“output” — 2019/2/11 — page 1 — #1
Bioinformatics
doi.10.1093/bioinformatics/xxxxxx
Advance Access Publication Date: Day Month Year
Manuscript Category
Subject Section
Implementations of the chemical structural and
compositional similarity metric in R and Python
Asker Brejnrod
1,∗
, Madeleine Ernst
2,3
, Piotr Dworzynski
1,4
, Lasse Buur
Rasmussen
1
, Pieter C. Dorrestein
2,3,6,
, Justin J.J. van der Hooft
5
,
Manimozhiyan Arumugam
1
1
Novo Nordic Foundation Center for Basic Metabolic Research, University of Copenhagen, Copenhagen, 2200, Denmar k
2
Collaborative Mass Spectrometry Innovation Center, Skaggs School of Pharmacy and Pharmaceutical Sciences, University of
California, San Diego, La Jolla, CA 92093, USA
3
Skaggs School of Pharmacy and Pharmaceutical Sciences, University of California, San Diego, La Jolla, CA 92093, USA
4
Department of Epidemiology Research, Statens Serum Institut, Copenhagen, Denmark
5
Bioinformatics Group, Plant Sciences Group, Wageningen University, Wageningen, The Netherlands
6
Center for Microbiome Innovation, University of California, San Diego, La Jolla, CA 92093, USA
∗
brejnrod@sund.ku.dk
Associate Editor: XXXXXXX
Received on XXXXX; revised on XXXXX; accepted on XXXXX
Abstract
Motivation: Tandem mass spectrometry (MS/MS) has the potential to substantially improve metabolomics
by acquiring spectra of fragmented ions. These fragmentation spectra can be represented as a molecular
network, by measuring cosine distances between them, thus identifying signals from the same or similar
molecules. Metrics that enable comparison between pairs of samples based on their metabolite profiles
are in great need. Taking inspiration from the successful phylogeny-aware beta-diversity measures used
in microbiome research, integrating chemical similarity information about the features in addition to
their abundances could lead to better insights when comparing metabolite profiles. Chemical Structural
and Compositional Similarity (CSCS) is a recently published similarity metric comparing the full set of
signals and their chemical similarity between two samples. Efficient, scalable and easily accessible
implementations of this algorithm is currently lacking. Here, we present an easily accessible and scalable
implementation of CSCS in both python and R, including a version not weighted by intensity information.
Results: We provide a new implementation of the CSCS algorithm that is over 300 times faster
than the published implementation in R, making the algorithm suitable for large-scale metabolomics
applications. We also show that adding chemical information enriches existing methods. Furthermore,
the R implementation includes functions for exporting molecular networks directly from the mass spectral
molecular networking platform GNPS for ease of use for downstream applications.
Contact: brejnrod@sund.ku.dk
Availability: github.com/askerdb/rCSCS, github.com/askerdb/pyCSCS
1 Introduction
Liquid chromatography tandem - mass spectrometry (LC-MS/MS) is
gaining more and more popularity in the metabolomics field with a
wide range of applications (e.g. [11, 9, 13, 12]). As an extension of
LC-MS, one of the most frequently used analytical platforms in mass
spectrometry-based metabolomics [27, 4], LC-MS/MS does not only
provide a high coverage and sensitivity towards semi-polar metabolites,
but also provides chemical structural information of the metabolites
investigated. In LC-MS/MS, metabolites are ionized and fragmented and
the resulting fragmentation fingerprints of mass-to-charge ratios (m/z) are
characteristic to the molecular structure of the metabolite. Furthermore,
metabolites resulting in similar fragmentation patterns presumably exhibit
similar chemical structures [24, 25]. The similarity of these fingerprints can
be calculated using the cosine score and represented as a graph, resulting
in so-called mass spectral molecular networks (recently reviewed in [14]).
This approach, popularized by the user-friendly Global Natural Products
Social Molecular Networking (GNPS) platform [24], has been highly
effective in visualizing structural chemical relatedness across samples
as well as in aiding the structural identification of metabolites. Several
studies have focused on deriving and applying new distance metrics to
1
.CC-BY-ND 4.0 International licenseis the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a
The copyright holder for this preprint (which was not certified by peer review)this version posted February 11, 2019. ; https://doi.org/10.1101/546150doi: bioRxiv preprint
“output” — 2019/2/11 — page 2 — #2
2 Sample et al.
metabolomics profiles to compare pairs of samples [16, 15, 8]. Among
these, the chemical structural and compositional similarity (CSCS) as
proposed by Sedio and collaborators [19, 17, 18], accounts for the chemical
structural similarity across metabolites by integrating the similarity of their
MS/MS fragmentation patterns through the cosine score. Even though
metabolites can be identified for only about 2 % of all signals in a typical
LC-MS/MS experiment [1], CSCS has the advantage to integrate chemical
structural information without depending on metabolite identification, thus
enabling a comparison of the entire range of detected molecules. CSCS can
be used to visualize chemical differences across samples by using Principal
Coordinate Analysis (PCoA) plots [5] or distance based hypothesis testing
of differences. Here, we implement CSCS in user-friendly python and R
packages, and demonstrate its performance by computing commonly used
metrics on nearly 500 LC-MS/MS samples of the American Gut Project
[9] as well as fecal samples obtained from children with Crohn’s disease
at different time points during nutritional therapy [22].
2 Methods
2.1 Mathematical exposition
Consider two samples A and B with the union of features [1, ..., C], and
intensity vectors A = [I
a1
, ..., I
aC
] and B = [I
b1
, ..., I
bC
]. Calculation of
the Chemical Structural and Compositional Similarity (CSCS) takes the
following inputs: a chemical similarity matrix, CSS, and the two vectors
of ion intensities, A and B. CSS is defined as a matrix of pairwise cosine
distances
CSS =
cosΘ
1,1
... cosΘ
1,C
cosΘ
2,1
... cosΘ
2,C
cosΘ
C,1
... cosΘ
C,C
In practice this matrix can be calculated from a file specifying pairwise
cosine distances across nodes in a mass spectral molecular network, which
can be downloaded from GNPS.
The CSCS as defined by Sedio and collaborators [19], referred to here as
weighted CSCS (CSCSw) is then defined as:
CSCSw =
CSS × AB
T
max(CSS × AA
T
, CSS × BB
T
)
where × is element-wise multiplication.
Analogously, we here define the unweighted CSCSu where all
elements of A and B are dichotomized into 1 or 0 upon presence or
absence of a signal in the sample, respectively. CSCS is a metric of
similarity. However, many downstream analyses within metabolomics
rely on a dissimilarity matrix (e.g. PCoA), thus our libraries return a
dissimilarity matrix, corresponding to 1-CSCS.
2.2 Implementation
Implementations are available at https://github.com/askerdb/
rCSCS as an R [21] package, which can be installed through devtools [26],
and at https://anaconda.org/askerdb/pycscs as a python
package, which can be installed through conda. The R package depends
on the packages foreach [2], igraph [3] and Rcurl[7], wheras the python
package depends on numpy [23], pandas [10], scikit bio (http://
scikit-bio.org) as well as the sparse matrix implementation in scipy
[6].
2.3 Data
Run time benchmarks were computed on 18 fecal samples obtained
from children with Crohn’s disease at different time points during
Implementation Time (s)
rCSCS 1.53
pyCSCS 1.66
Sedio et al. 507.31
Table 1. CSCS timing benchmarks for a dataset consisting of 18 fecal samples
obtained from children with Crohn’s disease at different time points during
nutritional therapy [22].
nutritional therapy [22]. This dataset is publicly available at https:
//massive.ucsd.edu/ under the MassIVE accession number
MSV000081120. A total of 1245 MS/MS features were retrieved
for this dataset using the GNPS networking parameters publicly
accessible at https://gnps.ucsd.edu/ProteoSAFe/status.
jsp?task=b0524246804a4b50a8a4ec6244a8be2e. Samples
from the American Gut Project data are publicly available under the
MassIVE accession number MSV000080179 [9]. This dataset consisted
of 489 samples and 16349 features using GNPS networking parameters
publicly accessible at https://gnps.ucsd.edu/ProteoSAFe/
status.jsp?task=a07557dc26cc4d3f8a2076d5ae0898a2.
3 Results
3.1 Benchmarks
We computed CSCS for a dataset consisting of 18 fecal samples obtained
from children with Crohn’s disease at different time points during
nutritional therapy [22] and a total of 1245 MS/MS features, a realistically
sized dataset on a single core of a Macbook Pro (2.8 GHz i7), and compared
it against the published implementation from Sedio and collaborators [19].
Results shown in Table 1 include run time overhead that should be minimal
in realistic applications. To demonstrate the utility of this implementation
on a scale that is relevant to high-throughput collection of data, we used
the American Gut Project that consists of 489 samples with 16349 MS/MS
features when downloaded from GNPS. Analysis of this dataset with our
python implementation finished in 7.5 hours on 40 CPU cores. We did not
compare run time of the original implementation [19], as this will likely
not finish in a reasonable time.
3.2 Chemical information produces distinct similarities
To evaluate how the CSCS metrics compare with other popular metrics, we
compared them with Bray-Curtis, Jaccard and Canberra metrics estimated
for a dataset of 594 metabolome samples from the American Gut Project
[9]. We used Procrustes analysis[20] to measure the similarity between
the distance matrices and used this information as input for a PCoA plot
(Figure 1). Metrics with chemical information are clearly separated from
those without, and on this dataset the weighting of chemical information
with ion intensities drives the separation on the axis that explains the most
variance, while the unweighted CSCS is distinctly separated on the second
principal component.
4 Conclusion
We have implemented the calculation of weighted and unweighted CSCS
distances in R and Python. Through benchmarking of publicly available
data we have demonstrated the highly significant run time improvement,
which expands the application of CSCS to large datasets.
.CC-BY-ND 4.0 International licenseis the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a
The copyright holder for this preprint (which was not certified by peer review)this version posted February 11, 2019. ; https://doi.org/10.1101/546150doi: bioRxiv preprint
“output” — 2019/2/11 — page 3 — #3
short Title 3
Fig. 1. Figure 1: PCoA of the procrustes distances between various distance metrics
commonly used in metabolomics. Metrics colored in green include chemical information,
either weighted by ion intensities or unweighted. Metrics with no chemical information are
colored red. information
Acknowledgements
We would like to acknowledge Brian Sedio and collaborators for proposing
and implementing the original chemical structural and compositional
similarity metric, and making software available for direct comparisons.
Funding
Asker Brejnrod was supported by Independent Research Fund Denmark
(DFF-6111-00471). JJJvdH was supported by an ASDI eScience grant
(ASDI.2017.030) from the Netherlands eScience Center (NLeSC).
References
[1]Alexander A Aksenov, Ricardo da Silva, Rob Knight, Norberto P Lopes, and
Pieter C Dorrestein. Global chemical analysis of biology by mass spectrometry.
Nature Reviews Chemistry, 1(7):0054, 2017.
[2]Revolution Analytics and Steve Weston. foreach: Foreach looping construct for
r. R package version, 1(1):2013, 2013.
[3]Gabor Csardi and Tamas Nepusz. The igraph software package for complex
network research. InterJournal, Complex Systems, 1695(5):1–9, 2006.
[4]Madeleine Ernst, Denise Brentan Silva, Ricardo Roberto Silva, Ricardo Z. N.
Vncio, and Norberto Peporine Lopes. Mass spectrometry in plant metabolomics
strategies: from analytical platforms to data acquisition and processing. Nat.
Prod. Rep., 31:784–806, 2014.
[5]John C Gower and Pierre Legendre. Metric and euclidean properties of
dissimilarity coefficients. Journal of classification, 3(1):5–48, 1986.
[6]Eric Jones, Travis Oliphant, and Pearu Peterson. SciPy: open source scientific
tools for Python. 2014.
[7]Duncan Temple Lang. Rcurl: General network (http/ftp/...) client interface for
r. R package version, 1(1), 2012.
[8]Kang Liu, Azian Azamimi Abdullah, Ming Huang, Takaaki Nishioka, Md Altaf-
Ul-Amin, and Shigehiko Kanaya. Novel approach to classify plants based on
metabolite-content similarity. BioMed research international, 2017, 2017.
[9]Daniel McDonald, Embriette Hyde, Justine W Debelius, James T Morton,
Antonio Gonzalez, Gail Ackermann, Alexander A Aksenov, Bahar Behsaz,
Caitriona Brennan, Yingfeng Chen, et al. American gut: an open platform for
citizen science microbiome research. mSystems, 3(3):e00031–18, 2018.
[10]Wes McKinney et al. Data structures for statistical computing in python. In
Proceedings of the 9th Python in Science Conference, volume 445, pp. 51–56.
Austin, TX, 2010.
[11]Don D Nguyen, Alexey V Melnik, Nobuhiro Koyama, Xiaowen Lu, Michelle
Schorn, Jinshu Fang, Kristen Aguinaldo, Jr Lincecum, Tommie L, Maarten G K
Ghequire, Victor J Carrion, Tina L Cheng, Brendan M Duggan, Jacob G Malone,
Tim H Mauchline, Laura M Sanchez, A Marm Kilpatrick, Jos M Raaijmakers,
René De Mot, Bradley S Moore, Marnix H Medema, and Pieter C Dorrestein.
Indexing the pseudomonas specialized metabolome enabled the discovery of
poaeamide b and the bananamides. Nature microbiology, 2:16197–16197, 10
2016.
[12]Florent Olivon, Pierre-Marie Allard, Alexey Koval, Davide Righi, Gregory
Genta-Jouve, Johan Neyts, Cécile Apel, Christophe Pannecouque, Louis-
Félix Nothias, Xavier Cachet, Laurence Marcourt, Fanny Roussi, Vladimir L.
Katanaev, David Touboul, Jean-Luc Wolfender, and Marc Litaudon. Bioactive
natural products prioritization using massive multi-informational molecular
networks. ACS Chemical Biology, 12(10):2644–2651, 10 2017.
[13]Daniel Petras, Louis-Felix Nothias, Robert A. Quinn, Theodore Alexandrov,
Nuno Bandeira, Amina Bouslimani, Gabriel Castro-Falcon, Liangyu Chen, Tam
.CC-BY-ND 4.0 International licenseis the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a
The copyright holder for this preprint (which was not certified by peer review)this version posted February 11, 2019. ; https://doi.org/10.1101/546150doi: bioRxiv preprint
“output” — 2019/2/11 — page 4 — #4
4 Sample et al.
Dang, Dimitrios J. Floros, Vivian Hook, Neha Garg, Nicole Hoffner, Yike Jiang,
Clifford A. Kapono, Irina Koester, Rob Knight, Christopher A. Leber, Tie-Jun
Ling, Tal Luzzatto-Knaan, Laura-Isobel McCall, Aaron P. McGrath, Michael J.
Meehan, Jonathan K. Merritt, Robert H. Mills, Jamie Morton, Sonia Podvin,
Ivan Protsyuk, Trevor Purdy, Kendall Satterfield, Stephen Searles, Sahil Shah,
Sarah Shires, Dana Steffen, Margot White, Jelena Todoric, Robert Tuttle, Aneta
Wojnicz, Valerie Sapp, Fernando Vargas, Jin Yang, Chao Zhang, and Pieter C.
Dorrestein. Mass spectrometry-based visualization of molecules associated with
human habitats. Analytical Chemistry, 88(22):10775–10784, 2016. PMID:
27732780.
[14]Robert A Quinn, Louis-Felix Nothias, Oliver Vining, Michael Meehan, Eduardo
Esquenazi, and Pieter C Dorrestein. Molecular networking as a drug discovery,
drug metabolism, and precision medicine strategy. Trends in pharmacological
sciences, 38(2):143–154, 2017.
[15]Anita Rácz, Filip Andri´c, Dávid Bajusz, and Károly Héberger. Binary
similarity measures for fingerprint analysis of qualitative metabolomic profiles.
Metabolomics, 14(3):29, 2018.
[16]Lora A. Richards, Lee A. Dyer, Matthew L. Forister, Angela M. Smilanich,
Craig D. Dodson, Michael D. Leonard, and Christopher S. Jeffrey.
Phytochemical diversity drives plant–insect community diversity. Proceedings
of the National Academy of Sciences, 112(35):10973–10978, 2015.
[17]Brian E. Sedio. Recent breakthroughs in metabolomics promise to reveal the
cryptic chemical traits that mediate plant community composition, character
evolution and lineage diversification. New Phytologist, 214(3):952–958, 2017.
[18]Brian E. Sedio, Cristopher A. Boya P., and Juan Camilo Rojas Echeverri.
A protocol for high-throughput, untargeted forest community metabolomics
using mass spectrometry molecular networks. Applications in Plant Sciences,
6(3):e1033, 2018.
[19]Brian E Sedio, Juan C Rojas Echeverri, P Boya, A Cristopher, and S Joseph
Wright. Sources of variation in foliar secondary chemistry in a tropical forest
tree community. Ecology, 98(3):616–623, 2017.
[20]Bernard W Silverman. Density estimation for statistics and data analysis.
Routledge, 2018.
[21]R Core Team et al. R: A language and environment for statistical computing.
2013.
[22]Justin J. J. van der Hooft, Joe Wandy, Francesca Young, Sandosh Padmanabhan,
Konstantinos Gerasimidis, Karl E. V. Burgess, Michael P. Barrett, and
Simon Rogers. Unsupervised discovery and comparison of structural families
across multiple samples in untargeted metabolomics. Analytical Chemistry,
89(14):7569–7577, 2017. PMID: 28621528.
[23]Stéfan van der Walt, S Chris Colbert, and Gael Varoquaux. The numpy
array: a structure for efficient numerical computation. Computing in Science
& Engineering, 13(2):22–30, 2011.
[24]Mingxun Wang, Jeremy J Carver, Vanessa V Phelan, Laura M Sanchez,
Neha Garg, Yao Peng, Don Duy Nguyen, Jeramie Watrous, Clifford A
Kapono, Tal Luzzatto-Knaan, et al. Sharing and community curation of mass
spectrometry data with global natural products social molecular networking.
Nature biotechnology, 34(8):828, 2016.
[25]Jeramie Watrous, Patrick Roach, Theodore Alexandrov, Brandi S. Heath,
Jane Y. Yang, Roland D. Kersten, Menno van der Voort, Kit Pogliano,
Harald Gross, Jos M. Raaijmakers, Bradley S. Moore, Julia Laskin, Nuno
Bandeira, and Pieter C. Dorrestein. Mass spectral molecular networking of
living microbial colonies. Proceedings of the National Academy of Sciences,
109(26):E1743–E1752, 2012.
[26]Hadley Wickham and Winston Chang. Devtools: Tools to make developing r
packages easier. R package version, 1(0), 2016.
[27]Jean-Luc Wolfender, Serge Rudaz, Young Hae Choi, and Hye Kyong Kim. Plant
metabolomics: from holistic data to relevant biomarkers. Current Medicinal
Chemistry, 20(8):1056–1090, 2013.
.CC-BY-ND 4.0 International licenseis the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under a
The copyright holder for this preprint (which was not certified by peer review)this version posted February 11, 2019. ; https://doi.org/10.1101/546150doi: bioRxiv preprint
![Table 1. CSCS timing benchmarks for a dataset consisting of 18 fecal samples obtained from children with Crohn’s disease at different time points during nutritional therapy [22].](/figures/table-1-cscs-timing-benchmarks-for-a-dataset-consisting-of-2vfkh0vf.webp)