A weighted profile based method for protein-RNA interacting residue prediction

Abstract

The prediction of putative RNA-interacting residues in proteins is an important problem in a field of molecular recognition. We suggest a weighted profile based method for predicting RNA-interacting residues, which utilizes the trained neural network. Most neural networks have a learning rule which allows the network to adjust its connection weights in order to correctly classify the training data. We focus on the network weights that are dependent on the training data set and give evidence of which inputs were more influential in the network. A large set of the network weights trained on sequence profiles is analyzed and qualified. We explore the feasibility of utilizing the qualified information to improve the prediction performance for protein-RNA interaction. Our proposed method shows a considerable improvement, which has been applied to the profiles of the PSI-BLAST alignment. Results for predictions using alternative representations of profile are included for comparison. © Springer-Verlag Berlin Heidelberg 2006.