The two primary steps in the Bis-SNP workflow are outlined in Figure 2a and include base quality re-calibration and local realignment followed by SNP calling. Bis-SNP accepts standard alignment files (.bam format), which can be generated by popular Bisulfite-seq mapping programs such as MAQ, Bismark, BSMAP, PASH, or Novoalign (reviewed in 10). This allows the user to decide which mapping criteria are most important for their specific application. This also makes Bis-SNP compatible with specialized mappers such as RRBSMAP 33 and any other program that can output (.bam) files.

The Bis-SNP model relies on the accuracy of base quality scores, which are initially estimated by the instrument-specific base caller. However, these initial base scores do not accurately represent true error probabilities, which are highly dependent on local sequence context 12. In the GATK workflow, empirical mismatch rates for each nucleotide at each sequencing cycle are calculated by comparing base calls to the reference genome, and these mismatch rates are used to recalibrate instrument-generated values 12. We cannot use this default implementation with bisulfite-seq data, because true C>T sequencing errors can not be identified when the underlying methylation state of each bisulfite-converted DNA fragment is unknown. Therefore, instead of treating Ts at reference cytosines as errors, we treat them as a 5th base X, and estimate these as a group separately from T>T, A>T, or G>T. The effect is that we can effectively recalibrate base call quality scores for all except the X nucleotide, improving our ability to accurately identify SNPs. Importantly, we are able to improve SNP calling at cytosines by recalibrating 'G-strand' Gs that are complementary to the cytosine.

The user can choose among several output files. For methylation levels, Bis-SNP can return a standard UCSC .bed or .wig file, and a separate output file is generated for each cytosine context specified by the user on the command line. Example cytosine contexts are CG, CH, or CHH (H is the IUPAC symbol for A,C, or T). The .wig output contains the methylation percentage for each methylated cytosine, while the .bed format also contains the number of C/T reads the percentage is based on, plus the strand of each cytosine relative to the reference genome. For SNPs, Bis-SNP can return a Variant Calling Format (.vcf) file, which contains all SNP calls and likelihood scores in addition to methylation percentages.

The core of the SNP calling algorithm is based on the Bayesian inference model of GATK 12, and implemented using GATK's LocusWalker class. For each locus, Bis-SNP evaluates one of ten possible diploid genotypes (G), as shown in Figure 2B (a diploid genotype is made up of two parental alleles, referred to as A and B). The prior probability of each genotype, π(G), is determined using population data from dbSNP (including 1000 genomes data) similar to SOAPsnp 13 (See Materials and Methods). In this model, the likelihood of observing all base calls at a particular locus, assuming a particular diploid genotype AB, is expressed as Pr(D|G = AB) and is the product of observing the base call at each individual read j (Equation 2 of Materials and Methods). As described below, Pr(D_j|G = AB) is calculated according to the strand of read j and several bisulfite-specific parameters, β,α and γ (Figure 2b).

In the GATK non-bisulfite SNP calling model, the probability of observing a base call different from the presumed genotype G is simply the base call quality score (defined as the probability of a base calling error). In the case of Bisulfite-seq, this is true for A:T genotypes but not C:G. For C:G genotypes, the probability of observing a T depends on the strand of the read, the methylation state, and the efficiency of bisulfite conversion. Reads on the G-strand opposite the cytosine are treated with the normal GATK model. Reads on the C-strand use an alternate model that considers C>T substitutions as either potential errors or bisulfite conversions (see Materials and Methods). The probability of observing a bisulfite conversion event depends on both the underlying methylation state and bisulfite conversion errors. While none of these are observed directly, they are included in the model as variables β,α and γ as described in Equation 5 in the Methods section.

After bisulfite treatment, an unmethylated C that fails to get converted to a T is referred to as an underconversion, while a methylated C that is converted to T is referred to as an overconversion. The underconversion rate, α, is often estimated using either a spike in control 4 or the unmethylated mitochondrial genome 6. This rate can be set manually by the user and has a value of 0.25% by default. While bisulfite overconversion can not be reliably measured using current Bisulfite-seq data, we include an additional parameter, γ, which is set to 0% by default. In the future, this could be estimated by spiking in fully-methylated control DNA.

The percentage of methylated reads at a given cytosine position can vary widely. Since C reads and T reads yield more information about the presence of a C>T SNP than T reads, the locus-specific methylation rate can strongly influence SNP calling. In mammalian genomes, CpG methylation levels are multimodal, with various classes of functional elements having distinct methylation patterns. At least four different classes exist with mean methylation rates ranging from around 0% to over 80% 424. Furthermore, methylation at particular di- or tri-nucleotide contexts is organism and even cell type specific. To better understand how methylation estimates could affect SNP calling performance, we implemented several different methods for estimating the methylation frequency parameter β, which we describe next.

First, we used a naive estimate for β where the probability of a read being methylated or unmethylated at any particular cytosine position was 0.5. Second, we used context-specific estimates which were determined in a two-round procedure as follows. In the first round, naive estimates were used as described above, and the resulting SNP calls were used along with dbSNP to select a set of high-confidence non-SNP homozygous cytosines (probability>99.99%). These homozygous cytosines were used to estimate average methylation levels for a set of cytosine sequence contexts that could be specified on the Bis-SNP command line (by default, set to β_CG and β_CH). In the third and final estimation method, β was estimated for each cytosine locus individually using the number of C and T reads ( c c + t ). The rationale for this locus-specific method was our concern that genome-wide estimates might be inappropriate CpGs, given the strongly bimodal nature of CpG methylation levels. Each of these three β estimation methods was run individually as described below. The default method for the public version of Bis-SNP is locus-specific estimation.

We evaluated Bis-SNP calling accuracy for each of the three different methylation estimation methods (naive, context-specific, and locus-specific). The latter two methods performed substantially better than naive estimation, so those are the only two discussed below. We evaluated accuracy using an actual whole-genome Bisulfite-seq dataset from a normal (male) human colon mucosa sample published previously by our lab 6 (sequence available via accession dbGap:phs000385). All reads were 75 bp long single-end, and generated using the Illumina Genome Analyzer IIx platform. The complete dataset had an average read depth of 32X. The Bisulfite-seq data were compared to Illumina Human1M-Duo BeadChip SNP array data from same sample.

The primary goal of bisulfite sequencing is the accurate determination of cytosine methylation levels, so we first investigated the ability of Bis-SNP to correctly identify homozygous cytosines. As the 'ground truth', we used 435,120 positions identified as homozygous cytosines on the 1 M SNP array, and examined false negative and false positive calls made by Bis-SNP (Figure 3a-c). Calls at varying stringencies were generated by adjusting the Bis-SNP score cutoff, which is defined as the odds ratio between the first and second most likely genotype (see Methods). Evaluating the different Bis-SNP methylation estimates with and without base quality recalibration showed that the locus-specific β estimation plus recalibration produced the most accurate results. Using the complete sequence dataset and the default score cutoff (Figure 3c, red circle), Bis-SNP was able to detect 95.22% of the true cytosines (414,327 features) with a false positive rate of 0.37% (2,461 features). We simulated lighter sequencing coverage by randomly picking reads from the full dataset to estimate accuracy at 8× (Figure 3a) and 16× (Figure 3b) genomic coverage. The reader should note that these false positive rates are not indicative of the genome-wide false positive rates, since most false positives come from heterozygous SNPs which are frequent on the SNP array but very infrequent in the genome.

For comparison, we determined the accuracy of homozygous cytosine calling using several published methods (Figure 3a-c). Bismark34 returns methylation estimates for all cytosines in the reference genome. It is thus not surprising that Bismark performs poorly for features on the 1 M SNP array, which were selected for their polymorphism and differences from the reference genome. Several other published studies use the same strategy and estimate methylation at all reference cytosines 3536. In our own earlier work 6, we also restricted methylation calling to reference cytosines. Thus it is not surprising that when we applied this method ('Berman2012') to the 1 M SNP array dataset, it achieved almost the same false negative rate as Bismark. However, 'Berman2012' filtered out positions where less than 90% of reads were C or T on the C-strand and G on the G-strand, resulting in a substantially lower false positive rates than Bismark, but not as low as Bis-SNP.

We next focused on the ability of Bis-SNP to determine heterozygous SNPs, which can be used both for improving methylation calling accuracy as well as allele-specific methylation analysis (see Figure 1b). Heterozygous SNPs are more difficult to identify than homozygous SNPs, due to the approximately 1/2 the read coverage for each allele. We excluded the haploid × chromosome, leaving 303,656 autosomal loci called as heterozygous by the 1 M SNP array. As before, the locus-specific β methylation estimation plus recalibration performed the best of all methods. Using the full dataset with the default Bis-SNP cutoff (Figure 3c, red circle), Bis-SNP was able to identify 93.18% of heterozygous SNPs (282,944 loci) with a false positive rate of 0.094% (755 loci). Of the 303,656 heterozygous loci examined, 242,347 (79.81%) were C/T heterozygotes. C>T is the most common SNP in mammals, arising from evolutionary deamination of methylated cytosines. It is also the most difficult SNP to detect in bisulfite-treated DNA, because the C-strand reads are often uninformative (see Figure 1). As expected, Bis-SNP (and other methods) performed more poorly on C/T heterozygous SNPs than others, due to C>T conversion ambiguity (Additional File 2).

We compared Bis-SNP results to heterozygous SNPs called using two alternate 'k-allele' techniques that used read count cutoffs without incorporating base quality scores. We implemented a generalized form of the method used by 2130 to use a variable read count cutoff. This cutoff, k, was defined as the minimum percentage of reads with a secondary allele necessary to call a heterozygous SNP. As in 30, we counted C and T as a single allele at reference cytosines (on the C-strand only). In addition to k-allele, we also tried the Shoemaker method 20, which does not evaluate C/T SNPs at all and requires observations of the less frequent allele on at least 20% of reads on each strand. Finally, we tried the bisReadMapper algorithm 3, which calls SNPs independently on each strand using a non-bisulfite SNP caller, SAMTOOLS 11, and reports only those SNPs that agree between strands. Figures 3d-f show that each variation of Bis-SNP performs better than other methods.

An important practical question is the minimum read depth required for accurate SNP identification. We addressed this problem by downsampling our 32× Bisulfite-seq genome to various coverage levels from 2× to 30× (Figure 4). For each coverage level, we determined the number of false positives and false negatives across a range of Bis-SNP stringency cutoffs using the 1 M SNP array data, as in Figure 3. At each coverage level, we then selected the least stringent cutoff that produced a False Discovery Rate (FDR) of less than 5%, and plotted the number of true positives (sensitivity). For both homozygous cytosines (Figure 4a) and heterozygous SNPs (Figure 4b), sensitivity increased dramatically up to about 10× coverage and then began to level off. Homozygous SNPs were almost fully detected (98% sensitivity) by 10× coverage, while heterozygous SNPs had a more gradual increase from 80% detected at 10× to 95% detected at 30×.

To verify the ability of Bis-SNP to correctly identify cytosines and improve methylation quantification genome-wide, we ran Bis-SNP across an entire chromosome for the OTB colon mucosa sample and four additional whole-genome bisulfite-seq samples (Table 1). TCGA normal lung and normal breast were generated by the USC Epigenome Center and aligned using BSMAP, while the two mouse methylomes were generated by UCSD and aligned using Novoalign 22. Runtimes for chromosome 1 were about 3 hours using a standard 12-core Intel server with 10 GB RAM (Intel, Santa Clara, CA, shown). The entire human genome takes about 30-40 hours on a single server (data not shown).

We used Bis-SNP to identify four classes of cytosines in the sample genome (Figure 5 and Table 2 'Sample Genotypes'), and separated these by their corresponding sequences in the reference genome (Figure 5 and Table 2 'Reference Genotypes'). As shown in Table 2 about 0.5-0.6% of reference CpGs were lost in the sample genome, and 0.5-0.6% of CpGs in the sample genome were lost in the reference. The two mouse samples had significantly higher SNP rates, presumably due to true strain differences between the crossed strains and the C57BL/6J strain sequenced for the mouse reference genome. In both F1 mice, about 2.5% of reference CpGs were lost in the sample genome, and about 1.1% of CpGs in the sample genome were lost in the reference.

We next compared average methylation levels across each sample genotype (Figure 5). As expected, homozygous CpHs were consistently low, while homozygous CpGs were consistently high, regardless of the corresponding reference sequence. Both mouse frontal cortex brain samples showed elevated levels of CpH methylation as described in the original publication 22. Interestingly, homozygous CpGs that represented SNPs (where the sample differed from the reference genome) had consistently higher methylation. This fits with what is known about mammalian genome evolution - evolutionary C>T changes occur much more frequently at methylated than unmethylated CpGs because the C>T deamination and deamination repair process is methylation-specific. We next looked at heterozygous CpGs (Figure 5, right). CpG/CpH positions had methylation about halfway between CpG homozygous and CpH homozygous positions. At CpG/ApG or CpG/GpG heterozygous positions, methylation can only be measured for the C allele, and the methylation state is about the same as homozygous CpGs. CpG/TpG heterozygous positions are not shown, because we can not accurately measure methylation at these positions. Together, these data show that Bis-SNP genotype calling produces accurate methylation quantification even when the sample genome differs from the reference genome.

Reads with mapping quality scores less than 30 and those mapped to multiple genomic regions were removed, as are PCR duplicates (optional). For paired-end reads, we remove read pairs that do not have the ProperlyPaired field set.

We use GATK to perform local multiple sequence realignment and sequence recalibration mostly as described 12. Since most of bisulfite sequencing mapping tools (e.g. Bismark, BSMAP, MAQ etc) do not provide correct CIGAR string in the BAM file for GATK's indel realignment, the CIGAR string is recalculated when necessary. We extend GATK's RealignerTargetCreator to count mismatch number but not count thymine as a mismatch when the reference genome position is cytosine. After we create a potential indel interval, we realign using a modified version of GATK's IndelRealigner. PCR duplicate reads are marked after indel realignment.

For base quality recalibration, we modify the GATK algorithm to account for bisulfite conversion by extending the GATK CountVariantWalker and TableRecalibrationWalker classes. The algorithm first tabulates empirical mismatches to the reference at all loci not known to vary in the population (i.e., not in dbSNP build 135). These counts are categorized by their reported instrument-reported quality score (R) and position (cycle) within the read (C). In tabulating mismatches, we do not count thymine as a mismatch when the reference genome position is cytosine (on the second end of a paired-end read, we instead don't count adenine as a mismatch when the reference is guanine).

By default, only positions with a recalibrated Base Calling Quality Score of greater than 5 are used for SNP calling. This quality cutoff can be set using a command line parameter (see User Manual in Additional File 3).

We begin with the bayesian likelihood model of GATK (12), and make a number of bisulfite-specific adaptations. Assuming the underlying genome is diploid, we let D = (D₁, D₂, ..., D_r) represent the base calls at a particular genomic position i that is covered by r sequencing reads. We then calculate the posterior probability by (1) as in GATK:

P r ( G | D ) = π ( G ) P r ( D | G ) P r ( D )

Here, G is the underlying diploid genotype, AB, with A and B being the two parental alleles. π(G) is a genotype prior probability for observing the given genotype based on the genotype of the reference genome and population frequencies, the same as discussed in Table 1 of SOAPsnp paper 13. Pr(D) is defined as the sum over all possible genotypes ∑_AB π(AB) Pr (D|AB), but is the same in each case and can generally be ignored since we are concerned with likelihood ratios. We assume that each of the two alleles are equally likely to be sequenced, and calculate the overall likelihood of D as the product of all individual reads (2),(3):

P r ( D | G ) = ∏ j = 1 r P r ( D j | G )

P r ( D j | G = A B ) = 1 2 P r ( D j | A ) + 1 2 P r ( D j | B )

The following steps are shown for single-end sequences. For paired end sequences, the first end is treated as described, but the second end is reverse complemented before performing these calculations (because the Illumina second end is the complementary strand of the same template as the first end). This changes G>A bisulfite substitutions, which occur on the second end, to the actual C>T substitutions present on the bisulfite-converted template. The recalibrated base quality scores are on a phred scale which represents the probability ε that the position is an error, which is used in the following calculation.

When the underlying allele is adenine (a), thymine (t), bisulfite conversion does not apply and the probability estimation is straightforward as shown for t:

Pr ( D j | B = t ) = ε j 3 if D j ≠ t 1 - ε j if D j = t

Here, ε_j is the probability of a sequencing or base calling error at position j, i.e. probability that the true allele B is a t, but base call D_j is observed as an a, c, or g. The likelihood function for a is equivalent to that of Equation (4). When the underlying allele is a c or a g, however, the probabilities are strand-specific since bisulfite conversion only affects one strand in the directional Bisulfite-seq protocol (Figure 1). The probability of seeing a t in the read depends on the probability that the position is methylated (β), as well as the bisulfite conversion efficiency (α and γ). Bisulfite treatment converts all unmethylated cytosines to thymine, but in practice it is not 100% efficient 4. The parameter α is the estimated frequency of unmethylated cytosines which are not converted (typically taken from unmethylated spiked in DNA 4 or the mammalian mitochondrial sequences, which we have found to be almost completely unmethylated 6. In this case, α = β_chrM). By default, α is set to 0.0025 but can be specified by the user. We also include a γ parameter for over-conversion, i.e. the rate at which methylated cytosines are converted. Although this is not routinely measured in practice, it could be estimated by including an enzymatically methylated control DNA 40, or a sequencing library without bisulfite conversion. By default, γ is set to 0 but can be specified by the user. The full likelihood calculation for cytosines is as follows:

P r ( D j | B = c ) = ( 1 - ε j ) [ β j ( 1 - γ ) + ( 1 - β j ) α ] if D j = c + ε j 3 + ( 1 - ε j ) [ β j γ + ( 1 - β j ) ( 1 - α ) ] if D j = t + 1 - ε j if D j = c - ε j 3 otherwise β j 1 - γ = methylated and  properly  not converted β j γ = methylated and  improperly  converted ( 1 - β j ) α = unmethylated and  improperly  not converted ( 1 - β j ) ( 1 - α ) = unmethylated and  properly  converted

The key to these calculations is that reads on the same strand as the inferred cytosine allele (denoted with +) are treated differently than reads from the opposite strand (denoted with -). As expected based on the example in Figure 1, a true allele of B = c results in a very high probability of seeing a t⁺ (a 't' read on the C-strand), but a very low probability of seeing a t^- (an 'a' read on the G-strand). The genotype G_best with the highest posterior probability Pr(G|D) is chosen, and the final output score is the odds ratio between the best (G_best) and the second best (G_nextbest), as in Equation (6). In practice, we optimize execution by evaluating only the subset of the 10 possible diploid genotypes that are possible given the sequences read.

s c o r e = l o g ( P r ( G b e s t | D ) P r ( G n e x t b e s t | D )

Bisulfite efficiency, i.e. α and γ typically vary by less than 1%, so the critical parameter included in Equation 5 is the methylation rate β. Since this rate varies by genomic context, organism, and even cell type, we allow the user to specify the possible contexts as a set of n nucleotides sequences specified by their IUPAC degeneracy codes (for instance, CH represents CC, CT, or CA). In mammalian genomes where typically only the single base 3' of the cytosine is considered relevant, the user would specify CG and CH (the Bis-SNP default). For Arabidopsis, one might specify CG, CHH, and CHG. Any arbitrary number of 5' and 3' bases may be specified in order to accommodate the full range of Bisulfite-seq assays. For instance a CCGG pattern could be specified for MspI restriction sites inherent to the RRBS protocol ( 41).

One methylation output file (BED6+2 format) is created for each cytosine context specified by the user. For each cytosine determined to have the particular sequence context, the percent methylated (the number of C reads on the C-strand divided by the number of C or T reads on the C-strand) is output as the score field. To aid in statistical analysis, a second field contains the total number of C/T reads.

Non-conversion of unmethylated Cs is known to preferentially affect the 5' end of Illumina-generated reads, most likely driven by the re-annealing of sequences adjacent to the fully methylated sequence adapters during bisulfite conversion. We control for this using a 5' non-conversion filter as implemented in our earlier work 6. For each read, we walk along the read from 5' to 3', and we remove any Cs on the C-strand until we reach the first reference C which is converted to a T. By applying this filter, early bisulfite conversion in early cycles is brought to levels very similar to those of late cycles, thus removing a potential source of methylation bias (data not shown). Notice that this filter should be turned off for RRBS data, which gleans most of its methylation data from the first cycle (see user manual).

Using the approach of GATK, we apply additional quality filters before SNP calling to avoid known sources of false positives. SNPs found in clusters (two or more within a ten-base-pair window) were filtered out. SNPs with coverage depth above 120, Strand Bias(SB) score more than -0.02, or Quality by Depth(QD) less than 1.0 are filtered out. All of these parameters are configurable (see User Manual). If BAM contains Mapping Quality scores, suspicious regions are filtered out when greater than 10% of aligned reads (minimum of 40 reads) have mapping quality of 0.

Bisulfite sequencing can have higher strand biases since high bisulfite concentration can lead to DNA degradation when the depurination step causes random strand breaks 4243. We calculated strand bias score as in GATK, but bisulfite converted reads have an apparent strand bias which is higher than the actual strand bias, since the G-strand contributes more than the C-strand at cytosines. For this reason, we used a substantially less stringent strand bias cutoff (-0.02) than the GATK default.

We downsampled the human colon mucosa Bisulfite-seq dataset into different mean coverages using GATK, which randomly picks z reads at each individual nucleotide locus. The following formula is used, where N is the mean coverage of total dataset before downsampling (32× in this case), n is the desired downsampling coverage, and m is the actual coverage at the particular locus.

z = m * n N