ゲノム情報科学研究教育機構  アブストラクト
Date March 30, 2006
Speaker Dr. Andrew J. Bordner, Oak Ridge National Laboratory, USA
Title Learning from Structure: Predicting Protein-Protein Interfaces and Peptide-MHC Binding Affinities
Abstract   Two recent applications of machine learning to biological problems that use structural information will be described.  The first is the prediction of protein interaction interfaces using a Support Vector Machine (SVM) trained on evolutionary conservation and residue propensity.  A large non-redundant set of 1494 true protein-protein interfaces was used for training and validation.  Five-fold cross validation showed that as much as 97% of the predicted patches overlapped with the true interface even though only 22% of the surface residues were included in an average predicted patch.  The model also identified many potential new interfaces and corrected mislabeled oligomeric states.  The second is the prediction of peptide binding affinity to class I MHC using the predicted geometry of the complex.  An efficient all-atom docking method using grid potentials accurately predicted the conformations of peptides bound to different MHC allotypes.  An SVM trained on binding energy terms provided good classification of binding peptides, even for a different MHC than used for training.  The generality of the model is important because of the lack of experimental data for most of the large number of MHC allotypes (> 1000 for human).
「セミナー」に戻る      
 ホーム