AFFINITY PREDICTION OF SPINOCEREBELLAR ATAXIA USING PROTEIN-PROTEIN INTERACTIONS AND DEEP NEURAL NETWORK WITH USER-DEFINED LAYER

Abstract

Binding affinity prediction for a rare genetic disorder like spinocerebellar ataxia is crucial in biomedical study. Numerous models for affinity prediction have been developed through machine learning and deep learning. The basic deep neural network architecture uses a linear stack of layers and sharing of layers is not feasible whereas the functional deep neural network uses sharing of layers but the models are affected, when there is a change in layer. Hence complex models cannot be constructed and cannot predict binding affinity efficiently. This problem can be overcome by customizing the layers in deep neural network architecture. In this research work, the network layers are defined by sharing features with several layers and weights are trained and updated for every iteration to obtain accurate prediction. The work is implemented with 626 protein structures for protein-protein interaction and 313 complexes are attained from the protein-protein interaction. Binding site is identified by passing the 3D protein structures into convolutional neural network. Features like energy calculations, physio-chemical properties and interfacial and non-interfacial properties are extracted from interacted complex for building the model. Feature representations are learned automatically by deep learning through trainable weights in customized layers. Deep neural network with user defined layers is modelled with three optimizers and the results are correlated with functional deep neural network based affinity prediction models. The result shows that the proposed deep neural network with customized layers and adam optimizer achieves the highest prediction rate of 0.98.