SNP discovery via 454 transcriptome sequencing.

A massively parallel pyro-sequencing technology commercialized by 454 Life Sciences Corporation was used to sequence the transcriptomes of shoot apical meristems isolated from two inbred lines of maize using laser capture microdissection (LCM). A computational pipeline that uses the POLYBAYES polymorphism detection system was adapted for 454 ESTs and used to detect SNPs (single nucleotide polymorphisms) between the two inbred lines. Putative SNPs were computationally identified using 260,000 and 280,000 454 ESTs from the B73 and Mo17 inbred lines, respectively. Over 36,000 putative SNPs were detected within 9980 unique B73 genomic anchor sequences (MAGIs). Stringent post-processing reduced this number to > 7000 putative SNPs. Over 85% (94/110) of a sample of these putative SNPs were successfully validated by Sanger sequencing. Based on this validation rate, this pilot experiment conservatively identified > 4900 valid SNPs within > 2400 maize genes. These results demonstrate that 454-based transcriptome sequencing is an excellent method for the high-throughput acquisition of gene-associated SNPs.

Introduction

SNPs (single nucleotide polymorphisms) are single base differences between haplotypes. Once discovered, SNPs can be converted into genetic markers that can be inexpensively assayed in a high-throughput manner (Gut, 2001; Kwok, 2001). Due to their abundance, it is possible to use SNP-based markers to generate very dense genetic maps (Rafalski, 2002). Such maps can be used to conduct marker-assisted selection (MAS) programs, construct the specific genotypes required for quantitative genetic studies, and to enhance our understanding of genome organization and function and address fundamental questions relating to evolution and meiotic recombination. SNPs can also be used for genome-wide linkage disequilibrium and association studies that assign genes to specific functions or traits. Furthermore, transcript-associated SNPs can be used to develop allele-specific assays for the examination of cis-regulatory variation within a species (Bray ; Cowles ; Guo ; Pastinen et al., 2004; Stupar and Springer, 2006).

Although SNPs can be identified by sequencing candidate genes from a set of individuals that represent diversity in the species of interest, this is neither high-throughput nor inexpensive. Alternative approaches used during construction of the human SNP map included identifying sequence polymorphisms within overlapping BAC clones derived from different individuals and shotgun sequencing of genomic fragments (Sachidanandam ). However, this approach is not always possible because many genome sequencing projects use DNA extracted from highly similar or inbred individuals. Instead, SNP-based markers are typically mined from whole-genome sequences or expressed sequence tags (ESTs) obtained from genetically diverse individuals. For example, SNPs have been identified by comparing genomic sequences from two or more genetically distinct inbred lines of mouse (Wiltshire ), the indica and japonica sub-species of rice (Feltus ), the Columbia and Landsberg ecotypes of Arabidopsis (Jander ), and different lines of maize (Yamasaki ). EST collections from genetically dissimilar individuals have similarly been mined for SNPs in humans (Marth ), pine (Dantec ), barley (Kota ), cassava (Lopez ) and maize (Batley ).

The latest maize genetic map (IBM_IDP_bd map, ver4) contains over 3000 gene-based PCR markers distributed across the 2.5 Gbp genome (Fu ). Even so, this map is not dense enough to support high-resolution mapping applications and association genetics, particularly given the decay of linkage disequilibrium outside of maize genes (Ching ; Tenaillon ). Additionally, because the maize inbred B73 line is being hierarchically sequenced, a higher density genetic map would be invaluable for anchoring each sequenced BAC contig to its proper place in the genome. Increasing the marker density of this crop therefore has applications in accurately assembling this highly complex genome and ultimately in improving agricultural traits.

Maize is genetically very diverse; SNP and indel polymorphism frequencies between inbred lines and landraces average one variation per 124 or 28 bases for coding regions (Ching ) or all associated regions (Tenaillon ), respectively. We were particularly interested in identifying SNPs between B73 and the inbred line Mo17. These two inbred lines represent two of the major heterotic groups, and historically are the parental lines of much of the commercial corn grown in the USA. These inbred lines are also the parents of the IBM RILs (recombinant inbred lines) that were used to develop the maize genetic community’s high-resolution genetic maps.

The size and complexity of the maize genome make it unlikely that a second inbred line will be sequenced in the immediate future. Although there are currently over 650 000 maize EST sequences available in GenBank, nearly all of these were drawn from a small subset of inbred lines, principally B73, W23 and Oh43A. Hence, the identification of B73/Mo17 SNPs requires the development of Mo17 EST sequence resources. Although genome sequencing technology has become progressively more efficient, EST projects require substantial investments in library construction and sequencing efforts to achieve the overall coverage required to locate SNPs.

Recently, 454 Life Sciences (http://www.454.com) reported a highly parallel DNA sequencing system that is 100 times faster than standard sequencing methods and is capable of providing over 20 Mbp of sequence in a single four-hour run (Margulies ). Increased throughput comes at the expense of read length (100 bp average length) and the absence of clone pair information, making it less attractive for whole-genome sequencing of complex genomes. However, 454 sequencing of maize cDNAs obtained from shoot apical meristem (SAM) tissue isolated by laser capture microdissection (LCM; reviewed by (Schnable ) has recently been shown to be an effective method for tagging tens of thousands of maize genes without cloning and its associated costs (Emrich ). Therefore, 454-based sequencing of the B73 and Mo17 SAM transcriptome was expected to provide a collection of diverse ESTs that could support high-throughput computational identification of gene-associated SNPs. Because 454 reads contain more sequence errors than do reads generated by traditional sequencing technology (Margulies ), it was not, however, clear whether 454-based ESTs could be used for SNP discovery.

Here, we describe the generation of over 280 000 Mo17 SAM ESTs using 454 sequencing technology, the development of an efficient computational SNP mining pipeline based on the POLYBAYES sequence polymorphism detection tool, and the subsequent identification of over 7000 putative Mo17/B73 SNPs within expressed sequences, a subset of which has been experimentally validated.

Results

The shoot apical meristem ultimately gives rise to all above-ground tissues. Thus, it is expected that many rare and developmentally important transcripts are present in the SAM transcriptome. Indeed, we have demonstrated that 454 sequencing of maize SAM cDNA captures fragments of thousands of genes, including many that may be expressed only rarely or only in the SAM (Emrich ).

Using 454 sequencing, we previously generated from the B73 inbred line a collection of 260 887 high-quality SAM ESTs with an average length of 101 bp (Emrich ). Using the same methodology, a collection of 454 SAM ESTs was generated from the maize inbred line Mo17 (Experimental procedures). After trimming polyA/T tails, the 287 917 resulting SAM ESTs from Mo17 had an average length of 100 bp, and consisted of 30.7 Mbp in total.

Assignment of Mo17 and B73 SAM ESTs to maize genomic anchor sequences

MAGIs are maize genomic sequence assemblies (Fu ) composed of gene-enriched B73 genomic survey sequences (Whitelaw ). Because these sequences are highly accurate (1 disagreement per 10 000 bp; Fu ) and comprehensive (> 75% of all maize genes are present), they provide an excellent collection of B73 reference sequences for SNP detection. Attempts were made to align each of the 260 887 B73 and 287 917 Mo17 454 ESTs to the MAGI version 3.1 partial maize B73 genome assembly using a two-step approach. The initial pre-processing step uses BLAST to save time and improve accuracy by grouping together individual 454 SAM ESTs that preferentially align to a single MAGI template (Experimental procedures). This analysis assigned 432 431 of the 454 ESTs (207 294 B73 and 225 137 Mo17) to 48 063 MAGIs (Table 1). Of these MAGIs, 20 432 aligned to both B73 (n= 120 662) and Mo17 (n= 135 249) ESTs. An additional 14 496 and 13 135 MAGIs aligned to only Mo17 (n= 89 888) or only B73 (n= 86 632) ESTs, respectively. The MAGI assembly sequences identified above served as templates upon which associated 454 ESTs were multiply aligned by CROSS_MATCH (P. Green, University of Washington, personal communication).

Table 1

Summary of multiple sequence alignments (MSAs) between MAGI 3.1 anchors and B73 and Mo17 454 ESTs

	Types of ESTs in MSAs
	All ESTs	All B73 ESTs	All Mo17 ESTs	Both B73 and Mo17 ESTs	Only B73 ESTs	Only Mo17 ESTs
Number of MAGIs aligned	48 063	33 567	34 928	20 432	13 135	14 496
Bases covered	8 897 508	4 989 045	5 798 933	1 890 459	3 098 586	3 908 463
Coverage depth	1.8 x	2.3 x	2.3 x	8.4 x	1.3 x	1.3 x

For this analysis, 454 sequences were initially mapped to individual MAGIs using BLAST, which later served as the template on which these MSA were computed using CROSS_MATCH (Experimental procedures). Coverage data are presented for all alignments, as well as alignments between individual subsets of ESTs.

Doing so produced 48 063 anchored multiple sequence alignments (MSAs) that covered a total of 8 897 508 MAGI template bases with 454 ESTs. Approximately 5 Million and 5.8 Million anchor template bases were sampled by B73 and Mo17, respectively, while slightly fewer than 1.9 Million bases were sampled by both inbred lines (Table 1). The relative proportions and average sequence depths (coverage) of the 1.9 Mbp MAGI nucleotides sampled by B73 and Mo17 454 ESTs are presented in Table 2. Although it is theoretically possible to identify putative B73/Mo17 SNPs across the entire region of the MAGI 3.1 sequence space that was simultaneously sampled by B73 and Mo17 454 ESTs (approximately 1.9 Mbp), analysis of those regions that contain deeper sequencing coverage for both inbred lines is expected to yield putative SNPs that are more likely to be valid. We therefore defined a high-confidence set of bases on the MAGI anchor that was sampled to ≥ threefold by Mo17 ESTs and to ≥ twofold by B73 ESTs. With the inclusion of the MAGI 3.1 anchor sequence (B73), these bases are sampled a minimum of three times for both inbred lines. This set comprises 42% (606 180) of the simultaneously sampled sequence space (Table 2).

Table 2

Average coverage of nucleotide sites represented within B73 454 ESTs, Mo17 454 ESTs and MAGI 3.1 anchored multiple sequence alignments

454 EST component depths			Number of nucleotides	Average coverage
Mo17		B73	1 092 570	3.2
1 x		≥1 x
	or
≥ 1 x		1 x	326 095	5.9
2 x		≥ 2 x
	or
2 x		≥ 2 x
≥ 3 x		2 x	134 386	6.7
≥ 3 x		≥ 3 x	471 794	22

Although the alignment of a single Mo17 EST to a B73-derived MAGI is sufficient to predict a SNP, increased sampling depth is expected to increase the accuracy of SNP calling by filtering out sequencing errors. Depth classes that are grouped together were pooled for analysis.

Polymorphism detection using POLYBAYES

Putative SNPs were identified from the MSA using the POLYBAYES polymorphism software package (Marth ). POLYBAYES uses a Bayesian statistical model that considers depth of coverage, sequence quality and an expected polymorphism rate to determine the probability that polymorphic sites within an MSA are SNPs rather than disagreements resulting from either sequencing errors or the alignment of paralogous (rather than allelic) sequences (Marth ).

454 sequencing technology is susceptible to indel-type errors (Margulies ), and the resulting ESTs exhibit an overall rate of sequencing error of approximately 1.5% (Emrich ). To address the issue of indel-type errors, we used MAGI assemblies as templates on which 454 SAM ESTS were aligned (Experimental procedures). Template-based MSAs such as these are often correct even in the presence of abundantly expressed or alternatively spliced transcripts (Marth ), and are therefore more likely to overcome the technical issues associated with 454 ESTs.

POLYBAYES identifies single base substitutions as well as single base insertions and deletions. However, because of the high number of indel errors associated with 454 technology (Margulies ), only base substitutions (i.e. SNPs) were considered in the current analysis. Initially, a total of 36 006 putative SNPs (P= 0.5) were detected within 9980 unique MAGI anchor sequences. This number of putative SNPs is expected to over-estimate the diversity present in SAM-expressed genes in the two maize inbred lines. Because Mo17 and B73 are inbred lines, they should be mono-allelic at every base position, with relatively rare exceptions caused by nearly identical paralogs (NIPs) (Emrich ). Hence, the observation that many of the putative SNPs discovered initially are multi-allelic within Mo17, B73 or both, suggests that many are false positives due to sequencing errors. With this in mind, we purposefully set the SNP probability low (P= 0.5) and filtered the putative SNPs using the following rules designed to substantially decrease the rate of false positives within the context of this study:

Polymorphic sites require a minimum of twofold representation in the Mo17 454 ESTs.

All Mo17 base calls at sites that were polymorphic between Mo17 454 ESTs and the B73 MAGI anchors were expected to be identical. This ensures mono-allelism within the Mo17 454 ESTs.

When B73 454 EST sequences also align across polymorphic sites that pass rules 1 and 2, all of the B73 454 ESTs and the MAGI 3.1 anchor base calls must agree. This avoids polymorphisms resulting from incorrect MAGI base calls or NIPs within B73.

To reduce the possibility of an erroneous base in the MAGI anchor mimicking a true SNP, regions of the MAGI assemblies composed of sequences from high-Cot selected clones that are not covered by B73 ESTs were avoided because 40% of high-Cot clones contain cloning artifacts that mimic SNPs (Fu ) (see Experimental procedures).

Applying these stringent rules to the raw SNP data returned 7016 putative B73/Mo17 SNPs distributed among 3403 MAGIs. The numbers of 454 ESTs that cover these polymorphic sites range from only two Mo17 454 ESTs to at least three B73 and three Mo17 ESTs (Table 3).

Table 3

Number of putative SNPs, depth at each SNP site by inbred line, and estimates of the total number of maize genes that contain at least one putative SNP between the B73 and Mo17 inbred lines in this SNP dataset

454 EST component depths of MSAs
Mo17		B73	Number of putative SNPs	Number of MAGI 3.1 anchorsa	Additive SNP number	Additive minimum estimate of SNP-containing genesb
1 x		1 x	1762	1154
	or
1 x		0
2 x		≥ 2 x	1648	1039
	or
≥ 2 x		2 x
≥ 3 x		≥ 3 x	1452	900	1452	900
≥ 3 x		2 x	565	404	2017	1205
≥ 3 x		1 x	717	513	2734	1570
2 x		≥ 3 x	537	372	3271	1821
2 x		2 x	546	363	3817	2053
2 x		1 x	1045	707	4862	2548
≥ 3 x		0	481	283	5353	2775
2 x		0	1673	830	7016	3403

Polymorphic bases sampled with low redundancy (rows 1 and 2) were not further analyzed. In contrast, rows 4 and 5 illustrate polymorphic sites with a minimum sampling depth of threefold for both inbred lines, and, as a result, have the highest confidence. The remaining rows summarize alignments that predict SNPs with decreasing confidence levels. Sub-categories that are grouped together were pooled for analysis.

MAGIs are gene-enriched maize genomic sequence assemblies that are likely to contain only a single gene or gene fragment (Emrich ; Fu ).

Numbers represent a non-redundant collection at each row.

For completeness, Table 3 presents all polymorphism data. The total numbers of polymorphic bases sampled by only one or two Mo17 454 ESTs and/or B73 454 ESTs are displayed in rows 1 and 2, respectively; these were removed from further consideration. The numbers of putative SNPs that pass the above rules and their associated MAGIs are presented in rows 3–12. Rows 3 and 4 illustrate the total number of polymorphic sites sampled simultaneously by a minimum of three Mo17 454 ESTs, two B73 454 ESTs and the B73 MAGI 3.1 anchor. This represents the highest-confidence data set, with a minimum sampling depth of threefold for both inbred lines. Rows 5–12 display putative SNPs at sites with decreasing depths of coverage, which are expected to represent decreasingly confident data sets. This expectation is supported by their corresponding POLYBAYES-assigned SNP probabilities (pSNP) (Supplementary Tables S1 and S2). The number of potential SNPs, the number of their associated MAGI anchors for each B73/Mo17 sampling depth, and the total number (additive) of potential SNPs and the number of unique MAGIs anticipated by systematically including data sets (starting with row 3) is also presented in Table 3. In summary, after single 454 GS-20 sequencing runs of B73 and Mo17 SAM cDNA, our computational polymorphism mining strategy identified over 7000 putative SNPs (Supplementary Table S1).

Validation of SNPs

A set of 110 putative B73/Mo17 SNPs were subjected to validation by sequencing (using Sanger technology) the corresponding alleles that had been PCR-amplified from B73 and Mo17 genomic DNA. Detailed results of these validation experiments are presented in Supplementary Table S2. The overall rate of validation was over 85% (94/110). Most of the SNPs selected for testing represent sites with at least moderate levels of B73/Mo17 coverage. Over 88% (85/96) of SNPs sampled by three or more Mo17 454 ESTs and two or more B73 454 ESTs (Table 3, rows 3 and 4) were validated. Fewer of the lesser-confidence SNPs were assayed; these exhibit a collective validation rate of 64% (9/14). Using the above validation rates, the number of SNPs that could be validated was estimated (Table 4); these data suggest that 4984 computationally identified B73/Mo17 SNPs represent ‘true’ polymorphisms, and that these are distributed within 2472 MAGIs. The average sizes of the MAGI assemblies suggest they contain only one (or a portion of one) maize gene. Because these polymorphisms were mined from cDNA sequences derived from mRNA and conservatively filtered, we estimate that this analysis identified at least 4900 valid SNPs within at least 2400 maize genes.

Table 4

Number of putative SNPs, depth at each SNP site by inbred line, and estimates of the potential number of polymorphic maize genes adjusted for validation rates

454 EST component depths
Mo17	B73	Number of putative SNPs (Table 3)	Validation rate	Estimated number of valid SNP sitesa	Number of MAGI 3.1 anchorsab	Number of additive SNPs	Additive minimum estimate of genes impactedc
≥ 3 x	≥ 2 x	2017	0.885	1785	1154	1785	1066
≥ 2 x	0–1 x	3916
1–2 x	≥ 2 x	1083	0.64	3199	1963	4984	2472

Validation of 110 putative B73/Mo17 SNPs divided into two groups was performed by sequencing the corresponding B73 and Mo17 alleles using Sanger technology. Using the validation rates obtained, the number of SNPs that could be validated was estimated. Because many MAGIs correspond to single genes (see Results), the number of non-redundant MAGI anchors was used to generate the estimate of the number of genes impacted. Depths that are grouped together were pooled for analysis.

Numbers are corrected for validation rate (see text).

MAGIs are gene-enriched maize genomic sequence assemblies that are likely to contain only a single gene or gene fragment (Emrich ; Fu ). These numbers represent a non-redundant collection of MAGIs (see text).

Numbers represent a non-redundant collection See comment above at each row.

Discussion

Once discovered, SNPs have a wide variety of applications in biological research. One means to discover SNPs is to align ESTs from more than one genotype. LCM 454 sequencing enables efficient deep sampling of ESTs obtained from specific cell types (Emrich ), but suffers from the disadvantage of higher error rates than Sanger sequencing. Even so, this study demonstrates that it is possible to use ESTs obtained via LCM 454 sequencing to achieve high-throughput SNP discovery. Over 260 000 Mo17 ESTs were obtained from a single GS-20 sequencer run on cDNA isolated from SAM tissue, and over 7000 putative SNPs were identified relative to B73 genomic and 454 EST sequences. A subset of these SNPs was validated via direct sequencing of PCR products amplified from B73 and Mo17 genomic DNA.

Putative SNPs are identified as mismatches between aligned sequences, and several computational tools for SNP identification are available (Manaster ; Marth ; Nickerson ; Wang and Huang, 2005; Weckx ; Zhang ). Our SNP discovery pipeline implements POLYBAYES, which has been used to identify SNPs in several studies (Dantec ; Marth ; Pavy ; Useche ). We assigned default values to 454 sequences based on an empirical evaluation of the base error rate rather than using the relatively new 454 quality scores. As a result, sequence depth and relative allele proportions have the greatest influence on polymorphism detection, and, based on this observation, potential SNPs were filtered by examining these statistics at each polymorphic site. The highest-confidence polymorphisms are those that are minimally covered by both Mo17 and B73 sequences to threefold. Experimentally, > 88% of these sites could be validated as being polymorphic, and are assigned prior probability scores (pSNP Check nomenclature, cf pSNP used above) of at least 0.997 by POLYBAYES.

POLYBAYES is designed to use template-driven MSAs, in which sequences are scaffolded across a high-quality template sequence that serves as an anchor. In addition to being highly accurate (Marth ), this approach eliminates the need to perform de novo assemblies of 454 ESTs, which are complicated by the short lengths of 454 reads. Furthermore, gaps and insertions in this template-driven multiple sequence alignment approach are propagated throughout all members, so 454 semi-random indels can be easily identified and ignored (Figure 1). Finally, the ability of POLYBAYES to use quality scores during SNP detection provides the option of utilizing 454 sequence calls once they are better accepted by the research community, or if Sanger sequences are also used, or if the base accuracy of the template is suspect. In all of these cases, the availability of accurate base quality data could improve the accuracy of SNP detection.

Figure 1

A portion of the CROSS_MATCH-produced, template-driven, padded alignment between B73 and Mo17 454 EST sequences and the high-quality MAGI_105195 sequence assembly constructed from the B73 maize genomic survey sequence that serves as an alignment template. A G/A polymorphism occurs at position 2846 of the template (green highlight), with the Mo17 allele (A) in red and the B73 allele (G) in blue. Two insertions have occurred (yellow), one within a Mo17 454 EST and the second within a B73 454 EST. Because these insertions are not supported by other sequences, they are easily identified as errors by the POLYBAYSE pipeline and are not called as polymorphisms.

We estimate that our SNP collection contains at least 4984 valid SNPs within 2472 genes (see Results). This estimate is based on an observed validation rate of > 0.88 for polymorphic sites minimally sampled to threefold by each inbred line, and the assumption that all other depth classes of polymorphism have a conservative validation rate of 0.64. A subset of 2017 high-confidence SNPs was detected within B73 genomic sequence that was sampled by a minimum of two B73 ESTS and a minimum of three Mo17 454 ESTs (Table 3). The size of this reduced sequence space is 621 956 bp (Table 2), providing an observed polymorphism rate of at least 1/300. This rate is only about half of that previously reported in maize coding sequence (Ching ); however, the published rate was based on only 18 genes and may not be representative of the genome. Furthermore, the conservative parameters used in this study are expected to under-estimate polymorphism rates. Specifically, in the absence of 454 quality information, we required that B73 and Mo17 inbred lines both be mono-allelic at each nucleotide before calling a putative SNP. In fact, 17 671 instances where either inbred line (or both) exhibits bi-allelism were initially ignored to simplify polymorphism detection and subsequent validation. These were further parsed to identify putative SNPs where the B73 and/or Mo17 major allele frequencies are ≥ 0.75, and each major allele is represented at least three times within the MSA. There are 879 such cases (Supplementary Table S1), which, if all were validated, would increase our polymorphic rate to at most 1/214 bp.

All of the polymorphic sites discussed in this study were detected by comparing the sequences obtained from single 454 GS-20 sequencer runs on cDNA obtained from Mo17 and B73 SAM tissue. Additional sequencing runs would be expected to increase the proportion of the transcriptome sequence space covered, and, perhaps most importantly, increase the overall depth of coverage. Consequently, additional sequencing runs would be expected to increase the confidence of at least a fraction of the putative SNPs that are currently poorly supported due to insufficient sampling depth (Table 3). Increased depth would also lend additional support to the identification of NIPs, a process that is particularly dependent on deep sampling.

Maize is a globally important crop and a model system for the study of genome structure, evolution and genetics. Between 5000 and 10 000 years ago, the wild grass teosinte was domesticated to produce modern maize. Domestication resulted in a population bottleneck that reduced allelic diversity in maize relative to teosinte. Over the past decade, analysis of DNA sequence polymorphism data to detect signatures of genes that were involved in domestication and subsequent selection has become a well-established approach (e.g. Wang ; Whitt ; Tenaillon ; Wright ; Yamasaki ).

The maize genome is composed of approximately 2.5 billion bases and contains an estimated 50 000 genes (Fu ). The vast majority of this genome is composed of a small number of highly repetitive retrotransposons (Bennetzen, 1996; Meyers ; SanMiguel ; Whitelaw ). Hence, it has not been economically feasible to conduct whole-genome scans for SNPs by sequencing multiple maize haplotypes. However, the 454 EST-based SNP mining procedure described here, which is focused on a specific transcriptome using LCM, provides the underpinning for a high-throughput SNP discovery platform than could be used to cost-effectively identify genes that exhibit signatures of having been involved in the domestication or improvement of maize and other large-genome crops, and that are therefore potential targets for improving agriculturally relevant traits.

Experimental procedures

Isolation of SAM mRNA and 454 sequencing

Maize SAM cDNA isolation, 454 sequencing and raw sequence processing were performed as previously described (Emrich ). A single GS-20 run produced 260 887 (28.8 Mbp) and 287 917 (30.7 Mbp) B73 and Mo17 SAM ESTs, respectively.

Identification of B73 reference sequences for 454 ESTs

Mo17 454 ESTs were initially mapped to a specific contig or singleton (217 773 total) from the MAGI 3.1 partial genome assembly of the maize inbred line B73 (Fu ) using best BLASTN matches (minimum E-value 1e-8). Although ‘best hit’ criteria were used, it is possible that some 454 ESTs align to paralogous genomic fragments, especially given the partial nature of the MAGI assembly. To compensate, we used POLYBAYES (see below), which includes an internal paralog filter and should identify and discard these instances. These ESTs were also aligned to MAGIs using GeneSeqer (http://deepc2.psi.iastate.edu/cgi-bin/gs.cgi) and its maize-specific splice models (Usuka ) for display on the MAGI website (http://magi.plantgenomics.iastate.edu). Only alignments consisting of at least 50 bp in length and with identity ≥ 95% over at least 80% of the length of the 454 EST were used to annotate genomic sequences.

Multiple sequence alignments and SNP detection of 454 sequence data

Custom PERL scripts were written to create a pipeline to process MAGI 3.1 anchor sequences and their associated B73 and Mo17 454 EST sequences for detecting SNPs. Anchored MSAs were produced by CROSS_MATCH with the following parameters: -discrep_lists -tags -masklevel 5 -gap_init -1 -gap_ext -1. Low initiation (-gap_init) and gap extension (-gap_ext) were used to increase alignment tolerance between the short 454 ESTs and the unplaced MAGI 3.1 genomic anchors. Sequence polymorphisms were detected by POLYBAYES using the following parameters:

anchorBaseQualityDefault 34 memberBaseQualityDefault 18

maskAmbiguousMatches nofilterParalogs -priorParalog 0.03 -thresholdNative 0.75 -screenSnps -considerAnchor -noconsiderTemplateConsensus -prescreenSnps -priorPoly 0.01 -thresholdSnp 0.5.

Default anchor quality values (34) were based on a previous assessment of sequence error rates within the MAGI 3.1 assembly (Fu ). Default quality values of 18 were assigned to the 454 reads. This corresponds to an error rate of approximately 1/65, which over-compensates for the error rate observed for current 454 sequencing (Emrich ; Margulies ). Although each base within the 454 sequence reads is given a quality score, these scores are only reliable when confirmed within independent sequences covering the same region. Because CROSS_MATCH aligns each sequence individually to the anchor during MSA construction, and POLYBAYES assesses base quality on an individual basis, use of a stringent default rather than the base quality information provided by 454 Life Sciences is expected to increase the accuracy of polymorphism detection.

SNP parsing

Mo17 and B73 are inbred lines, and thus should be mono-allelic at every base position. Custom PERL scripts were written to parse the POLYBAYES output (see Results). POLYBAYES identifies indel polymorphisms. Because indels are a common form of 454 sequencing error, only base substitutions were considered during this analysis. MAGI 3.1 assemblies contain a low frequency of base substitutions propagated during shotgun sequencing of the high-Cot selected maize genomic DNA (Fu ). High-Cot selected maize DNA sequences account for only a portion of the MAGI 3.1 assembly sequence, but unidentified base substitutions within these regions could increase the number of false polymorphisms detected. Strict parsing rules (see Results) ensured that potential MAGI 3.1 sequence errors were avoided when B73 454 EST sequences are present in the multiple alignment. In cases where B73 454 ESTs are not present in the multiple alignment, SNPs called within regions of the MAGI 3.1 assemblies containing high-Cot selected DNA were avoided.