Contributions to the study of the protein folding problem using deep learning and molecular dynamics

Abstract

The Protein Folding Problem (PFP) is one of the main challenges in the Computational Biology area. Globular proteins are believed to evolve from random initial conformations through folding pathways achieving, in almost all cases, to a functional native structure. Studies of the folding process are related to several abnormal events, such as misfolding and protein aggregation. Therefore, several computational approaches have been proposed in the literature for this problem. Deep Learning (DL) methods have been highlighted in studies in the Proteomics area, given their ability to extract features vectors and their efficiency after the training process. Recurrent Neural Networks (RNN) are cyclic DL methods that have achieved state-of-the-art performance for sequential and temporal problems. Therefore, this thesis presents contributions to studying the spatial-temporal pathways of the protein folding using RNN methods. To achieve these contributions, experiments of this thesis were organized in three steps: develop a framework to generate a massive amount of protein folding data using pure sequential and parallel Molecular Dynamics (MD) methods in the canonical ensemble; propose a Neighbourhood List (NL) approach to the parallel MD method; apply RNNs networks to the PFP. In the first step, we presented a package called PathMolD-AB to simulate and analyze folding data trajectories using the 3D-AB off-lattice model to represent the protein structure. The datasets generated from PathMolD-AB correspond to the MD evolution of 3,500 folding pathways, encompassing 35×106 states. The speedup analysis showed that the parallel approach obtained faster simulations when used protein sequences with more than 99 amino acids were used. In the second step, the NL approach with parallel MD showed higher improvement in the speedup performance than the purely parallel MD version with protein sequences between 99 to 1,000 amino acids, which covers 80% of the entire Protein Data Bank (PDB). In the last step of this thesis, a comparative analysis between RNNs architectures were carried out using the many-to-one model with datasets generated by the PathMold-AB. Results indicate that the Long Short-Term Memory ( obtained the best performance than other RNNs architectures in terms of prediction error. The biological analysis indicated that the LSTM predicted structures with similar features to the target (MD), in terms of hydrophobic and polar compactness, and also torsion and bond energies, suggesting that this approach is auspicious for the PFP study.