DNA methylation is an important epigenetic control that plays a major role in gene expression, splicing, and genomic imprinting. Current bisulfite conversion, coupled with next-generation sequencing (NGS)-based methods, e.g., whole-genome bisulfite sequencing (WGBS) and reduced representation bisulfite sequencing (RRBS), are able to profile genome-wide DNA methylation at single base-pair resolution. Subsequent analysis of NGS data proceeds with the alignment of the bisulfite reads. However, sodium bisulfite treatment converts each unmethylated cytosine (C) to uracil, so an aligner needs to allow a thymine (T) in the read to match to a C in the reference genome when an unmethylated C occurs.

Since the early 2010s, various algorithms have been proposed to accomplish such alignments; among them, BS-Seeker, Bismark, and BSMAP were the earliest developed and are the most commonly used [1–3]. The former two employ an “in silico bisulfite conversion” strategy that results in a three-letter genome, where all the Cs in both the reads and the reference are converted to Ts prior to alignment. In contrast, BSMAP aligns with a wildcard approach. The recent version of BS-Seeker, BS-Seeker2, is arguably one of the most versatile bisulfite aligners [4]; it can map reads from both WGBS and RRBS, allows gapped local alignment, and supports a suite of traditional DNA aligners. Additionally, the advances in high-throughput sequencing technologies in recent years had significantly lowered sequencing cost and affordability. Current bisulfite aligners need to be updated to process this out-burst of information in a timely manner.

Here, we introduce BS-Seeker3, an extensively improved and optimized implementation of BS-Seeker2 that leverages the available computational power of a standard bioinformatics lab. BS-Seeker3 adopts all alignment features of BS-Seeker2, some of them include the support for local and gapped alignment, RRBS mapping, and built-in adapter trimming [4].

BS-Seeker3 also incorporates a series of new features to achieve significantly faster speed and better accuracy compared to other available bisulfite aligners. It is 1.5X faster than BSMAP, 10X faster than Bismark and Brat-nova, and maps twice as many reads as either of those aligners [2, 3, 5]. BS-Seeker3 also offers downstream analysis of bisulfite read data to further investigate bisulfite conversion efficiency and to visualize the methylation pattern after alignment. It is also well integrated with downstream methylation analyzer MethGo to provide a variety of genomic and epigenomic analyses [6].

To evaluate BS-Seeker3 performance, we benchmarked its default setting against major bisulfite aligners, Bismark, BSMAP, and Brat-nova, using their default parameters (Additional file 1: Method S2) with 20 cores on an 80 cores server. The default settings were used, so the performance could be generalized to an arbitrary novel dataset where the optimal parameters would be unknown. We ran all aligners on real human reads to examine the mapping efficiency (user runtime), and on a series of synthetic HiSeq-2500 and HiSeq-1000 reads with different degrees of data complexity from the Arabidopsis library to compare the mapping accuracy and overall performance (Additional file 1: Method S3) [10, 11]. The different levels of complexity were simulated by varying the single base insertion/deletion rate when generating simulated reads. We recorded the percentage of reads mapped correctly, the user runtime, and the mapability (Additional file 1: Method S4).

BS-Seeker3 performed the fastest on average on synthetic read data (Fig. 1c, Additional file 1: Figure S5). Bseeker3 is 10X faster than Bismark and 9X faster than Brat-nova. Even though BSMAP’s performance parallels BSseeker3’s in both speed and accuracy (Fig. 1d) with HiSeq2500 reads at low indel rates, BSMAP mapped significantly fewer reads correctly at higher level of data complexity. When the indel rate rose to .025, BSMAP’s accuracy dropped sharply and mapped even fewer reads than Bismark or Brat-nova. On the other hand, BS-Seeker3 consistently mapped more than 90% of the reads correctly at all indel rates. In brief, BS-Seeker3 would be much more suitable to process data with high complexity where a high insertion/deletion is expected.

To showcase the feasibility of BS-Seeker3 on real-time data from a large genome, we downloaded human data sets to create two data sets of increasing sizes (.5X Human Genome and 1X Human Genome) (Fig. 1e). BS-Seeker3 achieved the fastest speed on these data, followed by BSMAP, Brat-nova and Bismark. As a matter of fact, on the human data set, BS-Seeker3 performed at least twice as fast as the other aligners, including BSMAP. Because BSMAP builds its index online, the performance gap between BS-Seeker3 and BSMAP widens as the genome size increases.

In conclusion, BS-Seeker3 is an accurate, versatile, ultra-fast pipeline for processing bisulfite-converted reads. It also helps the user better visualize the methylation data.