Integração de métodos computacionais clássicos e baseados em inteligência artificial para classificação tumoral e identificação de compostos vegetais com potencial anticâncer

Abstract

Cancer is one of the leading causes of morbidity and mortality worldwide, posing a significant challenge to public health and global economic systems. Despite therapeutic advances, many treatments still have limited efficacy and are associated with considerable adverse effects. This limitation is often related to the molecular complexity of the disease and the activation of oncogenic signaling pathways that sustain tumor progression. Among the main molecular mechanisms involved in tumorigenesis, the hyperactivation of the PI3K/AKT/mTOR signaling pathway stands out, often associated with increased cell proliferation, tumor survival, therapeutic resistance, and immune evasion. The PI3Kα isoform, encoded by the PIK3CA gene, is among the most frequently mutated proteins in various types of cancer, including endometrial cancer. Its activation promotes the modulation of multiple cellular processes through the activation of the mTORC1 and mTORC2 complexes, reinforcing its central role in maintaining the tumor phenotype. Given this context, it is essential to identify new molecules capable of acting as selective or dual inhibitors of PI3Kα and mTOR. To this end, computational methods such as structural bioinformatics, machine learning, and deep learning have established themselves as strategic tools in the drug discovery process. In addition to prospecting for new inhibitors, the development of computational models can also contribute to the identification and characterization of tumor burden in different types of cancer, allowing for better molecular and prognostic stratification. In this context, this thesis is structured around three complementary axes. The first chapter presents a review of the main computational approaches applied to rational drug discovery, discussing theoretical foundations, methodological advances, and limitations, with a focus on structural bioinformatics, machine learning, and deep learning in oncology. In the second chapter, structural bioinformatics analyses were performed to identify, characterize, and evaluate atropoisomeric plant compounds with the potential to inhibit PI3Kα and mTOR proteins, including virtual screening in databases, molecular interaction studies, and molecular dynamics. In the third chapter, a multimodal pipeline was developed for predicting TMB in endometrial cancer (TCGA-UCEC),

integrating transcriptomic and digital histopathological data to identify and characterize tumor burden. Endometrial cancer is one of the types of cancer with the highest frequency of PIK3CA mutations and with worrying growth in young women, which reinforces the clinical relevance of the proposed approach. The results obtained are expected to contribute to the rational development of new selective and dual inhibitors of PI3Kα and mTOR, in addition to providing computational support for the molecular characterization of endometrial cancer, broadening the prospects for more personalized and precise therapeutic strategies.