Final Year Project

Project Title

Predicting Protein Structures using Machine Learning Techniques

Description

Machine learning models are well-used today in classifying objects and in feedback systems. Current methods used in protein structure prediction include comparative or homology modelling (using Hidden Markov Models) or physical methods (energy, torsion, thermal fluctuations). Several such algorithms exist and are tested by CASP bi-annually, the most popular of which are MODELLER, and SWISSMODEL. We create a machine learning system (neural network), which learns from existing solved protein structures (from pDB library) to determine folds from the primary structure. The learned model then produces a 3-D Cartesian coordinate for each monomer to predict the folds and structure of a target protein. The dataset consisted of 138,888 proteins, used to predict single-chain protein structures. A Recurrent Neural Network Model was used for prediction and BioPython was used to parse and structure input and output files.