diSBPred: A machine learning based approach for disulfide bond prediction.

Area of Science:

• Biochemistry
• Computational Biology
• Bioinformatics

Background:

• Protein disulfide bonds are crucial covalent links formed by cysteine oxidation during post-translational modification.
• These bonds stabilize protein 3D structure, influence folding, and are vital for ab initio protein structure prediction (aiPSP) by constraining conformational searches.
• Experimental determination of disulfide bonds is costly and time-consuming, necessitating computational prediction methods.

Purpose of the Study:

• To develop an accurate, sequence-based computational method for predicting protein disulfide bonds.
• To improve the efficiency and accuracy of ab initio protein structure prediction (aiPSP) through better disulfide bond identification.
• To provide a tool for annotating cysteine bonding residues in proteins with unknown structures.

Main Methods:

• A stacking-based machine learning approach, diSBPred, was developed for disulfide bond prediction.
• Features extracted include conservation profiles, solvent accessibility, torsion angle flexibility, disorder probability, and sequential cysteine distance.
• A two-stage prediction process was employed: individual cysteine bonding prediction followed by cysteine-pair bonding prediction.

Main Results:

• diSBPred demonstrated improved feature relevance compared to existing methods, yielding a 7.39% increase in balanced accuracy via jackknife validation.
• The method achieved 82.29% balanced accuracy for individual cysteine prediction and 94.20% for cysteine-pair prediction using 10-fold cross-validation.
• Overall, diSBPred showed a 43.25% improvement in balanced accuracy over the nearest neighbor algorithm (NNA) based approach.

Conclusions:

• diSBPred offers a highly accurate and efficient computational tool for predicting protein disulfide bonds.
• The method can significantly aid in annotating cysteine bonding residues for proteins with unknown structures.
• Improved disulfide bond prediction accuracy using diSBPred is expected to enhance ab initio protein structure prediction and support experimental structure determination efforts.