Protein protein interaction prediction using deep learning- a critical review and development of a novel DenseNet based prediction strategy

Abstract

Protein-protein interactions(PPI) are crucial for understanding behaviour of living organisms and identifying causes of diseases. In this thesis, a novel image based deep learning method has been proposed for predicting PPI, which we call DensePPI. Novelty of our approach is that we represent PPI using images. PPI images were generated based on sequences of amino acids present in the two interacting proteins. Various colours have been used to represent each Amino Acid. Rectangular images so generated were further divided into square sub-images using horizontal and vertical strides. Square sub-images were then given as input to a Deep learning model (DenseNet201) for automatic feature extraction and prediction. A consensus-based model was also introduced to tackle the problem enormous data size and high model complexity. Model was trained using the Pan et al.’s dataset and S.Cerevisiae datset. Model’s performance was tested on independent datasets like Caenorhabditis elegans, Escherichia coli, Helicobacter Pylori, Homo sapiens and Mus Musculus PPI after removing sequence similarities. Maximum accuracies on those datasets were 99.95%, 100.00%, 99.90%, 99.90% and 100.00% respectively. The Consensus model achieved a high accuracy of 98.86% for external test sets. Improved performance of DensePPI shows that the image based DL classifier could be effective for PPI prediction. DensePPI can be used to predict cross-species interactions, based on the enhanced prediction accuracies obtained on separate test sets. It can also give researchers new insights into signalling pathway analysis, therapeutic target prediction, and disease pathophysiology.