Neural network pairwise interaction fields for protein model quality assessment and ab initio protein folding

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Artificial Intelligence in Biology

Background:

• Accurate assessment of predicted protein structure quality is crucial for its utility.
• Numerous Model Quality Assessment Programs (MQAP) exist, with performance evaluated in competitions like CASP.
• Existing methods vary in complexity and the representations of protein structures they utilize.

Purpose of the Study:

• To introduce a novel knowledge-based Model Quality Assessment Program (MQAP) for evaluating single protein structure models.
• To develop a fast and efficient method for assessing the global quality of protein structure models.
• To explore the potential of the developed model in *ab initio* protein structure prediction.

Main Methods:

• Utilized a tree representation of the Cα trace to train a Neural Network Pairwise Interaction Field (NN-PIF).
• NN-PIF was trained to predict the global quality of protein structure models.
• A separate version of the model was trained to assess *ab initio* protein folding capabilities.

Main Results:

• The developed NN-PIF model demonstrated superior performance in global model quality prediction compared to most other methods.
• NN-PIF achieved fast evaluation of multiple protein structure models for a single sequence.
• The system showed general improvements in protein structure quality and promising *ab initio* prediction capabilities.

Conclusions:

• The novel NN-PIF approach provides an efficient and effective method for protein structure quality assessment.
• NN-PIF shows potential for *ab initio* protein structure prediction, particularly when structural templates are unavailable.
• Further research is needed to fully realize the potential of NN-PIF in protein structure prediction and quality assessment.