Transcript Document
High Throughput Computing and Protein Structure Stephen E. Hamby Overview • • • • • • • • Introduction To Protein Structure Dihedral Angles Previous Work Support Vector Regression Optimisation Prediction Results Conclusions Introduction To Protein Structure Molecules with massive biological importance Structure determination gives insight into …. • Function, Dynamics, Potential drug targets. Experimental structure determination is…. • Expensive, Slow, Difficult Introduction To Protein Structure Primary Structure: Order of Amino Acids Secondary Structure: Building blocks Tertiary Structure: Complete 3D Structure Introduction To Protein Structure Secondary Structure Types α-helix β-sheet Random Coil Dihedral Angles Dihedral Angles Dihedral Angles Finding the secondary structure of a protein is a step towards finding its complete structure Predicting dihedral angles can help us to get the secondary structure How Can We Predict Dihedral Angles? Previous work Destruct Multiple neural networks. Iterative method. Predicts secondary structure and dihedral angles. Previous work Real Spine Twin neural networks give a consensus prediction. Predicts dihedral angles from various amino acid properties amino acid composition and predicted structure. Support Vector Regression Kernel machine learning raises the data to a higher dimension so a linear relationship can be found. Support Vector Regression Attempts to fit a linear function to the data in a high dimensional feature space Accurate but… Slow, needs optimisation, black box. Support Vector Regression Kernel Choice We tested the various kernels available through the PyML package. These the are linear, polynomial, and gaussian kernels. We tested them using the CASP4 dataset. Gaussian kernel produced the best results. Optimisation Three interdependent parameters Grid based optimisation on a the CASP4 dataset Around 10000 3 hour jobs. Run in blocks of 10 on Jupiter Accuracy assessed using the Pearson correlation coefficient Prediction Support vector machine using a Gaussian kernel and optimal parameters. Training on the CB513 dataset. Tested by 10 fold cross validation CASP 4 used as a test set. Results Results measured by cross validation Destruct Pearson 0.42 Correlation Coefficient Real Spine SVM Prediction 0.62 0.57 CASP4 Test set gives Pearson Correlation Coefficient of 0.56 Results Using Secondary structure predictions made by cascade correlation neural networks: Dihedrals assisted by predicted structure Pearson correlation coefficient 0.582. Subsequent iterations should lead to better predictions of both structure and dihedral angles. What Next? Using further iterations to improve accuracy. Current method is a black box. Can we use a program like Trepan to get some definite rules about secondary structure. Conclusions • Dihedral Angles define protein secondary structure • Using Support Vector Machines it is possible to predict dihedral angles • We (hopefully!) can use predicted dihedral angles to improve the accuracy of secondary structure prediction. Acknowledgements Jonathan Hirst Hirst group members BBSRC The University of Nottingham