Ferramenta baseada em aprendizagem multitarefa para a predição de mecanismos de desregulação endócrina

Abstract

Introduction: Endocrine disruptors are exogenous substances or mixtures that alter the functions of the endocrine system, causing adverse effects on the health of an intact organism, its offspring, populations, or subpopulations. Currently, studies on the toxicity of endocrine disruptors chemicals can be conducted through both in vitro and in vivo approaches. However, the implementation of these experimental assays faces numerous challenges, including the limited capacity for data processing given the vast number of commercial chemicals, the operational complexity of the tests, and ethical concerns related to the use of animal models. Objective: this work aims to develop artificial intelligence models that serve as alternative methods to predict the toxicity and adverse effect pathways of endocrine disruptors potentially harmful to both the male and female reproductive systems. Methods: Initially, datasets of compounds tested in vitro were compiled from the Tox21 and ToxCast databases, from which single-task models (Random Forest, SVM, LightGBM) and multi-task models were developed using ECFP4 fingerprints as molecular descriptors. Results and Discussions: The results showed predictive models, especially the multi-task models, which achieved an accuracy rate of 87% and balanced recall and specificity values (~75%) after task-to-task calibration developed in-house. Our second-layer models achieved an accuracy rate of over 70%, with specificity and recall for the female models exceeding 70%, while the best results for the male models were over 50%. Conclusion: In summary, this study offers promising methods for identifying endocrine disruptors, representing a valuable alternative or complement to in vitro and in vivo assays, highlighting the relevance and potential impact of the modeling tools developed in advancing toxicological research and reducing reliance on animal models.