Prediction of Proteins' Molecular Surfaces from their Corresponding Amino Acid Sequences

This paper presents a new approach for the prediction of proteins' molecular surfaces from their amino acid sequences, using a multilayer artificial neural network. Our novel approach allows one to learn, and to predict, a pose-invariant shape index describing the molecular surface of a protein, by starting from a descriptor of the amino acid sequence. The input layer of the neural network is formed by a set of probability density functions associated with the correlation in between the constituent residues, while the output layer is formed by a pose invariant shape index describing the molecular surface. Once the neural network is trained, our molecular surface search engine allows one to retrieve the closest known molecular surface associated with an unknown, so-called query protein. We test our method against a database of more than 45,000 molecular surfaces. The neural network is evaluated for various topologies and optimization methods, and yield promising results.