Collective judgment predicts disease-associated single nucleotide variants.

Area of Science:

• Genetics
• Bioinformatics
• Computational Biology

Background:

• Human genetic databases are rapidly expanding, with millions of validated Single Nucleotide Variants (SNVs).
• Non-synonymous SNVs (nsSNVs) alter protein sequences and can cause diseases.
• Existing methods for predicting nsSNV effects have room for improvement due to increased annotated data.

Purpose of the Study:

• To develop an improved computational approach for identifying disease-associated nsSNVs.
• To integrate existing prediction tools to enhance accuracy and reliability in variant classification.

Main Methods:

• Evaluated four established methods (PANTHER, PhD-SNP, SIFT, SNAP) on a large dataset of disease-annotated mutations.
• Developed a machine learning-based algorithm, Meta-SNP, integrating the outputs of these four methods.
• Tested Meta-SNP's performance against individual methods using accuracy and Matthew's Correlation Coefficient (MCC).

Main Results:

• Individual methods achieved 64%-76% accuracy and 0.38-0.53 MCC.
• Meta-SNP reached 79% accuracy and 0.59 MCC, outperforming the best single predictor by ~3% in accuracy.
• For difficult-to-classify variants, Meta-SNP showed an 8% accuracy improvement over the best single method.

Conclusions:

• The Meta-SNP algorithm demonstrates superior performance by combining orthogonal prediction methods.
• Integrating predictions from diverse resources is effective for selecting high-reliability variant-disease associations.
• Meta-SNP achieved 87% accuracy and 0.73 MCC for variants where all predictors agreed, highlighting its robustness.