Protein contact prediction using patterns of correlation

• DOI: 10.1002/prot.20160
• Publication Date: 2004-05-14T22:01:04Z
• Authors: Nicholas Hamilton; Kevin Burrage; Mark A. Ragan; Thomas Huber
• Keywords: Disjoint sets; Sequence (biology)

Abstract We describe a new method for using neural networks to predict residue contact pairs in protein. The main inputs the network are set of 25 measures correlated mutation between all residues two “windows” size 5 centered on interest. While individual pair‐wise correlations relatively weak predictor contact, by training windows correlation accuracy prediction is significantly improved. trained 100 proteins and then tested disjoint 1033 known structure. An average predictive 21.7% obtained taking best L /2 predictions each protein, where sequence length. Taking /10 gives an 30.7%. also 59 from CASP5 experiment. found be consistent across different lengths, but vary widely according secondary Predictive improve multiple alignments containing many sequences calculate correlations. Proteins 2004. © 2004 Wiley‐Liss, Inc.