ECOTOXICOLOGICAL PROFILING AND LEADS PRIORITIZATION THROUGH MACHINE LEARNING-BASED APPROACH FOR THE DESIGN OF SAFER ANTIPARASITIC AGENTS

Despite increasing evidence of their ecological impact, the environmental toxicity of pharmaceuticals remains largely overlooked in drug discovery. Many APIs, especially antiparasitics, are excreted unmetabolized and persist in the environment, threatening non-target species.1 This oversight contrasts with the growing emphasis on sustainability and One Health principles. To align with global sustainability goals, compound prioritization must integrate environmental safety, adopting ecotoxicological filters alongside traditional pharmacological criteria. The prioritization of compounds from diverse sources is a critical step in drug discovery, shaping the success of medicinal chemistry programs.2 Beyond potency and selectivity, incorporating ecotoxicological parameters enables the early identification of safer and more sustainable molecules. With this aim, in the framework of the COST Action CA21111 “OneHealthdrugs,” we developed an integrated cheminformatics workflow to identify environmentally safer and pharmacologically promising compounds. Focusing on Trypanosoma brucei, we assembled a curated dataset of both synthetic and natural products reported in the literature between 2019 and 2024, selecting the most active representatives from each series. Starting from this collection, we initiated a systematic evaluation to investigate how the inclusion of ecotoxicological parameters influences compound prioritization. Compounds were filtered based on predicted ADME and ecotoxicological profiles,3 and key molecular properties were analyzed to compare selection outcomes with and without ecotoxicology criteria. Among the ecotoxicological endpoints evaluated in our analysis, we included key parameters such as the Bioconcentration Factor (BCF), lethal concentration in aquatic organisms (LC50DM and LC50FM), and algal growth inhibition concentration (IGC50), enabling a comprehensive assessment of environmental risks associated with candidate compounds. To enhance the selection process, we implemented a supervised machine learning model trained to classify compounds as SAFE or UNSAFE based on molecular descriptors. Only compounds predicted as SAFE were retained for further analysis and ranked using the GreenDrugScore (GDS), a modular scoring system integrating a total of 33 endpoints spanning pharmacokinetics and toxicology (ADMETscore; e.g., hERG inhibition, Drug-Induced Liver Injury [DILI], Ames mutagenicity), environmental safety (ECOscore; described above), and drug-likeness (DLscore; e.g., QED, Lipinski’s Rule of Five, Topological Polar Surface Area [TPSA]).The integration of ecotoxicological parameters (GDS 2-block) led to a reprioritization of the compound set, with 60.9% of molecules improving their rank and 36.3% being downgraded compared to the ADMET-only scoring. Each score component combines normalized and weighted parameters, enabling transparent prioritization of compounds with favourable multiparametric profiles. Structure-activity relationships were explored through scaffold clustering using both topological (ECFP4) and pharmacophore-informed (PHARM) fingerprints, revealing scaffold families associated with improved ecotoxicological propert ies. This dual approach supports the identification of eco-friendly chemotypes with retained antiparasitic potential. Altogether, our study demonstrates how the ecotoxicological profiling integrated in early-phase drug discovery performed using machine learning technology can lead to a different scaffolds prioritization and deeply change the next step antiparasitic drug-discovery study. This approach may aligning compound selection with One Health principles and environmental sustainability of drug discovery and, finally, use.