RESUMEN
Increasing interest in constituents and dietary supplements has created the need for more efficient use of this information in nutrition-related fields. The present work aims to obtain optimal models to predict the total antioxidant properties of food matrices, using available information on the amount and class of flavonoids present in vegetables. A new dataset using databases that collect the flavonoid content of selected foods has been created. Structural information was obtained using a structural-topological approach called TOPological Sub-Structural Molecular (TOPSMODE). Different artificial intelligence algorithms were applied, including Machine Learning (ML) methods. The study allowed us to demonstrate the effectiveness of the models using structural-topological characteristics of dietary flavonoids. The proposed models can be considered, without overfitting, effective in predicting new values of Oxygen Radical Absorption capacity (ORAC), except in the Multi-Layer Perceptron (MLP) algorithm. The best optimal model was obtained by the Random Forest (RF) algorithm. The in silico methodology we developed allows us to confirm the effectiveness of the obtained models, by introducing the new structural-topological attributes, as well as selecting those that most influence the class variable.
RESUMEN
BACKGROUND: There are a great number of tools that can be used in QSAR/QSPR studies; they are implemented in several programs that are reviewed in this report. The usefulness of new tools can be proved through comparison, with previously published approaches. In order to perform the comparison, the most usual is the use of several benchmark datasets such as DRAGON and Sutherland's datasets. METHODS: Here, an exploratory study of Atomic Weighted Vectors (AWVs), a new tool useful for drug discovery using different datasets, is presented. In order to evaluate the performance of the new tool, several statistics and QSAR/QSPR experiments are performed. Variability analyses are used to quantify the information content of the AWVs obtained from the tool, by means of an information theory-based algorithm. RESULTS: Principal components analysis is used to analyze the orthogonality of these descriptors, for which the new MDs from AWVs provide different information from those codified by DRAGON descriptors (0-2D). The QSAR models are obtained for every Sutherland's dataset, according to the original division into training/test sets, by means of the multiple linear regression with genetic algorithm (MLR-GA). These models have been validated and compared favorably to several previously published approaches, using the same benchmark datasets. CONCLUSION: The obtained results show that this tool should be a useful strategy for the QSAR/QSPR studies, despite its simplicity.