Incorporating structural features to improve the prediction and understanding of pathogenic amino acid substitutions

Abstract

Background: The wide application of gene sequencing has accumulated numerous amino acid substitutions (AAS) with unknown significance, posing significant challenges to predicting and understanding their pathogenicity. While various prediction methods have been proposed, most are sequence-based and lack insights for molecular mechanisms from the perspective of protein structures. Moreover, prediction performance must be improved. Methods: Herein, we trained a random forest (RF) prediction model, namely AAS3D-RF, underscoring sequence and three-dimensional (3D) structure-based features to explore the relationship between diseases and AASs. Results: AAS3D-RF was trained on more than 14,000 AASs with 21 selected features, and obtained accuracy (ACC) between 0.811 and 0.839 and Matthews correlation coefficient (MCC) between 0.591 and 0.684 on two independent testing datasets, superior to seven existing tools. In addition, AAS3D-RF possesses unique structure-based features, context-dependent substitution score (CDSS) and environment-dependent residue contact energy (ERCE), which could be applied to interpret whether pathogenic AASs would introduce incompatibilities to the protein structural microenvironments. Conclusion: AAS3D-RF serves as a valuable tool for both predicting and understanding pathogenic AASs.

Keywords

• amino acid substitution
• single-nucleotide variant
• pathogenic
• protein structure
• machine learning