Converging XGboost Machine Learning and Molecular Docking Strategies to Identify Attractants for Ceratitis capitata: Molecular Characterization and Database Curation of Natural Ligands for In Vitro/In Vivo Tests

The Mediterranean fruit fly Ceratitis capitata (Wiedemann) (Diptera: Tephritidae) is one of the most critical agricultural pests, causing economic damage globally due to its wide range of fruit hosts. Conventional insecticides have brought environmental, human health, and resistance challenges, driving interest in semiochemicals as sustainable pest management alternatives. Potential molecular attractants can be assessed experimentally through methods such as electroantennography (EAG) or behavioral assays. Odorant Binding Proteins (OBPs) have been recognized as crucial mediators in detecting these chemical signals. Although isolated compounds can provide mechanistic insights, volatile blends more accurately reflect natural conditions and typically elicit stronger behavioral responses. However, designing effective blends is challenging due to their complexity and regulatory limitations. Therefore, curated molecular databases of potential attractants become essential to accelerate the discovery and reduce cost in research programs, both in vitro and in vivo tests. The in silico molecular approaches, including Molecular Docking, Molecular Dynamics (MD) and Quantitative Structure–Activity Relationships (QSAR), offer cost-effective methods to prioritize candidates and/or understand ligand-OBP interactions. In this study, computational methodologies including Machine Learning (ML) based QSAR, molecular docking and MD simulations were integrated to highlight molecular features of standard molecules and identify potential attractors for C. capitata, which are expected to be good OBP binders. Initially, was applied a Bee Colony Algorithm, combined with an final XGBoost Machine Learning model, enabled the identification of five essential molecular descriptors to explain the attractant effect of 20 standard compounds recognized in the literature. Applying this model to an online database of natural products from Brazil (NuBBE—Nuclei of Bioassays, Ecophysiology and Biosynthesis of Natural Products Database), 206 molecules were identified from over 2000 candidates. In a parallel front of investigation, docking-based virtual screening was performed using the same NuBBE database. Most promissory compounds were discussed based on binding energy, structure/geometry focusing on interactions and estimated volatility, through the evaluation of vapor pressure. MD simulations with the gold standard compound (E,E)-α-farnesene provided insights into ligand-protein interactions. Interestingly, 16 of the top 20 ranked compounds after dockings were predicted as attractors by the XGBoost model. Finally, the curated database of 206 compounds, the great contribution of this paper (beyond the model), can be used to assertively select molecules for experimental tests of future blends or isolated compounds.