To meet the need for a reference control in cell-free methylated DNA immunoprecipitation-sequencing (cfMeDIP-seq)1,2 experiments, we designed spike-in controls and ligated unique molecular indexes (UMI) to adjust for PCR bias, and immunoprecipitation bias caused by the fragment length, G+C content, and CpG density of the DNA fragments that are immunoprecipitated. This enables absolute quantification of methylated DNA in picomoles, while retaining epigenomic information that allows for sensitive, tissue-specific detection as well as comparable results between different experiments. We designed DNA fragments with 2x3x3x3=54 combinations of methylation status (methylated and unmethylated), fragment length in basepair (bp) (80 bp, 160 bp, 320 bp), G+C content (35%, 50%, 65%), and fraction of CpGs within a fragment (1 CpG/ 80 bp, 1 CpG/ 40 bp, 1 CpG/ 20 bp). Spiky was developed for analyzing DNA methylation of cell-free DNA obtained from cfMeDIP-seq method using reference ‘spike-in’ controls. This package will:
Install and load the spiky package from Bioconductor.
For each combination of parameters, we designed two distinct spike-in sequences. One to be methylated and one to be unmethylated. The allows us to assess non-specific binding of the monoclonal antibody on a sample-by-sample basis. To calculate methylation specificity we take the number of methylated reads divided by the total number of reads. It is our recommendation that if methylation specificity is <0.98, then the sample should be flagged or removed from analysis as the cfMeDIP performed inadequately.
This calculation is done by the ‘methylation_specificity’ function.
For each batch of samples, the coefficients used in the Gaussian generalized linear model will differ. The ‘model_glm_pmol’ will calculate these coefficients and output the model to be used to calculate molar amount (picomoles) on the user’s DNA sequences of interest. We assume that all DNA sequences of interest are methylated after undergoing cfMeDIP-seq. As such, we build the Gaussian generalized linear model on only the methylated spike-in control fragments. A generated Bland-Altman plot will visualize how well the model performs.
## Build the Gaussian generalized linear model on the spike-in control data
gaussian_glm <- model_glm_pmol(covg_to_df(ssb_res,spike=spike),spike=spike)
summary(gaussian_glm)
#>
#> Call:
#> glm(formula = conc ~ read_count + fraglen + GC + CpG_3, data = x)
#>
#> Deviance Residuals:
#> Min 1Q Median 3Q Max
#> -7.901e-04 -1.812e-04 5.277e-05 2.006e-04 4.682e-04
#>
#> Coefficients:
#> Estimate Std. Error t value Pr(>|t|)
#> (Intercept) 3.729e-03 3.779e-04 9.866 2.45e-09 ***
#> read_count -4.150e-07 9.332e-08 -4.447 0.000223 ***
#> fraglen -1.215e-05 9.877e-07 -12.303 4.59e-11 ***
#> GC 7.323e-06 5.698e-06 1.285 0.212698
#> CpG_3 3.523e-04 2.261e-04 1.558 0.134238
#> ---
#> Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
#>
#> (Dispersion parameter for gaussian family taken to be 1.056382e-07)
#>
#> Null deviance: 3.3927e-05 on 25 degrees of freedom
#> Residual deviance: 2.2184e-06 on 21 degrees of freedom
#> AIC: -337.41
#>
#> Number of Fisher Scoring iterations: 2
For the samples in which the Gaussian generalized linear model was built, we will calculate the molar amount (picomoles) for each DNA sequence of interest.
# Predict pmol concentration
# To select a genome other than hg38, use BSgenome::available.packages() to find valid BSgenome name
#library("BSgenome.Hsapiens.UCSC.hg38")
sample_data_pmol <- predict_pmol(gaussian_glm, ssb_res,bsgenome="BSgenome.Hsapiens.UCSC.hg38",ret="df")
#> Adjusting for bin-level biases...
#> Attempting to load BSgenome.Hsapiens.UCSC.hg38...
#> OK.
#> Done.
#> Starting pmol prediction...
#> Done.
head(sample_data_pmol,n=1)
#> chrom range.start range.end
#> chr22:10536001-10536300 chr22 10536001 10536300
#> sequence
#> chr22:10536001-10536300 ATACAACAGAGAAATGCATAATGTCCAATCAATTTATTAAATTTCCAAAGTCGGTCACGCGCAGTGGCTCACACCTGTAATCTGAACACTTCAGGAGGCCGAGACGTGTGGATCACGAGGTCAGGAGTTGGAGACTAGCCTGACCAACATGGTGAAACCCCGTCTATACTAAAAATACAAAAATTAGCCAGGCATGGTGGCACGTGGCTGTAATCCCCGCTACTCAGGAGGCTGAGGCAGGAGAATTGCTTGAACCTGGGAGGCAGAGGTTGCAGTGAGCCGAGATGGCGCCACCGCACT
#> read_count fraglen GC CpG_3 pred_conc
#> chr22:10536001-10536300 0.16 300 0.5033333 2.289428 0.0008930107
For our analyses, we binned the genome into 300 bp non-overlapping windows. We then look overlap between fragments in our data with each of the 300 bp genomic windows. We adjust the molar amount (picomoles) by a multiplier. This multiplier is the proportion of overlap between our fragment and the 300 bp window. This is done for every fragment in our sample.
sample_binned_data <- bin_pmol(sample_data_pmol)
head(sample_binned_data,n=1)
#> chrom range.start range.end
#> chr22:10536001-10536300 chr22 10536001 10536300
#> sequence
#> chr22:10536001-10536300 ATACAACAGAGAAATGCATAATGTCCAATCAATTTATTAAATTTCCAAAGTCGGTCACGCGCAGTGGCTCACACCTGTAATCTGAACACTTCAGGAGGCCGAGACGTGTGGATCACGAGGTCAGGAGTTGGAGACTAGCCTGACCAACATGGTGAAACCCCGTCTATACTAAAAATACAAAAATTAGCCAGGCATGGTGGCACGTGGCTGTAATCCCCGCTACTCAGGAGGCTGAGGCAGGAGAATTGCTTGAACCTGGGAGGCAGAGGTTGCAGTGAGCCGAGATGGCGCCACCGCACT
#> read_count fraglen GC CpG_3 pred_conc
#> chr22:10536001-10536300 0.16 300 0.5033333 2.289428 0.0008930107
#> adjusted_pred_con
#> chr22:10536001-10536300 0.0001428817
1. Shen, S. Y. et al. Sensitive tumour detection and classification using plasma cell-free dna methylomes. Nature 563, 579–583 (2018).
2. Shen, S. Y., Burgener, J. M., Bratman, S. V. & De Carvalho, D. D. Preparation of cfMeDIP-seq libraries for methylome profiling of plasma cell-free dna. Nature protocols 14, 2749–2780 (2019).