Adapterama II: universal amplicon sequencing on Illumina platforms (TaggiMatrix)

Introduction

Next-generation DNA sequencing (NGS) has facilitated a wide variety of benefits in the life sciences (Ansorge, 2009; Goodwin, McPherson & McCombie, 2016), and NGS instruments have an ever-growing capacity to generate more reads per run. Substantial progress has been made in developing new, lower-cost instruments, but much less progress has been made in reducing the cost of sequencing runs (cf., Glenn, 2011 vs. Glenn, 2016). Thus, the large number of reads from a typical NGS run comes with a relatively large buy-in cost but yields an extremely low cost per read. Frustratingly, within every NGS platform, the lowest-cost sequencing kits have the highest costs per read (Glenn, 2011; Glenn, 2016). This creates a fundamental challenge: how do we efficiently create and pool large numbers of samples so that we can divide the cost of high capacity NGS sequencing runs among many samples, thereby reducing the cost per sample?

It is well known that identifying DNA sequences (commonly called indexes, tags, or barcodes; we use the term “indexes” throughout) can be incorporated during sample preparation for NGS (i.e., library construction) so that multiple samples can be pooled prior to NGS, thereby allowing the sequencing costs to be divided among the samples (see Faircloth & Glenn, 2012 and references therein). When sufficient unique identifying indexes are available, many samples, including samples from multiple projects, can be pooled and sequenced on higher throughput platforms which minimizes costs for all samples in the pool.

In many potential NGS applications, the number of desired reads per sample is limited, so the cost of preparing samples for NGS sequencing becomes the largest component of the overall cost of collecting sequence data. Thus, it is desirable to increase the number of low-cost library preparation methods available. As the cost of library construction is reduced, projects requiring fewer DNA sequences per sample become effective to conduct using NGS (e.g., if sample preparation plus sequencing for NGS is less than sample preparation plus sequencing on capillary machines, then it is economical to switch).

Materials & Methods

Methodological approach

Illumina libraries require four sequences (i.e., P5 + Read1 sequencing primer [hereafter called Read1] and P7 + Read2 sequencing primer [hereafter called Read2]; Fig. 1), and can accommodate internal index sequences on each end, (i.e., P5 + i5 index + Read1 and P7 + i7 index + Read2; Fig. 1; Illumina Sequencing Dual-Indexed Libraries on the HiSeq System User Guide; Glenn et al., 2019). The Read1 and Read2 sequences can be of two types—TruSeq or Nextera. As in Adapterama I (Glenn et al., 2019), we have designed systems for both.

Our overall approach is to make amplicons with indexes and/or fusions (Fig. 2) that can use iTru or iNext primers described in Adapterama I (Glenn et al., 2019) to make full-length Illumina libraries (Fig. 3A; Figs. S1 and S2). The resulting libraries always contain dual-indexes in the standard indexing positions and may optionally contain additional internal indexes (Figs. 1, 2 and 3; Table 1; Illumina, 2018b). These indexes are recovered through the four standard separate sequencing reactions generated by Illumina instruments when doing paired-end sequencing (Fig. 3B).

Table 1:

Internal identifying index sequences.

All indexes have an edit distance of ≥ 3. Upper case letters are the indexes; lower case letters add length variation to facilitate sequence diversity at each base position of amplicon pools (see text for details). For Illumina MiSeq and HiSeq models ≤ 2500, adenosine and cytosine are in the red detection channel, whereas guanine and thymine are in the green channel. Indexes and spacers have balanced red (shown here in red) and green (shown here underlined and in blue) representation at each base position within each group of four indexes (i.e., count 1–4, 5–8, 9–12, 13–16, and 17–20).

Index count	Index label	Sequence	Length
1	A	GGTAC	5
2	B	cAACAC	6
3	C	atCGGTT	7
4	D	tcgGTCAA	8
5	E	AAGCG	5
6	F	gCCACA;	6
7	G	ctGGATG	7
8	H	tgaTTGAC	8
9	1	AGGAA	5
10	2	gAGTGG	6
11	3	ccACGTC	7
12	4	ttcTCAGC	8
13	5	CTAGG	5
14	6	tGCTTA	6
15	7	gcGAAGT	7
16	8	aatCCTAT	8
17	9	ATCTG	5
18	10	gAGACT	6
19	11	cgATTCC	7
20	12	tctCAATC	8

DOI: 10.7717/peerj.7786/table-1

Although iTru and iNext primers facilitate quick and low-cost additions of dual-indexed adapters, this still requires a separate PCR reaction (but, see Discussion). Thus, when hundreds of amplicons are to be sequenced, it becomes economical to use additional internal indexes (Table 1) so that amplicons can be pooled prior to the use of iTru or iNext primers (Figs. 1 and 2). This approach should work with a wide variety of primers (e.g., Table 2). Such combinatorial indexing is designed to work in 96-well plate arrays but can be modified for other systems. Typically, eight indexed fusion forward primers (A–H) and 12 indexed fusion reverse primers (1–12) are designed and synthesized (File S1). Then, each DNA sample in each well of the 96-well plate can be amplified with a different forward and reverse primer combination (File S1, PCR_Set_up). These PCR products can be pooled and amplified using a similar combinatorial scheme with indexed universal iTru/iNext primers in the second PCR (Table 3), enabling the large-scale multiplexing of samples in one Illumina run (Table 4). Finally, because Illumina MiSeq platforms have documented issues in the quality of Read 2, particularly in GC-rich regions (Quail et al., 2012), fusion primers can be designed to swap forward and reverse primers with Read1 and Read2 fusions (e.g., R1Forward + R2Reverse, vs. R1Reverse + R2Forward; “flipped” primers) to account for this issue (Fig. 2). It is also possible to do replicate amplification with both sets of primers (regular and flipped), to significantly increase base diversity in amplicon libraries.

Table 2:

Primer pairs used in the example projects presented.

Project, target locus, forward and reverse primer names and sequences, as well as the sources of the primer sequences are shown.

Project	Target locus	Forward primer	Reverse primer
Kissing Buga	cyt-b	L14816: CCATCCAACATCTCAGCATGATGAAA	H15173: CCCCTCAGAATGATATTTGTCCTCA
Pathogenic Fungib,c	ITS	ITS1-F_KYO2: TAGAGGAAGTAAAAGTCGTAA	ITS2_KY02: TTYRCTRCGTTCTTCATC
		ITS3-KYO2: AHCGATGAAGAACRYAG	ITS4: TCCTCCGCTTATTGATATGC
		ITS1-F_KYO2: TAGAGGAAGTAAAAGTCGTAA	ITS4: TCCTCCGCTTATTGATATGC
Salamander eDNA	12S	Pleth_12S_F: AAAAAAGTCAGGTCAAGG	Pleth_12S_R: GGTGACGGGCGGTGTGTG
Methylationd	p21-TSS	hp21-TSS F: ATAGTGTTGTGTTTTTTTGGAGAGTG	hp21-TSS R: ACAACTACTCACACCTCAACTAAC
	SIE-1	hp21-SIE1 F: TTTTTTGAGTTTTAGTTTTTTTAGTAGTGT	hp21-SIE1 R: AACCAAAATAATTTTTCAATCCC
Bacterial Communitye,f	16S	Bact-0341-b-S-17: CCTACGGGNGGCWGCAG	S-D-Bact-0785-a-A-21: GACTACHVGGGTATCTAATCC
	16S	515F: GTGCCAGCMGCCGCGGTAA	806R: GGACTACHVGGGTWTCTAAT
Wisteriag,h,i	nr824	w898-824F: CATGTTGCATTCAATCTTGG	w898-824R: GCCTCCATACAAGTTAGTTG
	nr997	w843-997F: GAATCAACGCTGAACGTT	w843-997AluR: GGTTCAATTTATTGATGTG
	trnL; trnL/F	WistmLF: AGTTGACGACATTTCCTTAC	WistmLR: GGAGTGAATGGTTTGATCAATG
	nad4	NAD4RSF1: CTACTAGACTACTAGAGGT	NAD4RSRl: GTTTGGCAACAAGCAAACG
	cyt-b	COBRSF1: CATATTGACTTTCTCTCGCC	COBRSR1: GAATAGGATGACTCAGCGTC

DOI: 10.7717/peerj.7786/table-2

Notes:

aParson et al., 2000.

bToju et al., 2012.

cWhite et al., 1990.

dKolli et al., 2019.

eKlindworth et al., 2013.

fCaporaso et al., 2011.

gTrusty et al., 2007a.

hTrusty et al., 2007b.

iTrusty et al., 2008.

Table 3:

General strategies for producing and indexing amplicon libraries for Illumina sequencing.

These examples use iTru primers, but as mentioned in the text, this can be implemented instead with iNext primers. Method 5 is illustrated below, but we are not including any dataset in the present manuscript that has implemented it (see Discussion). Note: this table does not include “flipped” primers.

Method 1	Method 2	Method 3	Method 4	Method 5
Standard primers	Standard primers	Indexed primers	Fusion primers	Indexed fusion primers
↓	↓	↓	↓	↓
PCR	PCR	PCR	PCR	PCR
↓	↓	[Pool]	↓	[Pool]
	Indexed fusion primers
	↓	↓		↓
Y-yoke	PCR	Y-yoke
↓	[Pool]	↓
iTru PCR	↓	iTru PCR	iTru PCR	iTru PCR
	iTru PCR
↓	↓	↓	↓	↓
Completed library	Completed library	Completed library	Completed library	Completed library
−	+	+	–	+	Base diversity in reads
−	+	+	−	+	Poolable to reduce library preparation costs
2	20	20	2	20	Number of primers
192	193	97	192	97	Minimum number of PCRs for 96 samples
−	−	+	−	+	PCR bias varies among samples
Low	Low	Med	Med	High	Optimization difficulty
Low	High	Med	Med	High	Relative primer cost
High	Med	Med	Med	Low	Relative library preparation cost

DOI: 10.7717/peerj.7786/table-3

Table 4:

Detailed information for example projects presented to validate our approach.

Summarized information for all example projects used to demonstrate Taggimatrix. The “Method” column refers to methods in Table 3; the “Pool name” column applies only to projects in which individual samples were pooled prior to the final pooling proportionate to targeted read number; the “Samples in pool” column cites the number of samples (including replicates) pooled together before final pooling for sequencing; the “Target reads” column cites the approximate number of reads per pool (i.e., not per individual sample) we targeted when pooling samples with other libraries. Note that these data were generated on many independent MiSeq runs.

#	Organism	Project Goal	Target Loci	Library type	Method	Pool name	Samples in pool	Target reads	Actual reads	Summary
1	Kissing bug	Diet analysis	cyt-b	iTru	1	N/A	1	100k (<1%)	916k	Identified five vertebrate sources of blood meals.
2	Fungal Community	Fungal identification	Full-ITS1 (standard & “flipped”)	iTru	2	Homokaryon	48	400k (2.7%)	515k	Identified the primary fungal OTU from each culture
						Het. multispore	48	400k (2.7%)	619k
						Het. tissue	47	400k (2.7%)	444k
			Full-ITS2 (standard & “flipped”)	iTru	2	Homokaryon	48	400k (2.7%)	268k
						Het. multispore	48	400k (2.7%)	310k
						Het. tissue	47	400k (2.7%)	257k
			Incomplete-ITS1&ITS2 (standard & “flipped”)	iTru	2	Homokaryon	48	400k (2.7%)	460k
						Het. multispore	48	400k (2.7%)	579k
						Het. tissue	47	400k (2.7%)	514k
3	Salamander	Environmental DNA	12S	iTru	4 & 5	Reference samples	7	10k (<1%)	8k	Detected 6/7 species of salamander expected in community
						eDNA samples	30	12M (48%)	4.4M
4	Human	Methylation	p21-TSS SIE-1	iTru	4	N/A	1	40k (0.3%)	121k	Compared methylation patterns between cell types
5	Peromyscus	Microbiome	16S	iTru	5	Ash Basin	44	1.5M (6%)	3.8M	Detected 90,862 bacterial OTUs
						Pond B	35	1.5M (6%)	2.8M
						Tim’s Branch	48	1.5M (6%)	0.7M
						Upper Three Runs	44	1.5M (6%)	2.9M
6	Wisteria	Population genetics	nr824 nr997 trnL; trnL/F nad4 cyt-b	iNext	5	N/A	1	150k (1.3%)	79k	Demonstrated mixed ancestry and no population structure in an introduced population

DOI: 10.7717/peerj.7786/table-4

Results

We used five methods that take advantage of iTru or iNext indexing primers developed in Adapterama I in six exemplar amplicon sequencing projects. These projects illustrate the range of methodological approaches that can be used to overcome challenges of amplicon library preparation and fulfill most of the characteristics of an ideal amplicon library preparation system.

In all but one project (Table 4, project 1), we designed fusion primers to generate amplicons that can be amplified by iTru5 and iTru7 (or iNext5 and iNext7) primers to create full-length Illumina TruSeq (or Nextera) libraries. The indexed fusion primers utilize 20 (i.e., 8 + 12) internal identifying sequences with an edit distance ≥ 3 (Table 1) to create up to 96 internally dual-indexed amplicon libraries which were used individually or pooled for additional outer indexing by iTru5 and iTru7 (or iNext5 and iNext7) primers. Sequential PCRs that start with internally indexed primers create quadruple-indexed amplicon libraries that achieve our design goals of cost reduction, facilitation of large-scale multiplexing, increased base-diversity for Illumina sequencing, and maximization of efficiency of library preparation.

In our project characterizing the blood meals of kissing bugs (Table 4, project 1), we obtained an average of 116,902 reads for each sample and identified a total of five unique vertebrate species as the source of the blood meals. In our project identifying fungal pathogens in tree tissues (Table 4, project 2), we obtained an average of 436,825 reads per pool (i.e., 96 samples) and characterized the diverse fungal communities found in these samples. In our project characterizing plethodontid salamander communities from environmental DNA samples (Table 4, project 3), we obtained an average of 163,555 reads for each PCR replicate and identified reads matching 6/7 species expected to be present in the streams. In our project comparing basal DNA methylation of p21 (Table 4, project 4), we obtained approximately 10,000 reads per sample and detected differences in methylation of CpG sites between embryonic kidney cells and human proximal tubule cell (Kolli et al., 2019). In our project characterizing bacterial gut microbiomes (Table 4, project 5), we rarified to 15,000 quality-filtered reads per sample and identified an average of 3,847 OTUs per sample. In our project focused on the fine-scale population genetic analysis of Wisteria (Table 4, project 6), we obtained an average of 1,697 reads per sample and discovered little evidence of population structure among samples. Variation in the average number of reads among projects reflects the intentional allocation of reads when pooling with genomic libraries for sequencing; for example, we pooled plates of libraries for the fungal pathogen project in relative quantities intended to generate approximately 4,000 reads per sample. Variation in the number of reads among samples within a given project likely reflects quantification error and variation in input DNA quantity and quality. Full results and associated figures for each project are detailed in File S3.

The costs associated with each method vary significantly, and which approach has the lowest cost depends on the number of samples processed (Fig. 4: note axis scales are not linear; Table 5; File S6). In all cases, we present the costs associated with targeting a single locus; for projects targeting multiple loci, these numbers can be adjusted to estimate the costs of purchasing necessary primers (i.e., locus-specific primers or fusion primers) and for more complicated pooling schemes. Methods 1 and 4 have the lowest buy-in cost, but the cost of library preparations are fixed, rather than decreasing as the number of samples increases. The constant cost per sample is due to the need for individual second round PCRs (e.g., iTru5/7). The other methods allow pooling of samples prior to second round PCR, which reduces costs. Because Method 1, with no use of fusion primers (non-indexed/indexed), has the highest library preparation costs per sample, it quickly becomes the most expensive method, more than doubling the cost of most other methods with as few as 96 samples. Method 4 remains economically reasonable for processing one or two plates of samples but becomes less reasonable as more plates of samples are used. Method 2 is never economically best, but it is sometimes necessary to achieve sufficient amplification to construct the desired libraries. Thus, Method 2 is only viable when the other methods fail. Method 3 has a moderate buy-in cost and the second-lowest cost per sample for large numbers of samples. Also, Method 3 has the lowest cost when ≤11 plates of samples will be processed, though the cost is very similar to Method 5 after ≥2 plates of samples are processed. Method 5 has the second highest buy-in costs, but the lowest costs per sample when large numbers of samples are processed. Method 5 is optimal when >12 plates of samples are processed. Because Methods 3 and 5 are similar in cost after a few plates of samples are processed, other considerations, such as workflow and personnel costs, are likely to drive decisions about the optimal method rather than the costs of reagents.

Discussion

In Adapterama I, we introduced a general approach to reduce the cost of genomic library preparations for Illumina instruments. Here, we made extensive use of the iNext and iTru primers described in Adapterama I and show that these can also be used to facilitate amplicon library construction at reduced cost with increased flexibility. As we did in Adapterama I, we focused mostly on iTru to simplify our presentation of the method, but iNext works identically in most situations.

Although we focused on Illumina, many of these approaches can be extended to other platforms following the design principles described here (e.g., to create amplicons for PacBio, use primers from sheet ITS_10nt_5′ tags in File S1 following Method 3). For platforms that sequence individual molecules (e.g., PacBio and Oxford Nanopore), there is no advantage to variable-length indexes and negligible penalty for longer indexes, but there are significant informatic advantages to equal-length indexes. Thus, for many other platforms, it will be better to use longer indexes of equal length.

In general, TaggiMatrix Method 5 achieves our design goals, in that it: (1) uses the universal Illumina sequencing primers; (2) minimizes costs (<$2 per library when prepping ≥18 plates of 96 samples, Fig. 4, File S6); (3) minimizes time and equipment needed for library preparations; (4) minimizes buy-in costs through the use of a limited number of fusion primers and universal iTru7 and iTru5 primers; (5) eliminates error-prone ligation steps; (6) allows for >thousands of samples to be pooled and run simultaneously; (7) allows users to vary amplicon representation from tiny to large fractions of a sequencing run (up to 91% have been validated for other projects, i.e., 1,827,086/1,998,538); (8) supports creating millions of samples (8 × 12 × 384 × 384 = 14, 155, 776) that can be tracked and multiplexed through quadruple-indexing. Our method is similar to the “nested tagging” approach proposed in Evans et al. (2016), but we use variable-length inner indexes to create base diversity among reads, use the larger number of dual indexes described in Adapterama I (Glenn et al., 2019), and empirically demonstrate results. TaggiMatrix Method 3 shares nearly all of these advantages and per sample costs are minimized for 2–11 plates of samples. The major disadvantage of method 3 is the requirement for ligation of a universal stub onto the amplicon pool, which is similar to starting with sheared DNA when using the methods in Adapterama I. One challenge for these methods that pool indexed samples before a shared iTru PCR is ensuring an even distribution of reads across samples. This may be especially problematic when using DNA sources that differ greatly in origin, quality, or quantity. If samples cannot be reliably normalized prior to library preparation, researchers may choose to either quantify and normalize products before pooling for an iTru PCR or use a method (e.g., Methods 1 or 4) that includes individual iTru PCRs, after which samples could be quantified and pooled proportionately for sequencing.

Similar to other Adapterama applications, TaggiMatrix offers several methods for combinatorial and hierarchical indexing of samples (Table 3), allowing users to optimize various criteria. For example, different indexes can be used at any combination of the four index positions in the TaggiMatrix library (Fig. 3). By using inner indexes in combination, 20 (8 + 12) indexes can be used to identify 96 (8 × 12) samples. By using inner and outer indexes hierarchically, 40 (8 + 12 + 8 + 12) indexes can identify 9,216 (8 × 12 × 8 × 12) samples. By using two sets of iTru5 and iTru7 primers, 36,864 (8 × 12 × [8 + 8] × [12 + 12]) samples can be identified. Varying indexes at all index positions is the most economical way to tag samples, especially as the number of samples increases (Table 5). By combining a single set of 20 (8 + 12) fusion primers with the full set of 384 iTru5 and 384 iTru7 primers from Adapterama I (Glenn et al., 2019), a total of 14,155,776 (8 × 12 × 384 × 384) samples can be multiplexed.

Our methods address the issue of base diversity through the incorporation of indexes with variable-length spacers that allow for diversity at each base position. This strategy is based on independently originating ideas implemented at the Broad Institute, our lab and others, such as the system developed by Fadrosh et al. (2014) where they introduced “heterogeneity spacers” for sequencing amplicons out of phase. Longer spacers (e.g., 0–7 nt) are advantageous over shorter spacers to compensate for longer repeats in the target amplicons. Mononucleotide repeats are particularly problematic in terms of base diversity. Mononucleotide repeats of ≥ five bp will not be addressed by our short spacers (Table 1). Because Illumina reads are of set length, longer spacers decrease the total amount of useful sequence obtained for downstream analyses. Thus, there is a trade-off in how long the heterogeneity spacers should be. Here, we implement a 0–3 nt long heterogeneity spacers, although this could be easily tuned to 0–7 nt for forward primers and 0–11 nt for reverse primers, to accommodate any researcher’s preferences and mononucleotide repeats known to occur in the target sequences.

Our approach does not deal with the limitation of read-length on Illumina platforms. For long amplicons where complete sequencing is desired, it is possible to construct shotgun libraries from the longer amplicons (e.g., using Illumina Nextera XT, Kapa Biosystems Hyper Prep Plus, NEB Ultra II FS or many other commercial kits). The methods used in Adapterama I may be helpful in those cases. Such libraries can take advantage of the reduced costs per read on higher capacity instruments. It is also possible to design internal locus-specific fusion primers that recover the entire desired DNA region through independent PCRs. It is important to note, however, that the recent introduction of the PacBio Sequel II along with sequencing chemistry v.6 makes circular consensus sequencing of long amplicons on PacBio an economically reasonable approach. Thus, use of the longer consistent-length indexes noted above to create amplicon pools for PacBio is likely to be increasingly attractive as their platform continues to improve.

TaggiMatrix provides an easy way to create indexed fusion primers for convenient ordering at any oligo vendor of your choice. However, the current web page and spreadsheets do not perform quality control of the primer sequences generated. Thus, before ordering, it is important to validate the fusion primers to ensure hairpins, dimers and other secondary structures that inhibit PCR are not created. Several programs exist to validate the primers designed and these should be used before ordering. It is also generally recommended that a small number of fusion primers should be obtained and tested prior to investing in large batches of long fusion primers. When deciding on the best method to use (i.e., Methods 1–5), the number of samples, comparability of samples, reagent cost, and time available to optimize the primers should be considered (Fig. 5).

While developing adapters and primers to make multiple libraries that will be pooled and sequenced, it is important to determine if the primers with different indexes have biased amplification characteristics. This can be accomplished by testing all primers via quantitative PCR using a common template pool to ensure that each primer was synthesized, aliquoted, and reconstituted successfully and has similar amplification efficiency. In practice, however, it will not be economical or necessary to conduct such rigorous quality control for many projects. It is important to note that because sequencing reads are so cheap (∼10,000 reads per $1 USD for PE300 reads on a MiSeq), being off by thousands of reads per sample is less expensive than precise quantification, especially when personnel time for such quantification is considered. Thus, it will often be less expensive to subsample reads from overrepresented samples and/or simply redo the small proportion of samples that do not generate a sufficient number of reads.

Another common concern with amplicon libraries involves “tag jumping”—the artificial creation of unintended index combinations through, for example, the formation of chimeras (e.g., Schnell, Bohmann & Gilbert, 2015). The great diversity of indexes from Adapterama I (Glenn et al., 2019) upon which TaggiMatrix is built allows for index redundancy (e.g., using unique dual outer indexes or building libraries with unique combinations of both inner and outer indexes), which can allow for the easy identification and removal of chimeric sequences. In practice, it is easy to leave negatives within plates of samples so that the frequency of tag jumping can at least be measured and reported (analogous to other measures of genotyping error). Finally, an additional concern with amplicon library preparation methods involving PCR is the introduction of bias due to PCR duplicates. Our method can be modified to incorporate 8N indexes similar to how we addressed this issue with RADcap libraries (Hoffberg et al., 2016). It is also possible to use internal N indexes of any length desired as molecular identifiers (i.e., Jabara et al., 2011; Kou et al., 2016). These modifications, in conjunction with long-amplicon sequencing on other platforms, are worthy of further work.

Conclusions

In summary, we demonstrate how several variants of TaggiMatrix solve common challenges for amplicon sequencing on NGS platforms. Our methods can be implemented in projects from a wide array of disciplines such as microbial ecology, molecular systematics, conservation biology, population genetics, and epigenetics, and we encourage others to further develop the tools we provide for solving additional challenges posed by these applications.

Supplemental Information