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

Abstract

Background: We previously reported on CERENKOV, an approach for identifying regulatory single nucleotide polymorphisms (rSNPs) that is based on 246 annotation features. CERENKOV uses the xgboost classifier and is designed to be used to find causal noncoding SNPs in loci identified by genome-wide association studies (GWAS). We reported that CERENKOV has state-of-the-art performance (by two traditional measures and a novel GWAS-oriented measure, AVGRANK) in a comparison to nine other tools for identifying functional noncoding SNPs, using a comprehensive reference SNP set (OSU17, 15,331 SNPs). Given that SNPs are grouped within loci in the reference SNP set and given the importance of the data-space manifold geometry for machine-learning model selection, we hypothesized that within-locus inter-SNP distances would have class-based distributional biases that could be exploited to improve rSNP recognition accuracy. We thus defined an intralocus SNP "radius" as the average data-space distance from a SNP to the other intralocus neighbors, and explored radius likelihoods for five distance measures.

Results: We expanded the set of reference SNPs to 39,083 (the OSU18 set) and extracted CERENKOV SNP feature data. We computed radius empirical likelihoods and likelihood densities for rSNPs and control SNPs, and found significant likelihood differences between rSNPs and control SNPs. We fit parametric models of likelihood distributions for five different distance measures to obtain ten log-likelihood features that we combined with the 248-dimensional CERENKOV feature matrix. On the OSU18 SNP set, we measured the classification accuracy of CERENKOV with and without the new distance-based features, and found that the addition of distance-based features significantly improves rSNP recognition performance as measured by AUPVR, AUROC, and AVGRANK. Along with feature data for the OSU18 set, the software code for extracting the base feature matrix, estimating ten distance-based likelihood ratio features, and scoring candidate causal SNPs, are released as open-source software CERENKOV2.

Conclusions: Accounting for the locus-specific geometry of SNPs in data-space significantly improved the accuracy with which noncoding rSNPs can be computationally identified.

Keywords: Data space; GWAS; Machine learning; SNP; noncoding; rSNP.

Fig. 1

The geometric idea behind the intralocus distance features that are used in CERENKOV2. Top panel, SNPs from the same locus form a data-space “cloud.” Triangles and circles, SNPs; black lines, distances between a central SNP and the other SNPs within the locus. Bottom panel, SNPs shown in their chromosomal context

Fig. 2

Distributions of intralocus radii computed using five different distance measures (Canberra, Euclidean, Manhattan, cosine, and Pearson) applied to scaled and unscaled feature data, conditioned on the type of reference SNP (rSNP or cSNP) for the intralocus radius calculation. Results shown are for all OSU18 SNPs (see “The OSU18 reference SNP set” section). Significant differences in the rSNP likelihoods vs. cSNP likelihoods are evident for Canberra, Canberra (scaled), Euclidean (scaled), Manhattan (scaled), cosine, and Pearson methods for computing intralocus radii. Modest differences in rSNP vs. cSNP likelihoods were evident for the cases of Euclidean and Manhattan methods for computing intralocus radii

Fig. 3

Empirically estimated log-likelihood ratios (rSNP/cSNP) based on intralocus radii computed using ten methods. Results shown are for all OSU18 SNPs (see “The OSU18 reference SNP set” section). LLR, log-likelihood ratio (natural logarithm); ln, natural logarithm

Fig. 4

Performance of GWAVA, CERENKOV and CERENKOV2 on the OSU18 reference SNP set, by three performance measures. Marks, sample arithmetic mean of validation-set performance; bars, estimated 95% confidence intervals (see “Gradient boosted decision trees” section); GWAVA, based on the GWAVA’s Random Forest model with 174 features [24]; CERENKOV, our previous model with the base 248-column feature matrix; CERENKOV2, our current model consisting of the base feature matrix plus ten log-likelihood features derived from intralocus radii and fitted using training data only; AUPVR, area under the precision-vs-recall curve (higher is better); AUROC, area under the receiver operating characteristic curve (higher is better); AVGRANK, intralocus average score rank (lower is better [22])

Fig. 5

Gini and permutation importance values of 258 features in 14 categories (colored marks). Feature category labels as follows: “LLR”, log-likelihood ratio (the new data-space geometric features); “repliseq”, replication timing; “geneannot”, gene-model annotation-based; “epigenome”, epigenomic segmentation [67, 68]; “featdist”, SNP location-related; “chrom”, the chromosome; “eigen”, based on the Eigen [21] score; “phylogenetic”, phylogenetic interspecies local sequence conservation [6, 80, 81]; “allelism”, allele and MAF-related; “DHS”, DNase I hypersensitive site; “DNAContent”, local nucleotide frequences; “eQTL”, expression quantitative trait locus [75]; “repeats”, genomic repeat annotation; “TFBS”, transcription factor binding site (see “Extracting the nongeometric features” section and Ref. [22] for details)