Probabilistic graphical models and deep learning techniques for gene prediction

• Ano de defesa: 2025
• Autor(a) principal: Caro, Waldir Edison Farfán
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Aprendizado computacional; Arcabouços probabilísticos; Computational genomics; Gene prediction; Genômica computacional; Machine learning; Neural networks; Predição de genes; Probabilistic frameworks; Redes neurais
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/45/45134/tde-19082025-123303/

Sequential data emerges in various research fields including computer vision, natural language processing, and biology, posing complex inference challenges like filtering, prediction, control, and labeling. Probabilistic Graphical Models (PGMs) are comprehensive mathematical frameworks that have been extensively utilized for these tasks with significant success. For gene prediction the best predictors currently focus on utilizing Markov Chains such as Augustus, SNAP, Genscan, or MYOP. Numerous studies have emphasized the efficacy of Deep Neural Networks (DNNs) in tackling complex problems. These models offer an advanced representation that facilitates inference tasks. For instance, Convolutional Neural Networks (CNNs) have been employed to infer biological signals, sequence motifs and classify pre-miRNAs, whereas Recurrent Neural Networks (RNNs) are used for predicting protein secondary structures. In this work, we introduce a framework that integrates DNN models with PGMs to enhance gene prediction from sequential data, surpassing current state-of-the-art methods. To achieve this, we developed DeepToPS, an extension of the ToPS tool, which is an object-oriented framework with efficient implementations of frequently used PGMs for sequential data, now incorporating DNNs. Consequently, DeepToPS facilitates the creation of gene predictors utilizing hybrid models that seamlessly blend deep learning techniques with graphical models.