Clustering Secondary Structure With Ramachandran Plots: a Classic Adaptation of Machine Learning

Abstract

By Vageesha Herath, Biochemistry

Advisor: Ruxandra Dima

Award: Excellence in Research Commuication

Presentation ID: 195

Abstract: Microtubules (MT) are large cellular biopolymers responsible for functions such as chromosomal segregation during cell division. MTs are regulated by MT-associated proteins, such as katanin, which is a severing enzyme. These enzymes, which function as oligomeric assemblies of protomers, cut MTs using the energy from ATP hydrolysis. An open question is whether the motor is stable in its hexameric state or in lower-order oligomeric states. To answer this question, we carried out molecular dynamics (MD) simulations of lower-order oligomers of katanin in four ligand states. Simulations revealed changes in a region of the protomers, which is a part of the network of allosteric positions responsible for the coordination of the ligand-binding and oligomerization actions in katanin. To characterize the structural changes in this region, we developed an unsupervised learning method. The algorithm describes amino acids based on their backbone angles and clusters the identified structures. The performance of our algorithm was compared with two established methods for clustering secondary structures. The GROMOS method clusters structures using the structural similarity captured by the RMSD. The Combinatorial Averaged Transient Structure Clustering Algorithm (CATS) method clusters structures based on Gaussian distributions of the dihedral angles for each peptide bond. We found that our method classifies better structures compared to the CATS and the GROMOS methods. Using our methodology, we characterized the role of ligand binding on the dynamics of this region, and we determined that katanin trimers are not the stable state of the severing enzyme.