DNA modifications such as 5-methylcytosine (5mC) and 5-hydroxymethylcytosine (5hmC), are key epigenic regulatory mechanisms. It is now possible to study DNA methylation at single-base resolution. However, if the coverage is less than the desired depth, the level of methylation of a given CpG is uncertain and is normally given as 'missing'. Imputation approaches have shown potential for addressing this missingness in the DNA methylation data, but comprehensive benchmarking studies that consider genomic regulatory context remain limited. Here, we benchmarked 10 imputation methods and compared their performance using the human placenta methylation data from the Pregnancy Outcome Prediction Study (POPS). We obtained the methylation data from 119 placenta samples using the Biomodal evoC platform that measures both 5hmC from 5mC. We generated ground truth datasets across four genomic regions (promoters, CpG islands, gene bodies and 5000bp tiling arrays) including only CpG sites with at least 10X— coverage across all samples. Imputation performance was evaluated separately for 5mC and 5hmC using Root Mean Square Error (RMSE). Our findings indicate that mean imputation method consistently outperformed other approaches evaluated, such as boostme, missForest and imputePCA. For example, within gene bodies, the RMSE of the mean method was the smallest for both 5mC (0.102) and 5hmC (0.051), while KNN and softImpute produced the highest RMSE values (0.176 at 5mC and 0.0746 at 5hmC) relative to the other methods. In summary, the mean approach offers a robust and accurate strategy for handling missing values in placental mC and hmC sequencing data.

Co-authors: Dafina Angelova, Gordon Smith, Steve Charnock-Jones, Sung Sam Gong