Available online at www.sciencedirect.
com
ScienceDirect
Advances in gap-filling genome-scale metabolic models
and model-driven experiments lead to novel metabolic
discoveries
Shu Pan1,2 and Jennifer L Reed1,2
With rapid improvements in next-generation sequencing
technologies, our knowledge about metabolism of many
organisms is rapidly increasing. However, gaps in metabolic
networks exist due to incomplete knowledge (e.g., missing
reactions, unknown pathways, unannotated and misannotated
genes, promiscuous enzymes, and underground metabolic
pathways). In this review, we discuss recent advances in gap-
filling algorithms based on genome-scale metabolic models
and the importance of both high-throughput experiments and
detailed biochemical characterization, which work in concert
with in silico methods, to allow a more accurate and
comprehensive understanding of metabolism.
Addresses
1 g., growth phenotypes). They then solve for a set of
Department of Chemical and Biological Engineering, University of
Wisconsin-Madison, 1415 Engineering Dr, Madison, WI 53706, United
States
2
Great Lakes Bioenergy Research Center, Madison, WI 53706, United
States
Corresponding author: Reed, Jennifer L (reed@engr.wisc.edu)
Current Opinion in Biotechnology 2018, 51:103–108
This review comes from a themed issue on Systems biology
Edited by Nathan Price and Eran Segal
For a complete overview see the Issue and the Editorial
Available online 23rd December 2017
https://doi.org/10.1016/j.copbio.2017.12.012
0958-1669/ã 2017 Elsevier Ltd. All rights reserved.
Introduction
A genome-scale metabolic model is a mathematical repre-
sentation of the metabolic capabilities of an organism,
which is inferred primarily from genome annotations.
Such models have shown great utility in predicting bio-
logical capabilities, metabolic engineering, and systems
medicine [1–4]. A draft model is often generated auto-
matically by software platforms, which use genome anno-
tations of a specific organism and connect genes to meta-
bolic reactions using reference databases [5,6]. A draft
model has to be further refined and evaluated in multiple
steps to ensure its quality [5,6]. This refinement and
evaluation process includes gap-filling, which improves
the network connectivity by modifying content of the
www.sciencedirect.com
metabolic model. Gap-filling analyses can lead to discov-
eries of missing reactions, unknown pathways, unanno-
tated and misannotated genes, as well as promiscuous
enzymes and underground metabolic pathways. Classic
gap-filling algorithms have been reviewed previously by
Orth and Palsson [7]. These algorithms generally include
three steps: detecting gaps, suggesting model content
changes (i.e., add/remove reactions, change biomass
compositions, or change reaction reversibility), and iden-
tifying genes responsible for the gap-filled reactions
(Figure 1). In the first step, gap-filling algorithms identify
dead-end metabolites (metabolites which cannot be con-
sumed or produced in the network), and/or inconsisten-
cies between model predictions and experimental data (e.
reactions from metabolic databases of potential reactions
that if added to the metabolic model ‘activate’ dead-end
metabolites or resolve the inconsistencies. In the third
step, some gap-filling algorithms discover genes that
could be responsible for these reactions, which can be
further tested biochemically or genetically. A simple gap-
filling example is illustrated in Figure 2. Here, we first
review recent gap-filling methods, which are more effi-
cient or operate under different assumptions. Then we
discuss how advances in experimental techniques have
significantly advanced gap-filling methods by identifying
model-data inconsistencies. Finally, we describe recent
studies that have used gap-filling analyses to discover the
promiscuous functions of enzymes.
Advances in gap-detection and reaction-
addition algorithms
Some recent algorithms aim to detect and fill gaps more
efficiently than earlier gap-filling algorithms [7]. For exam-
ple, FASTGAPFILL [8] is a scalable algorithm that com-
putes a near minimal set of added reactions for a compart-
mentalized model. Another method, GLOBALFIT [9],
reformulates the mixed integer linear programming prob-
lem of gap-filling into a simpler bi-level linear optimization
problem. It efficiently identifies the minimal set of net-
work changes needed to correct multiple in silico predic-
tions that are inconsistent with in vivo observations simul-
taneously. Meneco [10] and a hybrid metabolic network
completion algorithm [11] reformulate the reaction-addi-
tion problem using answer set programming, a declarative
programming paradigm intended to solve difficult combi-
natorial search problems. Their usage of answer set pro-
gramming allows for stoichiometry constraints to be
Current Opinion in Biotechnology 2018, 51:103–108
104 Systems biology

Figure 1

Gap-Filling & Promiscuous Enzyme Discovery Input Data &

Computational Algorithms

detect dead-end metabolites

gaps in vivo / in silico inconsistencies
metabolomics data
knockout phenotypes (e.g. Tn-seq)

databases of potential reactions

suggest
model reaction
network topology
13
content x reaction z C fluxomics data
reaction
changes
y2
metabolomics data

gap-filled reactions promiscuous enzymes alternative pathways

gene x gene y gene z
sequence similarity
genomic-context data
find genomic functional selections
genes & protein X protein Y protein Z knockout phenotypes (e.g. Tn-seq)
change GPR
biochemical characterization

reaction x reaction y1 reaction y2 reaction z

open gaps
closed gaps using gap-filled reactions
closed gaps using promiscuous enzymes
closed gaps using alternative pathways
Current Opinion in Biotechnology

Steps, input data, and computational algorithms of gap-filling. First, dead-end metabolites, in silico and in vivo inconsistencies, metabolomics
data, and knockout phenotypes allow detection of gaps in metabolic models. Then, the model content (i.e., reactions and biomass compositions)
is changed to resolve these inconsistencies. In this step, missing reactions can be added from databases, and network topology analysis can rank
these potential reactions. Metabolomics and 13C fluxomics data could also suggest reactions that should be included in the model. Finally, the
genes responsible for the filled gaps are identified using sequence similarity, genomic-context data, genomic functional selections, or knockout
phenotypes and are verified by biochemical characterization. Similarly, promiscuous enzymes and underground metabolic pathways can also be
identified when analyzing the gaps in the models.

violated, potentially resulting in solutions which are less
biased by the inaccurate stoichiometry of a model. The
hybrid metabolic network completion approach combines
answer set programming with linear stoichiometry con-
straints, and offers a better solution for restoring highly
degraded models [11].
Algorithms, such as GAUGE [12
], have been developed
to exploit alternative mechanisms for gap-identification.
GAUGE exploits flux coupling analysis (FCA) [13] that
detects how two reactions depend on each other. Using
FCA, GAUGE finds gaps involving genes that are associ-
ated with fully dependent reactions but show uncorre-
lated expression patterns. However, GAUGE can only
analyze a subset of a model where gene-protein-reaction
(GPR) associations are defined and isozymes or multi-
functional genes do not create possibilities for uncorre-
lated gene expression patterns.
Some algorithms exploit alternative mechanisms for add-
ing reactions. The novel algorithm DEF [14] is based
upon filling reactions to ‘activate’ dead-end metabolites
in a manner similar to eukaryotes engulfing mitochondria
to find the most efficient pathways for consuming oxygen.
Following this quasi-endosymbiosis theory, DEF aims to
add reactions that maximize production/consumption of
dead-end metabolites in the original model. A DEF
solution could contain more reactions that are biologically
reasonable compared to a parsimonious gap-filling solu-
tion, which often is a minimal set of reactions.
Inherently different from all algorithms mentioned
above, pattern-based gap-filling algorithms do not contain
an explicit gap-identification or reaction-addition step. In
a metabolite pattern and probabilistic method [15], fea-
ture propagation Markov models (HMMs) are used to
rank potential gap-filled reactions by how closely they are
related to the network. In MATBoost and BoostGAP-
FILL [16,17
], a training incidence matrix, S, with artifi-
cial gaps is created by deleting reactions randomly from a
network. Then a machine learning technique, matrix
factorization, completes the missing entries creating

Current Opinion in Biotechnology 2018, 51:103–108 www.sciencedirect.com

 Advances in gap-filling genome-scale metabolic models Pan and Reed 105

Figure 2

growth phenotyping draft model with missing reactions

and inaccurate gene-protein-reaction associations

metabolite reaction a
A
metabolite reaction b
A 1 0 0 gene a gene c gene d
Ca
reaction c B 0 0 0
metabolite C 0 1 0
C D 0 0 1 protein A protein C protein D
metabolite reaction d
optical density

D
reaction a reaction c reaction d
growth a b c d

time

improved model with gap-filled reactions biochemical characterization reaction addition from databases
and promiscuous enzyme added
atgCCTATCG
gene a gene b gene c gene d gene b ...ACGCCAG database
A 1 0 0 0 ...TGCCAAtta
B 0 1 0 0
C 0 0 1 0 database
D 0 0 0 1 atgTTCATTG
protein A protein B protein C protein D
gene d ...TTGTTAG
...CGCCATtta

reaction a reaction b reaction c reaction d

a b c d

Current Opinion in Biotechnology

An example of the iterative process of laboratory and computational experimentation in gap-filling. As revealed by high-throughput growth
phenotyping experiments, the wild type can grow on metabolites A, B, or D as a sole carbon source. Additionally, a knockout mutant of gene c
that is known to catalyze reaction c can grow on metabolite C. Contrarily, the model incorrectly predicts that the wild type cannot grow on B and
the gene c mutant cannot grow on C, indicating that reaction b is missing from the model and there exists another enzyme that can catalyze
reaction c. To fix these inconsistencies, gap-filling adds reaction b from databases and finds that gene d has sequence similarity to gene
c. Further biochemical characterization confirms that protein B is responsible for reaction b and protein D has weak activities towards reaction c.
The model is then updated and compared to additional experimental datasets to allow iterative metabolic discoveries and to improve prediction
accuracy.

matrix A. Finally, an integer least square optimization explored step which should be supported with additional
selects reactions from a database that best match A. These experiments.
steps, which are based on the network pattern alone,
cannot fill in all the gaps in the network. However, they Advances in gene-assignment algorithms and
can rank and set weights of reactions in databases and approaches
work in conjunction with other gap-filling algorithms such Knowing the gene sequence(s) responsible for a reaction
as FASTGAPFILL [8] to improve their gap-filling is extremely useful in applications using genetic manip-
accuracy. ulation or drug discoveries. Several algorithms that have
been previously reviewed [7] provide bioinformatics solu-
Unlike the above algorithms that add reactions to an tions to match gene sequences to gap-filled reactions.
incomplete metabolic model of any organism, SONEC These methods could incorporate data such as sequence
[18
] works for an organism in a metagenomics sample. It similarity, co-expression, chromosomal proximity, and
is initialized with a bin of contigs for each organism in the phylogenetic profiles. Recently, GLOBUS [19] has com-
sample and a bin for unassigned sequence fragments. bined these data with a global probabilistic approach to
Then it computes metabolite connectivity scores provide probable annotations including possible alternate
between the unassigned fragments and each organism functions. Another novel algorithm GO-MEP [20] iden-
to determine which organism these fragments belong to. tifies biologically relevant groups of proteins (e.g., pro-
teins in the same pathways, protein complexes, proteins
The abovementioned gap-detection and reaction-addi- with the same localization, and physically interacting
tion algorithms are used in the first two steps of gap-filling protein pairs) and has showed improved accuracy of
(Figure 1). They can hypothesize new network compo- identifying missing genes.
nents including gap-filled reactions. After the two steps,
genes are assigned to gap-filled reactions (Figure 1). This An alternative approach for gene assignment is to directly
last step of gap-filling is a very important yet under- incorporate gene assignment in the gap-filling procedure.

www.sciencedirect.com Current Opinion in Biotechnology 2018, 51:103–108

106 Systems biology
In MIRAGE [21], phylogenetic profiles and gene expres-
sion are first used to estimate the likelihood that a specific
reaction in the database can be added. In the next step,
starting from a model with all reactions in the database
added, a new model is created to determine the set of gap-
filled reactions by removing low likelihood reactions
while keeping the desired model properties. Another
likelihood-based gap-filling algorithm [22] and its web
service (ProbAnnoWeb) and standalone python package
(ProbAnnoPy) [23], have been developed to directly
incorporate gene assignment into gap-filling. In ProbAn-
noWeb and ProbAnnoPy, an organism-specific likelihood
score is assigned to each reaction in a database of potential
reactions. These scores are calculated based on the
BLASTp results of a gene in the organism genome
against high confidence functional annotations. Reactions
with higher scores are weighted heavier in the objective
function. Therefore, a ProbAnno solution might contain
more reactions than a parsimonious gap-filling solution.
Similar to ProbAnno, there exists another BLAST-
weighted gap-filling algorithm [24] in which reactions
are added from weighted biochemistry databases.
Ultimately, all gap-filled gene-reaction associations
should be evaluated in subsequent laboratory experi-
ments. For example, a change in gene expression might
be measured by quantitative PCR (qPCR) or growth
phenotypes of knockout mutants under different media
conditions. Or the product of an enzyme’s proposed
reaction could be measured by analytical chemistry meth-
ods. Recently an integrated computational and experi-
mental approach, MEGS [25

], has been developed that
allows the direct and rapid identification of genes encod-
ing for gap-filled reactions. In MEGS, gap-filled reactions
in a model are identified with a classic gap-filling proce-
dure by resolving model and experiment inconsistencies.
A genomic functional selection then can be designed
using a forced-coupling algorithm [26] such that a recipi-
ent strain from a well-characterized organism can only
grow in a selective medium if it acquires the gene(s)
encoding for the gap-filled reaction. Finally, a genomic
library from a donor strain catalyzing the proposed gap-
filled reaction is transformed into the recipient strain and
a genomic functional selection is performed to identify
the gene(s) of interest. Using MEGS, five genes were
successfully identified for the gap-filled reactions in the
draft model of the marine bacterium, Vibrio fischeri.
Identifying inconsistencies between model
predictions and high-throughput phenotyping
experiments
Laboratory experiments are not only important to validate
gap-filling solutions, but also are important for identifying
the metabolic gaps themselves. Cheaper and more avail-
able high-throughput experimental datasets, which can
be compared to in silico predictions, have also made gap-
filling analyses increasingly powerful [27,28]. For
Current Opinion in Biotechnology 2018, 51:103–108
example, gap-filling algorithms could improve the quality
of a metabolic model by resolving incorrect growth phe-
notype predictions for knockout mutants under different
medium conditions. Currently, phenotypic microarrays
and robotic instruments [29] can test growth of cells in
fifty 96-well plates for hundreds of different media con-
ditions. RB-TnSeq has also allowed rapid quantification
of genome-wide mutant fitness in 387 successful assays
[30]. With improved genome editing tools such as MAGE
[31,32], pORTMAGE [33], and CRMAGE [34], more
knockout libraries will be added to the current collections
[35–41] in the near future. Such libraries can be rapidly
tested under a variety of conditions to identify incon-
sistencies between model predictions and experimental
observations and to identify metabolic gaps.
Limitations in current gap-filling algorithms
Despite the increasing power of newer gap-filling meth-
ods to complete our knowledge in metabolic networks,
major challenges still exist. First, a prevalent problem
among these algorithms is that most cannot resolve false-
positive predictions (negative growth in vivo and positive
growth in silico). Simply removing reactions or limiting
reaction directionality to resolve false-positive predic-
tions may not be successful, since unknown regulatory
rules or essential biomass components could instead be
the source of the issue. Gap-filling algorithms might also
add reactions without genomic evidence to models and
risk overfitting when trying to activate dead-end metab-
olites or resolve false-negative model predictions (posi-
tive growth in vivo and negative growth in silico). Third,
solutions that reconcile experiment and model inconsis-
tencies might be non-unique, with different solutions
being hard to rank or prioritize. These solutions could
also change depending on which experiment and model
inconsistency is solved first. When one inconsistency is
fixed, another new inconsistency might be created. Unin-
tuitively, a global gap-filling procedure, where multiple
inconsistencies are solved simultaneously, may result in a
larger set of gap-filled reactions and may take significantly
longer to run as the number of inconsistencies increases
[42].
Discoveries of promiscuous enzymes and
underground metabolic pathways via gap-
filling analyses
In addition to discoveries of unannotated or misannotated
genes, gap-filling procedures can identify promiscuous
activities of pre-existing enzymes that are recruited by
‘underground’ metabolic pathways [43] to bypass the
more commonly used pathways (Figures 1 and 2). Tradi-
tionally, enzyme promiscuity is evaluated using multi-
copy suppression experiments where overexpression of a
single gene from a plasmid library rescues a conditionally
lethal knockout [44,45]. However, these experiments
have limited utility because it can be tricky to obtain
optimal expression levels and only a fraction of genes are
www.sciencedirect.com
 Advances in gap-filling genome-scale metabolic models Pan and Reed 107
conditionally lethal in common growth media [46

].
Recently, gap-filling analyses have been combined with
experimental characterization for discoveries of promis-
cuous enzymes. Predicted gaps in central metabolic path-
ways and cofactor biosynthesis have led to refined anno-
tations for Dehalobacter restrictus and discovery of a D.
restrictus promiscuous serine hydroxymethyltransferase
[47]. Experiments where knockout and synthetic lethal
mutants grow under different media conditions are useful
for discovering underground reactions that do not take
place in wild-type strains under standard media condi-
tions (e.g., rich media). An extensive study of Escherichia
coli multiple-knockout mutants on 13 carbon sources
confirmed underground reactions using metabolomics
analysis and in vitro enzymatic assays [48]. A novel work-
flow has been developed to resolve incorrect model pre-
dictions of gene essentiality using sequence similarity
analysis, knockout strain analysis, and adaptive evolution.
Such a workflow has revealed promiscuous functions of
several E. coli enzymes [49]. However, it cannot reveal
long underground metabolic pathways containing multi-
ple promiscuous enzymes and novel intermediate metab-
olites. A purely computational approach, PROPER [46

],
has been developed to predict promiscuous enzymes by
searching distant gene similarities and by analyzing phy-
logenetic trees to assign secondary functions to genes. A
related model-based approach (GEM-PROPER) [46

]
has been designed to identify underground metabolic
pathways. Unlike the analyses that have combined
experiments and gap-filling, the accuracies of PROPER
and GEM-PROPER predictions still need to be evalu-
ated. So far 10 out of the 63 PROPER predicted E. coli
promiscuous enzymes and 1 out of the 98 GEM-
PROPER predicted pathways have been validated.
Conclusions
Recent developments in gap-filling algorithms for
genome-scale metabolic models have showed great
potential for revealing missing knowledge of metabolic
reactions, gene-protein-reaction associations, and promis-
cuous enzymes. These algorithms are diverse in the terms
of experimental data input, formulation strategies, and
computational methods. Novel techniques from machine
learning, network topology analysis, likelihood modeling,
and bioinformatics research have been incorporated into
these methods. In the future, gap-filling algorithms could
incorporate kinetic, thermodynamics and regulatory con-
straints for more accurate predictions. With the advances
in high-throughput phenotyping technologies and omics
measurement, gap-filling algorithms will generate more
and better hypotheses. Meanwhile, biochemical charac-
terization that reveals the details and caveats of novel
metabolic discoveries should be performed to inform and
validate in silico predictions.
Conflict of interest
None.
www.sciencedirect.com
Acknowledgements
The authors wish to thank Matthew R. Long, Paul A. Adamczyk, and
Sanjan Gupta for their comments on the manuscript. This work was
supported by the National Science Foundation (CBET-1053712); and the
U.S. Department of Energy Great Lakes Bioenergy Research Center (DOE
BER Office of Science DE-FC02-07ER64494).
References and recommended reading
Papers of particular interest, published within the period of review,
have been highlighted as:

of special interest


of outstanding interest
1. Zhang C, Hua Q: Applications of genome-scale metabolic
models in biotechnology and systems medicine. Front Physiol
2016:6.
2. King ZA, Lloyd CJ, Feist AM, Palsson BO: Next-generation
genome-scale models for metabolic engineering. Curr Opin
Biotechnol 2015, 35:23-29.
3. O’Brien EJ, Monk JM, Palsson BO: Using genome-scale models
to predict biological capabilities. Cell 2015, 161:971-987.
4. McCloskey D, Palsson BO, Feist AM: Basic and applied uses of
genome-scale metabolic network reconstructions of
Escherichia coli. Mol Syst Biol 2014, 9 661-661.
5. Thiele I, Palsson BØ: A protocol for generating a high-quality
genome-scale metabolic reconstruction. Nat Protoc 2010,
5:93-121.
6. Hamilton JJ, Reed JL: Software platforms to facilitate
reconstructing genome-scale metabolic networks. Environ
Microbiol 2014, 16:49-59.
7. Orth JD, Palsson B: Systematizing the generation of missing
metabolic knowledge. Biotechnol Bioeng 2010, 107:403-412.
8. Thiele I, Vlassis N, Fleming RMT: FASTGAPFILL: efficient gap
filling in metabolic networks. Bioinformatics 2014, 30:2529-
2531.
9. Hartleb D, Jarre F, Lercher MJ: Improved metabolic models for
E. coli and Mycoplasma genitalium from GlobalFit, an
algorithm that simultaneously matches growth and non-
growth data sets. PLOS Comput Biol 2016, 12:e1005036.
10. Prigent S, Frioux C, Dittami SM, Thiele S, Larhlimi A, Collet G,
Gutknecht F, Got J, Eveillard D, Bourdon J et al.: Meneco, a
topology-based gap-filling tool applicable to degraded
genome-wide metabolic networks. PLoS Comput Biol 2017, 13:
e1005276.
11. Frioux C, Schaub T, Schellhorn S, Siegel A, Wanko P: Hybrid
metabolic network completion. LPNMR 2017: Logic
Programming and Nonmonotonic Reasoning. 2017:308-321.
12. Hosseini Z, Marashi S-A: Discovering missing reactions of

metabolic networks by using gene co-expression data. Sci
Rep 2017, 7:41774.
This algorithm, GAUGE, is the first algorithm that incorporates flux
coupling analysis for gap-identification. It allows the incorporation of
gene expression datasets.
13. Burgard AP, Nikolaev EV, Schilling CH, Maranas CD: Flux
coupling analysis of genome-scale metabolic network
reconstructions. Genome Res 2004, 14:301-312.
14. Liu L, Zhang Z, Sheng T, Chen M: DEF: an automated dead-end
filling approach based on quasi-endosymbiosis. Bioinformatics
2016, 33:btw604.
15. Ganter M, Kaltenbach H-M, Stelling J: Predicting network
functions with nested patterns. Nat Commun 2014, 5:3006.
16. M. Zhang, Z. Cui, T. Oyetunde, Y. Tang and Y. Chen, Recovering
metabolic networks using a novel hyperlink prediction
method. 2016. arXiv:1610.06941.
17. Oyetunde T, Zhang M, Chen Y, Tang Y, Lo C: BoostGAPFILL:

improving the fidelity of metabolic network reconstructions
Current Opinion in Biotechnology 2018, 51:103–108
108 Systems biology
through integrated constraint and pattern-based methods.
Bioinformatics 2016, 33:btw684.
BoostGAPFILL is a novel pattern-based gap-filling method that uses
matrix factorization, a machine learning technique. It can be used as a tool
for ranking reactions in databases.
18. Biggs MB, Papin JA: Metabolic network-guided binning of

metagenomic sequence fragments. Bioinformatics 2016,
32:867-874.
This work introduces a unique algorithm, SONEC, which is specially
designed for completing networks of organisms in metagenomics
samples.
19. Plata G, Fuhrer T, Hsiao T-L, Sauer U, Vitkup D: Global
probabilistic annotation of metabolic networks enables
enzyme discovery. Nat Chem Biol 2012, 8:848-854.
20. Chitale M, Khan IK, Kihara D: Missing gene identification using
functional coherence scores. Sci Rep 2016, 6:31725.
21. Vitkin E, Shlomi T: MIRAGE: a functional genomics-based
approach for metabolic network model reconstruction and its
application to Cyanobacteria networks. Genome Biol 2012, 13:
R111.
22. Benedict MN, Mundy MB, Henry CS, Chia N, Price ND:
Likelihood-based gene annotations for gap filling and quality
assessment in genome-scale metabolic models. PLoS Comput
Biol 2014, 10:e1003882.
23. King B, Farrah T, Richards M, Mundy M, Simeonidis E, Price ND:
ProbAnnoWeb and ProbAnnoPy: probabilistic annotation and
gap-filling of metabolic reconstructions. bioRxiv, 2017. 151258.
24. Krumholz EW, Libourel IGL: Sequence-based network
completion reveals the integrality of missing reactions in
metabolic networks. J Biol Chem 2015, 290:19197-19207.
25. Pan S, Nikolakakis K, Adamczyk PA, Pan M, Ruby EG, Reed JL:


Model-enabled gene search (MEGS) allows fast and direct
discovery of enzymatic and transport gene functions in the
marine bacterium Vibrio fischeri. J Biol Chem 2017, 292:10250-
10261.
MEGS is an integrated computational and experimental approach that
allows the identification of gap-filled reactions and the gene-protein-
reaction associations with a high level of confidence.
26. Tervo CJ, Reed JL: FOCAL: an experimental design tool for
systematizing metabolic discoveries and model development.
Genome Biol 2012, 13:R116.
27. Joyce AR, Reed JL, White A, Edwards R, Osterman A, Baba T,
Mori H, Lesely Sa, Palsson BØ, Agarwalla S: Experimental and
computational assessment of conditionally essential genes in
Escherichia coli. J Bacteriol 2006, 188:8259-8271.
28. Szappanos B, Kovács K, Szamecz B, Honti F, Costanzo M,
Baryshnikova A, Gelius-Dietrich G, Lercher MJ, Jelasity M,
Myers CL et al.: An integrated approach to characterize genetic
interaction networks in yeast metabolism. Nat Genet 2011,
43:656-662.
29. Shea A, Wolcott M, Daefler S, Rozak DA: Biolog phenotype
microarrays. In Microbial Systems Biology: Methods and
Protocols. Edited by Navid A. Humana Press; 2012:331-373.
30. Wetmore KM, Price MN, Waters RJ, Lamson JS, He J, Hoover CA,
Blow MJ, Bristow J, Butland G, Arkin AP et al.: Rapid
quantification of mutant fitness in diverse bacteria by
sequencing randomly bar-coded transposons. mBio 2015, 6:
e00306-e315.
31. Gallagher RR, Li Z, Lewis AO, Isaacs FJ: Rapid editing and
evolution of bacterial genomes using libraries of synthetic
DNA. Nat Protoc 2014, 9:2301-2316.
32. Wang HH, Isaacs FJ, Carr PA, Sun ZZ, Xu G, Forest CR,
Church GM: Programming cells by multiplex genome
engineering and accelerated evolution. Nature 2009, 460:894-
898.
33. Nyerges Á, CsörgÅ B, Nagy I, Bálint B, Bihari P, Lázár V, Apjok G,
Umenhoffer K, Bogos B, Pósfai G et al.: A highly precise and
portable genome engineering method allows comparison of
Current Opinion in Biotechnology 2018, 51:103–108
mutational effects across bacterial species. Proc Natl Acad Sci
USA 2016, 113:2502-2507.
34. Ronda C, Pedersen LE, Sommer MOA, Nielsen AT: CRMAGE:
CRISPR optimized MAGE recombineering. Sci Rep 2016,
6:19452.
35. Côté J, French S, Gehrke SS, MacNair CR, Mangat CS, Bharat A,
Brown ED: The genome-wide interaction network of nutrient
stress genes in Escherichia coli. mBio 2016, 7:e01714-e1716.
36. Giaever G, Nislow C: The yeast deletion collection: a decade of
functional genomics. Genetics 2014, 197:451-465.
37. Yamamoto N, Nakahigashi K, Nakamichi T, Yoshino M, Takai Y,
Touda Y, Furubayashi A, Kinjyo S, Dose H, Hasegawa M et al.:
Update on the Keio collection of Escherichia coli single-gene
deletion mutants. Mol Syst Biol 2009, 5:335.
38. Baba T, Ara T, Hasegawa M, Takai Y, Okumura Y, Baba M,
Datsenko KA, Tomita M, Wanner BL, Mori H: Construction of
Escherichia coli K-12 in-frame, single-gene knockout
mutants: the Keio collection. Mol Syst Biol 2006, 2 2006.0008.
39. Kobayashi K, Ehrlich SD, Albertini A, Amati G, Andersen KK,
Arnaud M, Asai K, Ashikaga S, Aymerich S, Bessieres P et al.:
Essential Bacillus subtilis genes. Proc Natl Acad Sci USA 2003,
100:4678-4683.
40. Deutschbauer A, Price MN, Wetmore KM, Shao W, Baumohl JK,
Xu Z, Nguyen M, Tamse R, Davis RW, Arkin AP: Evidence-based
annotation of gene function in Shewanella oneidensis MR-1
using genome-wide fitness profiling across 121 conditions.
PLoS Genet 2011, 7:e1002385.
41. Skerker JM, Leon D, Price MN, Mar JS, Tarjan DR, Wetmore KM,
Deutschbauer AM, Baumohl JK, Bauer S, Ibanez AB et al.:
Dissecting a complex chemical stress: chemogenomic
profiling of plant hydrolysates. Mol Syst Biol 2014, 9 674-674.
42. Biggs MB, Papin JA: Managing uncertainty in metabolic
network structure and improving predictions using
EnsembleFBA. PLoS Comput Biol 2017, 13:1-25.
43. Notebaart RA, Kintses B, Feist AM, Papp B: Underground
metabolism: network-level perspective and biotechnological
potential. Curr Opin Biotechnol 2018, 49:108-114.
44. Patra B, Kon Y, Yadav G, Sevold AW, Frumkin JP,
Vallabhajosyula RR, Hintze A, Stman B, Schossau J, Bhan A et al.:
A genome wide dosage suppressor network reveals genomic
robustness. Nucleic Acids Res 2017, 45:255-270.
45. Patrick WM, Quandt EM, Swartzlander DB, Matsumura I:
Multicopy suppression underpins metabolic evolvability. Mol
Biol Evol 2007, 24:2716-2722.
46. Oberhardt MA, Zarecki R, Reshef L, Xia F, Duran-Frigola M,


Schreiber R, Henry CS, Ben-Tal N, Dwyer DJ, Gophna U et al.:
Systems-wide prediction of enzyme promiscuity reveals a new
underground alternative route for pyridoxal 50 -phosphate
production in E. coli. PLoS Comput Biol 2016, 12:1-19.
This work introduces in silico comparative genomics tools to identify both
promiscuous enzymes and underground metabolic pathways. It gener-
ates unintuitive hypotheses that could not be obtained using traditional
multicopy suppression experiments.
47. Wang P-H, Tang S, Nemr K, Flick R, Yan J, Mahadevan R, F
Yakunin A, Löffler FE, Edwards EA: Refined experimental
annotation reveals conserved corrinoid autotrophy in
chloroform-respiring Dehalobacter isolates. ISME J 2017,
11:626-640.
48. Nakahigashi K, Toya Y, Ishii N, Soga T, Hasegawa M, Watanabe H,
Takai Y, Honma M, Mori H, Tomita M: Systematic phenome
analysis of Escherichia coli multiple-knockout mutants
reveals hidden reactions in central carbon metabolism. Mol
Syst Biol 2009, 5:306.
49. Guzmán GI, Utrilla J, Nurk S, Brunk E, Monk JM, Ebrahim A,
Palsson BO, Feist AM: Model-driven discovery of underground
metabolic functions in Escherichia coli. Proc Natl Acad Sci USA
2015, 112:929-934.
www.sciencedirect.com