Hyperparameter optimization of autoencoders for deeplearning of collective variables

Understanding the conformational space of proteins is integral to deciphering the mechanism of action of the proteins in drug-discovery studies. Molecular dynamics (MD) simulations have enabled the realistic simulation of protein dynamics. However, the computational demands of processing extensive MD data necessitate efficient di￾mensionality reduction techniques. An automated method for extraction of collective variables (reduced representations of protein movements) for the accurate representa￾tion of inherent dynamics of proteins is important. For this purpose, this thesis involves implementation of a semi-supervised autoencoder (SAE) for the extraction of collective variables from MD simulations of MurD and adenosine kinase (AdK) proteins, selected for their distinct dynamic states. This model, which incorporates extra output nodes coming from the latent layer and encoder hidden layers carrying the extra information such as opening angle, significantly outperforms the traditional vanilla autoencoder by achieving better results in reconstructing protein dynamics. The research also explores the optimization of the semi-supervised autoencoder architecture using genetic algo￾rithm and enhances the quality of the extracted collective variables. By focusing on features such as Cα coordinates and dihedral angles, and adjusting the model's hyperparameters, the study advances our ability to capture the essential movements within proteins. Moreover, the semi-supervised approach shows an important improvement in handling the highly flexible regions of proteins, particularly the C-terminal domain of MurD, which is crucial for its functional transitions. The findings also shows the potential of machine learning in bioinformatics to refine simulation data analysis.