1

MetaboliticsDB: A Database of Metabolomics

Analyses
M. Hasan Celik2 , Onurcan Ersen1 , Taj Saleh2 , Alper Dokay2 , and Ali Cakmak1 *
1
Istanbul Technical University, 2 Istanbul Sehir University
*
Corresponding Author: ali.cakmak@itu.edu.tr

Abstract—Web-based metabolomics databases enable researchers to disseminate metabolite concentration datasets measured
under different physiological conditions. In addition, many of these databases offer a number of tools to process (e.g., normalization,
outlier elimination, etc.) and analyze (e.g., clustering, enrichment studies, etc.) users’ metabolomics data sets. Nevertheless, none of
the existing metabolomics databases offer infrastructure and tools to store, manage, compare, and search metabolomics analysis
results. Besides, their pathway-level analysis capabilities are mostly limited to superimposing the measurements onto the pathways of
the measured metabolites. In this paper, we present MetaboliticsDB that features a database of metabolomics analyses and a set of
associated analytics tools. It enables users to store their metabolomics analysis results, and compare them against their own or other
publicly available analysis results to study, for instance, the progression of a disease, the effect of a drug, similarities between
well-known physiological conditions and the currently studied data, etc. Besides, MetaboliticsDB allows querying the metabolomics
analysis results database with flexible criteria, such as, listing all analyses where a certain pathway experiences a major
increase/decrease in activity to help researchers identify conditions sharing a similar metabolic mechanism. Moreover, MetaboliticsDB
offers a genome-scale metabolic network-based analysis tool that significantly extends the capabilities of the existing databases.
Finally, MetaboliticsDB employs AI-based as well as distance-based methods to associate the studied metabolomics data with
diseases stored in its database. To this end, it automatically trains, manages, and updates machine learning models based on the
stored metabolomics analysis data stored in its database for each disease. We demonstrate the use of MetaboliticsDB with a case
study on Hepatocellular Carcinoma. Our results show that MetaboliticsDB provides biologically relevant metabolic network-level
analysis results, disease association with high accuracy, and a scalable architecture supporting hundreds of simultaneous users.
Availability: MetaboliticsDB is available online at http://metabolitics.itu.edu.tr/.
Web interface source codes are available at https://github.com/itu-bioinformatics-database-lab/metabolitics-client.
Web API source codes are available at https://github.com/itu-bioinformatics-database-lab/metabolitics-api.
Source codes of the Metabolitics data analysis algorithm are available at
https://github.com/itu-bioinformatics-database-lab/metabolitics.

Index Terms—Metabolomics, Biological Databases, Personalized Medicine.

1 I NTRODUCTION

M ETABOLOMICS is the study of concentration changes

for a large number of metabolites in cells as well as
extracellular environments (e.g., blood). It provides invalu-
able insights regarding physiological conditions, as the phe-
notype of diseases are often reflected on the metabolome of
an organism. With the recent advancements in experimental
methods, researchers are now able to measure the amount
of many metabolites with high accuracy at a relatively
low cost. The main challenge has been interpreting these
measurements to understand the health and disease states
of cells, and accordingly, develop new diagnosis, treatment,
and prognosis methods. Many methods and algorithms
have been proposed to analyze metabolomics data at dif-
ferent levels of granularity (e.g., [1] [2], [3], [4], [5], [6], [7],
[8], [9], [10], [11], [12], [13], [14], [15], [16], [17], [18]). Despite
the large variety of metabolomics data analysis algorithms
• M.H. Celik, T. Saleh, and A. Dokay were with the Department of
Computer Science, Istanbul Sehr University, Istanbul, Turkey.
• O. Ersen and A. Cakmak are with the Department of Computer
Engineering, Istanbul Technical University, Istanbul, Turkey. E-mail:
ali.cakmak@itu.edu.tr Sehir University, Istanbul, Turkey.
in literature, relatively few of them ( [12], [13], [14], [15],
[16]) have been made available to the researchers over web-
based databases. These works at a high-level fall under two
categories: (i) metabolic data resources and (ii) database-
enabled metabolomics data analysis resources.
Metabolic data resources ( [19], [20], [21], [22], [23], [24],
[25], [26]) have been around for a relatively longer time,
and they usually act as data management and dissemination
hubs for the research community. The maintainers of such
resources compile their data by (i) generating (fully or
partially) in their own labs, (ii) collecting from literature
(manually or in a semi-automated manner), and (iii) inte-
grating data from multiple external sources. Such resources
usually feature basic browsing, searching, and visualization
capabilities. Among several, the most widely used pathway
resources in this category are KEGG [19], BioCyc [20], Re-
actome [21], etc. Besides, some other data resources focus
on metabolites rather than pathways, such as HMDB [22],
CheBI [23], PubChem [25], etc.
Database-enabled metabolomics data analysis resources
( [12], [13], [14], [15], [16], [27]) usually work on data that is
imported from one or multiple of the above metabolic data

resources. In addition to the basic searching and visualiza-
tion capabilities provided by the metabolic data resources,
this category of tools also allows users to upload their
own metabolomics data, and then the analysis results are
provided to the user in tabular and/or graphical form.
One of the most comprehensive analysis resources in this
category is MetaboAnalyst 5.0 [28]. In addition to basic
statistical significance and discrimination analysis tools at
the metabolite level, it also features pathway-level analysis
in the form of pathway enrichment and topology-based
assessment. Moreover, MetaboAnalyst 5.0 allows integrated
analysis of transcriptomics and metabolomics datasets at
the pathway level. Metabolomics Workbench [13] is another
comprehensive metabolomics data repository and analysis
resource. It offers a wide array of statistical analysis tools at
the metabolite level, ranging from running ANOVA analysis
to hierarchical clustering analysis. It also allows creation
of various forms of plots on top of user data, such as
boxplots, volcano plots, bar graphs, etc. However, it does
not provide any pathway-level analysis. MeltDB 2.0 [14]
offers a number of features to annotate the metabolites,
detect peaks, eliminate noise, etc. in user-submitted raw
data. Once the data is preprocessed, users may perform
statistical significance tests, classification and clustering, and
various forms of visualization on their data. One differ-
entiating feature of MeltDB 2.0 is that it provides built-in
support to form project groups, share particular data files
among group members or with different project groups, and
create and manage different experiments. MetabolomeEx-
press [15] offers similar capabilities to MeltDB 2.0 in the
categories of raw data processing and statistical analysis
of the processed data. XCMS Online [16] takes multi-omics
integration one step further than MetaboAnalyst 5.0, and
accommodates proteomics (in addition to transcriptomics)
along with metabolomics measurements. It also features
other common metabolite- and pathway-level analysis fea-
tures (e.g., raw data processing, statistical analysis, pathway
enrichment analysis, etc.) similar to the above tools. Caleydo
[17] allows mapping omics data on pathways and nicely
visualizes them so that users can see which pathways are be-
ing covered by the uploaded omics data, and to what degree
their activities change based on the concentration change
of metabolites in a metabolomics dataset, or the change in
mRNA levels in a gene expression dataset. WebSpecmine
[29] is a web-based metabolomics data analysis and min-
ing tool built on the specmine R package. WebSpecmine
enables users to upload datasets to process, visualize, and
perform various analyses. WebSpecmine also trains machine
learning models and predicts classes for future samples.
3Omics [30] is a web-based human metabolomics data anal-
ysis tool that offers visualization and analysis capabilities
on the uploaded datasets with an emphasis on combining
different omics data. Workflow4Metabolomics [31] offers
data preprocessing with retention time alignment and peak
extraction, and univariate analysis with nonparametric and
parametric tests. POMAShiny [32] provides data prepro-
cessing with missing value imputation, normalization, and
outlier detection, univariate analysis with t-test, ANOVA,
Mann-Whitney U-test, and Kruskal-Wallis test. It also of-
fers clustering with the k-means algorithm and classifica-
tion with the random forest algorithm. In addition to the
2
above web tools, several others offer somewhat similar
metabolomics analysis features, but are available only as
stand-alone desktop applications, such as Pathway Tools
[18], SIMCA-P+ (Umetrics, Umea, Sweden), etc.
Even though the above database-enabled metabolomics
analysis resources are quite extensive in the number and
variety of the raw data processing and statistical analysis
features that they offer, we note the following gaps: (i)
Although some of the existing resources allow users to
store their raw metabolomics data, none of them enable
users to store, manage, compare, and search metabolomics
analysis results. (ii) Their pathway-level analysis capabilities
are limited to superimposing measured metabolite changes
onto their corresponding pathways. Moreover, all the
above resources consider each pathway independently from
the metabolic network that they belong to. Even though
MetaboAnalyst computes network-level measures, such as
betweenness, centrality, etc., it only uses them for ranking
pathways. Hence, the assessment in the above resources is
limited to only those pathways whose metabolites overlap
with the user-submitted measurements to some extent. Even
then, the included pathways are evaluated individually by
ignoring the production/consumption/regulation relation-
ships among them. (iii) Existing resources do not attempt
to associate metabolomics analysis results with known dis-
eases or physiological conditions.
In this paper, we present a novel database-enabled web
resource, MetaboliticsDB, that performs metabolic activity
analysis of user-provided metabolomics data in a holis-
tic manner considering interconnections between pathways
with mass-balances preserved. To this end, it employs a
state of the art systems-level algorithm [2] under the hood.
MetaboliticsDB stores analysis results in its database, and
users may compare current analysis results with previously
stored analysis results of their own or other publicly avail-
able analysis results shared by other users. Furthermore,
MetaboliticsDB allows users to make a comparison between
different analysis methods (e.g., Metabolitics vs. pathway
enrichment) on the same dataset. Users may also flexibly
search the stored analysis results to list those where certain
pathways experience significant activity increase/decrease.
As another novel feature, MetaboliticsDB enables users to
associate their metabolomics datasets with diseases based
on AI models that it creates and maintains.
We evaluate the features of MetaboliticsDB on a real
metabolomics data set obtained from individuals with Hep-
atocellular Carcinoma (HCC). Our results demonstrate that
MetaboliticsDB provides (i) biologically relevant metabolic
network-level analysis results along with markings through
a metabolic graph, (ii) disease association to analysis results
with high accuracy, and (iii) a scalable architecture support-
ing hundreds of simultaneous users. This paper is orga-
nized as follows. The next section summarizes data man-
agement, metabolomics analysis features, and architecture
of MetaboliticsDB. Then, we discuss our results from the
evaluation of MetaboliticsDB on an HCC dataset as well as
the performance and accuracy of different features. Finally,
we conclude with a discussion on how MetaboliticsDB may
be employed in a wider scope with the proposed features.

2 M ETHODS
2.1 Database Model and Data Management
MetaboliticsDB employs various data sources, namely,
a genome-scale reconstructed human metabolic network
dataset, Recon3D [33], a human disease ontology dataset
[34], a customized metabolite synonym mapping dataset,
and a metabolomics analysis database. To manage its data,
MetaboliticsDB employs a hybrid approach that involves
the traditional relational model coupled with more recent
NoSQL implementations.
2.1.1 Genome Scale Metabolic Network Data
Recon3D dataset contains 10600 reactions, 5835 metabolites,
and 106 pathways/subsystems. The metabolic network data
is kept as a single JSON file on the server side, and loaded
to the client side only once at the beginning of the website
life cycle. It is stored in browser cache, named local-storage.
The original data size on the disk is 5.3 MB. In order to
increase the storage and network efficiency, the data file is
compressed with gzip, and the final data size is reduced to
877 KB.
To optimize the query performance, the original Re-
con3D data model is transformed into a carefully designed
JSON schema (see Fig. 1). In this schema, at the top level,
there are three main properties, namely, pathways, reac-
tions, and metabolites. Each of those properties holds a
hashmap whose key is the id of the component, and the
value is the object of the component. Each object has a set
of foreign keys referring to the related components such
as reaction id in pathways, metabolite id in reactions, and
reaction id in metabolites. Those relations are replaced with
object references of the related components in the client side
after retrieval of data. Then, all queries are executed at the
client side on this data model in the corresponding web
interfaces.
2.1.2 Human Disease Ontology Data
Presently, the database contains a large set of diseases along
with their parent diseases (if any) and synonyms. If known
to them, users may associate their metabolomics datasets
with a disease from the database. MetaboliticsDB periodi-
cally updates the list of diseases stored within the relational
database.
2.1.3 Metabolomics Analysis Data
MetaboliticsDB employs a relational database to store
users, metabolomics analysis methods, diseases, uploaded
datasets, metabolomics measurements, metabolomics anal-
ysis results, and trained machine learning models for dis-
eases. Currently, the database contains metabolomics mea-
surements and the corresponding analysis results for 2174
individuals associated with 40 distinct diseases. Our rela-
tional database is hosted on an instance of PostgreSQL,
which offers hybrid NoSQL features such as JSON fields.
The database schema (see Fig. 2) involves seven tables,
namely, User, Methods, Datasets, MetabolomicsData, Anal-
ysis, Diseases, and DiseaseModels. The User table stores
user information and has a one-to-many relation with the
3
Analysis table. The Methods table stores metabolomics anal-
ysis methods available at MetaboliticsDB, namely Metabol-
itics, Direct Pathway Mapping, and Pathway Enrichment.
The Datasets table holds details of uploaded datasets.
Metabolomics measurements uploaded with datasets are
stored in the MetabolomicsData table. Metabolomics anal-
ysis results for each sample uploaded with datasets are
stored in the Analysis table. This table is a hybrid table with
two JSON columns that store reaction and pathway-level
analysis results. Both JSON columns are of type hashmap
whose key is either pathway or reaction name and the value
is the computed flux score during the analysis. Diseases
known by MetaboliticsDB are stored in the Diseases table.
Machine learning models trained on metabolomics analysis
results are recorded in the DiseaseModels table.
2.1.4 Multi-source Metabolite Synonym Mapping
Since most of the metabolites have several synonyms and do
not have a standard name adopted by all the sources, user-
provided metabolite names first need to be matched to those
in MetaboliticsDB’s database. The original Recon3D data
does not contain synonym information for the metabolites.
MetaboliticsDB utilizes BiGG [35] IDs of metabolites in the
Recon3D dataset for metabolomics data analysis. Increas-
ing recognized metabolite synonyms is vital for a more
comprehensive metabolomics data analysis. MetaboliticDB
enhances metabolite name mapping in two distinct ways:
First, it combines alternative synonyms for metabolites from
datasets such as HMDB [22], KEGG [19], PubChem [25],
and CheBI [23] stores it as a synonym repository. Second, if
the local synonym repository fails to match a name, RefMet
nomenclature [36] is checked for that name through HTTP
POST requests. Then, the returned metabolite names from
RefMet are appended to the customized synonym mapping
dataset of MetaboliticsDB if a RefMet match is found. As a
result, MetaboliticsDB is able to map most of the metabolites
from different datasets to the metabolites in our database.
2.2 Metabolomics Data Analysis
One of the main functionalities of MetaboliticsDB is the
analysis of user-provided metabolomics datasets to facilitate
the interpretation of measured changes. MetaboliticsDB of-
fers three different types of metabolomics analysis methods.
In this section, we present the usage and working principles
of the analysis tool.
2.2.1 Uploading Datasets
MetaboliticsDB expects the metabolomics measurements in
the form of fold-change values for each metabolite. The
fold changes may be computed in reference to different
baselines, such as average measurements of healthy indi-
viduals, before-treatment measurements, etc. depending on
the nature of the corresponding study. Users may either
manually specify fold-change information for each metabo-
lite through the web interface, or they may choose to up-
load them in a file. Fold-change values are computed by
MetaboliticsDB in reference to average measurements of
healthy samples if users upload their datasets in the form
of MetaboliticsDB’s own custom file format. MetaboliticsDB
accepts four different types of input file formats including

Fig. 1. The customized schema for the MetaboliticsDB Version of Re-
con3D JSON file
Metabolomics Workbench mwTab, JSON file in the form
of a dictionary where keys are metabolite names and the
values are the corresponding fold changes, CSV file as a
matrix of samples and metabolites, and MetaboliticsDB’s
own custom file format. Examples of these different input
formats are provided online in the Documentation section
of the MetaboliticsDB. Once the user data is uploaded,
MetaboliticsDB employs the above-discussed approaches
to maximize the mapping of metabolites in the input file
to the genome-scale metabolic network (Recon3D) that it
maintains in its database. Once the mapping is completed,
the final list of metabolites is presented to the user along
with the percentage of the mapped metabolites as well as
the list of unmapped metabolites. The user may choose to
remove any of the entries before the analysis takes place.
2.2.2 Metabolic Flux Differentiation Analysis
The analysis feature of MetaboliticsDB dynamically builds
a linear programming model [37] on the whole metabolic
network based on user-provided measurements. To achieve
this, it integrates an algorithm that we have recently devel-
oped [2]. More specifically, the total production flux of each
measured metabolite is added as a term into the objective
function of the constructed linear model. The fold change
of each metabolite is registered to the objective function
as the coefficient of the corresponding term. Then, on the
constructed model, flux variability analysis (FVA) [38] is
performed to determine the upper and lower flux limits for
each reaction in the metabolic network. Next, a metabolic
flux diffentiation score is computed for each reaction based
on how its flux boundaries differ from that healthy/control
samples. For each pathway, the mean of its reactions’ diff
score is set as the pathway’s diff score. Pathway ”diff” scores
4
represent how much the activity of a pathway deviates in a
given individual in comparison to that of healthy/control
people. This allows a personalized evaluation for each indi-
vidual. Negative diff values indicate a decrease in activity,
while positive diff values signify increase in pathway activ-
ity. The computed diff scores are stored in the database as
part of the corresponding analysis run.
2.2.3 Pathway Enrichment Analysis
Another metabolomics data analysis approach offered by
MetaboliticsDB is pathway enrichment analysis. This ap-
proach specifies pathways that are significantly enriched
with the metabolites included in the user uploaded
metabolomics datasets. The hypergeometric distribution is
used to calculate a p-value for each pathway. Metabol-
iticsDB does not strictly apply a significance threshold.
Instead, it reports the computed p-values for all pathways
as a table.
2.2.4 Direct Pathway Mapping Analysis
Direct pathway mapping is a simple linear model that
directly maps the metabolites included in a metabolomics
dataset to the pathways that they participate. Each pathway
is assigned a score that is the sum of the concentrations of
metabolites that it contains.
2.3 Tabular and Visual Analysis Results
The analysis results are presented in multiple levels of
details in both visual and tabular forms (Fig. 3). A bar plot
(at the top) charts diff values for the top-20 pathways sorted
by the absolute value of their diff scores. Below the bar
plot, in tabular form, all the pathways are listed with their
computed diff values. Next to each pathway, two buttons
are placed. The left button visualizes the corresponding
pathway where each node represents a metabolite, and each
edge represents a reaction. Edges are colored and thickened
according to the corresponding reactions’ diff value (see Fig.
6 for an example). This form of visualization allows the user
to quickly inspect what parts of the pathway experience the
most change. The right button, on the other hand, lists all the
pathway reactions and their diff values in tabular form. Each
analysis result set is stored in the database, and the user may
access past analysis results from through user account menu
(visible after signing in) or ”Browse Analysis Results” menu
item. The user may mark an analysis result as ”public” to
make it available to all other users, or keep it ”private” to
their own access (default setting).
2.4 Similarity-based Disease Association
Another novel feature of the MetaboliticsDB is that for each
metabolomics analysis result, the closest set of diseases
and physiological conditions are computed based on the
similarity of the pathway-level metabolic behavior . More
specifically, MetaboliticsDB stores metabolomics analysis re-
sults of a number of diseases (pre-computed and recorded in
its database). Then, for each user submitted analysis run, the
user is presented the top 5 diseases (e.g., diabetes) that have
most similarity to the currently analyzed metabolomics data
in terms of the activity level distribution over the metabolic

Fig. 2. Relational database schema of MetaboliticsDB
pathways (see Fig. 3). In order to quantify the similarity,
MetaboliticsDB turns both the current analysis results and
disease analysis results stored in the database into vectors
of numbers where each number represents the diff value
for a particular pathway. Then, Pearson correlation is com-
puted between each disease vector and the current analysis
result vector. Finally, the diseases in the database are sorted
according to their correlation values, and top 5 of them are
presented to the user.
2.5 Machine Learning-based Disease Prediction
Alternative to similarity-based disease association, Metabol-
iticsDB also offers machine learning-based disease status
prediction for each individual. To this end, MetaboliticsDB
trains machine learning models based on previously stored
metabolomics data analysis results for each disease peri-
odically. More specifically, for each disease four different
types of models based on Logistic Regression, Random
Forest, Support Vector Machines, and XGBoost are trained.
10-fold cross validation is used to tune the parameters
of each model and evaluate its classification performance
based on f1-scores. Then, the best performing model is
chosen and stored in the database as binary files with the
pickle Python package. The models employ the computed
reaction metabolic differentiation scores as features. Then,
the reaction diff values of future metabolomics data analysis
results are exploited to predict potential diseases associated
with an individual. Next to each disease a value between
5
0 and 1 is provided, which reflects the probability that the
underlying AI model assigns to its diseases prediction for
that individual.
2.6 Comparison of Analysis Results
MetaboliticsDB allows users to compare and contrast their
metabolomics data analysis results to (i) their previous anal-
ysis results and (ii) other users’ publicly available analysis
results stored in its database. From the analysis results page,
the user may choose any number of analysis results that
they are interested in comparing by clicking on the checkbox
next to each study. Then, clicking on ”compare” button
at the top leads to the comparison page. The comparison
interface features a heatmap at the top where the rows
represent pathways with the highest variance in terms of
their diff values among the compared analysis results, and
the columns represent the compared analysis results (see
Fig. 4). Each cell in the heatmap is colored according to
the corresponding pathway diff values. The bottom part of
the comparison interface includes a table that lists all the
pathways with their computed diff values for each selected
study (similar to the bottom part of Fig. 3).
2.7 Advanced Search Interface
As different from similar tools, as a novel feature, Metabol-
iticsDB allows to search the metabolomics analysis results
in the database in terms of the metabolic activity changes

Fig. 3. Analysis results page: top-20 pathways with the highest absolute
diff values.
of involved pathways. For instance, users may search for
metabolomics analysis results where Urea Cycle experi-
ences a decreased activity, whereas Fatty Acid Synthesis
experiences increased activity. Optionally, the user may
also specify the magnitude of increase and decrease. They
cana flexibly add and remove pathways from consideration
leading to multiple conditions which are connected via SQL
AND semantics.
2.8 Basic Searching, Browsing, and Other Visualiza-
tion Features
Similar to many other metabolic databases, Metaboli-
ticsDB includes a search interface to locate metabolites,
reactions, and pathways that they are interested in. The
search interface features an auto-complete feature (similar to
Google Search) that automatically suggests names from the
database as user types in their search terms. The suggestions
are not a flat list of mixed names, but are categorized
into different groups based on the matching entity types
(i.e., metabolites, reactions, etc.). The search results are also
presented in a similar manner where the matching items
are categorized based on their entity types. Clicking on
6
Fig. 4. MetaboliticsDB Comparison interface featuring a heatmap that
compares three common cancer types: Lung, Breast, and Hepatocellu-
lar
any search result entry takes the user to the details and
visualizations of the clicked entity.
Users may also choose to browse the database pathway
by pathway. The browsing page lists all the pathways in the
database on the left, and clicking on each pathway displays
the reactions in the pathway and a graphical visualization
that shows a network view of the pathway (see Fig. 6 for
an example). The pathway visualizations in MetaboliticsDB
are drawn by using the Escher library [39] and saved in the
database.
2.9 Architecture
The architecture of MetaboliticsDB (Fig. 5) is carefully de-
signed to meet demanding computational and data man-
agement requirements of various features in an efficient and
flexible manner. All frontend interfaces are implemented in
Angular, which is a Javascript framework that allows the
development of sophisticated single page web applications.
Most of the rendering and application logic is developed
at client side to enhance the performance and user expe-
rience. The relational database stores analysis results and
user accounts, and it is hosted by a Postgres instance. The
frontend communicates with the database via the RESTful
API that is developed in Flask, a micro web framework
in Python. One advantage of adopting RESTful API imple-
mentation is that it offers programmatic API access to other
researchers who may want to programmatically utilize the
analysis algorithms and services of MetaboliticsDB in their
project implementations. This feature provides another in-
terface to MetaboliticsDB for users with programming skills.
Documentation for MetaboliticsDB’s RESTful APIs inter-
faces is available at http://metabolitics.itu.edu.tr/api/spec
in OPENAPI specification. In addition, MetaboliticsDB have
a sophisticated system to manage the analysis requests,
as each analysis task is computationally expensive, which
cannot be handled in the regular life cycle of HTTP requests.

More specifically, MetaboliticsDB employs Celery which is
a distributed task queue, and is used with Redis database to
store information about the tasks. Celery workers subscribe
to Redis and perform metabolomics analysis when a new
analysis is submitted to Redis.
Scalability is a major aspect that is greatly considered in
the design of MetaboliticsDB’s architecture. All components
of the architecture are dockerized. Moreover, each celery
worker can be distributed to multiple instances. Lastly, the
choice of Angular in frontend significantly reduces the load
on servers, since most of the application logic runs at client
side.
Fig. 5. Architecture overview diagram of the project.
3 R ESULTS
In this section, we evaluate MetaboliticsDB in different as-
pects, and demonstrate that it provides biologically relevant
insights via a use case study on Hepatocellular Carcinoma.
3.1 A Case Study on Hepatocellular Carcinoma
In order to demonstrate the capabilities of MetaboliticsDB,
we perform a case study on a Hepatocellular Carcinoma
(HCC) dataset [40] which contains metabolomics mea-
surements obtained from 177 individuals (71 healthy and
106 patients). Metabolomics measurements are converted
into fold changes based on the average measurement
of healthy samples during analysis. We upload this file
through MetaboliticsDB’s analysis interface and submit it
for analysis after choosing ’Metabolitics’ as the analysis
method. On the analysis result page, MetaboliticsDB pro-
vides pathways with the highest absolute diff values based
on the submitted measurements in a bar chart as well
as in tabular form ( available online on MetaboliticsDB
at http://metabolitics.itu.edu.tr/past-analysis/1761). For
brevity, here, we discuss the relevance of the top 10 path-
ways with the highest absolute diff scores from this list in
reference to literature to illustrate the effectiveness of the
MetaboliticsDB analysis tool.
Protein formation pathway has the largest absolute diff
value in the negative direction. PROTEIN BS reaction in the
Protein formation pathway yields Torasemide-M3 metabo-
lite. Activated metabolite Torasemide-M3 is formed from
the oxidation of Torasemide [41]. Liver disease has been
7
Fig. 6. Change in citric acid cycle
reported to affect the activities of Torasemide with the in-
creased recovery of Torasemide inspected in urine [42]. This
may explain the decreased activity of Torasemide activation
in HCC patients.
Fatty acid synthesis pathway has the second largest diff
value in the positive direction. Hepatocellular tumorigen-
esis has been reported to increase with abnormal activity
in Fatty acid synthesis, and treatments inhibiting Fatty acid
synthase enzyme might be utilized in HCC therapies [43].
Increased activity of Fatty acid synthase C180 reaction in
the Fatty acid synthesis pathway observed in HCC analysis
results supports this observation.
Nucleotide metabolism has the third largest diff value
in the positive direction. To keep up with the fast pace of
cell proliferation during tumorigenesis, increasing de novo
nucleotide synthesis is essential for large-scale RNA pro-
duction and DNA replication [44]. In addition, Nucleotide
interconversion has a substantial diff value in the positive
direction. This observation is plausible, as the essential
building blocks of Nucleotide metabolism are produced by
this pathway with a transformation of (d) NMP ↔ (d) NDP
↔ (d) NTP.
Another pathway with a large diff value in the positive
direction is Hippurate metabolism. Reduced amounts of
Hippurate are quantified in HCC patients due to decreased
Benzoate binding proficiency [45]. Even though this is not
fully aligned with our observation, the elevated quantity
of Hippurate is also related to fruit and whole grains con-
sumption. Therefore, nutrition content of individuals during
period of sample collection may be one factor regarding this
observation. [46].
Heme degradation is another pathway with a large diff
value in the positive direction. The enzyme that catalyzes
heme degradation, Heme oxygenase 1 has been reported to
be related to cancer progression [47]. Inhibiting the activity
of Heme oxygenase 1 has been claimed to decrease HCC
progression [48].
Another pathway with a large diff value in the positive
direction is Vitamin B6 metabolism. The amount of Vitamin
B6 compounds present in cancer cases has been reported
 8

to be less than that in control cases, and the activation disease. Presently, there are 40 distinct diseases stored in
levels of Pyridoxal kinase enzyme have been reported to MetaboliticsDB. Hence, the ground-truth clustering includes
help disease progression [49]. Increased activity of Pyridoxal 40 clusters. To compare the clusterings, we employ two
kinase reaction in the Vitamin B6 metabolism seen in HCC intuitive comparison metrics, i.e., homogeneity and com-
analysis results supports these statements. pleteness [55]. Briefly, homogeneity checks if each cluster
Another pathway with a large diff value in the positive contains patients with the same disease, and completeness
direction is Limonene and pinene degradation. Limonene deals with whether all patients with the same disease are
has been reported to inhibit the progression of HCC by assigned to the same cluster. Both measures may have
suppressing cell proliferation [50]. Pinene also has been a value between 0 and 1, where 1 represents best score,
reported to inhibit cancer cell development in vitro and in and 0 represents the worst score. In our evaluation, both
vivo [51]. Decreased levels of Limonene and Pinene due homogeneity and completeness are measured as 0.94, which
to activities of Limonene and pinene degradation might indicates that cluster assignments are mostly accurate.
contribute to HCC progression.
3.2.2 Machine Learning-based Disease Association
Cytochrome metabolism is another pathway with a large
diff value in the positive direction. Intrinsic clearance val- In this section, we present the prediction performance of the
ues indicating activity levels show an activity growth for machine learning models that MetaboliticsDB creates and
CYP2E1, CYP2D6, and CYP2C9 cytochrome P450 types in manages in its database. We employ k-fold cross validation
HCC samples [52]. Increased activity of Cytochrome P450 (with k = 10 or k = 5 depending on sample size) to test the
2E1, Cytochrome P450 2C9, and Cytochrome P450 2D6 classification accuracy. Table 2 summarizes precision, recall,
reactions in the Cytochrome metabolism pathway seen in and F1 scores for the disease prediction models stored in the
HCC analysis results supports these observations. database. The F1 score was calculated by the harmonic mean
Another pathway with a large diff value in the posi- of precision and recall values. The results show that healthy
tive direction is Thiamine metabolism. The activity levels and tumor samples are classified with high accuracy.
of enzymes that rely on Thiamine have been reported to Disease Precision Recall F1 Alg. K
increase in cancer cases [53]. Increased activity levels of Hepatocellular Carcinoma 0.89 0.91 0.90 LR 10
Thiamine diphosphokinase, Thiamine diphosphate kinase, Colon Carcinoma 0.96 0.99 0.98 RF 5
Breast Cancer 0.88 0.94 0.91 LR 10
and Thiamine-triphosphatase reactions in the Thiamine Stomach Cancer 0.94 0.99 0.96 LR 10
metabolism pathway seen in HCC analysis results support Ovarian Cancer 0.94 0.92 0.91 RF 10
these findings. Crohn’s Disease 0.81 0.91 0.84 RF 10
Finally, Fructose and mannose metabolism is the last Asthma 0.99 0.99 0.99 RF 10
Rheumatoid Arthritis 0.89 0.85 0.85 RF 10
among the top 10 pathways with a large diff score in the Steatotic Liver Disease 0.82 0.95 0.87 RF 10
positive direction. The development of HCC is increased Type 2 Diabetes Mellitus 0.86 0.96 0.91 LR 10
with diets rich in fructose since it enhances activity levels of Wilson Disease 0.80 0.90 0.83 RF 5
Adult Respiratory Distress
the lipogenic pathway and lipid accumulation [54]. Syndrome
0.88 1.00 0.93 LR 5
The above brief discussion illustrates that Metaboli- Androgenic Alopecia 0.81 0.94 0.86 LR 10
ticsDB is useful and effective in analyzing metabolomics Ankylosing Spondylitis 0.83 1.00 0.89 RF 5
datasets with insights on the underlying metabolic mech- Autistic Disorder 0.75 0.88 0.79 RF 5
Chronic Fatigue Syndrome 0.73 0.74 0.72 LR 10
anisms. Cystic Fibrosis 0.75 1.00 0.83 RF 5
Intermediate Coronary
0.77 1.00 0.85 RF 5
Syndrome
3.2 Disease Association Evaluation Peanut Allergy 0.83 0.92 0.83 RF 10
Placental Abruption 0.88 1.00 0.92 RF 5
In this section, we evaluate the disease association feature
Pre-eclampsia 0.75 0.88 0.75 RF 5
of MetaboliticsDB in two distinct dimensions. Sarcoidosis 0.81 0.86 0.78 SVM 10
Schizophrenia 0.90 1.00 0.95 SVM 10
3.2.1 Similarity-based Disease Association TABLE 1
Average results of k-fold cross validation
MetaboliticsDB reports diseases and physiological condi-
tions that are most similar to the analyzed metabolomics
dataset based on the correlation between the current We further evaluate the prediction performance of the
analysis results and the previously computed disease models by relaxing the true positive definition slightly. In
metabolomics data analysis results. In order to evaluate particular, since MetaboliticsDB provides a list of possible
the relevancy of the results provided by the proposed associated diseases sorted by their predicted likelihoods, we
scheme, we cluster all diseases in the database using the consider a disease association as true positive if the true
same proposed vector representation and similarity mea- disease appears in top 3 suggested diseases for patients,
sure (i.e., using agglomerative clustering with similarity and it does not appear at all for healthy individuals. The
measure: pearson correlation, linkage: complete). Then, we prediction is said to be accurate if the disease is listed in
compare the resulting disease clustering to a ”ground-truth top-3 predictions for patients or the disease isn’t listed in
clustering”. The ground-truth clustering that we employ predictions for healthy samples. The number of samples
in this evaluation includes one cluster per distinct disease. predicted accurately divided by all samples is given in the
That is, patient characteristics such as gender and age are precision column. The number of accurate predictions for
ignored, metabolomics samples are assigned to the same disease samples divided by all disease samples is given in
cluster as long as the corresponding patients have the same the recall column.
 9
Disease Precision Recall F1
Hepatocellular Carcinoma 0.89 0.84 0.87
Moreover, in order to test the effect of the metabolomics data
Colon Carcinoma 1.00 1.00 1.00 size, random metabolites are selected from each network
Breast Cancer 0.81 0.75 0.78 and random fold-change values are assigned to them. The
Stomach Cancer 0.98 0.98 0.98 number of metabolite measurements included in the tests
Ovarian Cancer 0.92 0.88 0.90
Crohn’s Disease 0.88 0.76 0.82 varied between 5 and 150 (incremented by 5 leading to 30
Asthma 0.99 0.99 0.99 different metabolomics data sets). Then, the analysis is run
Rheumatoid Arthritis 0.88 0.76 0.82 with those artificial metabolomics data sets on each evalu-
Steatotic Liver Disease 0.91 0.84 0.88
Type 2 Diabetes Mellitus 0.90 0.87 0.88
ated metabolic network. Fig. 7 charts the average analysis
Wilson Disease 0.92 0.83 0.87 running time (in seconds) over all metabolomics data sets
Adult Respiratory Distress Syndrome 0.91 0.90 0.91 for each metabolic network.
Androgenic Alopecia 0.88 0.77 0.82
Ankylosing Spondylitis 0.75 0.50 0.60
Autistic Disorder 0.92 0.83 0.87
Chronic Fatigue Syndrome 0.80 0.63 0.71
Cystic Fibrosis 0.89 0.67 0.76
Intermediate Coronary Syndrome 1.00 1.00 1.00
Peanut Allergy 0.91 0.82 0.86
Placental Abruption 0.92 0.83 0.87
Pre-eclampsia 1.00 1.00 1.00
Sarcoidosis 0.75 0.69 0.72
Schizophrenia 0.93 0.99 0.96
TABLE 2
Disease prediction results

3.3 Responsiveness Evaluation

In this section, we evaluate the responsiveness of Metaboli-
ticsDB with varying numbers of simultaneous users. Each
user is assumed to send one request per second, and in Fig. 7. Running time of MetaboliticsDB analysis feature on different
total, 100 requests in their browsing life cycle. Table 3 reports networks
the average response rate in seconds, and the percentage of
response success rate. For this simulation, an open source
load testing Python tool, called Locust, is used. The server
4 C OMPARISON
hosting MetaboliticsDB during these tests has the following
configuration: DELL R720 with 2 x XEON E5-2620v2 2.10 In this section, we compare MetaboliticsDB to some of
GHz CPU and 80 GB RAM running Linux Ubuntu. the well known tools in the field in terms of different as-
pects. In particular, the comparison includes MetaboAnalyst
Number of Users Average Response Time (sec) Success Rate (%) 5.0 [28], Metabolomics Workbench [13], MeltDB 2.0 [14],
10 0.28 100
100 0.36 100
MetabolomeExpress [15], XCMS Online [16], Caleydo [17],
1000 1.51 100 WebSpecmine [29], 3Omics [30], Workflow4Metabolomics
TABLE 3 [31], and POMAShiny [32]. Table 5 summarizes the consid-
Responsiveness evaluation results ered aspects for comparison.
MetaboliticsDB offers metabolite name mapping, fold
change scaling, and reaction diff conversion in terms of
data preprocessing. Data filtration, normalization, name
3.4 Load Test mapping of metabolites, and alignment and detection of
In this section, additional performance tests are performed peaks are some of the data preprocessing steps supported
for the MetaboliticsDB analysis feature. The running time by MetaboAnalyst 5.0. Normalization and scaling are also
of MetaboliticsDB’s analysis feature depends on two fac- available in Metabolomics Workbench for data preprocess-
tors: the size of the underlying metabolic network and the ing. Alignment and detection of peaks on raw data are also
number of metabolites in the input metabolomics data. As offered by MeltDB 2.0 and MetabolomeExpress. Data filtra-
part of this evaluation, we run MetaboliticsDB’s analysis tion is available on XCMS Online and Caleydo, however,
feature on several metabolic networks of different sizes data normalization is not supported. Data is processed with
from various organisms (obtained from BIGG [24]) (Table 4). the alignment and detection of peaks steps during data up-
load by WebSpecmine, also other steps such as normaliza-
tion and scaling are available. Normalization is supported
Metabolic Net. BIGG Id Num. of Reactions Num. of Metabolites
e coli core 95 72 by Workflow4Metabolomics whereas data filtration isn’t
iAB RBC 283 342 469 available. Detection and cleaning of outliers are available
iRC1080 1706 2191 on POMAShiny contrary to Workflow4Metabolomics and
RECON1 3742 2766
RECON2 7785 5324
MetaboAnalyst 5.0.
TABLE 4 Fold-change analysis is the univariate analysis method
Metabolic network size available in MetaboliticsDB. Volcano plots, t-tests, and fold-
change analysis are among the univariate analysis methods
 10
Genome-
Metabolic
Scale Python/R Advanced
Data Pre- Univariate Classi- Enrichment Pathway Disease Analysis AI Model Flux
Metabolic Package Analysis
processing Analysis fication Analysis Analysis Prediction Comparison Management Change
Network Availability Search
Prediction
Support
MetaboliticsDB 3 3 3 3 3 3 3 3 3 3 3 3
Metabo-
3 3 3 3 3 3 3 3 7 3 7 7
Analyst 5.0
Metabolomics
3 3 3 3 7 3 3 7 3 7 3 7
Workbench
MeltDB 2.0 3 3 3 3 3 3 7 7 3 7 7 7
Metabolome-
3 3 7 7 3 7 7 7 3 7 3 7
Express
XCMS Online 3 3 7 3 3 3 3 7 3 7 3 7
Caleydo 3 7 7 7 3 3 7 7 7 7 7 7
WebSpecmine 3 3 3 7 3 7 3 3 3 3 7 7
3Omics 7 7 7 3 3 3 7 7 3 7 7 7
Workflow4-
3 3 7 7 7 7 7 7 7 7 7 7
Metabolomics
POMAShiny 3 3 3 7 7 3 3 3 7 3 7 7
TABLE 5
Comparison of MetaboliticsDB with existing tools

offered by MetaboAnalyst 5.0 and MeltDB 2.0. Volcano
plots and ANOVA analysis are available in Metabolomics
Workbench. t-tests and fold-change analysis are provided
by MetabolomeExpress, but volcano plots are not sup-
ported. Fold-change analysis is also offered by XCMS On-
line. ANOVA analysis is also accessible along with t-tests
and fold change analysis in WebSpecmine. Non-parametric
and parametric tests and ANOVA analysis are provided by
Workflow4Metabolomics and POMAShiny.
MetaboliticsDB offers automatically managed machine
learning classification models of type Logistic Regression,
Support Vector Machines, Random Forest, and XGBoost.
MetaboAnalyst 5.0 provides Support Vector Machine, Ran-
dom Forest, and Partial Least Squares Discriminant Analysis
classification methods. Random Forest and Orthogonal Par-
tial Least Squares Discriminant Analysis classification meth-
ods are provided by Metabolomics Workbench. Support
Vector Machine and Random Forest Classification methods
are also supported by MeltDB 2.0. Linear Discriminant
Analysis and Support Vector Machine methods are provided
by WebSpecmine. Random Forest algorithm is also available
in POMAShiny for classification.
Both enrichment analysis and pathway analysis are
available in MetaboliticsDB. Enrichment analysis is sup-
ported by MetaboAnalyst 5.0, Metabolomics Workbench,
MeltDB 2.0, XCMS Online, and 3Omics. MetaboAnalyst 5.0,
MeltDB 2.0, MetabolomeExpress, XCMS Online, Caleydo,
WebSpecmine, and 3Omics provide Pathway Analysis.
MetaboliticsDB, MetaboAnalyst 5.0, Metabolomics
Workbench, MeltDB 2.0, XCMS Online, Caleydo, 3Omics,
and POMAShiny support genome-scale metabolic
networks. Python or R packages are available for
MetaboliticsDB, MetaboAnalyst 5.0, Metabolomics
Workbench, XCMS Online, WebSpecmine, and POMAShiny.
MetaboliticsDB, MetaboAnalyst 5.0, WebSpecmine, and
POMAShiny train machine learning models and use these
models for sample prediction. MetaboliticsDB periodically
trains and stores predictive models on analysis results for
disease prediction.
MetaboliticsDB, Metabolomics Workbench, MeltDB 2.0,
MetabolomeExpress, XCMSOnline, WebSpecmine, and
3Omics enable users to compare analysis results. Advanced
analysis search interface is available in MetaboliticsDB,
Metabolomics Workbench, MetabolomeExpress, and XCMS
Online. Metabolic flux change prediction is only available in
MetaboliticsDB.
5 D ISCUSSION
MetaboliticsDB is novel in a number of aspects in compar-
ison to the existing similar works. In particular, it offers
a data management platform for metabolomics analysis
results along with a set of associated web-based tools that
allow user to effectively query, visualize, and study the
analysis results at network level. Considering metabolomics
analysis results as the main object, and designing tools
around it facilitates the offering of, in particular, three useful
features:
(i) With the comparison feature, MetaboliticsDB enables
researchers to compare their datasets to the known dis-
eases or other users’ public analysis results representing
different physiological conditions. In particular, the compar-
ison feature allows researchers to make connection between
seemingly different conditions, and have some insights
about what kind of condition the current metabolomics
data set may belong to. In the former case, the recognition
of common mechanisms between two different conditions
may pave the way for sharing known/existing therapies
designed for each condition. In the latter case, it may pro-
vide researchers with pointers on where to look at in the
existing knowledge while interpreting a new and possibly
complex case. Besides, the comparison interface can be
utilized to understand the differences between sub-types of
a disease, progression of disease stages, and the effect of
possible drugs through before and after comparison. Finally,
existing or possible common patterns across the same class
of diseases may be observed (e.g., Fig. 4 compares different
cancers). For instance, Warburg Effect in different types of
cancers may be studied in depth.
(ii) With the disease and physiological condition as-
sociation tools, MetaboliticsDB may help clinicians to get

one step closer to the ”personalized medicine” goal. More
specifically, MetaboliticsDB may help to narrow down the
number of possible conditions considered for a particular
patient by her medical care provider.
(iii) Advanced search interface on metabolic analysis
results may help discovering commonly repeating patterns
of metabolic fluctuations across different conditions in terms
of pathway activity changes.
MetaboliticsDB integrates a powerful analysis interface,
which is central to its functioning. As illustrated with a
case study on HCC, it helps researchers to effectively make
biologically relevant interpretations of their metabolomics
data in a holistic manner over all the metabolic network.
This way, users are not anymore limited to biomarkers that
differentiate a physiological condition from others; they may
now drill down into metabolic mechanism differences, even
on parts of the metabolism where metabolite measurements
are not available in the data, owing to its state of the art
personolized metabolic analysis algorithm.
Furthermore, MetaboliticsDB may be useful for drug
design research. Based on the highlighted changes on the
metabolic networks, MetaboliticsDB may provide pointers
and ideas for possible drug targets. Similarly, it may explain
why a certain drug works or does not work through the
visualization of the relevant parts of the pathways.
Last but not the least, MetaboliticsDB is developed in a
way which is easy to generalize. In near future, Metaboli-
ticsDB will support multi-omics analysis with incorporation
of gene expression and proteomics measurement interpre-
tation with respect to changes on the metabolism using
the same underlying database and the associated analysis
tools with minimal development requirement. Given the
enormous amount of publicly available gene expression
and proteomics datasets, the impact of MetaboliticsDB will
multiply with this extension. Finally, combining multi-omics
data sets that belong to the same patient by utilizing the
MetaboliticsDB analysis interfaces is highly likely to provide
invaluable insights into different physiological conditions.
6 C ONCLUSION
In this paper, we present MetaboliticsDB which incorporates
a novel pathway-level metabolomics data analysis results
database, and a set of powerful associated tools running on
this database. In particular, MetabolomicsDB allows users
to analyze their metabolomics datasets with three different
methods, store them in their private user area or share
with other users, compare them with known diseases in
terms of the underlying metabolic mechanisms, visualize
the changes on the metabolic network, perform basic and
advanced search on metabolomics analysis results, and as-
sociate their datasets with different diseases (if any). We
evaluate the effectiveness of the MetaboliticsDB’s analysis
features on a real data set obtained from HCC patients,
and show that (i) it provides biologically relevant results as
supported by the existing literature, (ii) disease associations
have high homogeneity and completeness scores, and (iii) it
can scale to large numbers of simultaneous users.
11
ACKNOWLEDGEMENTS
We would like to thank Beyza Turk for contributing some of
the visualizations stored in MetaboliticsDB.
7 D ECLARATIONS
7.1 Ethics approval and consent to participate
Not applicable.
7.2 Consent for publication
Not applicable.
7.3 Availability of data and materials
MetaboliticsDB is available online at http://metabolitics.
itu.edu.tr/.
Web interface source codes are available at https://github.
com/itu-bioinformatics-database-lab/metabolitics-client.
Web API source codes are available at https://github.com/
itu-bioinformatics-database-lab/metabolitics-api.
Source codes of the Metabolitics data analysis
algorithm are available at https://github.com/
itu-bioinformatics-database-lab/metabolitics.
7.4 Competing interests
None.
7.5 Funding
This work was in part supported by the Scientific and Tech-
nological Research Council of Turkey (TÜBİTAK) [Grant
Number: 114E115] and the National Center for High Perfor-
mance Compututing (UHEM) [Grant Number: 1009742021].
7.6 Authors’ contributions
MHC implemented the metabolomics analysis features. OE
worked on associating datasets with diseases. AD worked
on the pathway mapping method. TS implemented the data
ingestion features. AC conceived the study. All authors
contributed to the writing of the manuscript.
R EFERENCES
[1] A. J. Carroll, R. M. Salek, M. Arita, J. Kopka, and D. Walther,
“Metabolome informatics and statistics: current state and emerg-
ing trends,” Frontiers in bioengineering and biotechnology, vol. 4,
2016.
[2] A. Cakmak and M. H. Celik, “Personalized metabolic analysis
of diseases,” IEEE/ACM Transactions on Computational Biology and
Bioinformatics, vol. 18, no. 3, pp. 1014–1025, 2021.
[3] J. Wang, D. Duncan, Z. Shi, and B. Zhang, “Web-based gene set
analysis toolkit (webgestalt): update 2013,” Nucleic acids research,
vol. 41, no. W1, pp. W77–W83, 2013.
[4] A. Subramanian, P. Tamayo, V. K. Mootha, S. Mukherjee, B. L.
Ebert, M. A. Gillette, A. Paulovich, S. L. Pomeroy, T. R. Golub, E. S.
Lander et al., “Gene set enrichment analysis: a knowledge-based
approach for interpreting genome-wide expression profiles,” Pro-
ceedings of the National Academy of Sciences, vol. 102, no. 43, pp.
15 545–15 550, 2005.
[5] Y. Drier, M. Sheffer, and E. Domany, “Pathway-based personalized
analysis of cancer,” Proceedings of the National Academy of Sciences,
vol. 110, no. 16, pp. 6388–6393, 2013.

[6] E. Lee, H.-Y. Chuang, J.-W. Kim, T. Ideker, and D. Lee, “Inferring
pathway activity toward precise disease classification,” PLoS com-
putational biology, vol. 4, no. 11, p. e1000217, 2008.
[7] P. Khatri, S. Sellamuthu, P. Malhotra, K. Amin, A. Done, and
S. Draghici, “Recent additions and improvements to the onto-
tools,” Nucleic Acids Research, vol. 33, no. suppl 2, pp. W762–W765,
2005.
[8] S. Draghici, P. Khatri, A. L. Tarca, K. Amin, A. Done, C. Voichita,
C. Georgescu, and R. Romero, “A systems biology approach for
pathway level analysis,” Genome research, vol. 17, no. 10, pp. 1537–
1545, 2007.
[9] A. L. Tarca, S. Draghici, P. Khatri, S. S. Hassan, P. Mittal, J.-s. Kim,
C. J. Kim, J. P. Kusanovic, and R. Romero, “A novel signaling
pathway impact analysis,” Bioinformatics, vol. 25, no. 1, pp. 75–82,
2008.
[10] C. J. Vaske, S. C. Benz, J. Z. Sanborn, D. Earl, C. Szeto, J. Zhu,
D. Haussler, and J. M. Stuart, “Inference of patient-specific path-
way activities from multi-dimensional cancer genomics data using
paradigm,” Bioinformatics, vol. 26, no. 12, pp. i237–i245, 2010.
[11] L. M. Heiser, A. Sadanandam, W.-L. Kuo, S. C. Benz, T. C. Gold-
stein, S. Ng, W. J. Gibb, N. J. Wang, S. Ziyad, F. Tong et al., “Subtype
and pathway specific responses to anticancer compounds in breast
cancer,” Proceedings of the National Academy of Sciences, vol. 109,
no. 8, pp. 2724–2729, 2012.
[12] J. Xia, I. V. Sinelnikov, B. Han, and D. S. Wishart, “Metaboana-
lyst 3.0—making metabolomics more meaningful,” Nucleic acids
research, vol. 43, no. W1, pp. W251–W257, 2015.
[13] M. Sud, E. Fahy, D. Cotter, K. Azam, I. Vadivelu, C. Burant,
A. Edison, O. Fiehn, R. Higashi, K. S. Nair et al., “Metabolomics
workbench: An international repository for metabolomics data
and metadata, metabolite standards, protocols, tutorials and train-
ing, and analysis tools,” Nucleic acids research, vol. 44, no. D1, pp.
D463–D470, 2015.
[14] N. Kessler, H. Neuweger, A. Bonte, G. Langenkämper, K. Niehaus,
T. W. Nattkemper, and A. Goesmann, “Meltdb 2.0–advances of the
metabolomics software system,” Bioinformatics, vol. 29, no. 19, pp.
2452–2459, 2013.
[15] A. J. Carroll, M. R. Badger, and A. H. Millar, “The metabolome-
express project: enabling web-based processing, analysis and
transparent dissemination of gc/ms metabolomics datasets,” BMC
bioinformatics, vol. 11, no. 1, p. 376, 2010.
[16] R. Tautenhahn, G. J. Patti, D. Rinehart, and G. Siuzdak, “Xcms
online: a web-based platform to process untargeted metabolomic
data,” Analytical chemistry, vol. 84, no. 11, pp. 5035–5039, 2012.
[17] M. Streit, A. Lex, M. Kalkusch, K. Zatloukal, and D. Schmalstieg,
“Caleydo: connecting pathways and gene expression,” Bioinfor-
matics, vol. 25, no. 20, pp. 2760–2761, 2009.
[18] P. D. Karp, S. M. Paley, M. Krummenacker, M. Latendresse, J. M.
Dale, T. J. Lee, P. Kaipa, F. Gilham, A. Spaulding, L. Popescu
et al., “Pathway tools version 13.0: integrated software for path-
way/genome informatics and systems biology,” Briefings in bioin-
formatics, vol. 11, no. 1, pp. 40–79, 2009.
[19] M. Kanehisa, M. Furumichi, M. Tanabe, Y. Sato, and K. Morishima,
“Kegg: new perspectives on genomes, pathways, diseases and
drugs,” Nucleic acids research, vol. 45, no. D1, pp. D353–D361, 2017.
[20] R. Caspi, R. Billington, L. Ferrer, H. Foerster, C. A. Fulcher,
I. M. Keseler, A. Kothari, M. Krummenacker, M. Latendresse,
L. A. Mueller et al., “The metacyc database of metabolic path-
ways and enzymes and the biocyc collection of pathway/genome
databases,” Nucleic acids research, vol. 44, no. D1, pp. D471–D480,
2015.
[21] A. Fabregat, K. Sidiropoulos, P. Garapati, M. Gillespie, K. Haus-
mann, R. Haw, B. Jassal, S. Jupe, F. Korninger, S. McKay et al., “The
reactome pathway knowledgebase,” Nucleic acids research, vol. 44,
no. D1, pp. D481–D487, 2015.
[22] D. S. Wishart, T. Jewison, A. C. Guo, M. Wilson, C. Knox, Y. Liu,
Y. Djoumbou, R. Mandal, F. Aziat, E. Dong et al., “Hmdb 3.0—the
human metabolome database in 2013,” Nucleic acids research,
vol. 41, no. D1, pp. D801–D807, 2012.
[23] J. Hastings, P. de Matos, A. Dekker, M. Ennis, B. Harsha, N. Kale,
V. Muthukrishnan, G. Owen, S. Turner, M. Williams et al., “The
chebi reference database and ontology for biologically relevant
chemistry: enhancements for 2013,” Nucleic acids research, vol. 41,
no. D1, pp. D456–D463, 2012.
[24] Z. A. King, J. Lu, A. Dräger, P. Miller, S. Federowicz, J. A. Lerman,
A. Ebrahim, B. O. Palsson, and N. E. Lewis, “Bigg models: A
platform for integrating, standardizing and sharing genome-scale
12
models,” Nucleic acids research, vol. 44, no. D1, pp. D515–D522,
2015.
[25] S. Kim, P. A. Thiessen, E. E. Bolton, J. Chen, G. Fu, A. Gindulyte,
L. Han, J. He, S. He, B. A. Shoemaker et al., “Pubchem substance
and compound databases,” Nucleic acids research, vol. 44, no. D1,
pp. D1202–D1213, 2015.
[26] B. Elliott, M. Kirac, A. Cakmak, G. Yavas, S. Mayes, E. Cheng,
Y. Wang, C. Gupta, G. Ozsoyoglu, and Z. Meral Ozsoyoglu, “Path-
case: pathways database system,” Bioinformatics, vol. 24, no. 21, pp.
2526–2533, 2008.
[27] A. E. Cicek, X. Qi, A. Cakmak, S. R. Johnson, X. Han, S. Alshalwi,
Z. M. Ozsoyoglu, and G. Ozsoyoglu, “An online system for
metabolic network analysis,” Database, vol. 2014, p. bau091, 2014.
[28] Z. Pang, J. Chong, G. Zhou, D. A. de Lima Morais, L. Chang,
M. Barrette, C. Gauthier, P.-É. Jacques, S. Li, and J. Xia, “Metabo-
analyst 5.0: narrowing the gap between raw spectra and functional
insights,” Nucleic acids research, vol. 49, no. W1, pp. W388–W396,
2021.
[29] S. Cardoso, T. Afonso, M. Maraschin, and M. Rocha,
“Webspecmine: A website for metabolomics data analysis and
mining,” Metabolites, vol. 9, no. 10, 2019. [Online]. Available:
https://www.mdpi.com/2218-1989/9/10/237
[30] T.-C. Kuo, T.-F. Tian, and Y. J. Tseng, “3omics: a web-based systems
biology tool for analysis, integration and visualization of human
transcriptomic, proteomic and metabolomic data,” BMC systems
biology, vol. 7, pp. 1–15, 2013.
[31] F. Giacomoni, G. Le Corguille, M. Monsoor, M. Landi, P. Pericard,
M. Pétéra, C. Duperier, M. Tremblay-Franco, J.-F. Martin, D. Jacob
et al., “Workflow4metabolomics: a collaborative research infras-
tructure for computational metabolomics,” Bioinformatics, vol. 31,
no. 9, pp. 1493–1495, 2015.
[32] P. Castellano-Escuder, R. González-Domı́nguez, F. Carmona-
Pontaque, C. Andrés-Lacueva, and A. Sánchez-Pla, “Pomashiny:
A user-friendly web-based workflow for metabolomics and pro-
teomics data analysis,” PLOS Computational Biology, vol. 17, no. 7,
p. e1009148, 2021.
[33] E. Brunk, S. Sahoo, D. C. Zielinski, A. Altunkaya, A. Dräger,
N. Mih, F. Gatto, A. Nilsson, G. A. Preciat Gonzalez, M. K. Aurich
et al., “Recon3d enables a three-dimensional view of gene variation
in human metabolism,” Nature biotechnology, vol. 36, no. 3, pp. 272–
281, 2018.
[34] J. A. Baron, C. S.-B. Johnson, M. A. Schor, D. Olley, L. Nickel,
V. Felix, J. B. Munro, S. M. Bello, C. Bearer, R. Lichenstein et al.,
“The do-kb knowledgebase: a 20-year journey developing the
disease open science ecosystem,” Nucleic acids research, vol. 52,
no. D1, pp. D1305–D1314, 2024.
[35] Z. A. King, J. Lu, A. Dräger, P. Miller, S. Federowicz, J. A. Lerman,
A. Ebrahim, B. O. Palsson, and N. E. Lewis, “Bigg models: A
platform for integrating, standardizing and sharing genome-scale
models,” Nucleic acids research, vol. 44, no. D1, pp. D515–D522,
2016.
[36] E. Fahy and S. Subramaniam, “Refmet: a reference nomenclature
for metabolomics,” Nature methods, vol. 17, no. 12, pp. 1173–1174,
2020.
[37] J. D. Orth, I. Thiele, and B. Ø. Palsson, “What is flux balance
analysis?” Nature biotechnology, vol. 28, no. 3, pp. 245–248, 2010.
[38] A. C. Müller and A. Bockmayr, “Fast thermodynamically con-
strained flux variability analysis,” Bioinformatics, vol. 29, no. 7, pp.
903–909, 2013.
[39] Z. A. King, A. Dräger, A. Ebrahim, N. Sonnenschein, N. E. Lewis,
and B. O. Palsson, “Escher: a web application for building, sharing,
and embedding data-rich visualizations of biological pathways,”
PLoS computational biology, vol. 11, no. 8, p. e1004321, 2015.
[40] T. Chen, G. Xie, X. Wang, J. Fan, Y. Qiu, X. Zheng, X. Qi, Y. Cao,
M. Su, X. Wang et al., “Serum and urine metabolite profiling
reveals potential biomarkers of human hepatocellular carcinoma,”
Molecular & Cellular Proteomics, vol. 10, no. 7, pp. M110–004 945,
2011.
[41] S. Sahoo, H. S. Haraldsdóttir, R. M. Fleming, and I. Thiele, “Mod-
eling the effects of commonly used drugs on human metabolism,”
The FEBS journal, vol. 282, no. 2, pp. 297–317, 2015.
[42] H. Knauf and E. Mutschler, “Clinical pharmacokinetics and phar-
macodynamics of torasemide,” Clinical pharmacokinetics, vol. 34,
pp. 1–24, 1998.
[43] L. Che, P. Paliogiannis, A. Cigliano, M. G. Pilo, X. Chen, and D. F.
Calvisi, “Pathogenetic, prognostic, and therapeutic role of fatty
 13

acid synthase in human hepatocellular carcinoma,” Frontiers in

oncology, vol. 9, p. 1412, 2019.
[44] X. Tong, F. Zhao, and C. B. Thompson, “The molecular determi-
nants of de novo nucleotide biosynthesis in cancer cells,” Current
opinion in genetics & development, vol. 19, no. 1, pp. 32–37, 2009.
[45] K.-x. Wang, G.-h. Du, X.-m. Qin, and L. Gao, “1h-nmr-based
metabolomics reveals the biomarker panel and molecular mech-
anism of hepatocellular carcinoma progression,” Analytical and
Bioanalytical Chemistry, vol. 414, no. 4, pp. 1525–1537, 2022.
[46] T. Pallister, M. A. Jackson, T. C. Martin, J. Zierer, A. Jennings, R. P.
Mohney, A. MacGregor, C. J. Steves, A. Cassidy, T. D. Spector et al.,
“Hippurate as a metabolomic marker of gut microbiome diversity:
Modulation by diet and relationship to metabolic syndrome,”
Scientific reports, vol. 7, no. 1, p. 13670, 2017.
[47] C.-S. Park, D.-W. Eom, Y. Ahn, H. J. Jang, S. Hwang, and S.-G. Lee,
“Can heme oxygenase-1 be a prognostic factor in patients with
hepatocellular carcinoma?” Medicine, vol. 98, no. 26, 2019.
[48] G. Sass, P. Leukel, V. Schmitz, E. Raskopf, M. Ocker, D. Neureiter,
M. Meissnitzer, E. Tasika, A. Tannapfel, and G. Tiegs, “Inhibition
of heme oxygenase 1 expression by small interfering rna decreases
orthotopic tumor growth in livers of mice,” International journal of
cancer, vol. 123, no. 6, pp. 1269–1277, 2008.
[49] L. Galluzzi, E. Vacchelli, J. Michels, P. Garcia, O. Kepp, L. Senovilla,
I. Vitale, and G. Kroemer, “Effects of vitamin b6 metabolism
on oncogenesis, tumor progression and therapeutic responses,”
Oncogene, vol. 32, no. 42, pp. 4995–5004, 2013.
[50] I. Kaji, M. Tatsuta, H. Iishi, M. Baba, A. Inoue, and H. Kasugai,
“Inhibition by d-limonene of experimental hepatocarcinogenesis
in sprague-dawley rats does not involve p21ras plasma membrane
association,” International journal of cancer, vol. 93, no. 3, pp. 441–
444, 2001.
[51] W. Chen, Y. Liu, M. Li, J. Mao, L. Zhang, R. Huang, X. Jin,
and L. Ye, “Anti-tumor effect of α-pinene on human hepatoma
cell lines through inducing g2/m cell cycle arrest,” Journal of
pharmacological sciences, vol. 127, no. 3, pp. 332–338, 2015.
[52] J. Zhou, Q. Wen, S.-F. Li, Y.-F. Zhang, N. Gao, X. Tian, Y. Fang,
J. Gao, M.-Z. Cui, X.-P. He et al., “Significant change of cytochrome
p450s activities in patients with hepatocellular carcinoma,” Onco-
target, vol. 7, no. 31, p. 50612, 2016.
[53] J. A. Zastre, R. L. Sweet, B. S. Hanberry, and S. Ye, “Linking
vitamin b1 with cancer cell metabolism,” Cancer & metabolism,
vol. 1, no. 1, pp. 1–14, 2013.
[54] L. Chávez-Rodrı́guez, A. Escobedo-Calvario, S. Salas-Silva, R. U.
Miranda-Labra, L. Bucio, V. Souza, M. C. Gutiérrez-Ruiz, and
L. E. Gomez-Quiroz, “Fructose consumption and hepatocellular
carcinoma promotion,” Livers, vol. 1, no. 4, pp. 250–262, 2021.
[55] A. Rosenberg and J. Hirschberg, “V-measure: A conditional
entropy-based external cluster evaluation measure.” in EMNLP-
CoNLL, vol. 7, 2007, pp. 410–420.