AmIActive (AIA) : uma ferramenta web preditiva de larga escala baseada em QSAR para identificação de alvos e polifarmacologia

Abstract

Computer-aided drug design (CADD) has become a cornerstone of modern drug discovery, with approaches such as ligand-based methods (e.g., QSAR-Quantitative Structure Activity Relationships modeling and pharmacophore analysis) and structure- based methods (e.g., molecular docking and molecular dynamics) driving innovation. Ligand-based strategies remain essential, particularly in light of the growing availability of large-scale biological activity data. As a result, over the years, several QSAR-based predictive platforms available on the internet have shaped the field, including PASS, PLATO, and SwissTargetPrediction, among others. Although this ecosystem has continuously expanded since its inception, the dynamic nature of computational methodologies and the wide diversity of possible approaches have consistently left room for the development of new tools and the optimization of existing ones. Accordingly, the aim of this study was to develop the first Brazilian tool of this type. In this context, we present AmIActive (AIA), a large-scale QSAR-based predictive web tool designed to facilitate target fishing studies and polypharmacology assessment for both molecular and non-molecular targets. Chapter 1 presents a review of the fundamental methodologies and theoretical concepts required for the technical and conceptual understanding of a QSAR system of this type, while Chapter 2 describes the development of the AIA system itself. The AIA system integrates 3,239 statistically validated QSAR models covering 2,277 distinct targets, including single proteins, protein complexes, protein families, cell lines, organisms, and tissues. Pharmacological data were extracted from ChEMBL30 and subjected to rigorous cleaning and standardization procedures, with the exclusion of unsuitable datasets and conversion of activity values to negative logarithmic scales (pIC50, pAC50). Classification thresholds separating active and inactive compounds were defined using a binary search algorithm to achieve balanced classes. ECFP4 molecular fingerprints (1024 bits) were used as input features, and models were built using the Random Forest algorithm, with Bayesian hyperparameter optimization and nested cross-validation. Only models with p-values < 0.05 and MCC ≥ 0.5 were retained, yielding robust average performance (accuracy ≈ 0.82, F1-score ≈ 0.82, MCC ≈ 0.65). Applicability domains were defined using a normalized Euclidean distance metric, in accordance with OECD (Organisation for Economic Co-operation and Development) QSAR guidelines. The system is freely available at https://amiactive.ccen.ufpb.br/ through an intuitive graphical interface based on ChemAxon’s MarvinJS, allowing users to draw chemical structures or input SMILES strings and export prediction results in CSV format, including predicted activity probabilities, classification thresholds, applicability domain reliability, and complete model validation metrics.