Two multi-classification strategies used on SVM to predict protein structural classes by using auto covariance.

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning in Structural Biology

Background:

• Accurate protein secondary structure prediction is crucial for understanding protein function.
• The representation of protein sequences as numeric features significantly impacts prediction quality.
• Machine learning, particularly Support Vector Machines (SVM), is widely used in this field.

Purpose of the Study:

• To evaluate two SVM multi-classification strategies, "one-against-one" (1-a-1) and "one-against-all" (1-a-a), for protein structural class identification.
• To assess the effectiveness of auto covariance (AC) for feature representation in this task.
• To compare the prediction accuracy and potential biases of the two SVM strategies.

Main Methods:

• Utilized Support Vector Machine (SVM) with two distinct multi-classification strategies: "one-against-one" (1-a-1) and "one-against-all" (1-a-a).
• Employed auto covariance (AC) to transform protein sequences and their physicochemical properties into a numerical data matrix, capturing residue interactions.
• Validated the prediction accuracy using the Jackknife test.

Main Results:

• The "one-against-one" SVM strategy achieved a high overall accuracy of 90.69% in predicting protein structural classes.
• This accuracy was over 10% higher than that obtained using the "one-against-all" strategy.
• The "one-against-one" approach demonstrated an ability to avoid biased prediction accuracy.

Conclusions:

• The combination of auto covariance (AC) feature representation and the "one-against-one" SVM strategy offers a highly accurate method for protein structural class identification.
• This approach provides a significant improvement over the "one-against-all" strategy, particularly in mitigating prediction bias.
• This method serves as a valuable complementary tool for various protein structure prediction applications.