Combinação de diferentes métodos de redução de dimensionalidade e de agrupamento para a detecção automática de conformações moleculares preferenciais

Abstract

Molecular dynamics simulation (MD) is a technique used to study atoms and molecules' movements, allowing the analysis of recurring conformations and transition states. However, many conformations are necessary for studies of the prediction of physical-chemical and geometric properties of molecules. Due to the number of parameters considered in the description of movements molecular (e. g. intra and inter-atomic distances, dihedral angles), the sets of trajectories present a high-dimensionality, this being the main factor that makes the analysis of long simulations by DM difficult. Use of techniques such as machine learning has been used to find a reduced dimensional space representing the essential movements of molecules, allowing them to identify representative movements and facilitate extended simulation analysis. In general, the principal component analysis (PCA), a linear transformation method, has often been used to reduce the problem's dimensionality in essential DM studies. However, the literature proposes the use of non-linear methods to detect the phase space of protein molecules. Thus, the objective of this thesis is to develop an automated work ow was developed to obtain the preferential conformations of proteins, bringing to discussion the methods of reducing the dimensionality of data (RDD): Autoencoder, Isomap, t-SNE, MDS, and Spectral. Additionally, we propose to combine these methods with algorithms of grouping to discover representative conformations of the DM trajectory. And finally, structural analysis and enzymatic inhibition of target-therapeutic proteins in the treatment of schistosomiasis. To select representative structures, a free energy profile (FEL) is generated using the Weighted Histogram Analysis Method (WHAM) method to check the energy surface obtained by each RDD and thus find the conformation with greater convergence. The atomic uctuation of proteins was represented by Euclidean distances between the Cα intra-molecular atoms in each conformation. The characteristic matrix obtained was used as an input for dimensionality reducers combined with clustering algorithms (K-means, Ward, Meanshift, and Affinity Propagation). The parameter defines the number of K-means groups, and Ward was predicted using the BIC, elbow, GAP, and maximum silhouette. And the quality analysis of the detected groups was evaluated by internal cluster validation metrics (e.g., Calinski-Harabasz (CH), Davies-Bouldin index (DBI) and Silhouette). As a set of tests, we used the DMs of the mini protein Trp-cage (PDB1L2Y) and calmodulin (PDB1CLL) as simulations in temperatures of 310K and 510K. According to the results, the Spectral and Isomap methods were able to generate dimensional spaces that provide a good insight into the separation of conformations classes. As they are non-linear methods, the space generated better represents protein movements than PCA and, therefore, can be considered promising alternatives for the analysis of MD by essential dynamics. We applied the work ow to HIV-1 protease conformations obtained by essential and accelerated MD simulations to validate these results. The results obtained again presented the Spectral and Isomap methods as the best approaches for separating classes of conformations. Finally, we apply these techniques in a case study with proteins evaluated by our research group as molecular targets for the treatment of schistosomiasis, isoforms 1 (smNTPDase1) and 2 (smNTPDase2) from ATP-Diphosphohydrolase de Schistossoma mansoni. For the lower energy structures obtained by the Spectral method, molecular docking studies against the LS1 compound synthesized and provided by Núcleo de Identificação e Pesquisa em Princípios Ativos Naturais da UFJF, previously studied experimentally and presented as a smNTPDase1 inhibitor. The results obtained were better than those previously published with the smNTDase1 model and point out that the compound LS1 has great potential for inhibition for both smNTPDases enzymes.