Mr Yusuf Dinah

Research project: Enhancing protein structure prediction by using contact maps generated by deep learning architectures

Abstract

The aim of the project is to design, develop and evaluate novel strategies to enhance the quality

of the 3D models produced by I-Tasser, a leading ab-initio protein structure prediction

framework. Those novel strategies will be based on the integration of accurate contact maps

that will be generated by deep learning architectures.

Successful completion of the project requires addressing the following scientific objectives:

1. Contact map prediction by enhancing standard Convolutional Neural Network (CNN)

architectures by integrating a multi-distance approach and autoencoder stacks to increase

the representation richness while reducing computational complexity.

2. Contact map prediction when dealing with small training set, e.g. for membrane proteins,

by adapting and enhancing Small Data-Driven CNN (SDD-CNN) architectures.

3. Binarisation of the previously developed architectures so that contact map predictions can

be performed much faster without requirement of specialised hardware.