Force field approximation using artificial neural networks

Protein structure prediction has been previously addressed using various computer modelling methods. For example, Chemistry at HARvard Molecular mechanics (CHARMm) version 22 has been used at the Air Force Institute of Technology to model protein potential energy when searching for good protein structures. Applying CHARMm is computationally expensive; therefore, an alternative to CHARMm is needed to expedite search results. In this study we report results of modelling CHARMm. with a multilayered perceptron neural network. In building a neural network to emulate the CHARMm many parameters settings are studied. One such parameter is the number of generations to train the neural network. Under and over training of the neural network using test data is a concern. In this study, special attention has been paid to the training of the neural network. Finally, the accuracy with which a neural network can mimic CHARMm and the time savings realized when using a neural network in place of CHARMm (effectiveness and efficiency) are investigated.