Presentation Title

Protein Folding Kernel Ridge Regression

Start Date

November 2016

End Date

November 2016

Location

HUB 302-94

Type of Presentation

Poster

Abstract

The objective of this research is to implement a mathematical method called kernel ridge regression (KRR) into a machine learning environment to accurately predict the formation of protein folds in their most natural state. This study will comparatively examine the results of KRR with other methods conducted during previous Critical Assessment of protein Structure Prediction (CASP) experiments. The goal is to improve the loss and error metrics by exploring a new computational technique. This is done by manipulating the parameters in KRR and employing a number of different approaches when implementing the method. These approaches include individual protein fold training and aggregation, feature selection and dataset manipulation, and uniform fold creation testing with respect to protein ID. The performance of these procedures has yielded an average median loss of 0.0489 for individual fold testing, 0.0382 for aggregation, 0.0353 for feature selection and dataset manipulation, and further experimentation is required to obtain results for uniform fold creation testing. This result rivals that of the Support Vector Machining (SVM) model performed at previous CASP competitions which produced a minimum loss of 0.042. To conclude, our methods are comparatively evaluated and disseminated for future experimentation to benefit.

This document is currently not available here.

Share

COinS
 
Nov 12th, 1:00 PM Nov 12th, 2:00 PM

Protein Folding Kernel Ridge Regression

HUB 302-94

The objective of this research is to implement a mathematical method called kernel ridge regression (KRR) into a machine learning environment to accurately predict the formation of protein folds in their most natural state. This study will comparatively examine the results of KRR with other methods conducted during previous Critical Assessment of protein Structure Prediction (CASP) experiments. The goal is to improve the loss and error metrics by exploring a new computational technique. This is done by manipulating the parameters in KRR and employing a number of different approaches when implementing the method. These approaches include individual protein fold training and aggregation, feature selection and dataset manipulation, and uniform fold creation testing with respect to protein ID. The performance of these procedures has yielded an average median loss of 0.0489 for individual fold testing, 0.0382 for aggregation, 0.0353 for feature selection and dataset manipulation, and further experimentation is required to obtain results for uniform fold creation testing. This result rivals that of the Support Vector Machining (SVM) model performed at previous CASP competitions which produced a minimum loss of 0.042. To conclude, our methods are comparatively evaluated and disseminated for future experimentation to benefit.