Thermopred : aplicação de inteligência artificial e química quântica na predição de propriedade termoquímica e espontaneidade de reações de degradação em insumos farmacêuticos ativos

Abstract

We developed artificial intelligence models to predict thermodynamic properties associated with the degradation of active pharmaceutical ingredients (APIs), focusing on Gibbs free energy (ΔG) and enthalpy (ΔH), in alignment with RDC No. 964/2025 issued by ANVISA, which reinforces the need for a deeper mechanistic understanding of forced degradation studies. A novel dataset comprising more than 14,000 chemical structures, including APIs and their degradation products, was constructed from the Brazilian DCB list and public databases such as PubChem. Thermodynamic properties were calculated using quantum chemical methods in Gaussian 16 at the M06-2X/6-31G(d) level of theory. Molecular representation was performed using Morgan fingerprints (RDKit), with structural similarity assessed through the Tanimoto coefficient. Thermodynamic variables were normalized to optimize model training and subsequently transformed back to their original scale to preserve physicochemical interpretability. Three machine learning algorithms—XGBoost, Random Forest, and Multilayer Perceptron (MLP)—were evaluated using R², Q², and RMSE as performance metrics. XGBoost achieved the best overall performance (Q² = 0.9947 and RMSE = 0.0137 for ΔG), with internal validation via StratifiedKFold and external validation confirming statistical robustness and strong generalization capability. The results demonstrate the potential of integrating computational chemistry and machine learning as a predictive framework for anticipating critical thermodynamic properties, supporting regulatory decision-making, and advancing data-driven pharmaceutical development strategies.