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

The structural information of a protein is pivotal to comprehend its functions, protein-protein and protein-ligand interactions. There is a widening gap between the number of known protein sequences and that of experimentally determined structures. The protein structure prediction has emerged as an efficient alternative to deliver the reliable structural information of proteins. However, it remains a challenge to identify the best model among the many predicted by one or a few structure prediction methods. Here we report ProFitFun-Meta, a neural network based pure single model scoring method for assessing the quality of predicted model structures by an effective combination structural information of various backbone dihedral angle and residue surface accessibility preferences of amino acid residues with other spatial properties of protein structures. The performance of ProFitFun-Meta was validated and benchmarked against current state-of-the-art methods on the extensive datasets, comprising a Test Dataset (n = 26,604), an External Dataset (n = 40,000), and CASP14 Dataset (n = 1200). The comprehensive performance evaluation of ProFitFun-Meta demonstrated its reliability and efficiency in terms of Spearman's (ρ) and Pearson's (r) correlation coefficients, GDT-TS loss (g), and absolute loss (d). An improved performance over the current state-of-the-art methods and leading performers of CASP14 experiment in quality assessment category demonstrated its potential to become an integral component of computational pipelines for protein modeling and design. The minimal dependencies, high computational efficiency, and portability to various Linux and Windows OS provide an additional edge to ProFitFun-Meta for its easy implementation and applications in various regimes of computational protein folding.