G Model
IJP 16548 No. of Pages 4
International Journal of Pharmaceutics xxx (2016) xxx–xxx
Contents lists available at ScienceDirect
International Journal of Pharmaceutics
journal homepage: www.elsevier.com/locate/ijpharm
Bitterness prediction in-silico: A step towards better drugs
Malkeet Singh Bahiaa,b,1 , Ido Nissima,b,1, Masha Y. Niva,b,*
a
Institute of Biochemistry, Food Science and Nutrition, The Robert H. Smith Faculty of Agriculture, Food, and Environment, The Hebrew University of
Jerusalem, Rehovot, 76100, Israel
b
The Fritz Haber Center for Molecular Dynamics, The Hebrew University of Jerusalem, 91904, Israel
A R T I C L E I N F O
Article history:
Received 13 February 2017
Received in revised form 23 March 2017
Accepted 27 March 2017
Available online xxx
Keywords:
Bitterness
GPCRs
Taste receptors
Promiscuity
in-silico
Prediction
Pediatric drugs
Tas2Rs
A B S T R A C T
Bitter taste is innately aversive and thought to protect against consuming poisons. Bitter taste receptors
(Tas2Rs) are G-protein coupled receptors, expressed both orally and extra-orally and proposed as novel
targets for several indications, including asthma. Many clinical drugs elicit bitter taste, suggesting the
possibility of drugs re-purposing. On the other hand, the bitter taste of medicine presents a major
compliance problem for pediatric drugs. Thus, efficient tools for predicting, measuring and masking
bitterness of active pharmaceutical ingredients (APIs) are required by the pharmaceutical industry. Here
we highlight the BitterDB database of bitter compounds and survey the main computational approaches
to prediction of bitter taste based on compound's chemical structure. Current in silico bitterness
prediction methods provide encouraging results, can be constantly improved using growing
experimental data, and present a reliable and efficient addition to the APIs development toolbox.
© 2017 Elsevier B.V. All rights reserved.
1. Introduction
Bitter taste is one of the basic taste modalities, the role of which
is typically linked to guarding against consumption of poisons
(Chandrashekar et al., 2006). However, not all bitter compounds
are toxic: many dietary phytonutrients commonly found in fruits
and vegetables (Drewnowski and Gomez-Carneros, 2000), as well
as many clinical drugs (Mennella et al., 2013) elicit bitter taste
sensation. The bitterness of drug molecules presents a major
problem of compliance for children (Mennella and Beauchamp,
2008). Sensory tasting of drug candidates by humans is not a trivial
matter, since it requires ethical approval achievable only after a
thorough toxicological study. Thus, efficient prediction of compounds’ bitterness in the initial stages of drug discovery is of great
interest. Several computational (in-silico) studies predicted bitter
taste of compounds based on their chemical structure, as described
below and summarized in Fig. 1. The cost-effectiveness and the
possibility to improve the prediction quality based on the growing
body of experimental data suggest that in-silico bitterness
* Corresponding author at: Institute of Biochemistry, Food Science and Nutrition,
The Robert H. Smith Faculty of Agriculture, Food, and Environment, The Hebrew
University of Jerusalem, Rehovot, 76100, Israel.
E-mail address: masha.niv@mail.huji.ac.il (M.Y. Niv).
1
These authors contributed equally to this work.
prediction could become a practical step in the process of
developing pediatric drugs.
2. Bitter compounds: databases and properties
We have recently established the BitterDB (Wiener et al., 2012),
a freely accessible repository of compounds that were reported to
be bitter for humans or to activate bitter taste receptors (Tas2Rs) in
cell-based functional assays. The bitter-tasting compounds are
recognized by a subfamily of G-protein coupled receptors, the
Tas2R (Chandrashekar et al., 2006). This family is comprised of 25
receptor subtypes in humans, activated by varying number of
partially activated ligands (Meyerhof et al., 2010). Most of the
information on the bitter compounds was extracted from the
Merck Index (O’Neil, 2006), Fenaroli’s handbook of flavor
ingredients (Burdock and Fenaroli, 2005) and research articles
retrieved from PubMed. Currently, the BitterDB comprises close to
700 bitter compounds, 120 of which were experimentally assigned
to their cognate Tas2Rs. BitterDB can be searched using words,
molecular identifiers (such as SMILES or CAS registry number),
chemical properties, or the associated bitter taste receptors.
Additionally, chemical structures may be submitted as queries to
find identical or similar compounds within the BitterDB. Notably,
bitter compounds, even those that activate the same bitter taste
receptor subtype, may vary dramatically in their chemical
structures and physicochemical properties (Di Pizio and Niv,
http://dx.doi.org/10.1016/j.ijpharm.2017.03.076
0378-5173/© 2017 Elsevier B.V. All rights reserved.
Please cite this article in press as: M.S. Bahia, et al., Bitterness prediction in-silico: A step towards better drugs, Int J Pharmaceut (2017), http://
dx.doi.org/10.1016/j.ijpharm.2017.03.076
G Model
IJP 16548 No. of Pages 4
2
M.S. Bahia et al. / International Journal of Pharmaceutics xxx (2016) xxx–xxx
There are several methods for predicting bitterness of compounds (see Fig. 1), as summarized below.
are applicable when biological data is available for a focused
chemical series. Such models were established for the prediction of
bitterness of several analogues of sesquiterpene lactones and for
some classes of peptides (Ley, 2008). Other well-established
techniques use ‘ligand-based pharmacophore’ (LBP) models which
are three-dimensional representations of structural features
conserved among known actives, and ‘shape-based screening’
which allows to filter compounds based upon shape and
electrostatics similarity to the query molecule. Roland and coworkers (Roland et al., 2013) described the structural requirements
of flavonoids for the activation of two bitter receptors – Tas2R14
and 39. Using a test set of 73 (for Tas2R14) and 77 (for Tas2R39)
compounds, the LBP models predicted activation of these receptors
by flavonoids. 68% sensitivity (the ratio between predicted true
positives and the total true positives, also known as true positives
rate) and 65% specificity (the ratio between predicted true
negatives and the total true negatives, also known as true
negatives rate) was achieved for Tas2R14 and 85% sensitivity,
78% specificity for Tas2R39. Additionally, favorable and unfavorable ligand molecular features for receptor activation were
highlighted.
Levit and co-workers screened the BitterDB compounds which
were not yet assigned to a particular taste receptor, and the
DrugBank dataset of clinical drugs (Wishart et al., 2006) for
prospective prediction of Tas2R14 activators using LBP and shapebased models. Subsequently, 9 out of 11 BitterDB predicted
compounds and 11 out of 23 DrugBank predicted compounds were
experimentally confirmed as Tas2R14 activators (Levit et al., 2014).
3.1. Ligand-based methods
3.2. Structure-based methods
Identification of previously unknown bitter compounds can
rely on similarity to known bitterants, an approach well
established in drug discovery and widely used when the structure
of the target protein is not available (Sliwoski et al., 2014). For
example, Quantitative Structure-Activity-Relation (QSAR) models
With the constant rise of structural information of the target
proteins, including GPCRs (Di Pizio et al., 2016; Yarnitzky et al.,
2010), structure-based methods which predict novel compounds
that favorably fit into the binding site of the target (Irwin and
Shoichet, 2016; Kitchen et al., 2004) become more and more
2015; Levit et al., 2014). For example, though hydrophobicity was
suggested to describe the bitterness of peptides (Ney's Q rule (Ney,
1971)), it was not predictive of bitterness of soy (Cho et al., 2004) or
cheese (Toelstede and Hofmann, 2008) peptides. Indeed, the value
of logP (octanol-water partition) of bitter compounds varies rather
widely, i.e. 6.292 for streptomycin or 4.836 for MgBr2 vs. 6.484
for adlupulone from beer or 6.417 for Polysorbate 60 (data obtained
via the BitterDB http://bitterdb.agri.huji.ac.il/dbbitter.php#compoundBrowse).
An early attempt to classify bitter compounds (Rodgers et al.,
2005) used a large clustering analysis to build a hierarchical
phylogenetic-like tree based on the maximal common substructures of over 800 proprietary bitter molecules. The final tree
contained 93 nodes with at least 8 members each, suggesting there
are at least 93 sub-structures that may be linked to bitterness.
Some bitter compounds are Tas2R-selective, meaning that they
activate a single Tas2R subtype. Others activate multiple Tas2R, and
can be termed Tas2R-promiscuous. Overall, we found that the
Tas2R-promiscous compounds tend to be smaller, more globular
and more hydrophobic than Tas2R-selective compounds. A linear
regression tool built using these features correctly classified an
external set of bitter molecules into Tas2R-selective and Tas2Rpromiscuous bitterants (Di Pizio and Niv, 2015).
3. Predicting bitter compounds
Fig. 1. Iterative usage of experimental data towards improvement of computational predictive tools.
Please cite this article in press as: M.S. Bahia, et al., Bitterness prediction in-silico: A step towards better drugs, Int J Pharmaceut (2017), http://
dx.doi.org/10.1016/j.ijpharm.2017.03.076
G Model
IJP 16548 No. of Pages 4
M.S. Bahia et al. / International Journal of Pharmaceutics xxx (2016) xxx–xxx
applicable. The ‘structure-based pharmacophore’ (SBP) approach
translates the receptor’s binding site into a three-dimensional
representation of structural features (Wolber and Langer, 2005).
Such techniques were shown to provide differentiation between
known activators and known non-activators of Tas2R39 (Roland
et al., 2015). The best model displayed 63% sensitivity (specificity
was not reported) when used on a test set of 176 molecules
containing 92 actives and 84 non-actives of Tas2R39. Docking of
ligands to proteins are methods that explore possible binding
poses of the ligands in the (usually pre-defined) binding site, and
return the orientation with the best score (a simplified estimate of
the most favorable free energy of binding). Docking to homology
models has worked well for some GPCRs, as assessed by success in
prospective prediction of new X-ray structures of ligand-receptor
complexes via community wide GPCRdock experiments (Kufareva
et al., 2014), and from the ability to predict novel ligands, i.e.
(Carlsson et al., 2011). Docking methods were used to elucidate
details of bitterant-Tas2R interactions and to rationalize activities
(Levit et al., 2012; Di&Pizio and Niv, 2014; Karaman et al., 2016) and
for discovering novel bitter ligands (unpublished data). In some
cases, atomistic molecular dynamics simulations were integrated
with ligand docking towards a more accurate investigation of
binding modes of bitter compounds (Marchiori et al., 2013; Sandal
et al., 2015).
While ligand- and structure-based methods successfully
discovered novel ligands of particular bitter taste receptors, they
are less suited towards the general classification of compounds as
either bitter or non-bitter: a compound that does not activate a
particular Tas2R subtype could still activate one or several of the
remaining 24 Tas2Rs. A labor-intensive approach to overcome this
problem is to develop ligand-based models for all subgroups of
active ligands, and supplement those with structure-based
predictors for each one of the 25 human bitter taste receptors.
Another, more practical alternative, is to develop general bitter/
non-bitter classifiers using modern machine learning (ML)
techniques.
3.3. Machine learning techniques
ML is a large field of research in computer science, dedicated to
developing algorithms that learn from data in supervised or
unsupervised manner. In practical terms, a training set of data is
provided as input for an ML algorithm (such as artificial neural
networks, Naïve Bayes classification, Support Vector Machine
(SVM), random forest, etc.). The algorithm iteratively learns how to
classify molecules as either positive or negative within the training
set, and then validated on test set(s). Model performance may vary
wildly, and critically depends on the training set provided.
Rodgers et al. used a dataset of 649 bitter and 13530 randomly
selected (and assumed non-bitter) molecules to develop a Naïve
Bayes classifier based on circular fingerprints (MOLPRINT 2D) and
information-gain feature selection (Rodgers et al., 2006). The
selected classifier had 70% sensitivity and 93% specificity in a 5-fold
cross-validation.
Bitterness predictor BitterX (Huang et al., 2016) combined
ligand-based descriptors with a SVM approach. The true positives
set was built from BitterDB and additional manually curated
compounds (500 compounds in total); the true negatives set was
comprised mainly of compounds not described as ‘bitter’ in the
literature. 500 representative structures obtained by a clustering
procedure were chosen in order to balance the bitter and the nonbitter sets. The predictor was trained and applied to an in-house
dataset of 220 compounds. Among compounds predicted as bitter,
five compounds (four of which appear in the BitterDB) were
experimentally tested in cell-based assays and confirmed to
activate at least one Tas2R.
3
Our lab is currently extending the BitterDB to include a set of
compounds that may serve as a reliable non-bitter set. Using this
set as true negatives and the BitterDB set as true positives, we have
developed an ‘AdaBoost-based’ ML classifier that is applicable to
predict both bitterness and non-bitterness of compounds (DaganWiener et al., submitted and ongoing work).
4. Summary and outlook
The encouraging results obtained so far, combined with further
growth of experimental data, can provide reliable and inexpensive
computational tools for developing active pharmaceutical ingredients (APIs). The ligand-based methods are most suitable when
activity data is available for a chemical series of compounds;
structure-based methods are most suitable when binding site of
the structural model of the receptor was confirmed by sitedirected mutagenesis (Levit et al., 2012), and machine learning
techniques are applicable when bitterness at large (not related to
particular receptor or subfamily of ligands) is investigated.
Importantly, allowing the training and prediction datasets to be
publicly available will enable benchmarking of different ML
methods and constant improvement of predictors. It is therefore
important to maintain and update public databases, such as
BitterDB. It is likely that computational prediction of bitterness
will become a useful addition to the pharmaceutical researcher
toolbox, alongside with additional techniques, such as animalbased and e-tongue models (Clapham et al., 2012).
Acknowledgements
The authors express their gratitude to the Planning and
Budgeting Committee (PBC) of the Council for Higher Education
in Israel for providing postdoctoral fellowship to Dr. Malkeet Singh
Bahia in the Robert H. Smith Faculty of Agriculture, Food and
Environment, The Hebrew University of Jerusalem, Rehovot, Israel.
Funding by the ISF494/16 and ISF-NSFC2463/16 grants to MYN, and
membership in COST actions CM1027 (GLISTEN) and CA15135 (Mu.
Ta.Lig), are gratefully acknowledged.
References
Burdock, G.A., Fenaroli, G., 2005. Fenaroli’s Handbook of Flavor Ingredients, 5th ed.
CRC Press.
Carlsson, J., Coleman, R.G., Setola, V., Irwin, J.J., Fan, H., Schlessinger, A., Sali, A., Roth,
B.L., Shoichet, B.K., 2011. Ligand discovery from a dopamine D3 receptor
homology model and crystal structure. Nat. Chem. Biol. 7, 769–778.
Chandrashekar, J., Hoon, M.A., Ryba, N.J.P., Zuker, C.S., 2006. The receptors and cells
for mammalian taste. Nature 444, 288–294.
Cho, M.J., Unklesbay, N., Hsieh, F.H., Clarke, A.D., 2004. Hydrophobicity of bitter
peptides from soy protein hydrolysates. J. Agric. Food Chem. 52, 5895–5901.
Clapham, D., Kirsanov, D., Legin, A., Rudnitskaya, A., Saunders, K., 2012. Assessing
taste without using humans: rat brief access aversion model and electronic
tongue. Int. J. Pharm. 435, 137–139.
Di Pizio, A., Niv, M.Y., 2015. Promiscuity and selectivity of bitter molecules and their
receptors. Bioorg. Med. Chem. 23, 4082–4091.
Di Pizio, A., Levit, A., Slutzki, M., Behrens, M., Karaman, R., Niv, M.Y., 2016. Comparing
class A GPCRs to bitter taste receptors: structural motifs, ligand interactions and
agonist-to-antagonist ratios. In: Shukla, A.K. (Ed.), G Protein-Coupled
Receptors: Signaling, Trafficking and Regulation, , pp. 401–427.
Drewnowski, A., Gomez-Carneros, C., 2000. Bitter taste, phytonutrients, and the
consumer: a review. Am. J. Clin. Nutri. 72, 1424–1435.
Huang, W., Shen, Q., Su, X., Ji, M., Liu, X., Chen, Y., Lu, S., Zhuang, H., Zhang, J., 2016.
BitterX: a tool for understanding bitter taste in humans. Sci. Rep. 6, 23450.
Irwin, J.J., Shoichet, B.K., 2016. Docking screens for novel ligands conferring new
biology. J. Med. Chem. 59, 4103–4120.
Karaman, R., Nowak, S., Di Pizio, A., Kitaneh, H., Abu-Jaish, A., Meyerhof, W., Niv, M.
Y., Behrens, M., 2016. Probing the binding pocket of the broadly tuned human
bitter taste receptor TAS2R14 by chemical modification of cognate agonists.
Chem. Biol. Drug Des. 88, 66–75.
Kitchen, D.B., Decornez, H., Furr, J.R., Bajorath, J., 2004. Docking and scoring in
virtual screening for drug discovery: methods and applications. Nat. Rev. Drug
Discov. 3, 935–949.
Please cite this article in press as: M.S. Bahia, et al., Bitterness prediction in-silico: A step towards better drugs, Int J Pharmaceut (2017), http://
dx.doi.org/10.1016/j.ijpharm.2017.03.076
G Model
IJP 16548 No. of Pages 4
4
M.S. Bahia et al. / International Journal of Pharmaceutics xxx (2016) xxx–xxx
Kufareva, I., Katritch, V., Participants of G.D, Stevens, R.C., Abagyan, R., 2014.
Advances in GPCR modeling evaluated by the GPCR Dock 2013 assessment:
meeting new challenges. Structure 22, 1120–1139.
Levit, A., Barak, D., Behrens, M., Meyerhof, W., Niv, M.Y., 2012. Homology modelassisted elucidation of binding sites in GPCRs. In: Vaidehi, N., KleinSeetharaman, J. (Eds.), Membrane Protein Structure and Dynamics: Methods
and Protocols. Humana Press, Totowa, NJ, pp. 179–205.
Levit, A., Nowak, S., Peters, M., Wiener, A., Meyerhof, W., Behrens, M., Niv, M.Y., 2014.
The bitter pill: clinical drugs that activate the human bitter taste receptor
TAS2R14. FASEB J. 28, 1181–1197.
Ley, J.P., 2008. Masking bitter taste by molecules. Chemosens. Percept. 1, 58–77.
Marchiori, A., Capece, L., Giorgetti, A., Gasparini, P., Behrens, M., Carloni, P.,
Meyerhof, W., 2013. Coarse-grained/molecular mechanics of the TAS2R38 bitter
taste receptor: experimentally-validated detailed structural prediction of
agonist binding. PLoS One 8, e64675.
Mennella, J.A., Beauchamp, G.K., 2008. Optimizing oral medications for children.
Clin. Ther. 30, 2120–2132.
Mennella, J.A., Spector, A.C., Reed, D.R., Coldwell, S.E., 2013. The bad taste of
medicines: overview of basic research on bitter taste. Clin. Ther. 35, 1225–1246.
Meyerhof, W., Batram, C., Kuhn, C., Brockhoff, A., Chudoba, E., Bufe, B., Appendino, G.,
Behrens, M., 2010. The molecular receptive ranges of human TAS2R bitter taste
receptors. Chem Senses 35.
Ney, K.H., 1971. Prediction of bitterness of peptides from their amino acid
composition. Z. Lebensm. Unters. For. 147 64-&.
O’Neil, M.J., 2006. The Merck Index: an Encyclopedia of Chemicals, Drugs, and
Biologicals, 14th ed. Merck.
Rodgers, S., Busch, J., Peters, H., Christ-Hazelhof, E., 2005. Building a tree of
knowledge: analysis of bitter molecules. Chem. Sens. 30, 547–557.
Rodgers, S., Glen, R.C., Bender, A., 2006. Characterizing bitterness: identification of
key structural features and development of a classification model. J. Chem. Inf.
Model. 46, 569–576.
Roland, W.S., van Buren, L., Gruppen, H., Driesse, M., Gouka, R.J., Smit, G., Vincken, J.
P., 2013. Bitter taste receptor activation by flavonoids and isoflavonoids:
modeled structural requirements for activation of hTAS2R14 and hTAS2R39. J.
Agric. Food Chem. 61, 10454–10466.
Roland, W.S., Sanders, M.P., van Buren, L., Gouka, R.J., Gruppen, H., Vincken, J.P.,
Ritschel, T., 2015. Snooker structure-based pharmacophore model explains
differences in agonist and blocker binding to bitter receptor hTAS2R39. PLoS
One 10, e0118200.
Sandal, M., Behrens, M., Brockhoff, A., Musiani, F., Giorgetti, A., Carloni, P., Meyerhof,
W., 2015. Evidence for a transient additional ligand binding site in the TAS2R46
bitter taste receptor. J. Chem. Theory Comput. 11, 4439–4449.
Sliwoski, G., Kothiwale, S., Meiler, J., Lowe Jr., E.W., 2014. Computational methods in
drug discovery. Pharmacol. Rev. 66, 334–395.
Toelstede, S., Hofmann, T., 2008. Sensomics mapping and identification of the key
bitter metabolites in Gouda cheese. J. Agric. Food Chem. 56, 2795–2804.
Wiener, A., Shudler, M., Levit, A., Niv, M.Y., 2012. BitterDB: a database of bitter
compounds. Nucleic Acids Res. 40, D413–419.
Wishart, D.S., Knox, C., Guo, A.C., Shrivastava, S., Hassanali, M., Stothard, P., Chang, Z.,
Woolsey, J., 2006. DrugBank: a comprehensive resource for in silico drug
discovery and exploration. Nucleic Acids Res. 34, D668–672.
Wolber, G., Langer, T., 2005. LigandScout: 3-D pharmacophores derived from
protein-bound ligands and their use as virtual screening filters. J. Chem. Inf.
Model. 45, 160–169.
Yarnitzky, T., Levit, A., Niv, M.Y., 2010. Homology modeling of G-protein-coupled
receptors with X-ray structures on the rise. Curr. Opin. Drug Discov. Dev. 13,
317–325.
Please cite this article in press as: M.S. Bahia, et al., Bitterness prediction in-silico: A step towards better drugs, Int J Pharmaceut (2017), http://
dx.doi.org/10.1016/j.ijpharm.2017.03.076