A dynamic Bayesian network approach to protein secondary structure prediction

Area of Science:

• Computational Biology
• Bioinformatics
• Structural Biology

Background:

• Probabilistic models like hidden Markov models (HMMs) are used for protein secondary structure prediction.
• However, HMM-based methods show lower accuracy compared to machine learning approaches like neural networks (NN) and support vector machines (SVM).

Purpose of the Study:

• To develop a novel probabilistic method for protein secondary structure prediction.
• To improve upon the accuracy limitations of existing HMM-type methods.

Main Methods:

• Developed a dynamic Bayesian network (DBN) model for protein secondary structure prediction.
• Modeled PSI-BLAST profiles using multivariate Gaussian distribution, incorporating dependencies between profiles and secondary structures, and neighboring residues.
• Introduced segment length distributions for secondary structure states.

Main Results:

• The DBN method demonstrated significant accuracy improvements over pure HMM-type methods.
• A combined DBN and NN approach (DBNN) achieved superior Q3 accuracy, competitive with state-of-the-art methods.
• DBN/DBNN showed improved accuracy when combined with other methods via consensus.

Conclusions:

• The DBN method, utilizing Gaussian distributions and high-ordered residue dependencies, outperforms other HMM-type probabilistic methods.
• The DBNN method, a hybrid of DBN and NN, offers enhanced accuracy due to their complementary nature.
• Future research may involve integrating DBNN with SVM methods for further prediction improvements.