AFFIPred: AlphaFold2 structure-based Functional Impact Prediction of missense variations

Area of Science:

• Genomics and Bioinformatics
• Computational Biology
• Molecular Biology

Background:

• Predicting protein pathogenicity is crucial for understanding genetic diseases.
• Structure-based prediction methods are limited by the scarcity of experimental protein structures, creating a "structure knowledge gap."
• Existing sequence-based predictors often lack the detailed structural insights necessary for precise pathogenicity assessment.

Purpose of the Study:

• To introduce AFFIPred, an ensemble machine learning classifier for missense variant pathogenicity prediction.
• To utilize highly accurate, full-length protein structures predicted by AlphaFold2 (AF2) to bridge the structure knowledge gap.
• To combine sequence and AF2-based structural features for improved pathogenicity prediction accuracy.

Main Methods:

• Developed AFFIPred, an ensemble machine learning model.
• Integrated protein sequence data with structural characteristics derived from AlphaFold2 (AF2) predicted structures.
• Utilized full-length, unbound state AF2 structures for more precise Solvent Accessible Surface Area (SASA) calculations.

Main Results:

• AFFIPred demonstrated performance comparable to state-of-the-art predictors like AlphaMissense on unseen datasets.
• AF2 structures provided a more comprehensive view of structural characteristics, capturing all variants.
• AFFIPred achieved high accuracy without the limitations associated with Protein Data Bank (PDB)-based classifiers.

Conclusions:

• Leveraging AF2-predicted structures significantly enhances missense variant pathogenicity prediction.
• AFFIPred offers a robust and accurate method for pathogenicity prediction, overcoming limitations of previous approaches.
• The AFFIPred predictions for over 210 million human proteome variations are publicly accessible.