key: cord-0284671-y04yxmnj authors: Margulis, Eitan; Slavutsky, Yuli; Lang, Tatjana; Behrens, Maik; Benjamini, Yuval; Niv, Masha Y. title: BitterMatch: Recommendation systems for matching molecules with bitter taste receptors date: 2022-01-15 journal: bioRxiv DOI: 10.1101/2022.01.13.476205 sha: cffc475a0e91773faee93caf8422f382c619393a doc_id: 284671 cord_uid: y04yxmnj Bitterness is an aversive cue elicited by thousands of chemically diverse compounds. Bitter taste may prevent consumption of foods and jeopardize drug compliance. The G protein-coupled receptors for bitter taste, TAS2Rs, have species-dependent number of subtypes and varying expression levels in extraoral tissues. Molecular recognition by TAS2R subtypes is physiologically important, and presents a challenging case study for ligand-receptor matchmaking. Inspired by hybrid recommendation systems, we developed a new set of similarity features, and created the BitterMatch algorithm that predicts associations of ligands to receptors with ~80% precision at ~50% recall. Associations for several compounds were tested in-vitro, resulting in 80% precision and 42% recall. The encouraging performance was achieved by including receptor properties and integrating experimentally determined ligand-receptor associations with chemical ligand-to-ligand similarities. BitterMatch can predict off-targets for bitter drugs, identify novel ligands and guide flavor design. Inclusion of neighbor-informed similarities improves as experimental data mounts, and provides a generalizable framework for molecule-biotarget matching. The sense of taste is a key driver in food choice and consumption 1 , and therefore has major implications for nutrition and health. The ability to taste is impacted in several diseases 2,3 , including COVID-19 infection 4 . Similarly to smell loss, taste loss or dysfunction can negatively affect the quality of life 5 and promote changes in body weight 6 . Sweet, umami and bitter tastes are mediated by G protein-coupled receptors (GPCRs) 7 . Furthermore, taste GPCRs were shown to be expressed in many extraoral tissues, suggesting physiological roles beyond taste perception 8 , such as mediation of hormone secretion 9 , regulation of upper respiratory innate immunity 10 and more. Bitter taste receptors (TAS2Rs or T2Rs) 11 present a particularly interesting case, where some TAS2Rs may be activated by tens of diverse ligands, whereas others are very selective and can be activated by only a few known ligands 12, 13 . In addition, the number of TAS2R subtypes varies across species, with 25 in humans and ~30 in rodents 14 . Some bitter molecules activate several TAS2Rs while others are specific for individual TAS2Rs 13, 15 . Bitter molecules have highly variable chemical structures, and include alkaloids, polyphenols, peptides, salts, fatty acids, and saponins 16 . Addressing the dire challenge of predicting TAS2R targets for bitter molecules has several important implications. First, it can assist in evaluating off-targets for bitter medicines; since TAS2Rs have additional physiological roles in extraoral tissues, unintentional activation of ectopic TAS2Rs may promote unwanted biological processes as side effects 17 . Second, specific agonists for TAS2Rs provide tools for studying the extraoral effects of these receptors, or even as potential drug candidates targeting TAS2Rs for gastrointestinal 18 and asthma indications 19 . Third, since numerous drugs and food compounds are intensely bitter, antagonists are needed to reduce bitterness and improve drug compliance, hence, designing antagonists for effective masking of bitterness relies on being able to identify the activated TAS2Rs [20] [21] [22] . Identification of TAS2R targets for a given molecule currently requires extensive in-vitro assays that consume time and resources, though some computational attempts to predict TAS2R targets began to appear 23 . Computational methods are highly desired due to their high speed and low cost, as well as their potential of improvement as the experimental knowledge base grows. Specifically for bitter taste receptors, computational tools were developed for bitterness prediction using docking to homology models of selected receptors 21, 24, 25 and ligand-based methods 16, 26 . For the GPCR family at large, several machine-learning methods were developed to predict potential GPCR targets for small molecules, including machine learning algorithms 27, 28 . The problem of predicting associations of TAS2Rs and ligands can be viewed as a recommendation problem. Generally, recommendation systems are aimed at rating pairings of items to categories 29, 30 . These correspond to ligands and receptors in our case. Contentbased recommendation systems 31 rely on attributes describing the items and the categories. However, collaborative recommendation systems rely on similarity measures from known associations or ratings 32 . Hybrid recommendation systems 33 combine the two approaches by incorporating both content-based and collaborative information. Here we develop BitterMatch, a classifier inspired by hybrid recommendation systems, to match bitter molecules to human and mice TAS2Rs. BitterMatch uses a novel set of features, which weight known associations between ligands and receptors by ligand-to-ligand and receptor-to-receptor similarities. Two BitterMatch scenarios are presented: "filling the gaps" for ligands for which some associations to TAS2Rs are known from cell-based experiments but some are missing, and "new ligands" for molecules for which bitter taste was established in human sensory tests, but no associations with individual TAS2Rs were measured. Data on associations between bitter molecules and 21 human and 20 mouse TAS2Rs was collected from the literature to create the association matrix. The matrix is sparse, since some ligands were tested only on a subset of TAS2Rs. Out of 4501 known associations for 303 ligands (36% of possible associations) 3761 are negative (ligands not activating the receptor, ~84%) and 740 are positive (ligands activating the receptor, ~16%). In agreement with previous observations 13 , TAS2R14 is the most broadly tuned human receptor, with 171 known agonists, followed by TAS2R39, TAS2R46 and TAS2R10 with 85, 79 and 48 ligands respectively. The most selective is TAS2R3 with only one known ligand. Mouse receptors (Tas2rs) also range from broadly tuned receptors, such as Tas2r105 (47 known agonists) to narrowly tuned Tas2rs, such as 122 and 139, each with one known ligand. In general, less screening experiments were performed for mouse Tas2s. Orphan TAS2Rs (that have no known ligands so far, four in humans and fourteen in mice) were excluded from training and inference. Tas2r113 was excluded due to having a single agonist that was not tested on human TAS2Rs. For ligands, 2-dimensional (2D) and 3-dimensional (3D) chemical features including MW, AlogP, QPlogHERG were calculated from the chemical structures of the molecules. For receptors, three types of chemical features were calculated : 1. Sequence-based features for the full protein sequence of the receptors and separately for the second extracellular loop, an important region for binding ligands in GPCRs 34 . 2. Binding site features that were calculated for the main site for ligands (the orthosteric binding site). 3. Structural features of the receptor, such as hydrophobicity and volume. For ligands, we compute chemical similarity of the molecules using Tanimoto scores of the linear fingerprints, as well as Tanimoto scores of the MOLPRINT2D fingerprints. For receptors, we compute sequence similarities based on the percent of identical positions in the protein sequences (percentage of sequence identity), and in the sub-sequences that consist of the orthosteric binding site residues. Additionally, since substitution between nonidentical residues are well studied 35 , we calculate similarities between the sequences of the receptors and the sub-sequences of the orthosteric binding site based on BLOSUM62 substitution matrix 36 (percentage of sequence similarity). All similarities are schematically represented in Figure 1A . Collaborative similarities between pairs of ligands (or receptors) were calculated as Jaccard similarities over their known associations, and used to construct ligand-to-ligand and receptor-to-receptor collaborative similarity matrices ( Figure 1B ). To avoid dependence of the learning algorithm on the size of the dataset, we devised new neighbor-informed features, which are based on the similarity matrices and the known associations. The features are coded separately for positive and negative associations. Specifically, from a ligand similarity matrix, we annotate each ligand ( )-receptor ( ) pair with four features: two summarize the similarities of to ligands with positive associations to the receptor and two summarize the similarities of to ligands with negative associations to . The features differ by their granularity: the first measures the similarity to the closest ligand that activates ; this feature represents positive examples in the local neighborhood. The second measures the sum of similarity values to all ligands that activate . The two negative features measure the similarity to the nearest ligand that does not activate , and the sum of 7 similarities of ligands that do not activate ( Figure 1C ). We repeat this feature extraction process for each ligand similarity matrix. We also extract neighbor-informed features from each receptor similarity matrix, reversing the roles of ligand and receptor: similarity to the closest receptor that is activated by l, the sum of similarities to receptors that are activated by l, similarity to the closest receptor that is not activated by l, and the sum of similarities to receptors that are not activated by l ( Figure 1D ). A -Ligand similarities include chemical similarities based on linear fingerprints, chemical similarities based on MOLPRINT2D fingerprints, and collaborative similarities. Receptor similarities include collaborative similarities, similarities calculated from sequence identity matrices, and from sequence similarities derived from BLOSUM62 substitution matrix. Sequence similarities were calculated based on the full protein sequence and based on the binding site sequences. Colored circle: red -the ligand does not activate the receptor, green -the ligand activates the receptor. Unknown associations are represented as blank spaces in the matrix. B -Collaborative similarities for pairs of ligands and pairs of receptors, calculated based on known associations using Jaccard similarity. C -Four similarity features are computed for ligands: the highest similarity to the ligand that activates , sum of similarity values to all ligands that activate , the highest similarity to the ligand that does not activate , and the sum of similarity values to all ligands that do not activate . D-Similarity based features are computed for the receptor: the highest similarity to the receptor that is activated by , the sum of similarities to receptors that are activated by , the highest similarity to the receptor that is not activated by , and the sum of similarities to receptors that are not activated by . The association matrix between bitter ligands (n = 303) and human or mouse TAS2Rs ( =21+20) has about 1/3 known and 2/3 unknown associations. In "filling the gaps" scenario we consider cases in which at least one association (positive or negative) is known for each ligand and for each receptor, and therefore, neighbor-informed features from the similarity matrices can be extracted for ligands and receptors. We model the problem as a binary classification task, in which each ligand ( )-receptor ( ) pair is considered an observation and is annotated with (a) features describing chemical properties of ( = 250), (b) features describing the chemical properties of ( = 235), (c) features derived from the similarities between and other ligands ( = 4 features ⋅ 3 types of similarities) and features derived from the similarities between r and other receptors ( = 4 features ⋅ 5 types of similarities). We train a classifier from these features using a gradient boosting algorithm with decision-tree learners (XGBoost 37 package) optimizing a binary logistic objective for predicting whether the ligand-receptor pair associates. We sample 80% of the known positive and negative associations to be used as a training set, and use the remaining 20% as a test set. We repeat this sampling process 100 times. The performance of BitterMatch is compared with a naive model that predicts for each ligandreceptor pair whether they associate according to the prior. In the prior, the prediction score is fixed per receptor for all ligands and is set to the proportion of ligands in the training set known to associate with the receptor. We further compare the full model with three submodels that contain different subsets of the features, as illustrated in Supplementary Figure S1 Recall and precision levels per receptor averaged over 100 repetitions, are shown in Figure 2B . The performance is generally better for receptors with more known (positive and negative) associations. Performance is higher for human receptors than for mice receptors, in accordance with a higher number of known positive associations. Nevertheless, mouse receptors 112, 126 and 105 all achieve recall above 49% and precision above 76%. Since different species have varying numbers of TAS2R subtypes 38 , assigning TAS2R functional analogs is not trivial. This task is especially important for humans and mice, since many taste examinations are performed on rodents, but aiming to reflect on humans 39 . In order to find the nearest functional receptor ("functional analogs"), we used the completed association matrix to calculate the Jaccard similarity between the receptors, based on their positive and negative, known and predicted associations. No score above 0.5 between human and mouse TAS2Rs was found (see supplementary data), suggesting relatively low overlap and no clear functional analogs. In this scenario we predict associations with human TAS2Rs for bitter molecules that do not have any known association with bitter taste receptors. Collaborative similarities, as well as features based on neighbors-informed receptor similarities, can not be calculated in this case. Therefore, we develop a version of BitterMatch that uses only the chemical properties of ligands and of receptors, and neighbors-informed ligand similarity features. For evaluation of the performance, we sample 80% of the ligands in the dataset into a training set, and consider the remaining 20% as a test set. For the ligands in the test set we remove the associations with all the receptors, marking them as unknown, and repeat this process 100 times. We compare our model to a prior model (as in the previous section), and to a "nearest-neighbor" model that predicts association between a ligand and a receptor based on the known association of with the ligand that has the highest chemical similarity with . Here chemical similarity is calculated according to linear fingerprints (using MOLPRINT2D fingerprints yielded similar results, not shown). The average precision-recall curves over 100 repetitions and corresponding confidence intervals are reported in Figure 4A (prediction intervals are shown in Supplementary Figure S3 ). BitterMatch achieves an average precision of 70% ± 5% , outperforming the prior model (average precision of 48% ± 4%) and the nearest-neighbor model (44% ± 5%). An accurate predictor of positive association can reduce the number of cell-based experiments needed to identify cognate bitter taste receptors for a bitter molecule. Therefore, for each ligand we count how many receptors should be tested, according to the prediction score, until the first TAS2R is activated. Figure 4B shows the number of tests required by the prior method and by BitterMatch. To quantify the differences, we fitted a linear regression that predicts the number of tests required by BitterMatch as a function of the tests required by the prior. On average, BitterMatch requires 3.5 times less cell-based experiments in order to match at least one TAS2R to a bitter-tasting molecule without any known bitter taste receptors. Table S1 . We note that for butein, 3,2′-Dihydroxychalcone and apigenin activation of TAS2R39 was not detected in our experiments but was found by Roland et al. 40 Figure 5 ). Comparing the proportion of drugs matched to TAS2R with the proportion of positives among all tested compounds for that TAS2R (hit rate) suggests that the proportion of drugs activating similar to hit rate for TAS2R14 but much lower for TAS2Rs 10, 39 and 46. The results strengthen the notion of importance of TAS2R14 as a major target of bitter pharmaceuticals 43 ( Figure 5 ). Very bitter predicted drugs from DrugBank (version 5.1.5) were assigned to human TAS2Rs using BitterMatch. The proportion of drugs that were predicted to activate each receptor represented in red. In grey-the proportion (hit rate) of known ligands per receptor. Normalization for the drugs was performed by dividing the number of drugs predicted to activate the receptor by the total number of drugs. The hit rate was calculated by dividing the number of ligands by the number of compounds that were tested for each specific receptor. To obtain intuitive insights into the model, we examine the average gain across all XGBoost splits (Methods: "Feature importance"). This analysis shows few features showing considerably higher importance values than others ( Figure 6 and Supplementary Figure S4 ). In both "filling the gaps" and "new ligands" scenarios, neighbor-informed features are most important: similarities of the ligand to both activating and non-activating ligands of the receptor. Receptor neighbor-informed features could not be calculated in "new ligands" scenario, but found to be important for "filling the gaps". In both scenarios, receptor chemical properties are important, top ones include hydrophobicity, charge and buried area properties. Ligand properties have low gain in both scenarios (Supplementary Figure S4) . In this work we presented BitterMatch, an algorithm designed to predict ligand-TAS2R associations. BitterMatch was modeled as a binary classification task, with a custom feature set composed of chemical descriptors as well as neighbor informed features extracted from multiple similarity matrices. We chose a tree boosting algorithm (XGBoost) as the learning algorithm due to its relative success with unbalanced classes and sparse. Despite the challenges, we achieved promising results with average precision of 70-76% for both "filling the gaps" and "new ligands" scenarios. Importantly, the high precision and recall were not only established in the train-test divisions of the dataset, but also confirmed in prospective predictions. Three associations predicted as positives for TAS2R39, were found positive in 40 but negative in our experiments. This could be due to stable vs. transient expression of the TAS2R gene, which might affect the sensitivity of the cells toward activation. Excluding these 3 associations from the analysis resulted in precision of 76% and recall of 37% instead of 80% and 42%. The novel neighbor-informed similarity-based features, which incorporate chemical similarities and information from known positive and negative associations, dramatically improve the performance of the algorithm. In contrast, chemical features of the ligands do not provide high gain, in accordance with the high chemical diversity of bitter molecules, which makes it difficult to associate specific chemical attributes to the ability of a molecule to activate specific TAS2Rs. On the other hand, chemical features of the receptors, in particular those related to hydrophobicity and net charge, substantially contribute to the model. This is in accord with hydrophobicity of the binding site importance for enabling recognition of multiple ligands 13 . The fact that orthosteric binding-site similarities play a dominant role in "fill the gap" scenario, suggests that this is the site most of the ligands bind to 44, 45 . The importance of the overall similarities of receptors suggests that access to the binding site is likely to play a key role as well 45 . In addition, we show that combining the chemical features of both ligands and receptors together with neighbor-informed ligand similarity features leads to much higher performance than relying only on the known associations of the most similar ligand, or on the number of receptor's positive associations. Rodents are used as a model for bitterness assessment, sharing a similar repertoire of bitter compounds 46, 47 . Our results suggest that at the individual receptor subtype level, activation of a rodent receptor is not predictive of the human receptor and vice versa. This is particularly important when using animal models to elucidate the physiological roles of extra-oral TAS2Rs 48 . Computational matching of TAS2Rs and their agonists by BitterMatch can assist in identifying specific agonists for TAS2Rs, elucidating functional relations between TAS2Rs and evaluating the potential TAS2R targets of food and drug compounds. We used the model to predict associations of intensely bitter drugs to human TAS2Rs. The results revealed that TAS2R14 is the main receptor in pharmaceutical drugs bitterness. TAS2R14 is known to be activated by antibiotics 49 and other diverse drugs 43 . It was also suggested that TAS2R14 regulates resveratrol transport across the human blood-cerebrospinal fluid barrier 50 . While BitterMatch yields good performance overall, limitations of the method should be kept in mind. The performance of BitterMatch is worse for receptors for which less associations are known, and in particular those with less positive ones. Indeed, since the similarity-based features also depend on known associations, in "filling the gaps" task we achieved poor results for ligands with few experimentally known positive associations. We expect further improvement as more associations are gathered experimentally. Additionally, since each TAS2R was represented by a single protein sequence (according to the common allele), mutations or variations in the sequence that might alter ligand recognition or receptor activation were not taken into account. Unraveling the associations of bitter compounds with TAS2Rs will advance studies of binding sites were shown in Maestro as a set of site points at or near the surface of the receptor that are contiguous or are separated in solvent-exposed region by short gaps that could plausibly be spanned by ligand functionality 57 Additional information on the receptor: we added the chromosome number and the organism (human/mouse) to the feature set. This data was also taken from BitterDB 12 . We compute multiple similarity matrices for ligands and receptors. Each ligand is compared to another ligand (and receptor to receptor) creating and matrices. Ligand similarities: Two similarity matrices were generated for the bitter molecules In addition to pre-computed similarities, we construct two collaborative similarity matrices, one for the ligands (S Lig ∈ ℝ L×L ) and one for the receptors (S Rec ∈ ℝ R×R ), based on the associations. Denote by A∈ {0,1,na} L×R the matrix of associations between L ligands (rows) and R receptors (columns). An entry Alr equals to 1 if the ligand activates the receptor r, 0 if it is known not to activate the receptor r, and "na" if the association is unknown or removed for training purposes. We define the collaborative similarity in each of these two matrices as proportion of matching associations. For ligands and ′ we define the collaborative similarity as where ′ is the number of receptors that their association with both ligands , ′ is known. Similarly, the collaborative similarity between two receptors, and ′ is defined as where ′ is the number of ligands that their association with both receptors , ′ is known. Feature extraction from similarities: The ligand similarity matrices determine the neighbors (or high similarity ligands) of a given ligand. We developed neighbor-informed features that summarize the abundance of positive and negative associations in the neighborhood of a given ligand to a receptor of interest. From each ligand similarity matrix, we annotate each ligand ( )-receptor ( ) pair with four features: two summarize the similarities of the ligand to ligands with positive associations to the receptor and two summarize the similarities of to ligands with negative associations. We measure the similarity to the nearest ligand that activates : and the sum of similarity values to all ligands that activate receptor Similarly, we measure the similarity to the closest ligand that does not activate receptor : and the sum of similarity values to all ligands that do not activate receptor : Likewise, we extract neighbor-informed features from each receptor similarity matrix, reversing the roles of ligand and receptor. We measure the similarity to the closest receptor that is activated by the ligand : and the sum of similarities to all receptors that are activated by the ligand : Similarly, we measure the similarity to the closest receptor that is not activated by ligand and the sum of similarities to receptors that are not activated by the ligand : Feature matrices: We construct a feature matrix X ∈ ℝ RL×m in which each row corresponds to a ligand-receptor pair, and each column corresponds to a feature. We denote by m the number of features. All models include Cl=235 chemical descriptors of the ligand (Methods: "Ligand features") and To avoid data leakage, before constructing the collaborative similarity matrices, we treat the values corresponding to test pairs as missing values in the association matrix A. We model the problem of predicting the TAS2Rs that associate with a ligand as a binary classification task where each ligand-receptor pair is given a prediction. We train a gradient boosting classifier with decision tree learners using the XGBoost classifier on the train feature matrix X Tr and the corresponding train label vector Y Tr to predict the outcome for each ligand-receptor pair. The hyper-parameters used for the XGBoost classifier were not chosen using the data, and are detailed in the Table S2 . the nearest ligand to , whose association with is known. Then the nearest-neighbor model uses the association of the nearest ligand as * its prediction for the ligand . Equivalent results were achieved when predictions are given according to MOLPRINT2D similarity (not shown). In order to evaluate performance of BitterMatch models (described in Methods: "Filling the gaps scenario" and Methods: "New ligands scenario"), we performed 100 repetitions of the experiment. In each repetition, train and test feature matrices X Tr ,X Te and label vectors Y Tr ,Y Te were randomly sampled (as described in Methods: "Feature matrices"). bounds of the prediction bands at , as the 5-th and 95-th percentiles of ( ) . Percentiles were computed using the "percentile" function, and the interpolation was performed using the "interp" function, both from the numpy 1. 16 Table S1 ). Additional to the maximal compound concentration, a tenfold dilution of the maximal concentration were tested to judge the apparent potencies. As positive controls we included the known agonists aristolochic acid (1 and 10 µM, TAS2R14) 63 and strychnine (30 and 300 µM, TAS2R10) 64 We used BitterMatch to predict the associations of intensely bitter drugs from DrugBank (version 5.1.5) 42 . 2406 intensely bitter predicted drugs were taken from Margulis, E. et al. 16 and inputted into BitterMatch. Analysis was performed at a threshold of 0.52. All the code for this work was implemented in Python 3. Figure 4B was implemented using seaborn 0.9.0. Figure 1 and S1 were Created with BioRender.com. The code is available at: https://github.com/YuliSl/BitterMatch Figure S1 -BitterMatch sub-models. A Venn diagram describing the four examined BitterMatch submodels. Model (1) includes only chemical properties, model (2) includes chemical properties and neighbor-informed collaborative features, model (3) includes chemical properties and neighborinformed chemical features, model (4) is the augmented one and it includes chemical properties, neighbor-informed collaborative features and neighbor-informed chemical features. Neighborinformed collaborative features are computed directly from the known associations that were also used to calculate collaborative similarities. However, neighbor-informed chemical features are computed from the known associations and chemical and sequence similarities. Table S2 -XGBoost hyper-parameters. To avoid overfitting 1000 trees were used accordingly the following hyper-parameters were adjusted. The rest of the parameters were set to their default values. Learning rate 0.001 Maximal depth of a tree 4 Column sample by tree 0.3 -minimum loss reduction for partition on a leaf of a node 2 Minimum sum of instance weight needed in a child 0.45 Subsample -ratio of the training instances used prior to growing trees 0.7 Table S3 -Prospective prediction results per ligand. 1 -represents an activation of the receptor that was confirmed experimentally and blank space means no activation was detected. * -positive activation that was detected in another publication but not in our in-vitro experiment. Bitter taste , phytonutrients , and the consumer : a review 1 -3. Am Bitter and sweet taste receptors in the respiratory epithelium in health and disease Taste and smell in disease More than smell -COVID-19 is associated with severe impairment of smell, taste, and chemesthesis Smell and taste disorders Dietary evaluation of patients with smell and/or taste disorders A novel family of mammalian taste receptors From machine learning to deep learning: progress in machine intelligence for rational drug discovery Bitter taste receptors and α-gustducin regulate the secretion of ghrelin with functional effects on food intake and gastric emptying Bitter and sweet taste receptors regulate human upper respiratory innate immunity T2Rs function as bitter taste receptors BitterDB: taste ligands and receptors database in 2019 Promiscuity and selectivity of bitter molecules and their receptors Lineage-Specific Loss of Function of Bitter Taste Receptor Genes in Humans and Nonhuman Primates The molecular receptive ranges of human TAS2R bitter taste receptors Intense bitterness of molecules: Machine learning for expediting drug discovery Extraoral bitter taste receptors as mediators of off-target drug effects Targeting extra-oral bitter taste receptors modulates gastrointestinal motility with effects on satiation Therapeutic potential and challenges of bitter taste receptors on lung cells Discovery and Development of S6821 and S7958 as Potent TAS2R8 Antagonists Rational design of agonists for bitter taste receptor TAS2R14: from modeling to bench and back Amino Acid Derivatives as Bitter Taste Receptor (T2R) Blockers* BitterX: A tool for understanding bitter taste in humans Dual binding mode of "bitter sugars" to their human bitter taste receptor target Bitter or not? BitterPredict, a tool for predicting taste from chemical structure Machine learning and AI-based approaches for bioactive ligand discovery and GPCR-ligand recognition Predicting Drug-Target Interaction Networks Based on Functional Groups and Biological Features The bellkor solution to the netflix prize. KorBell Team's Rep. to Netflix Item-Based Collaborative Filtering Recommendation Algorithms Content-Based Recommendation Systems A survey of collaborative filtering techniques Combining content-based and collaborative recommendation Understanding the common themes and diverse roles of the second extracellular loop (ECL2) of the GPCR super-family Selecting the Right Similarity-Scoring Matrix Where did the BLOSUM62 alignment score matrix come from? Xgboost: A scalable tree boosting system Vertebrate Bitter Taste Receptors: Keys for Survival in Changing Environments Development of a model for robust and exploratory analysis of the rodent brief-access taste aversion data Bitter taste receptor activation by flavonoids and isoflavonoids: modeled structural requirements for activation of hTAS2R14 and hTAS2R39 Numerous Compounds Orchestrate Coffee's Bitterness DrugBank 5.0: A major update to the DrugBank database for 2018 The bitter pill: Clinical drugs that activate the human bitter taste receptor TAS2R14 Modulation of Bitter Taste Perception by a Small Molecule hTAS2R Antagonist Structural requirements of bitter taste receptor activation Rats can predict aversiveness of Active Pharmaceutical Ingredients Assessment of bitter taste of pharmaceuticals with multisensor system employing 3 way PLS regression Comprehensive analysis of mouse bitter taste receptors reveals different molecular receptive ranges for orthologous receptors in mice and humans Chemosensory bitter taste receptors (T2Rs) are activated by multiple antibiotics The bitter taste receptor TAS2R14 regulates resveratrol transport across the human blood-cerebrospinal fluid barrier Soy Isoflavones and Other Isoflavonoids Activate the Human Bitter Taste Receptors hTAS2R14 and hTAS2R39 Amino acids and peptides activate at least five members of the human bitter taste receptor family Bitter substances from plants used in traditional Chinese medicine exert biased activation of human bitter taste receptors I-TASSER server: new development for protein structure and function predictions Protein Analysis Tools on the ExPASy Server The TOPCONS web server for consensus prediction of membrane protein topology and signal peptides Identifying and Characterizing Binding Sites and Assessing Druggability New Method for Fast and Accurate Binding-site Identification and Analysis Multiple Sequence Alignment Using ClustalW and ClustalX Jalview Version 2-a multiple sequence alignment editor and analysis workbench Comparing Class A GPCRs to bitter taste receptors: Structural motifs, ligand interactions and agonist-to-antagonist ratios Blends of Non-caloric Sweeteners Saccharin and Cyclamate Show Reduced Off-Taste due to TAS2R Bitter Receptor Inhibition Reengineering the ligand sensitivity of the broadly tuned human bitter taste receptor TAS2R14 The human bitter taste receptor TAS2R10 is tailored to accommodate numerous diverse ligands Broad tuning of the human bitter taste receptor hTAS2R46 to various sesquiterpene lactones, clerodane and labdane diterpenoids, strychnine, and denatonium We thank Lior Wolf, Ayana Dagan-Wiener and Fabrizio Fierro for insightful discussions, and Eva Boden for excellent technical assistance. MYN is supported in part by ISF #1129/19 funding. EM is the recipient of CIDR, Smith, Zehavi and Nano fellowships. YS is supported by the Israeli Council For Higher Education Data Science fellowship. MYN conceived the study, EM, YS, YB and MYN designed the study and wrote the manuscript.EM designed the data set and features, YS developed the learning algorithm. EM and YS analyzed the results. TL and MB performed in-vitro experiments. All authors read and approved the manuscript.