Mechanistic role of plant-based bitter principles and bitterness prediction for natural product studies II: prediction tools and case studies

2.1.1 Training set

The dataset of compounds used to build the model. Like the test set, this is often composed of compounds with a certain characteristic and those that do not have, e. g. bitter compounds versus non-bitter compounds.

2.1.2 Test set

The dataset of compounds used to prove or validate the model.

2.1.3 Decision trees

A DT uses a tree-like model of decisions and their possible consequences. They are used commonly used in operations research to arrive at decisions but are also popular in ML. Typically, a DT is a flowchart-like structure in which each internal node represents a “test” on an attribute (e. g. either an event occurs or not), each branch represents the outcome of the test, and each leaf node represents a class label (decision taken after computing all attributes). The paths from the root to the leaf represent classification rules. DTs can easily become unstable, i. e. a small change in the data could lead to a large change in the structure of the optimal DT. As a result, predictors derived by other ML methods would perform better with similar data. One way to solve this problem could be by replacing a single DT with a random forest (RF) of DTs. However, RF is not often easy to interpret when compared with a single DT.

2.1.4 Artificial neural network

Artificial neural networks (ANN) are inspired by the neural networks in the biological system that form animal brains since the original goal of the ANN approach was to solve problems as the human brain would do. However, the neural network is not an algorithm in itself, but a framework where several different ML algorithms work together in order to treat the data inputs. As such, an ANN is based on a collection of connected units (nodes) called artificial neurons (AN), with each connection functioning like the synapses in the brain, processing and transmitting signals from one AN to another. At a connection, the signal between ANs is a real number, and the output is computed by some non-linear function of the sum of its inputs. The connections between ANs are called “edges”, ANs and edges typically having a weight that adjusts as learning proceeds. The weight increases or decreases the strength of the signal at a connection. ANs may have a threshold such that the signal is only sent if the aggregate signal crosses that threshold. Typically, ANs are aggregated into layers. Different layers may perform different kinds of transformations on their inputs. Signals travel from the first layer (the input layer) to the last layer (the output layer), possibly after traversing the layers many times.

2.1.5 Deep neuron network

Deep neuron network or deep neural network (DNN) is a neural network with more than one hidden layer between the input and output layers. In DNN, thousands of neurons in each layer can be extensively applied to the dataset with thousands of features, and more advanced regularization technique such as the dropout can be used to prevent the overfitting problem. Nevertheless, DNN requires users to adjust a variety of parameters.

2.1.6 k-nearest neighbors

The k-nearest neighbors (k-NN) algorithm is among the simplest of all ML algorithms. Both for classification and regression, a weight is assigned to the contributions of the neighbors, so that the nearer neighbors contribute more to the average than the more distant ones. For example, a common weighting scheme consists in giving each neighbor a weight of 1/d, where d is the distance to the neighbor. This can be thought of as the training set for the algorithm, though no explicit training step is required.

2.1.7 Random forest

An RF algorithm is a type of ensemble learning method that constructs a large number of decisions trees (usually greater than 100), and outputs predictions based on a collection of the votes of the individual trees. A subset of the training dataset is chosen to grow individual trees, while the remaining samples are used to estimate the optimal fit. The constructed trees are grown by splitting the training set (subset) at each node according to the value of the random variable, which is sampled independently from a subset of variables.

2.1.8 Support vector machine

Support vector machine (SVM) or support vector network is a popular supervised ML technique that is used for classification and regression. Given a set of training examples, each marked as belonging to one of two categories, an SVM training algorithm builds a model that predicts whether a new example falls into one category or the other. The algorithm performs the classification by constructing the hyper-planes in the multi-dimensional space that separates the different classes.

2.1.9 Validation or performance assessment

The performance of statistical learning methods like ML is often measured by the number of true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN). In this scenario, TP, TN, FP, and FN would represent true bitterant, true non-bitterant, false bitterant, and false non-bitterant compounds, respectively.

Precision (P_re) is a measure of accuracy for a specific, predicted class.

Accuracy (A_cc) is another frequently used index for the overall classification performance, but it may be misleading due to the highly unbalanced class distribution in the used datasets.

Sensitivity (S_e) or recall and specificity (S_p) can assess a model’s ability to correctly identify TPs and TNs, respectively. These two parameters are usually interpreted in combination with each other.

The indices in eqs. (1) to (4) are often used for model validation and comparison.

In addition, the F1 measure (or F1-score) and non-error rate (NER) are defined, respectively, as:

F1-score (cross-validation) is evaluated on the internal validation dataset during the cross-validation. Meanwhile, F1- score (test test) is when F1-score is assessed on the test set. ΔF1-score is the absolute value of the difference between F1- score (cross-validation) and F1-score (test set), i. e.:

ΔF1-score is used to monitor the potential overfitting (or underfitting), i. e. if ΔF1-score is small, it means that the model performances are similar on the internal-validation dataset and test set.

2.1.10 Area under the curve (AUC)

This prediction metric is derived from a Receiver Operator Characteristics (ROC) plot. The ROC curve is a plot of the FP rate (1− S_p) on the y-axis against the TP rate (S_e) on the x-axis while varying the decision threshold. The area under the curve (AUC) of the ROC plot provides a convenient way of comparing classifiers. An AUC value of 0.5 represents a random classifier, while an ideal classifier has an area of 1.0.

2.1.11 Matthews correlation coefficient

F1-score and Matthews correlation coefficient (MCC), eq. (8), are commonly used to measure the quality of binary classifications.

2.1.12 Y-randomization

This is a tool used in the validation of statistical models, whereby the performance of the original model in data description (r²) is compared to that of models built for permuted (randomly shuffled) responses, based on the original descriptor pool and the original model building procedure [16].

2.1.13 Principal component analysis (PCA)

PCA is a mathematical technique that captures the linear interactions between the underlying attributes in a dataset. Every principal component can be expressed as a combination of one or more existing variables. All principal components are orthogonal to each other, and each one captures some amount of variance in the data.

3.2.1 Defensive plant-based bitter principles

SMs involved in plant defense (collectively known as antiherbivory compounds) are often classified into three sub-groups: nitrogen-containing compounds (including alkaloids, cyanogenic glycosides, glucosinolates, and benzoxazinoids), terpenoids, and phenolics. The chemical structures of some plant-based bitter principles which are known to play a defensive role are shown in Figure 6. Among them, plant phenols are the most common and widespread group of defensive compounds. Their major role is in resistance against insects, microorganisms, and competing plants. Cucurbitacins, for example, are bitter compounds whose hostility to a wide range of herbivores, including lepidopteran larvae, beetles, mites, and vertebrate grazers, can be explained by their taste [32]. The known defensive functions of some bitter SMs have been summarized in Table 2.

Figure 6:

Chemical structures of some bitter principles involved in diverse defensive roles in plants.

Popular plant-based alkaloids include nicotine, caffeine, morphine, cocaine, colchicine, ergolines, strychnine, and quinine. Their (mostly) aversively bitter taste is a natural deterrent to herbivores. While alkaloids mostly act on receptors of neurotransmitters, others (such as phenolics and terpenoids) are less specific and attack a multitude of proteins by building hydrogen, hydrophobic, and ionic bonds, thus modulating their 3D structures and in consequence their bioactivities.

Cyanogenic glycosides (often stored in inactive forms in plant vacuoles) become toxic when consumed by herbivores. This is because the rupture of the plant cell membranes releases the glycosides which bring them into contact with HCN-releasing enzymes in the cytoplasm (Figure 5). HCN is known to be highly toxic by blocking cellular respiration [55]. Glucosinolates are activated in a much similar manner, but the products of their breakdown rather cause milder effects like gastroenteritis, salivation, diarrhea, and mouth irritation [40]. Benzoxazinoids are also stored as inactive glucosides in plant vacuoles. Upon plant tissue disruption when herbivores feed, these antiherbivory compounds get into contact with β-glucosidases from the chloroplasts and lead to the enzymatic release of toxic aglucones [56]. Some benzoxazinoids are only synthesized when herbivores start feeding. Such SMs are considered to act by an induced plant defense mechanism [57].

Among the terpenoids, diterpenoids are widely distributed in latex and resins and can be quite toxic. The high toxicity of Rhododendron leaves can be attributed to the presence of diterpenoids [58]. Saponins are complex triterpenoids which are known to break down the red blood cells of herbivores [59]. Among the limonoids (a sub-class of terpenoids), azadirachtin is a well-known naturally occurring insecticide, which is active as a feeding inhibitor towards the desert locust (Schistocerca gregaria), acting mainly as an antifeedant and growth disruptor [48, 49]. Iridoid glycosides (another sub-class terpenoids), e. g. aucubin and catalpol, prevent the invasion of plants by insects and microorganisms [45].

Phenolics are known to exhibit antiseptic properties, while others disrupt endocrine activity. From simple tannins to the more complex flavonoids which confer on plants much of their colorful pigments, polyphenols are known for several activities, e. g. antioxidant activity. Some polyphenols are involved in plant defense mechanisms, e. g. lignin, silymarin, and cannabinoids [46, 47]. Condensed tannins (e. g. proanthocyanidins), including 2 to >50 flavonoid molecules, inhibit herbivore digestion by binding to the consumed plant proteins, rendering digestion difficult or almost impossible. This is done by the SMs interfering with protein absorption and digestive enzymes [50, 51, 52, 53, 60].

3.2.2 Case study: investigating human taste receptors of phenolic compounds

Soares et al. undertook an investigation of 6 polyphenols against several human taste receptors, with the view of identifying which receptors are activated by these compounds [61]. The compounds included; the hydrolyzable tannin pentagalloylglucose (PGG), the precursor of condensed tannins (-)-epicatechin, two procyanidin oligomers or condensed tannins (the dimer B3 and the trimer C2), and the anthocyanins malvidin-3-glucoside and cyanidin-3-glucoside, which are commonly found in plant-based foods and drinks, e. g. red wine, beer, tea, and chocolate. The chemical structures of the investigated compounds have been shown on Figure 7.

Figure 7:

Chemical structures of some plant-based bitter principles investigated in the study [61].

The observed EC₅₀ values for the different compounds vary 100-fold, the lowest values being for PGG and malvidin-3-glucoside. The compounds were shown to activate different combinations of the 25 hTas2Rs, e. g. (-)-epicatechin activated 3 of the receptors (i. e. hTas2R4, hTas2R5, and hTas2R39), while PGG activated only two receptors (i. e. hTas2R5 and hTas2R39). Meanwhile, malvidin-3-glucoside and procyanidin trimer only stimulated one receptor each (i. e. hTas2R7 and hTas2R5, respectively). The authors remarkably discovered tannins to be the first selective natural agonists for the receptor hTas2R5, with a high potency only toward this receptor. The authors also suggested the catechol and/or galloyl groups to be important structural determinants for mediating the interaction of these polyphenolic compounds with this receptor. This hypothesis could be verified by docking the compounds against the receptor site. The conclusions of this study would lead to the suggestion that the presence of these polyphenols in the food items could explain the bitterness of fruits, vegetables, and derived products even when in low concentrations.