A QSTR-Based Expert System to Predict Sweetness of Molecules

Abstract

This work describes a novel approach based on advanced molecular similarity to predict the sweetness of chemicals. The proposed Quantitative Structure-Taste Relationship (QSTR) model is an expert system developed keeping in mind the five principles defined by the Organization for Economic Co-operation and Development (OECD) for the validation of (Q)SARs. The 649 sweet and non-sweet molecules were described by both conformation-independent extended-connectivity fingerprints (ECFPs) and molecular descriptors. In particular, the molecular similarity in the ECFPs space showed a clear association with molecular taste and it was exploited for model development. Molecules laying in the subspaces where the taste assignation was more difficult were modeled trough a consensus between linear and local approaches (Partial Least Squares-Discriminant Analysis and N-nearest-neighbor classifier). The expert system, which was thoroughly validated through a Monte Carlo procedure and an external set, gave satisfactory results in comparison with the state-of-the-art models. Moreover, the QSTR model can be leveraged into a greater understanding of the relationship between molecular structure and sweetness, and into the design of novel sweeteners.

BitterSweet: Building machine learning models for predicting the bitter and sweet taste of small molecules

“…BitterPredict 19 and BitterX 13 have used unverified non-bitter molecules as a significant part of their training sets (55.6% and 50% respectively), which potentially adds noise to their models. While e-Bitter 21 , Rojas et al 16 and BitterSweetForest 22 mitigated this problem by utilizing only experimentally verified data, this significantly reduced the size of their datasets and might have led to insufficient representation of the bitter-sweet chemical space. Hence we surmise that towards an effective model for prediction of bitter-sweet taste, an exhaustive compilation of bitter, non-bitter, sweet and nonsweet compounds to span the chemical space is essential while not compromising the accuracy of taste information of the molecules.…”

“…Tasteless compounds were included as important controls for both bitter and sweet taste prediction. The datasets were split into training and testing sets such that the latter corresponded to the external validation/test sets established by BitterPredict 19 for bitter/non-bitter prediction and Rojas et al 16 for sweet/non-sweet prediction ( Supplementary Table S1). The curated training dataset is structurally diverse when seen in comparison to random bioactive molecules from ChEBI 24 , as evident in the 2D t-SNE plot generated using the physicochemical features ( Figure 1), with molecules from different sources incrementally capturing subsets of the general chemical space.…”

“…Rojas et. al 16 relied on 2D molecular descriptors and ECFPs from Dragon. Both e-Bitter 21 and BitterSweetForest 22 used binary fingerprints, with the former using ECFPs and the latter resorting to an amalgamation of Morgan, Atom-Pair, Torsion, and Morgan Feat fingerprints.…”

“…For the sweet-prediction task, BitterSweet models were only compared with the QSTR-based expert system developed in Rojas et al 16 due to the unavailability of BitterSweetForest 22 model as well as differences in external test sets ( Table 2). While their QSTR-system achieved an impressive NER of 0.848, it was unable to make predictions for 30 (19.3%) molecules in the test set.…”

“…As opposed to the use of 2D fingerprints, BitterX used physicochemical features of molecules (from Handbook of Molecular Descriptors 15 ) towards the training of Support Vector Machine (SVM) classifier. Rojas et al 16 instead tackled the problem of sweet-taste prediction using experimentally verified data from the literature, 2D molecular descriptors and ECFPs (Extended Connectivity Fingerprints) from Dragon 17 towards the development of a QSTR-based expert system 16,18 . BitterPredict 19 enhanced the training data of BitterX 13 , by including molecules curated by Rojas et al 18 and those identified as being 'probably non-bitter' from Fenaroli's Handbook of Flavor Ingredients 20 , while excluding the random molecules.…”