Quality assessment of modeled protein structure using physicochemical properties

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning

Background:

• Protein structure quality assessment is crucial for distinguishing native structures from predicted ones.
• Physicochemical properties offer valuable insights into protein structure characteristics.

Purpose of the Study:

• To develop a fast and cost-effective method for predicting protein structure quality metrics.
• To evaluate the efficacy of various machine learning methods and physicochemical properties for predicting Root Mean Square Deviation (RMSD), Template Modeling (TM-score), and Global Distance Test (GDT_TS-score).

Main Methods:

• Utilized six physicochemical properties: total surface area, Euclidean distance (ED), total empirical energy, secondary structure penalty (SS), sequence length (SL), and pair number (PN).
• Employed nine machine learning algorithms, with feature importance determined by a Self-adaptive Differential Evolution (SaDE) algorithm.
• Validated model robustness using K-fold cross-validation on a dataset of 95,091 modeled structures from 4896 native targets.

Main Results:

• The Random Forest (RF) model demonstrated superior performance compared to other machine learning methods.
• Achieved high prediction accuracy for RMSD (78.82%), TM-score (86.56%), and GDT_TS-score (87.37%) on the testing dataset.
• Reported low Root Mean Square Error (RMSE) values of 1.20 for RMSD, 0.06 for TM-score, and 0.06 for GDT_TS-score, with strong correlation scores (0.96, 0.92, 0.91 respectively).

Conclusions:

• Physicochemical properties combined with machine learning, particularly Random Forest, provide an efficient approach for protein structure quality assessment.
• This method accelerates and reduces the cost of predicting key protein structure quality indicators.
• The developed models offer reliable predictions for RMSD, TM-score, and GDT_TS-score, aiding in the evaluation of modeled protein structures.