Improving Single-Nucleotide Polymorphism-Based Fetal Fraction Estimation of Maternal Plasma Circulating Cell-Free DNA Using Bayesian Hierarchical Models

Abstract

The recent advances in next-generation sequencing (NGS) technologies have enabled the development of effective high-throughput noninvasive prenatal screening (NIPS) assays for fetal genetic abnormalities using maternal circulating cell-free DNA (ccfDNA). An important NIPS quality assurance is quantifying the fetal proportion of the sampled ccfDNA. For methods using allelic read count ratios from targeted sequencing of single-nucleotide polymorphisms (SNPs), systematic biases and errors may reduce accuracy and diminish assay performance. We collected ccfDNA NIPS MiSeq sequencing data from an amplicon-based 92 SNP panel along with complementary low-depth whole-genome sequencing (WGS) on 243 normal male fetus pregnancies along with additional 144 nonpregnant female donor samples. Using fetal fraction estimates based on X and Y chromosome WGS coverage as gold standard, we compared an existing SNP-based approach, FetalQuant, to a more flexible Bayesian hierarchical modeling strategy that borrows information across interrogated SNPs to character SNP-level error rates and biases to improve fetal fraction estimates. Posterior distributions for SNP-level model parameters indicate most SNPs exhibited modest to moderate extrabinomial variation and a consistent underrepresentation of fetal alleles, with some extreme outliers in both regards. Fetal fraction estimates using FetalQuant, naive to these SNP properties, were relatively poor (R = 0.14, root mean squared error [RMSE] = 0.050), particularly when the true fetal fraction was low (<5%). In contrast, by quantifying SNP-level biases and error rates, our proposed approach demonstrated improved performance by reducing the bias and variability in fetal fraction estimates (R = 0.794, RMSE = 0.025). Using high-depth targeted SNP sequencing data, we identified a high degree of variability in distributional properties across SNP allelic read counts. These results highlight the benefits of leveraging hierarchical modeling for SNP-based fetal quantification assays (FQAs) and the need to properly calibrate FQAs dependent on NGS allelic ratio data.

Uncertainty of fetal fraction determination in Non-Invasive Prenatal Screening by highly polymorphic SNPs

“…The median values of BAF for the SNPs go from around 25% to around 75%; see Fig. 1 in (Larson et al, 2018).…”

“…Already the first variant of snpFF considered by (Chu et al, 2010a) led the individual informative snpFF i values ranging from around 5% to almost 30%, when the true fetal fraction was known to be 10%. (Larson et al, 2018) report the values of BAF for 57 SNPs in 120 non-pregnant samples, where the BAF is expected to be 50%. The median values of BAF for the SNPs go from around 25% to around 75%; see Fig.…”

“…However, the papers on various variants of snpFF usually include a comparison of the studied variant of snpFF with some other method of fetal fraction determination. Usually, the chromosome Y-based ffY method is used in this role; see (Larson et al, 2018;Song et al, 2016). Kim et al compared the seqFF method with their variant of snpFF (Kim et al, 2015).…”

“…The model involves a mixture of distributions. The model was recently extended by (Larson et al, 2018). The authors also proposed a more elaborated method of estimating the parameters of the model.…”

“…Thus, snpFF incurs extra costs to NIPS. Several variants of snpFF were considered in the literature (Chu et al, 2010a;Jiang et al, 2012;Kang et al, 2016;Song et al, 2016;Larson et al, 2018). They differ, among other things, in complexity of the method for determining the set of informative SNPs.…”