Cyto-Safe: uma ferramenta de aprendizado de máquina para identificação precoce de compostos citotóxicos na descoberta de fármacos

Abstract

Introduction: Cytotoxicity is the ability of a substance to cause irreversible damage to living cells, leading to cell death. Evaluating cytotoxicity is crucial in the early stages of drug development, allowing for the early identification of toxic compounds, mitigating risks, and reducing animal testing. Quantitative Structure-Activity Relationship (QSAR) models that use Artificial Intelligence (AI) algorithms can predict cytotoxicity based on the chemical structure of compounds. Objective: To develop and validate QSAR models to predict the cytotoxicity of drug candidates and make them available as a free web application. Methods: A dataset from the literature with approximately 90,000 compounds tested on mouse embryonic fibroblasts (3T3) and human embryonic kidney cells (HEK 293) was used. After data cleaning and curation, Extended-Connectivity Fingerprints descriptors were generated. The models were created using the Light Gradient Boosting algorithm with 80% of the data for training and 20% for validation. Results: The generated models showed good performance, with a balanced accuracy (BACC) of 0.91 after applying data balancing techniques. The best models are available in the web application Cyto-Safe (http://cytosafe.labmol.com.br/), which incorporates elements of explainable AI, allowing visualization of molecular regions associated with cytotoxicity. Conclusion: The models were effective in classifying compounds regarding cytotoxicity in the 3T3 and HEK 293 cell lines. Cyto-Safe is a technological product that offers the scientific community a fast and reliable tool to evaluate the cytotoxicity of chemical compounds without experimental data, accelerating drug discovery.