Detecting Proline and Non-Proline Cis Isomers in Protein Structures from Sequences Using Deep Residual Ensemble...

Area of Science:

• Biochemistry
• Structural Biology
• Bioinformatics

Background:

• Cis conformations of amino acid residues are biologically significant but rare in protein structures.
• Existing prediction methods for cis isomers are limited by outdated data and lack independent validation.

Purpose of the Study:

• To update the statistical understanding of cis isomers using a large dataset of high-resolution protein structures.
• To develop a novel, sequence-based prediction technique for identifying cis isomers in proteins.

Main Methods:

• Utilized a database of over 10,000 high-resolution protein structures to re-evaluate cis isomer statistics.
• Developed a prediction model employing an ensemble of residual convolutional and long short-term memory bidirectional recurrent neural networks.
• The neural network architecture enables learning from the entire protein sequence.

Main Results:

• The ensemble of eight neural network models achieved a Matthews correlation coefficient of approximately 0.35 for cis-Pro isomers.
• The model demonstrated a Matthews correlation coefficient of approximately 0.1 for cis-nonPro residues.
• The developed method provides updated statistics on cis isomer occurrence.

Conclusions:

• The novel prediction method accurately identifies cis isomers based on protein sequence information.
• This tool can prioritize functionally relevant cis residues for experimental validation.
• The method has the potential to enhance the accuracy of *ab initio* protein structure prediction by improving the sampling of rare conformations.