Bayesian probabilistic approach for predicting backbone structures in terms of protein blocks.

Area of Science:

• Biochemistry
• Structural Biology
• Computational Biology

Background:

• Understanding protein structure is crucial for biological function.
• Predicting protein structure from amino acid sequences remains a challenge.

Purpose of the Study:

• To develop a novel method for predicting local protein structural patterns.
• To improve the accuracy of protein structure prediction using sequence-structure relationships.

Main Methods:

• Utilized an unsupervised cluster analyzer to define a local structural alphabet of 16 "protein blocks" (five consecutive C(alpha) atoms).
• Incorporated dependencies between successive protein blocks.
• Employed a Bayesian approach linking protein blocks to amino acid propensity for prediction.
• Grouped sequence windows into "sequence families" to enhance prediction accuracy.
• Developed two prediction strategies based on user-defined accuracy or block number.

Main Results:

• Achieved an initial prediction success rate of approximately 35%.
• Improved prediction accuracy by 6% through the use of "sequence families".
• Exceeded 75% prediction accuracy by considering the first four predicted protein blocks at each site.
• Demonstrated that 91% of sites in ubiquitin conjugating enzymes can be predicted with >77% accuracy using only three blocks per site.

Conclusions:

• The proposed prediction strategies enhance understanding of sequence-structure dependence.
• These methods offer significant utility for *ab initio* protein modeling.
• The identified protein blocks and prediction strategies provide a new framework for structural prediction.