Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

BalestraWeb: efficient online evaluation of drug–target interactions

BalestraWeb: efficient online evaluation of drug–target interactions Vol. 31 no. 1 2015, pages 131–133 BIOINFORMATICS APPLICATIONS NOTE doi:10.1093/bioinformatics/btu599 Systems biology Advance Access publication September 5, 2014 BalestraWeb: efficient online evaluation of drug–target interactions 1,2 1,3 1,2 1, Murat Can Cobanoglu ,Zoltan N. Oltvai ,D.Lansing Taylor and Ivet Bahar Department of Computational and Systems Biology, University of Pittsburgh, School of Medicine, Pittsburgh, PA, 2 3 The University of Pittsburgh Drug Discovery Institute, Pittsburgh, PA 15260 and Department of Pathology, University of Pittsburgh, School of Medicine, Pittsburgh, PA 15213, USA Associate Editor: Jonathan Wren ABSTRACT 2 METHOD Summary: BalestraWeb is an online server that allows users to in- BalestraWeb is built by training a latent factor model, as described in our stantly make predictions about the potential occurrence of interactions previous work (Cobanoglu et al., 2013), on approved drugs and their interactions data from DrugBank v3 (Knox et al., 2011). To build the between any given drug–target pair, or predict the most likely inter- latent factor model, we use the GraphLab collaborative filtering toolkit action partners of any drug or target listed in the DrugBank. It also implementation (Low et al., 2010). We mapped all the known names, permits users to identify most similar drugs or most similar targets brand names and synonyms of the drugs and targets to the relevant based on their interaction patterns. Outputs help to develop hypoth- latent factor using a pre-computed hash table that allows constant time eses about drug repurposing as well as potential side effects. access and enables maximal efficiency. Availability and implementation: BalestraWeb is accessible at http:// The server allows users to submit three types of queries: drug–target balestra.csb.pitt.edu/. The tool is built using a probabilistic matrix fac- interaction, drug–drug similarity and target–target similarity. In the torization method and DrugBank v3, and the latent variable models are former case (Fig. 1), the input is mapped to the corresponding drug trained using the GraphLab collaborative filtering toolkit. The server is latent vector (LV) and target LV, and the dot product of these vectors yields a score for the probabilistic occurrence of the queried drug–target implemented using Python, Flask, NumPy and SciPy. interaction. Alternatively, the user can enter a single type of input, either Contact: [email protected] a drug or a target. If a single drug is entered, the server retrieves the LV Received on April 16, 2014; revised on July 3, 2014; accepted on for that drug and screens it against the entire set of LVs corresponding to August 31, 2014 all targets, so as to identify known and newly predicted targets. Drug– drug and target–target similarity queries provide information on drugs (or targets) similar to the query drug (or target) based on the cosine similarity of their LVs. 1 INTRODUCTION The output is an interactive graph (that can be downloaded in JSON format) and a table displaying both the known drug–target interactions Contemporary drug discovery faces important challenges: bring- for the query drug (target) and the top N predicted targets (drugs), rank- ing a new molecular entity to the market is estimated to cost ordered by their score. Users can select to view a second layer of inter- upward of 1.8 billion US$ (Paul et al., 2010), and the rate of actions beyond the immediate neighbors of the query drug/target in the new drug discovery has steadily halved every 9 years for the past bipartite network of drugs and targets. The resulting subnet of inter- 60 years (Scannell et al., 2012). One of the common suggestions actions thus provides a more complete picture of the investigated drug/ brought forth to explain and remedy this trend is a paradigm target in the context of the interactions of their known targets/drugs. shift in drug discovery efforts from high-affinity binding on a In addition to providing information on the distribution of scores in single target toward modulation of cellular network states general in the tutorial, we provide query-specific histograms in the output files: the distribution of the predicted confidence score (for each member through multiple interactions (Csermely et al., 2005, 2013; of the drug–target pairs) or the histogram of cosine similarities (for each Hopkins, 2008; Keskin et al., 2007; Korcsmaros et al., 2007; member of the drug–drug or target–target pairs). These histograms fa- Mencher and Wang, 2005; Zimmermann et al., 2007). The de- cilitate the interpretation of the specific score released for the query pair velopment of computational methods that can efficiently assess in the context of the complete distribution of scores for the investigated potential new interactions for drug repurposing thus became an drug/target, and help make a better assessment of the significance of the important goal in quantitative systems/network pharmacology outputted score. research. We recently introduced a probabilistic matrix factorization 3 DISTINCTIVE FEATURES (PMF) method that can be applied to known drug–target inter- action graphs for predicting new interactions (Cobanoglu et al., There are three important distinguishing features of 2013). Here, we introduce BalestraWeb, a Web server that has BalestraWeb: its ability to provide information on the likelihood been developed for allowing users to efficiently obtain results of interaction between any drug and any target; its reliance on from the PMF analysis, and assess the likelihood of interaction interaction profiles instead of chemical similarity; and its ability between any drug–target pair. to compare drugs with drugs and targets with targets. The widely used similarity ensemble approach (SEA) (Keiser et al., 2007) *To whom correspondence should be addressed. uses ligand similarity to identify new ligands based on chemical The Author 2014. Published by Oxford University Press. This is an Open Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (http://creativecommons.org/licenses/by-nc/4.0/), which permits non-commercial re-use, distribution, and reproduction in any medium, provided the original work is properly cited. For commercial re-use, please contact [email protected] M.C.Cobanoglu et al. Fig. 1. BalestraWeb architecture and underlying methodology. The user input (lower left) is mapped onto the latent factor vector(s) u v (for targets), i j learned by minimizing squared error regularized by Frobenius norm (see equation at top left). The output (right) contains a score R representative of predicted interaction confidence along with a graphical representation of the close neighborhood of the query drug (Sunitinib) and/or query target (SCN5A) in the drug-target association network, along with a table of known and predicted interactions. Similar features hold for drug-drug and target- target similarity searches and outputs similarity. STITCH (Kuhn et al., 2014) is an extensive repository Web server to reflect changes as new data become available. of protein and chemicals (1.07 billion interactions), including Free, fast and easy-to-use BalestraWeb enables researchers to predictions based on chemical similarities of compounds. help eliminate improbable drug–target interactions and effi- However, the interactions predicted by BalestraWeb are not ciently focus their limited resources on selected drugs. based on a particular chemical or genomic similarity method, Funding: Support from the NIH (U19 AI068021 and PO1 but on the assessment of comparable interaction patterns, and DK096990) is gratefully acknowledged by I.B. as such they differ from, or complement, those predicted by SEA or listed by STITCH. Conflict of interest: none declared. 4CONCLUSION REFERENCES BalestraWeb provides users the ability to predict the most likely Cobanoglu,M.C. et al. (2013) Predicting drug-target interactions using probabilistic interaction partners of any drug or target beyond those known matrix factorization. J. Chem. Inf. Model., 53, 3399–3409. and compiled in the DrugBank. The technology used to build the Csermely,P. et al. (2005) The efficiency of multi-target drugs: the network approach might help drug design. Trends Pharmacol. Sci., 26, 178–182. Web server scales linearly with the number of drugs or targets Csermely,P. et al. (2013) Structure and dynamics of molecular networks: a novel and is therefore easily scalable to larger datasets as they become paradigm of drug discovery: a comprehensive review. Pharmacol. Ther., 138, available. Our plan is to regularly update the underlying engine 333–408. and optimized parameters by using the newly released data. The Hopkins,A.L. (2008) Network pharmacology: the next paradigm in drug discovery. modular architecture of the software allows us to update the Nat. Chem. Biol., 4, 682–690. 132 BalestraWeb Keskin,O. et al. (2007) Towards drugs targeting multiple proteins in a systems Low,Y. et al. (2010) Graphlab: a new framework for parallel machine learning. In: biology approach. Curr. Top. Med. Chem., 7, 943–951. Conference on Uncertainty in Artificial Intelligence. Catalina Island, CA. Keiser,M.J. et al. (2007) Relating protein pharmacology by ligand chemistry. Nat. Mencher,S.K. and Wang,L.G. (2005) Promiscuous drugs compared to selective Biotechnol., 25, 197–206. drugs (promiscuity can be a virtue). BMC Clin. Pharmacol., 5,3. Knox,C. et al. (2011) DrugBank 3.0: a comprehensive resource for ‘omics’ research Paul,S.M. et al. (2010) How to improve R&D productivity: the pharmaceutical on drugs. Nucleic Acids Res., 39, D1035–D1041. industry’s grand challenge. Nat. Rev. Drug Discov., 9, 203–214. Korcsmaros,T. et al. (2007) How to design multi-target drugs: target search options Scannell,J.W. et al. (2012) Diagnosing the decline in pharmaceutical R&D effi- in cellular networks. Discovery, 2, 1–10. ciency. Na.t Rev. Drug Discov., 11, 191–200. Kuhn,M. et al. (2014) STITCH 4: integration of protein-chemical interactions with Zimmermann,G.R. et al. (2007) Multi-target therapeutics: when the whole is greater user data. Nucleic. Acids. Res., 42, D401–D407. than the sum of the parts. Drug Discov. Today, 12,34–42. http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Bioinformatics Pubmed Central

BalestraWeb: efficient online evaluation of drug–target interactions

Cobanoglu, Murat Can; Oltvai, Zoltán N.; Taylor, D. Lansing; Bahar, Ivet
Bioinformatics , Volume 31 (1) – Sep 5, 2014
Download PDF
3 pages

Loading next page...
 
/lp/pubmed-central/balestraweb-efficient-online-evaluation-of-drug-target-interactions-3gYfl9gBqg

References (19)

Publisher
Pubmed Central
Copyright
© The Author 2014. Published by Oxford University Press.
ISSN
1367-4803
eISSN
1367-4811
DOI
10.1093/bioinformatics/btu599
Publisher site
See Article on Publisher Site

Abstract

Vol. 31 no. 1 2015, pages 131–133 BIOINFORMATICS APPLICATIONS NOTE doi:10.1093/bioinformatics/btu599 Systems biology Advance Access publication September 5, 2014 BalestraWeb: efficient online evaluation of drug–target interactions 1,2 1,3 1,2 1, Murat Can Cobanoglu ,Zoltan N. Oltvai ,D.Lansing Taylor and Ivet Bahar Department of Computational and Systems Biology, University of Pittsburgh, School of Medicine, Pittsburgh, PA, 2 3 The University of Pittsburgh Drug Discovery Institute, Pittsburgh, PA 15260 and Department of Pathology, University of Pittsburgh, School of Medicine, Pittsburgh, PA 15213, USA Associate Editor: Jonathan Wren ABSTRACT 2 METHOD Summary: BalestraWeb is an online server that allows users to in- BalestraWeb is built by training a latent factor model, as described in our stantly make predictions about the potential occurrence of interactions previous work (Cobanoglu et al., 2013), on approved drugs and their interactions data from DrugBank v3 (Knox et al., 2011). To build the between any given drug–target pair, or predict the most likely inter- latent factor model, we use the GraphLab collaborative filtering toolkit action partners of any drug or target listed in the DrugBank. It also implementation (Low et al., 2010). We mapped all the known names, permits users to identify most similar drugs or most similar targets brand names and synonyms of the drugs and targets to the relevant based on their interaction patterns. Outputs help to develop hypoth- latent factor using a pre-computed hash table that allows constant time eses about drug repurposing as well as potential side effects. access and enables maximal efficiency. Availability and implementation: BalestraWeb is accessible at http:// The server allows users to submit three types of queries: drug–target balestra.csb.pitt.edu/. The tool is built using a probabilistic matrix fac- interaction, drug–drug similarity and target–target similarity. In the torization method and DrugBank v3, and the latent variable models are former case (Fig. 1), the input is mapped to the corresponding drug trained using the GraphLab collaborative filtering toolkit. The server is latent vector (LV) and target LV, and the dot product of these vectors yields a score for the probabilistic occurrence of the queried drug–target implemented using Python, Flask, NumPy and SciPy. interaction. Alternatively, the user can enter a single type of input, either Contact: [email protected] a drug or a target. If a single drug is entered, the server retrieves the LV Received on April 16, 2014; revised on July 3, 2014; accepted on for that drug and screens it against the entire set of LVs corresponding to August 31, 2014 all targets, so as to identify known and newly predicted targets. Drug– drug and target–target similarity queries provide information on drugs (or targets) similar to the query drug (or target) based on the cosine similarity of their LVs. 1 INTRODUCTION The output is an interactive graph (that can be downloaded in JSON format) and a table displaying both the known drug–target interactions Contemporary drug discovery faces important challenges: bring- for the query drug (target) and the top N predicted targets (drugs), rank- ing a new molecular entity to the market is estimated to cost ordered by their score. Users can select to view a second layer of inter- upward of 1.8 billion US$ (Paul et al., 2010), and the rate of actions beyond the immediate neighbors of the query drug/target in the new drug discovery has steadily halved every 9 years for the past bipartite network of drugs and targets. The resulting subnet of inter- 60 years (Scannell et al., 2012). One of the common suggestions actions thus provides a more complete picture of the investigated drug/ brought forth to explain and remedy this trend is a paradigm target in the context of the interactions of their known targets/drugs. shift in drug discovery efforts from high-affinity binding on a In addition to providing information on the distribution of scores in single target toward modulation of cellular network states general in the tutorial, we provide query-specific histograms in the output files: the distribution of the predicted confidence score (for each member through multiple interactions (Csermely et al., 2005, 2013; of the drug–target pairs) or the histogram of cosine similarities (for each Hopkins, 2008; Keskin et al., 2007; Korcsmaros et al., 2007; member of the drug–drug or target–target pairs). These histograms fa- Mencher and Wang, 2005; Zimmermann et al., 2007). The de- cilitate the interpretation of the specific score released for the query pair velopment of computational methods that can efficiently assess in the context of the complete distribution of scores for the investigated potential new interactions for drug repurposing thus became an drug/target, and help make a better assessment of the significance of the important goal in quantitative systems/network pharmacology outputted score. research. We recently introduced a probabilistic matrix factorization 3 DISTINCTIVE FEATURES (PMF) method that can be applied to known drug–target inter- action graphs for predicting new interactions (Cobanoglu et al., There are three important distinguishing features of 2013). Here, we introduce BalestraWeb, a Web server that has BalestraWeb: its ability to provide information on the likelihood been developed for allowing users to efficiently obtain results of interaction between any drug and any target; its reliance on from the PMF analysis, and assess the likelihood of interaction interaction profiles instead of chemical similarity; and its ability between any drug–target pair. to compare drugs with drugs and targets with targets. The widely used similarity ensemble approach (SEA) (Keiser et al., 2007) *To whom correspondence should be addressed. uses ligand similarity to identify new ligands based on chemical The Author 2014. Published by Oxford University Press. This is an Open Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (http://creativecommons.org/licenses/by-nc/4.0/), which permits non-commercial re-use, distribution, and reproduction in any medium, provided the original work is properly cited. For commercial re-use, please contact [email protected] M.C.Cobanoglu et al. Fig. 1. BalestraWeb architecture and underlying methodology. The user input (lower left) is mapped onto the latent factor vector(s) u v (for targets), i j learned by minimizing squared error regularized by Frobenius norm (see equation at top left). The output (right) contains a score R representative of predicted interaction confidence along with a graphical representation of the close neighborhood of the query drug (Sunitinib) and/or query target (SCN5A) in the drug-target association network, along with a table of known and predicted interactions. Similar features hold for drug-drug and target- target similarity searches and outputs similarity. STITCH (Kuhn et al., 2014) is an extensive repository Web server to reflect changes as new data become available. of protein and chemicals (1.07 billion interactions), including Free, fast and easy-to-use BalestraWeb enables researchers to predictions based on chemical similarities of compounds. help eliminate improbable drug–target interactions and effi- However, the interactions predicted by BalestraWeb are not ciently focus their limited resources on selected drugs. based on a particular chemical or genomic similarity method, Funding: Support from the NIH (U19 AI068021 and PO1 but on the assessment of comparable interaction patterns, and DK096990) is gratefully acknowledged by I.B. as such they differ from, or complement, those predicted by SEA or listed by STITCH. Conflict of interest: none declared. 4CONCLUSION REFERENCES BalestraWeb provides users the ability to predict the most likely Cobanoglu,M.C. et al. (2013) Predicting drug-target interactions using probabilistic interaction partners of any drug or target beyond those known matrix factorization. J. Chem. Inf. Model., 53, 3399–3409. and compiled in the DrugBank. The technology used to build the Csermely,P. et al. (2005) The efficiency of multi-target drugs: the network approach might help drug design. Trends Pharmacol. Sci., 26, 178–182. Web server scales linearly with the number of drugs or targets Csermely,P. et al. (2013) Structure and dynamics of molecular networks: a novel and is therefore easily scalable to larger datasets as they become paradigm of drug discovery: a comprehensive review. Pharmacol. Ther., 138, available. Our plan is to regularly update the underlying engine 333–408. and optimized parameters by using the newly released data. The Hopkins,A.L. (2008) Network pharmacology: the next paradigm in drug discovery. modular architecture of the software allows us to update the Nat. Chem. Biol., 4, 682–690. 132 BalestraWeb Keskin,O. et al. (2007) Towards drugs targeting multiple proteins in a systems Low,Y. et al. (2010) Graphlab: a new framework for parallel machine learning. In: biology approach. Curr. Top. Med. Chem., 7, 943–951. Conference on Uncertainty in Artificial Intelligence. Catalina Island, CA. Keiser,M.J. et al. (2007) Relating protein pharmacology by ligand chemistry. Nat. Mencher,S.K. and Wang,L.G. (2005) Promiscuous drugs compared to selective Biotechnol., 25, 197–206. drugs (promiscuity can be a virtue). BMC Clin. Pharmacol., 5,3. Knox,C. et al. (2011) DrugBank 3.0: a comprehensive resource for ‘omics’ research Paul,S.M. et al. (2010) How to improve R&D productivity: the pharmaceutical on drugs. Nucleic Acids Res., 39, D1035–D1041. industry’s grand challenge. Nat. Rev. Drug Discov., 9, 203–214. Korcsmaros,T. et al. (2007) How to design multi-target drugs: target search options Scannell,J.W. et al. (2012) Diagnosing the decline in pharmaceutical R&D effi- in cellular networks. Discovery, 2, 1–10. ciency. Na.t Rev. Drug Discov., 11, 191–200. Kuhn,M. et al. (2014) STITCH 4: integration of protein-chemical interactions with Zimmermann,G.R. et al. (2007) Multi-target therapeutics: when the whole is greater user data. Nucleic. Acids. Res., 42, D401–D407. than the sum of the parts. Drug Discov. Today, 12,34–42.

Journal

Bioinformatics – Pubmed Central

Published: Sep 5, 2014

There are no references for this article.