Variant effect predictions capture some aspects of deep mutational scanning experiments.

Area of Science:

• Genomics
• Proteomics
• Bioinformatics

Background:

• Deep mutational scanning (DMS) systematically assays the effects of single amino acid variants (SAVs) in proteins.
• SAVs are also known as missense mutations or non-synonymous Single Nucleotide Variants (SNVs).
• This study compiled SAV annotations from 22 DMS experiments to evaluate variant effect prediction methods.

Purpose of the Study:

• To assess the performance of various variant effect prediction methods using DMS data.
• To compare traditional methods (PolyPhen-2, SIFT, SNAP2), a DMS-optimized method (Envision), and a conservation-based approach (PSI-BLAST).

Main Methods:

• Assembled and normalized SAV effect scores from 22 DMS experiments.
• Evaluated prediction methods including PolyPhen-2, SIFT, SNAP2, Envision, and PSI-BLAST conservation.
• Compared method performance on 32,981 SAVs, assessing prediction of deleterious and beneficial effects.

Main Results:

• All tested methods captured aspects of experimental effect scores, with varying degrees of success.
• Traditional methods like SNAP2 showed better correlation with measurements and binary classification.
• A simple conservation approach using PSI-BLAST unexpectedly outperformed other methods in many cases.
• All methods struggled to predict beneficial (gain-of-function) effects compared to deleterious ones.
• Independent experimental measurements for the same proteins showed substantial differences but agreed better with each other than with predictions.

Conclusions:

• DMS is a powerful experimental approach for understanding protein sequence space dynamics.
• DMS data is crucial for improving variant effect prediction methods.
• Challenges remain in data diversity, hindering simplification and generalization of prediction models.