Structure based function prediction of proteins using fragment library frequency vectors.

Area of Science:

• Structural Bioinformatics
• Computational Biology
• Machine Learning in Proteomics

Background:

• Protein function is intrinsically linked to its three-dimensional structure.
• Predicting protein function from structure is crucial for biological research.
• Existing methods may not fully capture functional insights from global protein architecture.

Purpose of the Study:

• To develop a novel Support Vector Machines (SVM) based prediction model for protein functional classification.
• To utilize features extracted from the global protein structure using fragment libraries.
• To assess the model's accuracy in predicting protein function, exemplified by Cell Adhesion Molecules (CAMs).

Main Methods:

• Protein structures were represented by collections of short backbone fragments.
• Fragment libraries were used to discretize and represent protein structures.
• Input feature vectors were generated based on fragment frequency counts from all-to-all comparisons.
• Support Vector Machines (SVM) models were trained and optimized using 10-fold cross-validation.

Main Results:

• The optimal prediction model utilized a library of 400 fragments of length 10.
• The model achieved a 10-fold cross-validation accuracy of 95.25% for classification.
• An independent test dataset of Cell Adhesion Molecules (CAMs) and Non-CAMs yielded an accuracy of 87.5%.

Conclusions:

• Global protein structure can be effectively described using a representative set of fragments.
• The developed SVM model provides an accurate and unique method for protein functional classification.
• This approach holds significant promise for advancing our understanding of protein structure-function relationships.