CERENKOV2: improved detection of functional noncoding SNPs using data-space geometric features

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Regulatory single nucleotide polymorphisms (rSNPs) are crucial for understanding gene regulation and disease association.
• The CERENKOV approach utilizes 246 annotation features and the xgboost classifier to identify causal noncoding SNPs within genome-wide association study (GWAS) loci.
• Previous work established CERENKOV's state-of-the-art performance using the OSU17 reference SNP set.

Purpose of the Study:

• To investigate the hypothesis that the geometric distribution of SNPs within loci influences rSNP identification accuracy.
• To develop and evaluate novel features based on inter-SNP distances within loci to enhance rSNP recognition.
• To release an improved open-source tool, CERENKOV2, for more accurate causal SNP discovery.

Main Methods:

• Expanded the reference SNP set to OSU18 (39,083 SNPs) and extracted CERENKOV feature data.
• Defined an 'intralocus SNP radius' based on average data-space distances to neighboring SNPs.
• Computed radius likelihoods and densities for rSNPs and control SNPs, fitting parametric models to derive ten log-likelihood features.
• Combined these new features with the original CERENKOV feature matrix for classification.

Main Results:

• Significant differences in likelihood distributions were observed between rSNPs and control SNPs.
• The addition of the ten distance-based log-likelihood features significantly improved rSNP recognition performance.
• Performance enhancements were validated using AUPVR, AUROC, and the novel AVGRANK measure on the OSU18 set.

Conclusions:

• Incorporating locus-specific SNP geometry in data-space substantially enhances the computational identification accuracy of noncoding rSNPs.
• The CERENKOV2 software, including feature extraction and scoring capabilities, is released as open-source, facilitating broader application in genetic research.