An integrated protein structure fitness scoring approach for identifying native-like model structures

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning in Biochemistry

Background:

• Protein structure is crucial for understanding protein function and interactions.
• A significant gap exists between known protein sequences and experimentally determined structures.
• Protein structure prediction offers a solution but requires accurate model selection.

Purpose of the Study:

• To develop and validate ProFitFun-Meta, a neural network-based method for assessing protein model quality.
• To address the challenge of identifying the best models from multiple structure predictions.
• To provide a reliable tool for computational protein modeling and design.

Main Methods:

• ProFitFun-Meta utilizes a neural network to combine structural features like dihedral angles and surface accessibility.
• It analyzes spatial properties of protein structures for quality assessment.
• Performance was evaluated on large datasets including Test, External, and CASP14 datasets.

Main Results:

• ProFitFun-Meta demonstrated high reliability and efficiency, validated by Spearman's (ρ) and Pearson's (r) correlation coefficients.
• The method showed superior performance compared to state-of-the-art methods on extensive datasets.
• It outperformed leading performers in the CASP14 quality assessment category.

Conclusions:

• ProFitFun-Meta is a robust and efficient single-model scoring method for predicted protein structures.
• Its performance suggests it can be a key component in computational protein modeling and design.
• Minimal dependencies, high efficiency, and portability enhance its practical application.