be specified to enable modular usage
of the XomicsToModel pipeline to implement model extraction priorities reflective of a wide variety of
scenarios. For example, one can switch from thermoKernel to a different model extraction algorithm,
such as FastCore[33], as desired. As such, it is envisaged that the XomicsToModel pipeline will have
widespread applicability.
Comparison with other methods

A systematic comparative evaluation of model extraction methods by Opdam et al. [21] has shown
that using various data types, such as multi-omics data, for model construction and validation,
as well as careful selection of gene expression cut-offs, lead to higher model accuracy. Table 1
compares other algorithms for extraction of a context-specific model, given a generic model of a
single organism, with the XomicsToModel pipeline. Overall, in comparison with similar methods, the
XomicsToModel pipeline provides several additional capabilities in terms of its ability to represent
additional constraints, its flexibility concerning data integration and its ability to suggest relaxations
to recover from the application of inconsistent constraints.
Table 1: Comparison of technical features of algorithms for extraction of a context-specific model, given
a generic model of a single organism. MILP, mixed integer linear programming; FVA, flux variability ana-
lysis, fastFVA[12], computational acceleration of flux variability analysis; QC-LP, quasiconcave sequence of
linear programs; DCA-LP, difference of convex function sequence of linear programs. ∗ denotes subsequently
implemented in the COBRA Toolbox [14].
GIMME iMAT MBA INIT mCADRE FastCore XomicsToModel

Citation [4] [36] [16] [1] [35] [33] N/A
Year 2008 2010 2010 2012 2012 2014 2021
Active metabolite list × × × × × × ✓
Metabolite weights ∈ R × × × × ✓ × ✓
Active reaction list ✓ ✓ ✓ ✓ ✓ ✓ ✓
Reaction weights ∈ R × × ✓ ✓ ✓ × ✓
Coupling constraints × × × × × ✓* ✓
Thermodynamic × × × × × × ✓
consistency
Constraint relaxation × × × × × × ✓
Algorithmic approach MILP MILP FVA MILP fastFVA QC-LP DCA-LP
6
Experimental Design
To run the XomicsToModel pipeline, users must first obtain and prepare a generic reconstruction or
model and the context-specific data used to extract the model. Furthermore, one must either accept
the default set of technical parameters or specify the parameters to suit a particular context (as
detailed below).
Generic reconstruction or model: A generic genome-scale reconstruction or model represents a

metabolic network constructed from the amalgamation and manual literature curation of metabolic
reactions that occur in various cell types or organisms.
Context-specific data: Context-specific data represents the genotype or phenotype of a specific

biological system. It can be obtained through a review of the existing literature or derived experiment-
ally from a biological system or both. The information can be entered manually or using the function
preprocessingOmicsModel described in the supplementary tutorial. The following context-specific in-
formation is currently supported by the XomicsToModel pipeline:
• Bibliomic data: Data derived from a manual reconstruction following a review of the existing
literature. This includes data on the genes, reactions, or metabolites known to be present
or absent in the studied biological system. Furthermore, bibliomics data can define a set of
coupled reactions or a set of constraints for model reactions based on phenotypic observations.
• Transcriptomic data: Measured gene expression levels in the studied biological system. It is
used to estimate the activity of reactions associated with the detected genes. Transcriptomic
data can be provided in fragments per kilobase million (FPKM) or raw counts.
• Proteomic data: Measured protein levels in the studied biological system. Similarly to tran-
scriptomics data, it can be used to estimate which reactions in the metabolic model should be
considered active based on the gene-protein-reaction association (Figure 2).
• Metabolomic data: The average and standard deviation of metabolite concentrations meas-
ured in cell media, biofluids, tissues, or organisms translated into flux units umol/gDW /h.
Metabolites detected experimentally can be assigned to be present in the metabolic model.
Measured uptakes and secretions in growth media or biofluids can be used to constrain the up-
takes and secretions of the model quantitatively. Furthermore, growth conditions information
such as growth media composition can also be provided to constrain available uptakes in the
model.
Technical parameters: With these options, technical constraints can be added to the model, as
well as setting the parameters for model extraction or debugging. If they are not specified, default
values are used.
• Bounds: Parameters that define the minimum and maximum flux value in the model or the
minimum non-zero flux value tolerance.
• Exchange reactions: Instructions to close or leave open demand and sink reactions and
whether exchange reactions should be added based on the metabolomic data.
• Extraction options: Parameters required by solvers and for the context-specific model extrac-
tion algorithms.
• Data-specific parameters: Parameters that define the minimum level of transcript/protein
to be considered as present in the model (threshold) and whether the transcripts below the
defined threshold should be removed from the generic reconstruction or model.
• Debugging options: Debugging parameters that allow the user to evaluate the output of
consecutive steps within the pipeline after they have been completed.
7
Figure 2: The gene-protein-reaction association or GPRs are boolean operators that describe the interactions
of genes, transcripts, proteins, and reactions. For the reaction glucose-6-phosphate isomerase (PGI) to occur,
the gene PGI must be activated (Entrez ID: 2821) leading to the transcription of the PGI mRNA (NCBI
Reference Sequence: NM 000175.5). The mRNA is then translated into the glycolytic enzyme PGI (UniProt
ID: P06744). Finally, the PGI enzyme (Enzyme Commission Number: 5.3.1.9) catalyse a reaction that
interconverts glucose-6-phosphate (VMH ID: g6p) and fructose-6-phosphate (VMH ID: f6p)
Required expertise
The pipeline is written in MATLAB, a programming language that is easy to learn but also power-
ful due to the numerous toolboxes for numerical and symbolic computing. This protocol can be
implemented by anyone who understands basic MATLAB programming and the fundamentals of
constraint-based modelling. The computational load associated with this protocol is determined
by the network’s size. A desktop personal computer is sufficient to generate a stoichiometrically,
thermodynamically and flux consistent, context-specific, genome-scale metabolic model.
Limitations
This protocol focuses on extracting a context-specific genome-scale model from an existing recon-
struction by integrating multi-omic data from a biological system. It does not cover generation of a
high-quality genome-scale metabolic reconstruction from scratch Norsigian et al. [20] and Thiele and
Palsson [28], or the analysis of existing models [14]. Additionally, the integration of multi-omics data
presents several challenges since they are generated using a variety of platforms, affecting storage and
data formats significantly. The integration of multi-omic data necessitates data in specific formats.
Therefore, individual omics data must be pre-processed. Furthermore, experimental errors, such as
data processing or measurement errors, can propagate through the extraction of a metabolic network
that is not a faithful representation of the original biochemical system. In this situation, a significant
loss of predictive capacity will be evident. Therefore, it is not recommended to extract a context-
specific model and claim that it has high predictive accuracy without commensurate comparison with
some independent experimental data.
8
MATERIALS
Equipment
Input data
The COBRA Toolbox supports commonly used data formats for model description such as Systems
Biology Markup Language (SBML) and Excel sheets (.xlsx). MATLAB supports a wide range of
text and spreadsheet formats that can be used to provide context-specific data for the XomicsToModel
pipeline.
Required hardware
• A computer with at least 8 GB of RAM and any 64-bit Intel or AMD processor. CRITICAL
Depending on the size of the model more processing power and more memory might be needed.
• A hard drive with at least 10 GB of space available.
Required software
• An operating system that is MATLAB qualified (https://mathworks.com/support/sysreq.
html) CRITICAL To make sure that the operating system is compatible with the MATLAB
version check the requirements at https://mathworks.com/support/sysreq/previous_
releases.html.
• MATLAB (MathWorks, https://mathworks.com/products/matlab.html), version R2021+.

Install MATLAB and its license by following the official installation instructions (https:
//mathworks.com/help/install/ug/installmathworks-software.html). ! CAUTION
Version R2021+ or above is recommended for running MATLAB live scripts (.mlx files).
CRITICAL No support is provided for versions older than R2021. MATLAB is released on
a twice-yearly schedule, and the latest releases of MATLAB may not be compatible with the
existing solver interfaces, necessitating an update of the MATLAB interface provided by the
solver developers, an update of the COBRA Toolbox, or both.
• The COBRA Toolbox version 3.4 [14], or above. To install the COBRA Toolbox follow
the instructions on https://github.com/opencobra/cobratoolbox. ! CAUTION If an
installation of COBRA Toolbox is already present, update the repository from MATLAB,
via terminal, or Git Bash. CRITICAL Check that all of the system requirements in
https://opencobra.github.io/cobratoolbox/docs/requirements.html are met.
Solvers
Table 2 provides an overview of the optimisation solvers supported by the XomicsToModel pipeline.
Table 2: An overview of the optimisation solvers XomicsToModel supports.
Name Version Interface

GLPK 2.7+ glpk
GUROBI 7.0+ gurobi
ILOG CPLEX 10-12.10 ibm_cplex
MATLAB R2014b+ matlab
Equipment setupCOBRA Toolbox

If one is not a developer of COBRA Toolbox code, update the COBRA Toolbox from within MATLAB
by running the following command:
>> updateCobraToolbox
9
Or, update from the terminal (or Git Bash) by running the following from within the cobratoolbox
directory.
$ cd c o b r a t o o l b o x # change to the l o c a l c o b r a t o o l b o x d i r e c t o r y
$ g i t checkout master # s w i t c h to the master branch
$ g i t p u l l o r i g i n master # r e t r i e v e changes
In the case that the update of the COBRA Toolbox fails or cannot be completed, clone the repository
again. CRITICAL The COBRA Toolbox can be updated as described above only if no changes
have been made to the code in one’s local clone of the official online repository. To contribute any
local edits to the COBRA Toolbox code into the official online repository please follow the MAT-
LAB.devTools guidelines [14] (https://opencobra.github.io/MATLAB.devTools/stable/) .
PROCEDURE
The XomicsToModel pipeline is implemented in a sequence of twenty-one steps (Figure 3). To gener-
ate a thermodynamic-flux-consistent, context-specific, genome-scale metabolic model, the pipeline
requires three inputs: a generic COBRA reconstruction or model, context-specific information and
technical parameters. Below, for each step, all the user-defined options used are listed and described.
Figure 3: It is possible to create genome-scale models iteratively and systematically using the XomicsToModel
pipeline. The generated models can be used to design or interpret the information based on bibliomic data
from manual literature curation (red), as well as metabolomic (blue), proteomic (green) or transcriptomic
(purple) data. Furthermore, experimental data can be used to validate or refine the context-specific model.
The XomicsToModel pipeline (grey) extracts a new model by integrating a generic model with context-specific
information based on the technical parameters such as the extraction algorithm, method for identifying active
genes, or maximum reaction rates.
To extract a context-specific model from a generic model using multi-omic data, first launch the CO-
BRA Toolbox [14] and specify the optimisation solver to be used. The XomicsToModel pipeline can be
run, if at least one input variable, a generic model is provided ,which must have the properties spe-
cified in https://github.com/opencobra/cobratoolbox/blob/master/docs/source/notes/
COBRAModelFields.md. The user-defined context-specific information, specificData, and technical
parameters, param, are optional, and if not specified, default values for the required fields will be
assigned.
>> initCobraToolbox
>> changeCobraSolver ( ' g u r o b i ' , 'LP ' ) ;
>> changeCobraSolver ( ' g u r o b i ' , 'QP ' ) ;
>> [ c o n t e x t S p e c i f i c M o d e l , modelGenerationReport ] = XomicsToModel ( genericModel , ...
s p e c i f i c D a t a , param )
? TROUBLESHOOTING
10
The name and types of the fields for the specificData and param variables must be identical to how they
are described in each step for the XomicsToModel pipeline to identify them. For user’s convenience,
supplementary information describes a tool for automatically generating the specificData variable,
as well as a tool for creating multiple models varying the user-defined options and for testing the
accuracy of the extracted model.
1 | Preparation of data TIMING ∼ 5 s
Description:
The first step performs data harmonisation of the user-defined options in specificData and param, and
generic model’s fields to comply with naming conventions and formats expected by the XomicsToModel
pipeline. Missing specificData and param fields are identified and default values are assigned. The
COBRA Toolbox function getCobraSolverParams is used to identify the primal feasibility tolerance.
Furthermore, the draft model is created based on the generic model. In addition, if a filename for
the diary has been specified in param.diaryFilename the diary file is opened, where many intermediate
results of the pipeline are printed for debugging.
Usable variables:
• param.diaryFilename - The name (and location) of a diary file with the printed pipeline output (Default: 0).
2 | Generic model checks TIMING ∼ 1 s
Description:
In this section, the draft model fields that may cause inconsistency when adding or removing reactions
are removed. For example, Recon3D [7], includes a cell compartment representing the mitochondrial
intra-membrane space, but this is not necessary when using an algorithm to compute thermody-
namically feasible fluxes. Therefore, if 'thermoKernel' is chosen as the tissue-specific solver, this
compartment is removed and metabolites are reassigned to the cytosol.
Regardless of the optimisation objective selected, every solution to Equation 1 must satisfy the steady
state constraints Sv = 0, reaction rate bounds lb ≤ v ≤ ub, and optionally coupling constraints
Cv ≤ d, implying that the system of inequalities is feasible; if the model is infeasible in this step, an
error will be generated ? TROUBLESHOOTING .
Usable variables:
• param.tissueSpecificSolver - The name of the tissue-specific solver to be used to extract the context-specific model
(Possible options: 'thermoKernel' and ' fastcore '; Default: 'thermoKernel').
• param.printLevel - Level of verbose that should be printed (Default: 0).
3 | Add missing reactions TIMING ∼ 30 s
Description:
Reactions specified in the specificData.rxns2add, are added to the draft model using the function
addReaction. Multiple data can be included but only the reaction identifier and the reaction formula
are mandatory. CRITICAL STEP When a metabolic reaction is added, default bounds are set
based on the characters used in the reaction formula to separate the substrates and products.
• Forward (->): lb = 0; ub =param.TolMaxBoundary.
• Reverse (<-): lb =param.TolMinBoundary; ub = 0.
• Reversible (<=>) lb =param.TolMinBoundary; ub =param.TolMaxBoundary.
11
The default values for missing reaction names, metabolic pathways, and gene rules are Custom
reaction, Miscellaneous, and an empty cell, respectively. ! CAUTION If a reaction is already present
in the draft model, it will be replaced by the new reaction. By associating a new gene ID with a
reaction, a new gene is added to the draft model.
Given a generic reconstruction or model, the findStoichConsistentSubset function approximately solves
the problem to extract the largest subset of it that is stoichiometrically consistent [14]. Inconsistent
metabolites and reactions are removed; if this affects feasibility, an error is generated ? TROUBLESHOOTING
. A set of reactions is stoichiometrically consistent if every reaction in that set is mass balanced [11].
In the split S := [N, B], the matrix N represents stoichiometrically consistent internal reactions,
while the matrix B represents stoichiometrically inconsistent external reactions. It is vital to appre-
ciate that the problem to extract the largest stoichiometrically consistent subset is combinatorially
complex, and therefore, at genome-scale, only approximation to the actual solution is achievable
in practice. The quality of the approximation is improved if heuristics are used to warm-start the
algorithm with a specification of the reactions that are accepted as being stoichiometrically incon-
sistent, e.g., external reactions, as specified as false in model.SIntRxnBool. If model.SIntRxnBool is
not provided, the function findSExRxnInd attempts to heuristically distinguish internal and external
reactions, based on reaction naming conventions (e.g. ’EX_’ prefix for an exchange reaction) and
reaction stoichiometry (any reaction with only one nonzero stoichiometric coefficient in the corres-
ponding row is an external reaction). Also in this step, the gene-protein-reaction rules are updated
for the newly added reactions, and the reaction-gene-matrix is regenerated. In addition, different
boolean vectors indicating stoichiometrically consistent and inconsistent metabolites and reactions
are added to the draft model. If param.printLevel is greater than zero, a table with a summary of the
draft model’s stoichiometric consistency will be printed (Table 3). If the param.debug is active (true),
all results generated up to this point are saved in the file 3.debug_prior_to_identification_of_act-
ive_genes.mat.
Usable variables:
• specificData.rxns2add - Table containing the identifier of the reaction to be added, its name, the reaction formula,
the metabolic pathway to which it belongs, the gene rules to which the reaction is subject, and the references. (Default:
empty).
rxnID rxnNames rxnFormulas subSystems rxnGrRules rxnReferences

newRxn1 Oxidation A + O -> AO Glycolysis gene1 or gene2 PMID: ****
newRxn2 Reduction AH + B <=> A + BH Glycolysis gene1 PMID: ****
: : : : : :
• param.TolMinBoundary - The reaction flux lower bound minimum value (Default: -1000).
• param.TolMaxBoundary - The reaction flux upper bound maximum value (Default: 1000).
• param.debug - Logical, should the pipeline save its progress for debugging (Default: false).
• model.SIntRxnBool - Logical vector, true if a reaction is heuristically considered an internal reaction, false for an
external reactions (Default: false).
4| Set limit bounds TIMING ∼ 1 s
Description:
Considering the constraints lb ≤ v ≤ ub shown in Equation 1, where lb is the lower bound and
ub the upper bound, each lower bound in the draft model such that lbi = min(lb) and lbi ≤
T olM inBoundary is set to param.TolMinBoundary and each upper bound such that ubi = max(ub)
and ubi ≥ T olM axBoundary is set to param.TolMaxBoundary. ! CAUTION There should be no
lower bound that exceeds an upper bound. The model’s feasibility is tested with the new bounds
? TROUBLESHOOTING . If param.printLevel is greater than zero, the modified bounds will be
printed.
12
Usable variables:
• param.TolMinBoundary - The reaction flux lower bound minimum value (Default: -1000).
5 | Identify active genes TIMING ∼ 102 s
Description:
The list of active genes is specified by NCBI Entrez gene identifiers, which are unique integers for
genes or other loci. ! CAUTION A gene must be present in the draft model to be considered as active.
Missing genes can be added using the gene-protein-reaction rules in specificData.rxns2add.rxnGrRules
(Step 3). The active genes list is composed based on the data from multiple sources:
• Bibliomics: Genes classified as active or inactive based on published data.
• Proteomics: Genes that, based on the list of detected proteins, are active based on their (norm-
alised) peak areas. If peakArea ≥ param.thresholdP, the gene is considered active. ! CAUTION
Proteomics data is usually annotated with UniprotKB. However, for data integration with the
model, UniprotKBs need to be converted into corresponding Entrez IDs.
• Transcriptomics: Genes that, based on the list of detected transcripts, are active based on
their expression levels (FPKM, Fragments Per Kilobase of transcript per Million mapped reads
or raw counts). If expressionLevel ≥param.thresholdT, the gene is considered active.
Furthermore, the levels of gene expression in specificData.transcriptomicData are used to link gene
expression levels to associated reactions using the function mapExpressionToReactions. In short, the
resulting model.expressionRxns field is based on the gene relationship described in the gene-protein-
reaction rules: an AND in the gene-protein-reaction rules is assigned min(lb), and an OR is as-
signed a max(ub). For genes that are not found in the transcriptomics data, no expression value
is added (NaN is default). If param.inactiveGenesTranscriptomics is used, the genes that fall below
the specified threshold are added to the list specificData.inactiveGenes. In the event of a discrep-
ancy, for example, an inactive gene according to manual literature curation but active according
to proteomics or transcriptiomics, preference will be given based on the option specified in the
param.curationOverOmics.Genes not found in the draft model are removed from the list of active genes.
Two fields are added to the model to indicate the gene expression value of the genes present in the
draft model model.geneExpVal and the expression of the corresponding reactions (model.expressionRxns)
based on the function mapExpressionToReactions. If param.printLevel is greater than zero, a histogram
of reaction expression is printed.
Usable variables:
• specificData.transcriptomicData - Table with a column with Entrez ID’s and a column for the corresponding
transcriptomics expression value in FPKM or raw counts (Default: empty).
gene expVal
entrezGene1 220
entrezGene2 14916
: :
• param.thresholdT - The transcriptomic cutoff threshold for determining whether or not a gene is active (Default:
log2(1)).
• specificData.proteomics - Table with a column with Entrez ID’s and a column for the corresponding protein levels
(Default: empty).
gene expVal
entrezGene1 24
entrezGene2 78
: :
13
• param.thresholdP - The proteomic cutoff threshold for determining whether or not a gene is active (Default:
log2(1)).
• specificData.activeGenes - List of Entrez ID of genes that are known to be active based on the bibliomic data
(Default: empty).
• param.curationOverOmics - Logical, indicates whether curated data should take priority over omics data (Default:
false).
• param.inactiveGenesTranscriptomics - Logical, indicate if inactive genes in the transcriptomic analysis should be
added to the list of inactive genes (Default: true).
• specificData.inactiveGenes - List of Entrez ID of genes known to be inactive based on the bibliomics data (Default:
empty).
6 | Identify active reactions TIMING ∼ 10 s
Description:
The XomicsToModel pipeline uses several fields from the specificData variable to identify the set of
metabolic reactions that must be active in the draft model. These optional fields include the re-
action specified as the objective function, the metabolic reactions to add, the reactions identified
as active based on bibliomic data, the reactions whose bounds are restricted based on bibliomic
data, the coupled reactions (See Step 12), the exchange reactions for metabolites present in the
growth media, and the exchange reactions for metabolites shown to be secreted or uptaken based
on exo-metabolomic data. All of these reactions are added to the active reaction list, which will be
appended or shortened if necessary to keep the model stoichiometrically, thermodynamically and flux
consistent.
Usable variables:
• param.setObjective - Cell string indicating the linear objective function to optimise (Default: none).
• specificData.rxns2add - Table containing the identifier of the reaction to be added, its name, the reaction formula,
the metabolic pathway to which it belongs, the gene rules to which the reaction is subject, and the references. (Default:
empty).
rxnID rxnNames rxnFormulas subSystems rxnGrRules rxnReferences

newRxn1 Oxidation A + O -> AO Glycolysis gene1 or gene2 PMID: ****
newRxn2 Reduction AH + B <=> A + BH Glycolysis gene1 PMID: ****
: : : : : :
• specificData.rxns2constrain - Table containing the reaction identifier, the updated lower bound (lb), the updated
upper bound (ub), a constraint description, and any notes such as references or special cases (Default: empty).
rxnID lb ub constraintDescription Notes

rxn1 0 Close uptake PMID: ****
rxn2 0.0.956 1000 Turnover constraints PMID: ****
: : : : :
• specificData.coupledRxns - Table containing information about the coupled reactions. This includes the identifier
for the coupling constraint (couplingConstraintID), the list of coupled reactions (coupledRxnsList), the coefficients of
those reactions (c, given in the same order), the right hand side of the constraint (d), the constraint sense or the
directionality of the constraint (dsence), and the reference (Default: empty).
couplingConstraintID coupledRxnsList c d dsence reference

cRxn1 crxn1, crxn2 11 2.0250 G PMID: ****
cRxn2 crxn1, crxn3 1 -1 0.2265 G PMID: ****
: : : : :
• specificData.mediaData - Table containing the fresh media concentrations. Contains the reaction identifier, the
maximum uptake (umol/gDW /h) assuming it is equal to the concentration of the metabolite in fresh medium divided
by the length of time before medium exchange, and the medium concentration (umol/l; Default: empty).
14
rxnID mediumMaxUptake mediumConcentrations

EX_met1 -220 1345
EX_met2 -14916 191021
: : :
• specificData.exoMet - Table with measured exchange fluxes, e.g., obtained from a exometabolomic analysis of fresh
and spent media. It includes the reaction identifier, the reaction name, the measured mean flux, standard deviation of
the measured flux, the flux units, and the platform used to measure it.
rxnID rxnNames mean SD units platform

−4
EX_mMet1 Exchange of mMet1 100 8x10 umol/gDW /h LC-MS
EX_mMet2 Exchange of mMet2 4x10−3 3x10−4 umol/gDW /h GC-MS
: : :
• specificData.inactiveReactions - List of reactions known to be inactive based on bibliomic data (Default: empty).
7 | Close ions TIMING ∼ 1 s
Description:
Reactions involving ion exchange can be closed in this step if param.closeIons = true, in which case the
lower and upper bounds for ion exchange reactions are closed, preventing net exchange of ions with
the environment. By doing so, it is possible to defer modelling of ion exchange, e.g., to represent
action potentials in neurons since they represent a dynamic trajectory of ion concentrations, while
prioritising metabolic activities. If the debugging option is set, all results generated up to this point
are saved in the file 7.debug_prior_to_exchange_constraints.mat.
Usable variables:
• param.closeIons - Logical, it determines whether or not ion exchange reactions are closed. (Default: false).
8 | Close exchange reactions TIMING ∼ 1 s
Description:
If the variable param.closeUptakes = true, the lb of the exchange reactions in the draft model are
closed (set to 0). This will allow the draft model to only take up the metabolites specified by
the bibliomics, growth media, or metabolomics data. It is advised to set param.closeUptakes to true
for biological systems derived from cell cultures. If param.printLevel is greater than zero, the new
constraints will be printed along with statistics such as the size of the stoichiometric matrix, the
number of exchange reactions closed, the number of exchange reactions in active reactions, and the
number of exchange reactions in reactions to constrain based on manual literature curation.
Usable variables:
• specificData.rxns2constrain - Table containing the reaction identifier, the updated lower bound, the updated upper
bound, a constraint description, and any notes such as references or special cases (Default: empty).

rxn2 0.0.956 1000 Turnover constraints PMID: ****
: : : : :
• param.closeUptakes - Logical, decide whether or not all of the uptakes in the draft model will be closed (Default:
false).
15
9 | Close sink and demand reactions TIMING ∼ 10 s
Description:
The demand and sink reactions are closed based on the value option indicated in
param.nonCoreSinksDemands. Only the demand and sink reactions in specificData.activeReactions will
remain open. If param.printLevel is greater than zero, the number and names of closed demand or
sink reactions are printed.
The solution of a feasible model must satisfy the constraints Sv = 0 and lb ≤ v ≤ ub and optionally
coupling constraints Cv ≤ d. The newly added constraints may not all be consistent with a steady
state flux at the same time, implying that the system of inequalities and the model are infeasible,
which could be caused by an incorrectly specified reaction bound. To solve the infeasibility we use
relaxFBA, which is an optimisation problem that approximately minimises the number of reaction
bounds to relax in order to make a Flux Balance Analysis problem feasible. The relaxation options
are set in param.relaxOptions. The old bounds from spent media metabolites are saved in two new
fields in the draft model, along with a field identifying the exchange reactions that were relaxed.
relaxFBA is a flexible function for relaxing. If the debugging option is set, all results generated up to
this point are saved in the file 9.debug_prior_to_metabolomicsBeforeReactionRemoval.mat.
Usable variables:
• specificData.activeReactions - List of reactions known to be active based on bibliomic data (Default: empty).
• param.nonCoreSinksDemands - The type of sink or demand reaction to close is indicated by a string (Possible
options: ' closeReversible ', 'closeForward', 'closeReverse', ' closeAll ' and 'closeNone'; Default: 'closeNone').
• param.relaxOptions - A structure array with the relaxation options (Possible options are described in detail in Cobra
Toolbox 3.0 protocol, Step 23 [14], Default: param.relaxOptions.steadyStateRelax = 0).
10 | Set metabolic constraints TIMING ∼ 10 s
Description:
The metabolic constraints can be obtained from two different sources: cell culture information
and quantitative metabolomic measurements. Based on the overall goal, each of these data-
sets can be added either before or after the extraction algorithm step (Step 20) by using the
param.growthMediaBeforeModelExtraction and param.metabolomicsBeforeModelExtraction parameters re-
spectively.
Cell culture data: The uptake rates for exchange reactions corresponding to metabolites present
in the media are adjusted based on their concentration, modifying only the lower bounds lb. The field
mediumConcentrations in specificData.mediaData can be measured or directly obtained from the product
specifications. The function preprocessingOmicsModel, described in the supplementary information,
can be used to calculate the maximum medium uptake rates based on the fresh media metabolite
concentrations.
Metabolomics: The metabolomic constraints are obtained from quantitative targeted metabolo-
mics experiments using biological samples from two time-point measurements. The quality of the
data should be assessed carefully prior to the integration with the model by following the general
community standards (measurement accuracy, precision, blank effect, linearity of the calibration line,
relative errors of calibration line fit etc.) [26]. Furthermore, outlier samples should be identified (for
example, based on the median of detected metabolite levels) and excluded. Next, obtained data
should be transformed into metabolic rates and normalised for the biomass content (g dry weight)
by the following formula:
¯
metConcentrationEnd(umol/L) − metConcentrationStart(umol/L)
¯
vmet =
interval(hr) ∗ proteinConcentration(g/L)
¯ ∗ proteinF ractionInDW
16
It should be noted that the biomass content can also be estimated based on the cell count and an
estimate of the dry weight mass of a cell or by direct measurements of dry weight mass (tissues).
Furthermore, the error of the flux estimation should also be calculated by propagating the uncertainty
of all the non-exact values:
√
σ∆metConcentration = (σmetConcentrationEnd)2 + (σmetConcentrationStart)2
√ ( )2
( σ∆metConcentration )2 σproteinConcentration
∆metConcentration + ¯
proteinConcentration
· |vmet|
σν =
interval · proteinF ractionInDW
Last but not least, measured metabolites should be mapped into the model namespace where meta-
bolite IDs and/or exchange reaction IDs for each metabolite should be found (ideally by cross-mapping
standard metabolite IDs such as InChI, ChEBI, HMDB etc.), and included in the final metabolomics
input data.
The metabolomic constraints are set by fitting the bounds of the draft model to the exometabolomic
reaction rates while allowing relaxation of net flux bounds based on the quadratic optimisation
problem presented in Preciat et al. [24],
min (vexp − v)T diag(wexp )(vexp − v) + pT diag(wl )p + q T diag(wu )q

v,p,q∈Rn
s.t. Sv = 0,
Cv ≤ d, (2)
lb − p ≤ v ≤ ub + q,
0 ≤ p,
0 ≤ q,
where p ∈ Rn≥0 and q ∈ Rn≥0 are non-negative variables that allow the lower (lb) and upper bound
(ub) constraints to be relaxed. This problem always returns a steady-state flux v ∈ Rn and allows
for different information to be input as parameters, including the penalisation of deviation from
experimental fluxes (wexp ∈ Rn≥0 ), penalising relaxation of lower bounds (wl ∈ Rn≥0 ), and penalising
relaxation of upper bounds (wu ∈ Rn≥0 ). The weights for penalising the experimental fluxes are set
in param.metabolomicWeights and
and are derived from the mean ( 'mean', default), standard deviation ( 'SD'), or relative standard
deviation ( 'RSD') of the experimental reaction flux.
To avoid numerical errors in the model analysis, a default flux value is used if the experimental
data exceeds the specified boundary limits: the minimum reaction rate (param.TolMinBoundary), the
maximum reaction rate (param.TolMaxBoundary) as well as the absolute minimum flux allowed defined
in param.boundPrecisionLimit, i.e. any |lb| or |ub| < param.boundPrecisionLimit is considered as zero. If
param.printLevel is greater than zero, the new metabolomics-based constraints are printed. Finally,
the feasibility of the draft model is tested, and if it is not feasible, e.g., if there are inconsistencies
or errors in the exometabolomic data, some reaction bounds will be relaxed using relaxFBA. If the
debugging option is set, all results generated up to this point are saved in the file 10.debug_prior_-
to_custom_constraints.mat.
Usable variables:
• specificData.mediaData - Table containing the fresh media concentrations. Contains the reaction identifier, the
maximum uptake (umol/gDW /h) based on the concentration of the metabolite and the concentration (umol/l;
Default: empty).
rxnID mediumMaxUptake mediumConcentrations

EX_met1 -220 1345
EX_met2 -14916 191021
: : :
17
• specificData.exoMet - Table with the fluxes obtained from exometabolomic experiments. It includes the reaction
identifier, the reaction name, the measured mean flux, standard deviation of the measured flux, the flux units, and the
platform used to measure it.
rxnID rxnNames mean SD units platform

EX_mMet1 Exchange of mMet1 100 8x10−4 umol/gDW /h LC-MS
EX_mMet2 Exchange of mMet2 4x10−3 3x10−4 umol/gDW /h GC-MS
: : :
• param.metabolomicWeights - String indicating the type of weights to be applied for metabolomics fitting (Possible
options: 'SD', 'mean' and 'RSD'; Default: 'mean').
• specificData.essentialAA - List exchange reactions of essential amino acids (Default: empty). Must never be
secreted, even in a relaxed FBA model.
• param.TolMinBoundary - The reaction flux lower bound minimum value (Default: -1000)
• param.TolMaxBoundary - The reaction flux lower bound maximum value (Default: 1000)
• param.boundPrecisionLimit - Precision of flux estimate, if the absolute value of the lower bound or the upper bound
is lower than the boundPrecisionLimit but higher than 0 the value will be set to the boundPrecisionLimit (Default:
primal LP feasibility tolerance).
>> param.boundPrecisionLimit = getCobraSolverParams ( 'LP ' , ' f e a s T o l ' ) ;
• param.growthMediaBeforeReactionRemoval - Logical, should the cell culture data be added before (true) or after
(false) the model extraction (Default: true).
• param.metabolomicsBeforeModelExtraction - Logical, should the metabolomics data be added before (true) or
after (false) the model extraction (Default: true).
11 | Add custom constraints TIMING ∼ 10 s
Description:
Internal and external constraints on reactions can be added using the data in table
specificData.rxns2constrain, which will be used to change the lower and upper bounds of the reactions
in the draft model. Demand reactions will be ignored to ensure thermodynamic consistency if
'thermoKernel' is chosen as the tissue-specific solver. No change is made if the bound is empty. If the
bounds in the table exceed the established limit bounds, they will be adjusted to what the user or
default data specifies. If param.printLevel is greater than zero, the number of closed reactions, as well
as the number of open demand or sink reactions, will be printed. Finally, the feasibility of the draft
model is tested, and if it is not feasible some reactions will be relaxed via relaxFBA. If the debugging
option is set, all results generated up to this point are saved in the file 11.debug_prior_to_setting_-
coupled_reactions.mat.
Usable variables:
bound, a constraint description, and any notes such as references or special cases (Default: empty).
rxnID lb ub constrainDescription Notes

rxn2 0.0.956 1000 Turnover contraints PMID: ****
: : : : :
• param.TolMinBoundary - The reaction flux lower bound minimum value (Default: -1000)
• param.TolMaxBoundary - The reaction flux upper bound maximum value (Default: 1000)
• param.boundPrecisionLimit - Precision of flux estimate, if the absolute value of the lower bound or the upper bound
is lower than the boundPrecisionLimit but higher than 0 the value will be set to the boundPrecisionLimit (Default:
primal feasibility tolerance).
>> param.boundPrecisionLimit = getCobraSolverParams ( 'LP ' , ' f e a s T o l ' ) ;
18
12 | Set coupled reactions TIMING 10 s
Description:
Often published biochemical data specifies metabolites that are known to be produced or metabolised
by the biological system. However, a model frequently includes multiple production or degradation
pathways. To ensure that the net production or consumption of the metabolite is represented
correctly in the model the concept of coupled reactions can be used [30]. In non-growing, non-
dividing cells, such as neurons, these constraints can be used to replace the biomass reaction to
set constraints that specify specific biochemical requirements for cell maintenance [24]. In short, all
reactions known to produce (c = 1) or degrade (c = −1) the metabolite are identified and listed
in the specificData.coupledRxns (column coupledRxnsList), and the sum of their net fluxes is set to
be greater than (csence = G) the known net production/degradation of the metabolite (d). This
ensures that the specified reactions are active in the model and the total net flux through them
satisfies the set conditions, but the exact flux distribution between the reactions is not specified. To
this end, addCOBRAConstraints is used to add these constraints to the draft model via the inequality
c1 ∗ v(1) + c2 ∗ v(2) ∗ ... ∗ ci ∗ v(i) ≥ d or c1 ∗ v(1) + c2 ∗ v(2) ∗ ... ∗ ci ∗ v(i) ≤ d based on the
specification in model.csense. That is, csense ( 'E', equality, 'G' greater than, 'L' less than. The vector
c ∈ {−1, 1}i contains the coefficients indicating the directionality of each coupled reaction (1 for
production, −1 for degradation), and d is the constraint’s right-hand side C · v ≤ d specifying the
net value of the production/degradation of the metabolite ? TROUBLESHOOTING . The coupled
constraints are added to the draft model by including fields such as the constraint matrix containing
reaction coefficients, the constraint IDs, the constraint senses, and the constraint right-hand side
values. If param.printLevel is greater than zero, the information about the added coupled reactions
will be printed. If the debugging option is set, all results generated up to this point are saved in the
file 12.debug_prior_to_removing_inactive_reactions.mat.
Usable variables:
• param.addCoupledRxns - Logical, should the coupled constraints be added (Default: false). ! CAUTION If it is
TRUE and the table coupledRxns is empty, the step is not performed.
• specificData.coupledRxns - Table containing information about the coupled reactions. This includes the coupled
reaction identifier, the list of coupled reactions, the coefficients of those reactions, the constraint, the sense or the
directionality of the constraint, and the reference (Default: empty).
couplingConstraintID coupledRxnsList c d csence reference

cRxn1 crxn1, crxn2 11 2.0250 G PMID: ****
cRxn2 crxn1, crxn3 1 -1 0.2265 G PMID: ****
: : : : :
13 | Remove inactive reactions TIMING 10 s
Description:
The metabolic reactions assigned as inactive by manual literature curation are removed from the
draft model. In the event of a discrepancy between the datasets, for example, a reaction assigned
to be inactive based on manual literature curation but active according Step 6, preference will be
given based on the value of param.curationOverOmics (true - prioritise manual literature curation, false -
prioritise experimental data). After reaction removal, the feasibility of the model is tested, and if
it fails, some reaction bounds will be relaxed via relaxFBA. If param.printLevel is greater than zero,
the number of reactions removed and the reactions that were assigned as inactive and were removed
will be printed. If the debugging option is set, all results generated up to this point are saved in
13.debug_prior_to_removing_inactive_genes.mat.
19
Usable variables:
• specificData.inactiveReactions - List of reactions known to be inactive based on manual literature curation (Default:
empty).
• specificData.rxns2remove - List of reactions to remove (Default: empty).
false).
14 | Remove inactive genes TIMING ∼ 102 s
Description:
Genes assigned as inactive in bibliomic or transcriptome data are removed from the draft model,
together with the reactions affected by their deletion (identified using the function deleteModelGenes).
However, any reaction assigned to be active in the Step 6 will not be deleted at this step. Furthermore,
if removing reactions linked to inactive genes compromises the draft model’s feasibility, the bounds
of the fewest number of those reactions are relaxed via relaxFBA , and the rest are removed. To
ensure that the model is feasible, the relaxed reactions are added to the list of active reactions. In
the event of a discrepancy between the datasets, for example, a gene assigned to be inactive based
on the manual literature curation but active based on the proteomics data, or transcriptiomics data,
or both (Step 5), preference will be given based on the value of param.curationOverOmics variable. If
param.printLevel is greater than zero, the deleted reactions and genes will be printed, as well as those
that were not deleted.
Usable variables:
false).
• specificData.inactiveGenes - List of Entrez ID of genes known to be inactive based on the bibliomics data (Default:
empty).
15 | Set objective function TIMING 1 s
Description:
The linear objective φ(v) := cT v used in Flux Balance Analysis [22] can be set in this step (but is
not required), where c ∈ Rn represents the biologically inspired linear objective to find the optimal
flux vector. The objective function in param.setObjective represents the reaction whose flux is to be
maximised, such as ATP or biomass production, or minimised, such as energy consumption. However,
alternative biological objectives can also be used as shown in Preciat et al. [24].
Usable variables:
• param.setObjective - Cell string indicating the linear objective function to optimise (Default: none).
16 | Test feasibility TIMING ∼ 10 s
Description:
Several constraints have been integrated into the draft model up to this point; in this step, the final
feasibility check is performed before the consistency checks. If the draft model is not feasible, some
reactions will be relaxed via relaxFBA. If the debugging option is set, all results generated up to this
point are saved in the file 16.debug_prior_to_flux_consistency_check.mat.
20
17 | Find flux consistent subset TIMING ∼ 102 s
Description:
The goal of this step is to find and extract the largest subset of the draft model that is flux con-
sistent, i.e. each metabolite and reaction can carry a non-zero steady-state flux. The options
param.fluxCCmethod and param.fluxEpsilon indicate the algorithm that will be used to identify the flux
consistent subset of the draft model and to determine the minimum flux magnitude accepted. All
reactions that are flux inconsistent, and all metabolites that exclusively participate in flux inconsist-
ent reactions, are removed. If a flux inconsistent metabolite or reaction is identified as active in
specificData.presentMetabolites or in the list of active reactions from Step 6, it is removed from the list.
If param.printLevel is greater than zero, it prints the dimensions of the stoichiometric matrix before and
after checking flux consistency, as well as the constraints for flux consistent reactions and metabol-
ites. In addition, different boolean vectors indicating the flux consistent and inconsistent metabolites
and reactions are added to the draft model. If the debugging option is set, all results generated up
to this point are saved in the file 17.debug_prior_to_thermo_flux_consistency_check.mat.
Usable variables:
• specificData.presentMetabolites - List of metabolites known to be active based on the bibliomics data (Default:
empty).
• param.fluxCCmethod - String with the name of the algorithm to be used for the flux consistency check (Possible
options: ' swiftcc ', ' fastcc ' or 'dc', Default: ' fastcc ').
• param.fluxEpsilon - Minimum non-zero flux value accepted for tolerance (Default: Primal feasibility tolerance).
>> f e a s T o l = getCobraSolverParams ( 'LP ' , ' f e a s T o l ' ) ;
>> p a r a m . f l u x E p s i l o n = f e a s T o l ∗ 1 0 ;
18 | Find thermodynamically consistent subset TIMING ∼ 10 s
Description:
This step is only performed if param.tissueSpecificSolver= 'thermoKernel' is chosen as the extraction
algorithm since the performance of 'thermoKernel' is accelerated if it is provided a thermodynam-
ically consistent model prior to extracting a subset of it. An approximation to the largest ther-
modynamically consistent subset is implemented by 'findThermoConsistentFluxSubset', which requires
model.S, model.lb and model.ub as essential inputs. To accelerate this step one may optionally spe-
cify the metabolites and reactions that are flux consistent using model.fluxConsistentMetBool and
model.fluxConsistentRxnBool.
Any stoichiometric matrix S may be split into one subset of columns corresponding to internal and
external reactions, S = [N, B], where internal reactions are stoichiometrically consistent, that is
∃ℓ ∈ Rm T
>0 such that N ℓ = 0, and external reactions are not stoichiometrically consistent, that is
∄ℓ ∈ Rm T
>0 such that B ℓ = 0, as they represent net exchange of mass across the boundary of the
system. Previously, Desouki et al. [8] introduced the CycleFreeFlux algorithm, which reduces a given
flux vector to its thermodynamically feasible part, using a linear optimisation post-processing step.
Inspired by this incisive result, we observed that a thermodynamically feasible flux may be computed
as a single linear optimisation problem
min ∥z∥1 + cT · w
z,w
s.t. N z + Bw = 0 : r (3)
l≤z≤u :s
y≤w≤x :t
where ∥z∥1 denotes the one-norm of internal reaction fluxes, cT · w is a linear objective of external
reaction fluxes, while l ∈ {0, −∞}m and u ∈ {0, ∞}m denote lower and upper bounds on internal
21
reaction fluxes, y, x ∈ Rk denote lower and upper bounds on external reaction fluxes and the rest
are dual variables. The optimality conditions of Problem 3 are
N z + Bw = 0
∇ ∥z∥1 = sign(z) = N T y + s
c = BT y + t
Since l ∈ {0, −∞}m and u ∈ {0, ∞}m then sj, the dual variable to the inequality constraints on
internal reaction j, is non-zero if and only if zj⋆ is zero, that is s⋆j ̸= 0 ⇐⇒ zj⋆ = 0. Therefore zj⋆ ̸=
0 ⇐⇒ sign(z ⋆ ) = N T y ⋆ , which enforces energy conservation and the second law of thermodynamics
on the optimal vector of nonzero internal reaction fluxes [10]. It is important to reiterate that, for
internal reactions only, positive lower bounds and negative upper bounds are eschewed to prevent a
set of bounds from enforcing thermodynamically infeasible flux around a stoichiometrically balanced
cycle [8]. Therefore, internal reactions forced to be active, i.e., the reactions for which lb > 0 or
ub < 0 are assumed to be external reactions.
Problem 3 may be used to compute a single flux vector that is thermodynamically feasible. Specific-
ally, we define thermodynamically feasibility as the requirement that any net flux to be driven by a
change in chemical potential for the corresponding reaction, that is
vj > 0 ⇒ NjT y < 0,

vj < 0 ⇒ NjT y > 0,
vj = 0 ⇒ NjT y = 0,
where y is a vector of chemical potentials. However, the change of chemical potential for a reaction
may be non-zero, without a corresponding nonzero flux, representing the absence of an enzyme for
the corresponding reaction. This is a relaxation of the constraint sign(v) = −sign(N T y), since
NjT y ̸= 0 ⇏ v ̸= 0. However, only reactions with non-zero thermodynamically feasible net flux are
assigned to be thermodynamically consistent.
One must compute multiple thermodynamically feasible net flux vectors in order to approximate
the largest thermodynamically consistent subset of a given flux consistent model, because different
combinations of reactions can be active in thermodynamically feasible fluxes with disjoint sparsity
patterns. Inspired by the use of randomly weighted objectives to efficiently explore the set of flux
consistent reactions [27], by default, we employ a similar strategy to bias toward thermodynamically
feasible flux vectors that have non-zero flux in random subsets of reactions. A greedy sequence of
optimisation problems is then used to iteratively increase the number of reactions in the thermody-
namically consistent subset.
At each iteration the following cardinality optimisation problem is solved
min g T ∥z∥0 + β1T (p + q)

z,w,p,q
s.t. N z + Bw = 0
z−p+q =0
l≤z≤u (4)
y≤w≤x
0≤p
0≤q
where g T ∥z∥0 denotes the weighted zero-norm of internal reaction fluxes, where the zero-norm for
each reaction is weighted individually, that is
∑
n
g T ∥z∥0 := gjT ∥zj ∥0 .
j=1
Here g ∈ Rn is standard random vector generated from the zero-mean normal distribution in the open
interval (−0.5, 0.5) (the random part), except gj = 0 if the j th reaction has already been identified
22
as part of the thermodynamically consistent set (the greedy part). Minimisation of the one-norm of
internal net reaction flux, to promote thermodynamic feasibility [8], is achieved by constraining net
flux to be equal to the difference between two non-negative vectors, that is z = p − q and minimising
the sum of forward net reaction flux p ∈ Rn≥0 and reverse net reaction flux q ∈ Rn≥0 . Bounds on
z and w are the same as in Problem 3. Each iteration of Problem 4 is efficiently solved using a
sequence of linear optimisation problems, as described in detail elsewhere.
Given a solution Problem 4, thermodynamic consistency is checked using a linear optimisation step [8].
The scalar weight β > 0 is chosen to trade off between cardinality optimisation of internal reaction
fluxes, to explore the thermodynamically consistent subset of the generic model, and minimising the
one norm of the sum of forward and reverse reaction rates, to promote thermodynamic consistency.
By default, β = 0.1, which computational experiments determined to be approximately optimal for
genome-scale models tested. If β is too small, too few of the reaction fluxes are thermodynamically
feasible, while if β is too large, exploration of the thermodynamically consistent set is impeded.
All reactions that are not thermodynamically consistent are removed from the draft model. All
metabolites that are exclusively involved in thermodynamically inconsistent reactions are also removed
from the draft model. If a thermodynamically inconsistent metabolite or reaction is identified as active
in specificData.presentMetabolites or in the list of active reactions from Step 6, it is removed from the
list. Additionally, all orphan genes are removed from the model and the reaction-gene-matrix is
regenerated. Boolean vectors indicating thermodynamically consistent metabolites and reactions are
added to the draft model, model.thermoFluxConsistentMetBool and model.thermoFluxConsistentRxnBool
respectively.
If param.printLevel is greater than zero, it prints the dimensions of the stoichiometric matrix before and
after checking thermodynamic consistency, as well as the constraints for thermodynamic consistent
reactions and metabolites. If the debugging option is set, all results generated up to this point are
saved in the file 18.debug_prior_to_create_dummy_model.mat.
Usable variables:
• param.thermoFluxEpsilon - Flux epsilon used in thermoKernel (Default: feasibility tolerance).
• param.iterationMethod - Enables different iteration methods to be employed when exploring the thermodynamically
consistent set.
19 | Identify active reactions from genes TIMING ∼ 10 s
Description:
The list of active reactions up to this point has been based on the context-specific reactions and
flux consistency data. The relationship between context-specific genes and metabolic reactions is
established in this step. Each identified reaction is added to the final list of active reactions used
to extract the context-specific model. The pipeline determines the gene-reaction relationship using
either of the two options in param.activeGenesApproach:
'oneRxnPerActiveGene': At least one reaction per active gene is included (default).
'deleteModelGenes': Find a list of reactions, whose rates are affected by the deletion of active
genes and include them all as active reactions.
Changing this parameter has a large effect on the size of the extracted models. Per active gene,
'oneRxnPerActiveGene' adds at least one corresponding reaction to the extracted model, whereas
'deleteModelGenes' adds all corresponding reactions, where the presence of the gene product is essential
for nonzero flux through each of the reactions corresponding. Experience with generation of whole-
body metabolic models [31] and extraction of a neuronal model [24] from Recon3D supports the use
of 'oneRxnPerActiveGene' as the default. If the debugging option is set, all results generated up to
this point are saved in the file 19.debug_prior_to_create_tissue_specific_model.mat.
23
Usable variables:
• param.activeGenesApproach - String with the name of the active genes approach will be used (Possible options;
'oneRxnPerActiveGene', 'deleteModelGenes', Default: 'oneRxnPerActiveGene').
• param.printLevel - Level of verbose that should be printed (Default: 0).
• param.debug - Logical, should the function save its progress for debugging (Default: false).
20 | Model extraction TIMING ∼ 102 s
Description:
In this step, given a draft model from step 3.19, a specific model is extracted using the
createTissueSpecificModel interface. Two extraction algorithms may be used for this step, as indicated
in the param.tissueSpecificSolver: 'fastCore', which extracts a flux consistent context-specific model [33],
and 'thermoKernel', the default algorithm, which extracts a stoichiometrically, thermodynamically and
flux consistent context-specific model. The 'FastCore' algorithm has been previously described [33].
Below the approach of the 'thermoKernel' is described.
thermoKernel: Extraction of a thermodynamically consistent context-specific model Be-

sides defining lists of present metabolites (core metabolites) and active reactions (core reactions),
with 'thermoKernel' one has the option, to set weights on metabolites and reactions. Negative weights
promote inclusion of metabolites and reactions in a specific model, positive weights do the opposite,
while a zero weights does not penalise for exclusion or inclusion. By default, 'thermoKernel' sets
a fixed penalty for inclusion of any metabolite or reaction not in a core set. If gene expression
data is provided, this fixed penalty is equal to the median of the log of the reaction expression
value model.expressionRxns, otherwise it is one. Importantly, a weight of zero is set for any meta-
bolite ranked in the top 100 for metabolite connectivity. This is to avoid exclusion or inclusion of
a reaction if it contains cofactors, and rather focus on determining exclusion or inclusion based on
metabolites that participate in a small number of reactions. In addition, if param.activeGenesApproach
= oneRxnPerActiveGene, all metabolites and reactions in the core set are set a weight of equal to
the negative of the median of the log of the reaction expression value model.expressionRxns, or -1
if model.expressionRxns is not provided. If model.expressionRxns is provided, but a reaction is missing
gene expression data, it is assumed the corresponding reaction weight of equal to the negative of the
median of the log of the reaction expression values model.expressionRxns.
We assume that a metabolite is present if it is either produced or consumed at a non-zero rate
by a corresponding net flux vector, and absent otherwise. Any stoichiometric matrix may be split
into N = R − F were Fi,j and Ri,j are the stoichiometric numbers of the ith molecule consumed
and produced in the j th directed reaction, respectively. In terms of forward net reaction flux and
reverse net reaction flux, the rate of consumption of each metabolite is F p + Rq and the rate of
production of each metabolite is then F q + Rp. Therefore, letting N̄ := F + R, we observe that
s := (F + R)p + (F + R)q = N̄ p + N̄ q ≥ 0 represents the sum of the rate of production and
consumption of each metabolite. If metabolite i is not present in a network, then si = 0.
To compute a thermodynamically feasible flux that simultaneously quantitatively balances incentives
for presence of metabolites and activity of reactions, with disincentives for absence of metabolites
and inactivity of reactions following optimisation problem
min g T ∥z∥0 + β1T (p + q) + hT ∥s∥0 + f T ∥r∥0

z,w,p,q,s,r
s.t. N z + Bw = 0
z−p+q =0
N̄ p + N̄ q − s = 0
N̄ z − r = 0 (5)
l≤z≤u
y≤w≤x
0≤p
0≤q
24
which is an extension of Problem 4. When hi > 0, the objective term hTi ∥si ∥0 represents minim-
isation of the zero-norm of the rate of production and consumption of metabolite i. However, if
si > 0 then a metabolite may not be produced or consumed at a non-zero rate by a corresponding
net flux vector since p = q ̸= 0 implies s > 0. Therefore, we introduce the variable r := N̄ z as an
approximation to the sum of production and consumption due to net reaction flux, which is N̄ |z|,
but is non-trivial to implement. If rj := N̄j z > 0 then N̄j |z| > 0 as desired. Therefore, when
fi < 0, the objective term fjT ∥rj ∥0 approximates maximisation of the weighted zero-norm of the
sum of production and consumption of metabolite i due to non-zero net reaction flux. In practice,
we set f = −min(h, 0) so that optimisation of the cardinality of s and r is concordant.
Problem 5 computes a thermodynamically feasible flux that simultaneously quantitatively balances
incentives for presence of metabolites and activity of reactions, with disincentives for absence of
metabolites and inactivity of reactions. This step is also compatible with a requirement that the
extracted model be able to optimise a linear combination of net fluxes (not shown), as in flux
balance analysis.
However, it is usually the case that more than one thermodynamically feasible flux is required to
ensure activity of a set of desired reactions and presence of a set of metabolites. Therefore, we adopt
a randomly greedy strategy inspired by Tefagh and Boyd [27] whereby we choose a random vector
from a uniform rectangular set, d ∈ {Rn | unif(0, 1)}, then update weights on cardinality optimisation
of reactions in the next iterate g(n +1) based on those of the previous iterate g(n), with the heuristic


 gj (0) × (n + 1) gj (n) < 0, zj (n) = 0, d ≥ 0.5,




0 gj (n) < 0, zj (n) = 0, d < 0.5,
gj (n + 1) := 0 gj (n) < 0, zj (n) ̸= 0,



 gj (0) gj (n) > 0,



0 gj (n) = 0.
This approach successively increases the incentive for activity of internal reactions that are desired
to be active, but have not yet been active in a previous iterate. The strategy is the same to
incentivise presence of metabolites. The iterations either conclude when all incentivised reactions
and metabolites are active and present, respectively, or when a pre-specified maximum number of
iterations is reached. Each iteration of Problem 4 is efficiently solved using a sequence of linear
optimisation problems, as described in detail elsewhere.
Any model extraction algorithm: If param.printLevel is greater than zero, the size of the stoi-
chiometric matrix, the metabolites and reactions removed by the extraction solver, and the core
metabolites and reactions removed by the extraction solver are all printed. If the debugging option
is set, all results generated up to this point are saved in the file 20.debug_after_create_tissue_spe-
cific_model.mat.
Usable variables:
• param.activeGenesApproach - String with the name of the active genes approach will be used (Possible options:
'oneRxnPerActiveGene', 'deleteModelGenes', Default: oneRxnPerActiveGene).
(Possible options: 'thermoKernel' and 'fastCore'; Default: 'thermoKernel').
• param.weightsFromOmics - Should gene weights be assigned based on the omics data (Default: 0).
• param.thermoFluxEpsilon - Flux epsilon used in thermoKernel (Default: feasibility tolerance).
• param.fluxEpsilon - Flux consistency tolerance value (Default: primal feasibility tolerance).
>> f e a s T o l = getCobraSolverParams ( 'LP ' , ' f e a s T o l ' ) ;
>> p a r a m . f l u x E p s i l o n = f e a s T o l ∗ 1 0 0 ;
21 | Final adjustments TIMING ∼ 1 s
25
In the last step, the flux and thermodynamic consistency of the model is checked again. Also
specific data used in the extraction of the specific model are added to the model structure. If a
field with metabolic reaction formulas did not already exist in the draft model, it is now added
(using printRxnFormula function [5]). Also, two vectors are added that indicate which reactions were
specified to be active or inactive, as well as which metabolites were specified to be present or absent.
Finally, the fields of the context-specific model are reordered into a standard format.
26
ANTICIPATED RESULTS
Preparation of data
1 | The parameters that will be used by the 'XomicsToModel' pipeline are printed, including the
default values for options that have not been declared.
Generic model check

2 | Any modifications made at this step are indicated; for example, if the generic model contains
DM_atp_c_ reaction it is re-named to ATPM as it represents ATP maintenance and to ensure that
it is not closed together with standard demand reactions for the model extraction with 'thermoKernel'.
Furthermore, if the compartment representing mitochondrial intramembrane space ([i]) is removed
(if 'thermoKernel' is chosen) the modified reactions are also printed out.
Add missing reactions

3 | The reactions and formulas added in this step are printed, together with the results of the
stoichiometric consistency test. An example of the stoichiometric consistency test for Recon3D [7]
can be seen in Table 3
Table 3: Recon3D [7] stoichiometric consistency summary
Summary of stoichiometric consistency

5835 10600 totals
0 1809 heuristically external
5835 8791 heuristically internal
5835 8791 ... of which are stoichiometrically consistent
0 0 ... of which are stoichiometrically inconsistent.
0 0 ... of which are of unknown consistency.
0 0 heuristically internal and stoichiometrically inconsistent or unknown consistency.
0 0 ... of which are elementally imbalanced (inclusively involved metabolite).
0 0 ... of which are elementally imbalanced (exclusively involved metabolite).
5835 8791 Confirmed stoichiometrically consistent by leak/siphon testing.
Set limit bounds

4 | A draft model with the updated maximum upper bounds and minimum lower bounds.
Identify active genes

5 | A set of active genes was identified using bibliomic, transcriptomic, and proteomic data. In
addition, a message is displayed indicating the number of genes that have no expression information
due to a lack of transcriptomic data.
Identify active reactions

6 | A set of active reactions identified based on bibliomic, metabolomic, and cell culture data.
Close ions
7 | A draft model in which the upper and lower bounds of ion exchange reactions such as calcium,
potassium, or sodium, among others, are set to zero.
27
Close exchange reactions

8| A draft model where all the lower bounds of exchange reactions are set to zero to allow only
fresh media metabolites to be taken up. A message is shown reporting model statistics such as
the stoichiometric matrix dimensions, the number of closed reactions, the number of all exchange
reactions, and the number of exchange reactions in the active reactions.
Close sink and demand reactions

9 | The sink and demand reactions of the draft model are set to zero and a message is shown
reporting the number of closed non-core sink/demand reactions and the method used in
param.nonCoreSinksDemands.
Set metabolic constraints

10 | A draft model with cell culture and/or exometabolomic constraints added. In addition, a
message is displayed reporting the number of exometabolomic and media uptake constraints. It
includes exometabolomic data statistics, the fitting procedure for each exometabolomic reaction, the
relaxed reactions and the summary of the fitting.
Add custom constraints

11 | A draft model with custom constraints based on bibliomic data. If 'thermoKernel' is being
used, demand reactions included in specificData.rxns2constrain, i.e. with prefix DM_, will be ignored
and the message indicating this is being printed. Furthermore, a message is given showing a list of
reaction IDs that cannot be constrained because they are not present in the model, the number of
constrained reactions, and whether or not the model was still feasible after the addition of custom
constraints.
Set coupled reactions

12 | A draft model that includes new fields that correspond to the coupled reactions. A message
is shown including the details of the coupled constraints imposed on the draft model.
Remove inactive reactions

13 | A draft model, with inactive reactions removed (based on bibliomic data). A message including
the number of reactions that were removed from the model, a list of reactions that were not present
in the draft model prior to the reaction removal, as well as the results of the model feasibility check.
Remove inactive genes

14 | A draft model with inactive genes removed and a detailed description of the number of
genes that were removed or missing in draft model prior to this step. Based on the value of
param.curationOverOmics a message is given about the genes that were kept or removed from the
manually curated set of inactive genes due to the conflict with the provided omics data (measured
as active in transcriptomics and/or proteomics). Furthermore, information is given about the num-
ber of genes that were not removed since their removal would lead to an infeasible model. Last,
model feasibility check results are given, including any reaction that was relaxed if the model was
not feasible.
Set objective function

15 | A message including the specified linear objective function, or a lack of thereof.
Test feasibility
16 | This step generates a draft model with the bounds relaxed if the draft model is not feasible.
28
Find flux consistent subset

17 | A draft model with flux consistent reactions and metabolites is printed, along with a report
containing model statistics such as stoichiometric matrix dimensions, flux consistent and inconsistent
reactions and metabolites.
Find thermodynamically consistent subset

18 | A thermodynamically consistent draft model. Additionally, a message is given with the para-
meters used to identify the thermodynamically consistent subset. It is followed by a detailed report
from the cardinality optimisation. Last, the stoichiometric matrix dimensions, thermodynamically
consistent and inconsistent reactions, and metabolites are printed out.
Identify active reactions from genes

19 | A set of reactions that must be active based on the active genes.
Model extraction
20 | An extracted genome-scale metabolic model. If 'fastCore' is selected as the extraction al-
gorithm a message is printed out with the extraction summary and the size of the new model. If
'thermoKernel' is selected as the extraction algorithm, a message is printed out with the parameters
used in thermoKernel, the optimizeCardinality procedure, an extraction summary, and statistics for
the new model.
Figure 4: Comparison of specified and actual metabolites and reactions in an extracted model.
Output from the 'thermoKernel' algorithm for an example where there is a trade off between presence
and absence of metabolites and reactions in an extracted model.The predicted class is the specification
of metabolites by weights to be present (active) or absent (inactive) while the target class refers to
the presence (active) or absent (inactive) metabolites in the extracted model. The predicted class is
the specification of reactions by weights to be present (active) or absent (inactive) while the target
class refers to the presence (active) or absent (inactive) reactions in the extracted model.
Final
21 | The final context-specific flux consistent genome-scale metabolic model. If 'thermoKernel' was
chosen, then the model is also thermodynamically consistent.
To test the predictive capacity and size of alternative versions of the iDopaNeuro1 model, four mod-
els were generated with the function XomicsToMultipleModels (See supplementary information) while
excluding context-specific information such as bibliomics, metabolomics, transcriptomics, and fresh
medium concentrations (Figure 5). On average, the iDopaNeuro1 model with all constraints cor-
rectly predicts uptakes and secretions with the highest frequency based on the result performed by
29
the function modelPredictiveCapacity (See supplementary information). It can also be seen how differ-
ent context-specific information influences the number of independent variables in the stoichiometric
matrixFor example, culture medium information has the biggest influence on the rank(S) because it
determines which nutrients are available to the neuron, requiring activity of corresponding specific
metabolic pathways. This showcases the importance of the inclusion of high-quality, highly context-
specific data from multiple origins for an accurate description of the biological system of interest
[21, 25]. More extensive validation of the iDopaNeuro1 model and its application to study neuronal
energy metabolism has been described in Preciat et al. [24].
Figure 5: Predictive capacity, with all omics data (blue), and selected types of omics data omitted in the
iDopaNeuro1 model. Predictions based on eight cellular objectives as detailed in Preciat et al. [24]. 1) Inclusion
of all omics data leads to the highest qualitative (ratio of correct/total prediction of uptake/secretion/neither)
and quantitative predictive accuracy (Spearman rank of predicted versus measured exchange reaction rates).
2) Qualitative accuracy (ratio of correct/total prediction of uptake/secretion/neither) is not simply a function
of model flexibility, as measured by the rank of the stoichiometric matrix, rank(S).
30
? TROUBLESHOOTING
Step Problem Possible reason Solution

2| Infeasible draft The draft model may The
input model. be infeasible for a XomicsToModel pipeline aims to generate a
variety of reasons, feasible stoichiometrically-flux consistent
including flux or model; if this is not accomplished, either
stoichiometric the model or the included information
inconsistency, or needs to be modified.
because its bounds The
are over-constrained. fastGapFill function can add missing
reactions to make the draft model feasible.
Reactions can be added using the
addReaction function after additional
manual literature curation.
If it is over-constrained,
relaxFBA relaxes the minimum number of
reaction bounds in the draft model to make
it feasible.
3| RXN1 in New metabolic Place the metabolic reaction formula in the
specificData. reactions in the draft specificData.rxns2add.rxnFormulas for each
rxns2add.rxnID model require the reaction in the following format:
requires a reaction addition of the
formula in spe- metabolic reaction
'met1 + 2 met2 -> met3 + met4'
cificData.rxns2add. formula.
rxnFormulas where the type of reaction is indicated with
the proper characters (Forward (− >);
Reverse (< −); Bidirectional (<=>))
11 | Lower bounds A lower bound of a Check the bounds included in the
greater than upper reaction cannot be options variable including:
bounds greater than an
upper bound specificData.rxns2add,
lb ≯ ub.
param.TolMinBoundary,
param.TolMaxBoundary,
specificData.rxns2constrain and
specificData.exoMetdescribed in
Sections 3, 4, 10 and 10.
12 | Missing fields in To add the coupled Fill in the empty cells with the missing
the table. reactions, no element information as described in the Section 12.
of the table
specificData.
coupledRxns can be
left empty.
31
TIMING
Step 1, Preparation of data: ∼ 5 s

Step 2, Generic model check: ∼ 1 s
Step 3, Add missing reactions: ∼ 30 s
Step 4, Set limit bounds: ∼ 1 s
Step 5, Identify active genes: ∼ 102 s
Step 6, Identify active reactions: ∼ 10 s
Step 7, Close ions: ∼ 1 s
Step 8, Close exchange reactions: ∼ 1 s
Step 9, Close sink and demand reactions: ∼ 10 s
Step 10, Set metabolic constraints: ∼ 10 s
Step 11, Add custom constraints: ∼ 10 s
Step 12, Set coupled reactions: 10 s
Step 13, Remove inactive reactions: 10 s
Step 14, Remove inactive genes: ∼ 102 s
Step 15, Set objective function: 1 s
Step 16, Test feasibility: ∼ 10 s
Step 17, Find flux consistent subset: ∼ 102 s
Step 18, Find thermodynamically consistent subset: ∼ 10 s
Step 19, Identify active reactions from genes: ∼ 10 s
Step 20, Model extraction: ∼ 102 s
Step 21, Final: ∼ 1 s
ACKNOWLEDGEMENTS The authors would like to thank Hanneke Leegwater for her helpful
feedback while revising the manuscript. GP, AW, TH and RF received funding from the European
Union’s Horizon 2020 research and innovation programme, for the SysMedPD project, under grant
agreement No. 668738, and from the Dutch National Institutes of Health (ZonMw) TKI-LSH
Neuromet project (LSHM18092). AW, TH and RF received funding from Dutch Research Council for
the The Netherlands X-Omics Initiative, project 184.034.019. IT received funding from the European
Research Council under the European Unions Horizon 2020 research and innovation programme
(grant agreement No 757922), and by the National Institute on Aging grants (1RF1AG058942-01
and 1U19AG063744-01). RF received funding from the U.S. Department of Energy, Office of Science,
Biological and Environmental Research Program, award DE-SC0010429.
AUTHOR CONTRIBUTIONS CRediT (Contributor Roles Taxonomy) author statement
Author Contributions
German Preciat Conceptualisation; Methodology; Software; Tutorial;
Writing - Original draft.
Agnieszka Wegrzyn Conceptualisation; Methodology; Software; Data Curation
Writing - review and editing;
Ines Thiele Conceptualisation; Methodology; Software; Funding acquisition
Thomas Hankemeier Conceptualisation; Resources; Supervision; Funding acquisition
Ronan M.T. Fleming Conceptualisation; Methodology; Software;
Writing - Original Draft; Writing - review and editing;
Supervision; Funding acquisition
32
COMPETING FINANCIAL INTERESTS The authors declare that they have no competing
financial interests.
References
[1] Rasmus Agren et al. “Reconstruction of Genome-Scale Active Metabolic Networks for 69 Hu-
man Cell Types and 16 Cancer Types Using INIT”. In: PLOS Computational Biology 8.5 (May
2012), e1002518. doi: 10.1371/journal.pcbi.1002518.
[2] R. Balestrino and A. H. V. Schapira. “Parkinson Disease”. In: European Journal of Neurology
27.1 (Jan. 2020), pp. 27–42. doi: 10.1111/ene.14108.
[3] D. A. Beard, S.-D. Liang and H. Qian. “Energy Balance for Analysis of Complex Metabolic
Networks.” In: Biophysical Journal 83.1 (2002), pp. 79–86. doi: 10.1016/S0006-3495(02)
75150-3.
[4] Scott A. Becker and Bernhard O. Palsson. “Context-Specific Metabolic Networks Are Con-
sistent with Experiments”. In: PLOS Computational Biology 4.5 (May 2008), e1000082. doi:
10.1371/journal.pcbi.1000082.
[5] Scott A. Becker et al. “Quantitative Prediction of Cellular Metabolism with Constraint-Based
Models: The COBRA Toolbox”. In: Nature Protocols 2.3 (Mar. 2007), pp. 727–738. doi:
10.1038/nprot.2007.99.
[6] Stephen P Boyd and Lieven Vandenberghe. Convex Optimization. Cambridge, UK; New York:
Cambridge University Press, 2004.
[7] Elizabeth Brunk et al. “Recon3D Enables a Three-Dimensional View of Gene Variation in
Human Metabolism”. In: Nature Biotechnology 36 (Feb. 2018), p. 272.
[8] Abdelmoneim Amer Desouki et al. “CycleFreeFlux: Efficient Removal of Thermodynamically
Infeasible Loops from Flux Distributions”. In: Bioinformatics 31.13 (July 2015), pp. 2159–2165.
doi: 10.1093/bioinformatics/btv096.
[9] Dill, Ken and Bromberg, Sarina. Molecular Driving Forces Statistical Thermodynamics in Bio-
logy, Chemistry, Physics, and Nanoscience. Garland Science, 2011.
[10] R. M. T. Fleming et al. “A Variational Principle for Computing Nonequilibrium Fluxes and
Potentials in Genome-Scale Biochemical Networks.” In: Journal of Theoretical Biology 292
(2011), pp. 71–77. doi: 10.1016/j.jtbi.2011.09.029.
[11] Albert Gevorgyan, Mark G. Poolman and David A. Fell. “Detection of Stoichiometric In-
consistencies in Biomolecular Models.” In: Bioinformatics 24.19 (2008), pp. 2245–51. doi:
10.1093/bioinformatics/btn425.
[12] Steinn Gudmundsson and Ines Thiele. “Computationally Efficient Flux Variability Analysis”. In:
BMC Bioinformatics 11.1 (Sept. 2010), p. 489. doi: 10.1186/1471-2105-11-489.
[13] Hulda S. Haraldsdóttir et al. “CHRR: Coordinate Hit-and-Run with Rounding for Uniform
Sampling of Constraint-Based Models”. In: Bioinformatics 33.11 (Jan. 2017), pp. 1741–1743.
doi: 10.1093/bioinformatics/btx052.
[14] Laurent Heirendt et al. “Creation and Analysis of Biochemical Constraint-Based Models Using
the COBRA Toolbox v.3.0”. In: Nature Protocols 14.3 (Mar. 2019), p. 639. doi: 10.1038/
s41596-018-0098-2.
[15] Christopher S. Henry, Linda J. Broadbelt and Vassily Hatzimanikatis. “Thermodynamics-Based
Metabolic Flux Analysis”. In: Biophysical Journal 92.5 (Mar. 2007), pp. 1792–1805. doi:
10.1529/biophysj.106.093138.
33
[16] Livnat Jerby, Tomer Shlomi and Eytan Ruppin. “Computational Reconstruction of Tissue-
Specific Metabolic Models: Application to Human Liver Metabolism”. In: Molecular Systems
Biology 6.1 (Jan. 2010). doi: 10.1038/msb.2010.56.
[17] Daniel Machado et al. “Fast Automated Reconstruction of Genome-Scale Metabolic Models for
Microbial Species and Communities”. In: Nucleic Acids Research 46.15 (Sept. 2018), pp. 7542–
7553. doi: 10.1093/nar/gky537.
[18] Gregory L. Medlock, Thomas J. Moutinho and Jason A. Papin. “Medusa: Software to Build
and Analyze Ensembles of Genome-Scale Metabolic Network Reconstructions”. In: PLOS Com-
putational Biology 16.4 (Apr. 2020), e1007847. doi: 10.1371/journal.pcbi.1007847.
[19] Alberto Noronha et al. “The Virtual Metabolic Human Database: Integrating Human and Gut
Microbiome Metabolism with Nutrition and Disease”. In: Nucleic Acids Research 47.D1 (Jan.
2019), pp. D614–D624. doi: 10.1093/nar/gky992.
[20] Charles J. Norsigian et al. “A Workflow for Generating Multi-Strain Genome-Scale Metabolic
Models of Prokaryotes”. In: Nature Protocols 15.1 (Jan. 2020), pp. 1–14. doi: 10 . 1038 /
s41596-019-0254-3.
[21] Sjoerd Opdam et al. “A Systematic Evaluation of Methods for Tailoring Genome-Scale Meta-
bolic Models”. In: Cell Systems 4.3 (Mar. 2017), 318–329.e6. doi: 10.1016/j.cels.2017.
01.010.
[22] Jeffrey D. Orth, Ines Thiele and Bernhard Ø Palsson. “What Is Flux Balance Analysis?” In:
Nature Biotechnology 28.3 (Mar. 2010), pp. 245–248. doi: 10.1038/nbt.1614.
[23] Bernhard Ø Palsson. Systems Biology: Constraint-Based Reconstruction and Analysis. Cam-
bridge, England: Cambridge University Press, Jan. 2015.
[24] German Preciat et al. Mechanistic Model-Driven Exometabolomic Characterisation of Human
Dopaminergic Neuronal Metabolism. July 2021. doi: 10.1101/2021.06.30.450562.
[25] Jonathan L. Robinson and Jens Nielsen. “Integrative Analysis of Human Omics Data Using
Biomolecular Networks”. In: Molecular bioSystems 12.10 (Oct. 2016), pp. 2953–2964. doi:
10.1039/c6mb00476h.
[26] Vinod P. Shah et al. “Bioanalytical method validation–a revisit with a decade of progress”. In:
Pharmaceutical Research 17.12 (2000), pp. 1551–1557. issn: 07248741. doi: 10.1023/A:
1007669411738. url: http://link.springer.com/10.1023/A:1007669411738 (visited
on 08/11/2021).
[27] Mojtaba Tefagh and Stephen P. Boyd. “SWIFTCORE: A Tool for the Context-Specific Recon-
struction of Genome-Scale Metabolic Networks”. In: BMC Bioinformatics 21.1 (Apr. 2020),
p. 140. doi: 10.1186/s12859-020-3440-y.
[28] I. Thiele and B. Ø. Palsson. “A Protocol for Generating a High-Quality Genome-Scale Meta-
bolic Reconstruction.” In: Nature Protocolss 5.1 (2010), pp. 93–121. doi: 10.1038/nprot.
2009.203.
[29] Ines Thiele, Nikos Vlassis and Ronan M. T. Fleming. “fastGapFill: Efficient Gap Filling in
Metabolic Networks”. In: Bioinformatics (Oxford, England) 30.17 (Sept. 2014), pp. 2529–
2531. doi: 10.1093/bioinformatics/btu321.
[30] Ines Thiele et al. “Functional Characterization of Alternate Optimal Solutions of Escherichia
Coli’s Transcriptional and Translational Machinery”. In: Biophysical Journal 98.10 (May 2010),
pp. 2072–2081. doi: 10.1016/j.bpj.2010.01.060.
[31] Ines Thiele et al. “Personalized Whole-Body Models Integrate Metabolism, Physiology, and
the Gut Microbiome”. In: Molecular Systems Biology 16.e8982 (2020), p. 24.
[32] Kalyan C. Vinnakota et al. “Network Modeling of Liver Metabolism to Predict Plasma Meta-
bolite Changes During Short-Term Fasting in the Laboratory Rat”. In: Frontiers in Physiology
10 (2019), p. 161. doi: 10.3389/fphys.2019.00161.
34
[33] Nikos Vlassis, Maria Pires Pacheco and Thomas Sauter. “Fast Reconstruction of Compact
Context-Specific Metabolic Network Models”. In: PLoS Comput Biol 10.1 (Jan. 2014), e1003424.
doi: 10.1371/journal.pcbi.1003424.
[34] Hao Wang et al. “Genome-Scale Metabolic Network Reconstruction of Model Animals as a
Platform for Translational Research”. In: Proceedings of the National Academy of Sciences
118.30 (July 2021). doi: 10.1073/pnas.2102344118.
[35] Yuliang Wang, James A. Eddy and Nathan D. Price. “Reconstruction of Genome-Scale Meta-
bolic Models for 126 Human Tissues Using mCADRE”. In: BMC Systems Biology 6.1 (Dec.
2012), p. 153. doi: 10.1186/1752-0509-6-153.
[36] Hadas Zur, Eytan Ruppin and Tomer Shlomi. “iMAT: An Integrative Metabolic Analysis Tool”.
In: Bioinformatics 26.24 (Dec. 2010), pp. 3140–3142. doi: 10 . 1093 / bioinformatics /
btq602.
35
Supplementary information
Supplementary Manual 1 - Additional Tools

A brief explanation of how to use the additional tools for the XomicsToModel pipeline.
Supplementary Manual 2 - XomicsToModel tutorial

A MATLAB live-script tutorial on how to use the XomicsToModel pipeline.
36
Supplementary Information 1 - Additional tools
Preprocessing Omics Data

Description:
In order to facilitate data integration, the function preprocessingOmicsModel prepares the context-
specific and technical information that will be used by the XomicsToModel pipeline. It is used by
running:
>> s p e c i f i c D a t a = preprocessingOmicsModel ( inputData , setMinActiveFlux , ...
setMaxActiveFlux , s p e c i f i c D a t a )
The variable inputData is a string describing the location of the file (path) and the file name for the
context-specific data. Each sheet in the table is added as a field in the specificData variable and, if
necessary, converted to the corresponding format. CRITICAL STEP The names of the sheets
must be identical to the fields described in each step and as shown in Figure S1, otherwise, the data
will be ignored by the XomicsToModel pipeline.
Figure S1: An example of the inputData file.
If Force activity is indicated in the column specificData.rxns2constrain.constrainDescription, but no lb or

ub values are provided, the parameters setMinActiveFlux and setMaxActiveFlux are used to make the
flux of a reaction nj to be vj > 0 (reaction will be forced to be active). The lower bound of the
reaction is then set to the feasibility tolerance multiplied by 100, while the upper bound is changed
to the maximum feasibility tolerance (1000, or param.TolMaxBoundary if is present). If lb or ub values
are provided in the data file they will not be changed.
Furthermore, the media uptakes can also be calculated in this step based on the initial metabolite
concentration in the media (defined either by the manufacturer or from measurements). It is necessary
to have the cell culture information provided in table specificData.cellCultureData for this step. The
following formula shows how the uptakes are calculated, and which variables need to be included:
metaboliteConcentration(umol/L) ∗ uptakeSign ∗ volume(L) ∗ proteinF raction

interval(hr) ∗ averageP roteinConcentration(gDW /L) ∗ assayV olume(L)
Lastly, it is also possible to include an older version of the specificData variable in order to update the
context-specific data.
1
Used variables:
bound, a description for the constraint and notes such as references or special cases (Default: empty).

rxn1 0.001 Force activity PMID: ****
• specificData.cellCultureData - Table containing the cell culture data used to calculate the uptake flux. Includes well
volume (mL), time interval cultured (hrs), average protein concentration (g/L), assay volume (L), protein fraction
(dimensionless), and the sign for uptakes (Default: empty).
volume interval averageProteinConcentration assayVolume proteinFraction uptakeSign
0.002 48 0.3989 0.00045 0.706 -1
XomicsToMultipleModels
Description:
Taking advantage of the flexibility of the XomicsToModel pipeline, the XomicsToMultipleModels function
can be used to generate an ensemble of context-specific genome-scale models by varying the options
used in the function. The function is used to represent experiments with varying time points or
measurement platforms, different cases (i.e. mutant vs WT), use different generic models, or identify
the conditions that are most similar to a training set. For each model, a directory with the names
of the different parameters will be created in which the diary and/or variables from debugging can
be saved. The parameters include different optimization solvers, transcriptomic threshold or active
genes approach.
To run the XomicsToMultipleModels function the command used is:
>> d i r e c t o r i e s = XomicsToMultipleModels ( modelGenerationConditions , param , ...
replaceModels )
Used variables:
• modelGenerationConditions.activeGenesApproach - The different approached described in Step 19 (Possible op-
tions: 'deleteModelGenes' and 'oneRxnPerActiveGene'; default: 'oneRxnperActiveGene');
• modelGenerationConditions.boundstoRelaxExoMet - The type of bounds that can be relaxed, upper bounds,
lower bounds or both ('b'; possible options: 'u', ' l ' and 'b'; default: 'b');
• modelGenerationConditions.closeIons - Indicate whether the ions are open or closed (Possible options: true and
false; default: false);
• modelGenerationConditions.cobraSolver - Optimisation solvers supported by the pipeline. Possible options: 'glpk',
'gurobi', 'ibm_cplex', 'matlab'; default: 'gurobi');
• modelGenerationConditions.genericModel - Generic COBRA model(s);
• modelGenerationConditions.inactiveGenesTranscriptomics - Use inactive transcriptomic genes or not (Possible
options: true and false; default: false);
• modelGenerationConditions.specificData - Specific data to be used (Default: empty);
• modelGenerationConditions.limitBounds - Boundary on the model (Default: 1000);
• modelGenerationConditions.tissueSpecificSolver - Extraction solver (Possible options: 'fastCore'[33] and
'thermoKernel'; default: 'thermoKernel');
• modelGenerationConditions.outputDir - Directory where the models will be generated (Default: current directory);
• modelGenerationConditions.transcriptomicThreshold - Transcriptomic thresholds that are defined by the user
(Default: log2(2));
• param - Variable with fixed parameters (Default: empty struct array);
• replaceModels - Logical, It is used to determine whether or not the models of an existing directory should be replaced
(Default: false).
2
modelPredictiveCapacity
Description:
The function modelPredictiveCapacity can be used to verify a model’s predictive capacity when evalu-
ating one or more models. Three tests can be performed by the function: flux consistency, thermo-
dynamic flux consistency, and predictions against a training set.
During flux consistency and thermodynamic flux consistency tests, it is determined whether active or
inactive metabolites and reactions have a consistent flux. On the other hand, the predictive capacity
of the model to qualitatively and quantitatively predict fluxes is determined by comparing it to a
training data set with constraints for different reactions. The function modelPredictiveCapacity relaxes
the model’s upper and lower bounds for the reactions in the training set, then the objective functions
supported by the function are used to predict the directionality of the fluxes as well as their euclidean
distance and Spearman correlation to the training set.
modelPredictiveCapacity returns two variables; the variable comparisonData contains detailed informa-
tion about the comparisons between predicted fluxes and the training dataset. It displays a table
containing the consistency of metabolites and reactions, the predicted and measured fluxes (training
dataset), and the final comparison. Furthermore, the variable summary contains a summary of the
results.
The data used for Figure 5 was obtained using modelPredictiveCapacity.
>> [ comparisonData , summary ] = m o d e l P r e d i c t i v e C a p a c i t y ( model , param )
Used variables:
• model - A Cobra model to be tested;
• param.tests - Array with the tests run on the model (Possible options: 'fluxConsistent': Flux consistency;
'thermoConsistentFlux': Thermodynamic flux consistency; ' flux ': Objective function comparison based on a training
set; ' all ': Do all tests; Default: ' all ').;
• param.activeInactiveRxn - nx1 with entries {1, −1, 0} depending on whether a reaction must be active, inactive,
or unspecified respectively;
• param.presentAbsentMet - nx1 with entries {1, −1, 0} depending on whether a metabolite must be present, absent,
or unspecified respectively;
• param.trainingSet - Table with the training set. It includes the reaction identifier, the reaction name, the measured
mean flux and standard deviation of the flux (Required for param.tests = 'flux').
rxnID rxnNames mean SD
EX_mMet1 Exchange of mMet1 100 8x10−4
EX_mMet2 Exchange of mMet2 4x10−3 3x10−4
: :
• param.objectives - List of objective functions to be tested (Required for param.tests = 'flux'; Default: ' all ').
3
Supplementary Information 2 - XomicsToModel pipeline tutorial


Xo Mics To Model

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Xo Mics To Model

Uploaded by

Copyright:

Available Formats

bioRxiv preprint doi: https://doi.org/10.1101/2021.11.08.467803; this version posted November 10, 2021.

The copyright holder for this

XomicsToModel: Multiomics data integration and generation of

8th November 2021

German Preciat1∗ , Agnieszka B. Wegrzyn1∗ , Ines Thiele2,3,4,5 ,

Constraint-based modelling, multi-omics, data integration, thermodynamic consistency, systems bio-

Introduction to genome-scale metabolic model extraction

Development of the protocol

Ensemble modelling To exploit the XomicsToModel pipeline flexibility, a supplementary

Applications of the XomicsToModel pipeline

Comparison with other methods

GIMME iMAT MBA INIT mCADRE FastCore XomicsToModel

Generic reconstruction or model: A generic genome-scale reconstruction or model represents a

Context-specific data: Context-specific data represents the genotype or phenotype of a specific

• A hard drive with at least 10 GB of space available.

• MATLAB (MathWorks, https://mathworks.com/products/matlab.html), version R2021+.

Table 2: An overview of the optimisation solvers XomicsToModel supports.

Name Version Interface

Equipment setupCOBRA Toolbox

1 | Preparation of data TIMING ∼ 5 s

2 | Generic model checks TIMING ∼ 1 s

3 | Add missing reactions TIMING ∼ 30 s

• Forward (->): lb = 0; ub =param.TolMaxBoundary.

• Reverse (<-): lb =param.TolMinBoundary; ub = 0.

• Reversible (<=>) lb =param.TolMinBoundary; ub =param.TolMaxBoundary.

rxnID rxnNames rxnFormulas subSystems rxnGrRules rxnReferences

4| Set limit bounds TIMING ∼ 1 s

5 | Identify active genes TIMING ∼ 102 s

• Bibliomics: Genes classified as active or inactive based on published data.

6 | Identify active reactions TIMING ∼ 10 s

rxnID rxnNames rxnFormulas subSystems rxnGrRules rxnReferences

rxnID lb ub constraintDescription Notes

couplingConstraintID coupledRxnsList c d dsence reference

rxnID mediumMaxUptake mediumConcentrations

rxnID rxnNames mean SD units platform

7 | Close ions TIMING ∼ 1 s

8 | Close exchange reactions TIMING ∼ 1 s

rxnID lb ub constraintDescription Notes

9 | Close sink and demand reactions TIMING ∼ 10 s

10 | Set metabolic constraints TIMING ∼ 10 s

min (vexp − v)T diag(wexp )(vexp − v) + pT diag(wl )p + q T diag(wu )q

rxnID mediumMaxUptake mediumConcentrations

rxnID rxnNames mean SD units platform

11 | Add custom constraints TIMING ∼ 10 s

rxnID lb ub constrainDescription Notes

12 | Set coupled reactions TIMING 10 s

couplingConstraintID coupledRxnsList c d csence reference

13 | Remove inactive reactions TIMING 10 s

14 | Remove inactive genes TIMING ∼ 102 s

15 | Set objective function TIMING 1 s

16 | Test feasibility TIMING ∼ 10 s

17 | Find flux consistent subset TIMING ∼ 102 s

18 | Find thermodynamically consistent subset TIMING ∼ 10 s

vj > 0 ⇒ NjT y < 0,

min g T ∥z∥0 + β1T (p + q)

19 | Identify active reactions from genes TIMING ∼ 10 s

'oneRxnPerActiveGene': At least one reaction per active gene is included (default).

20 | Model extraction TIMING ∼ 102 s

thermoKernel: Extraction of a thermodynamically consistent context-specific model Be-

min g T ∥z∥0 + β1T (p + q) + hT ∥s∥0 + f T ∥r∥0

21 | Final adjustments TIMING ∼ 1 s

Generic model check

Add missing reactions