Predicting the evolution of number of native contacts of a small protein by using deep learning approach

Abstract

Native contacts (NCs) are one of the most vital parameters in order to define the resemblance of a protein conformation with its native state. Prediction of number of native contacts in a protein is useful in protein folding mechanism. In this work, we focused to predict the time series of the number of NCs of a small protein, insulin monomer by using three neural network based models, namely; Multi-Layer Perceptron (MLP), Long Short Term Memory (LSTM) and Gated Recurrent Unit (GRU). The input data used in the study was the time evolution of NC values of the folded and unfolded protein conformations computed from the equilibrated trajectories of atomistic molecular dynamics (MD) simulations performed with the aqueous solution of the protein at ambient as well as at an elevated temperature. The evolutionary prediction accuracy of the three models was tested by calculating two error parameters; Root Mean Square Error (RMSE) and Mean Absolute Error (MAE). Our study revealed that, although these three models are successful in forecasting the time evolutions of the NCs in terms of lower RMSE and MAE, the prediction through static memoryless artificial neural network, MLP was relatively less precise as compared to other two recurrent units, LSTM and GRU. The study infers that by using the available input data generated from the MD trajectories; these neural network based models could be used to predict the complex evolution pattern of distanced based structural parameters of a protein with a satisfactory level.