Droplet barcoding for massively parallel single-molecule deep sequencing

Abstract

The ability to accurately sequence long DNA molecules is important across biology, but existing sequencers are limited in read length and accuracy. Here, we demonstrate a method to leverage short-read sequencing to obtain long and accurate reads. Using droplet microfluidics, we isolate, amplify, fragment and barcode single DNA molecules in aqueous picolitre droplets, allowing the full-length molecules to be sequenced with multi-fold coverage using short-read sequencing. We show that this approach can provide accurate sequences of up to 10 kb, allowing us to identify rare mutations below the detection limit of conventional sequencing and directly link them into haplotypes. This barcoding methodology can be a powerful tool in sequencing heterogeneous populations such as viruses.

ARTICLE

Received 29 Oct 2015 | Accepted 28 Apr 2016 | Published 29 Jun 2016

The ability to accurately sequence long DNA molecules is important across biology, but existing sequencers are limited in read length and accuracy. Here, we demonstrate a method to leverage short-read sequencing to obtain long and accurate reads. Using droplet microuidics, we isolate, amplify, fragment and barcode single DNA molecules in aqueous picolitre droplets, allowing the full-length molecules to be sequenced with multi-fold coverage using short-read sequencing. We show that this approach can provide accurate sequences of up to 10 kb, allowing us to identify rare mutations below the detection limit of conventional sequencing and directly link them into haplotypes. This barcoding methodology can be a powerful tool in sequencing heterogeneous populations such as viruses.

DOI: 10.1038/ncomms11784 OPEN

Droplet barcoding for massively parallel single-molecule deep sequencing

Freeman Lan1,2, John R. Haliburton1,3, Aaron Yuan1,4 & Adam R. Abate1,2,3,w

1 Department of Bioengineering and Therapeutic Sciences, California Institute for Quantitative Biosciences (QB3), University of California, San Francisco, California 94158, USA. 2 UC Berkeley - UCSF Bioengineering Graduate program, University of California, San Francisco, California 94158, USA. 3 Integrative Program in Quantitative Biology (iPQB) Biophysics Graduate program, University of California, San Francisco, California 94158, USA. 4 Department of Electrical Engineering and Computer Sciences (EECS), Computer Science Division (CS), University of California, Berkeley, California 94720, USA. w Present address: Department of Bioengineering and Therapeutic Sciences, University of California, San Francisco, 1700 4th Street, San Francisco, California 94158,

USA. Correspondence and requests for materials should be addressed to A.R.A. (email: mailto:[email protected]

Web End [email protected] ).

NATURE COMMUNICATIONS | 7:11784 | DOI: 10.1038/ncomms11784 | http://www.nature.com/naturecommunications

Web End =www.nature.com/naturecommunications 1

ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/ncomms11784

Next-generation sequencing (NGS) has tremendously impacted biomedical research due to its ability to acquire massive amounts of sequence data1,2. Currently,

the most widely adopted sequencing platform produces billions of short (o250 bp) reads at a low cost of B$50 per billion bases.

However, short NGS reads pose challenges for many applications. For instance, piecing together short reads into long contiguous sequences can be challenging when assembling new genomes, particularly when repetitive sequences are present3,4. When sequencing metagenomes comprising thousands of species, it is often impossible to assemble the short reads into longer sequences that allow discovery of useful information, such as identication of the species to which a sequence belongs, or detection of gene clusters encoding useful molecules or phenotypes57. Furthermore, NGS is error-prone, generating an error in every thousand bases; this is often above the rate of biological variation, and consequently, prevents detection of true variants within the cloud of sequencing error8,9. The ability to obtain massive amounts of long and accurate reads would thus be a major step forward in our ability to characterize genomes accurately, and to study the impact of sequence variation in a variety of systems, such as in rapidly evolving virus populations10, rare polymorphisms in human populations11, and diverse and uncultivable species in microbial communities12.

To obtain longer and more accurate reads, one approach is to directly improve the sequencing instrument13,14. In addition to providing accurate reads, the instrument must be widely available, easy to use and cost-competitive. Currently, no platform can match short-read NGS in these aspects and as such, short-read sequencers dominate the market. Rather than inventing a new sequencing instrument, an alternative is to synthetically reconstruct long reads from short-read data, leveraging the widespread popularity of short-read NGS. An elegant approach is using unique molecular barcodes, which were rst used to detect duplicated NGS reads for error correction, and digital counting of molecules15,16. To reconstruct long reads using molecular barcodes, long template molecules are broken into short fragments and labelled with barcode sequences identifying the template from which they originate1720. All short fragments can then be pooled and sequenced, and fragments of individual templates grouped by barcode. The reads in each group are then used to reconstruct synthetic long reads. Implementations of this approach rely on intramolecular reactions to attach barcodes to the fragments; however, this reaction becomes inefcient for templates above 3 kb. Alternatively, molecules can be physically isolated into wells,

followed by fragmentation and barcoding. This approach can theoretically be extended to molecules of any length, but is limited in the number of templates that can be sequenced due to the limitations in throughput of liquid handling in well plates. Throughput can be increased by barcoding multiple templates in each well, but then single-molecule identity is lost19,20. To enable long and accurate DNA sequencing, an optimal approach would combine physical isolation of molecules with ultrahigh-throughput uid handling.

In this paper, we describe single-molecule droplet barcoding (SMDB), an ultrahigh-throughput method to barcode long molecules for short-read sequencing. Using droplet microuidics, we isolate and barcode single molecules in aqueous droplets B1 million times smaller than conventional well plates. To validate the method, we sequence a library of known DNA templates of 35 kb long and reconstruct long reads fully covering the templates. Furthermore, to demonstrate the ability to sequence large DNA molecules, we apply the method to theE. coli genome, obtaining synthetic read-lengths up to 10 kb in length. Finally, to illustrate the power of the method for detecting variants below the detection limit of conventional sequencing, we apply it to a library of b-glucosidase genes mutated by PCR. While SMDB detects 457 SNPs in 81 haplotypes in the library, conventional short-read sequencing detects only one SNP and cannot generate haplotypes. The ability to characterize variants and haplotypes below the inherent detection limit of the sequencer should be powerful for studying systems in which rare variants have an important role, such as in microbial community dynamics and viral quasispecies.

ResultsOverview of the method. Droplet microuidics has recently been used to barcode the transcriptomes of single cells2123. In SMDB, we use it to barcode fragments of single DNA molecules, performing all steps of template amplication, fragmentation and barcoding in a microuidic workow (Fig. 1). DNA barcodes uniquely tag all reads derived from a template, which allows the reads to be unambiguously clustered to generate a long and accurate consensus sequence for the template.

Droplet microuidic workow for single-molecule barcoding. We leverage ultrahigh-throughput droplet microuidics to amplify, fragment and barcode large numbers of individual DNA templates. The rst step is to isolate and amplify the template molecules, accomplished by introducing them into a microuidic

Amplify, fragment and barcode templates Clustering barcoded reads

Reconstruct template

DNA templates

Template amplification

De novo assembly

Variant calling

Bar

Bar

Bar

Bar

Single template encapsulation

Fragmentation, barcoding

Sequencing

Barcode-based read clustering

Figure 1 | Schematic overview of SMDB. DNA templates are encapsulated into droplets such that most droplets contain zero or one template. Templates are clonally amplied to produce multiple copies in each droplet. Templates are fragmented inside drops, and barcodes are attached to fragments such that each droplet gets a unique barcode sequence. All fragments are sequenced in parallel and resulting reads are clustered based on barcode. Clustered reads are used to reconstruct the sequence or accurately detect SNPs for the template encapsulated in each droplet.

2 NATURE COMMUNICATIONS | 7:11784 | DOI: 10.1038/ncomms11784 | http://www.nature.com/naturecommunications

Web End =www.nature.com/naturecommunications

NATURE COMMUNICATIONS | DOI: 10.1038/ncomms11784 ARTICLE

ow focus droplet generator that encapsulates them in B50 mm diameter droplets of PCR reagent (Fig. 2a). The template concentration is controlled so that B1 in 10 droplets contains a single molecule, in accordance with Poisson statistics24. The droplets are collected into a PCR tube and thermal cycled for

amplication, generating within each droplet a clonal population of the single molecules so that, once fragmented and barcoded, we can obtain multi-fold coverage of each template.

Following amplication, the templates must be fragmented to a length compatible with short-read sequencing. Importantly,

a

Encapsulation and amplification

Fragmentation

DNA

Outlet

PCR

OR MDA

PCR/

MDA

Oil

Thermocycling

b

Nextera transposase

Transposase

Merger

Electrode

Templates

Outlet

Split waste

(1)

(3)

ConstA ConstB

Fragment

(2)

c

Barcoding

Merger

Electrode

PCR mix

Frag. DNA Barcodes

Oil

Splitter

ConstB

Fragment

ConstA

Barcode

(3) (4) (5)

Fragment

Barcode

(1) (2)

Bar

Thermocycling

Figure 2 | Microuidic workow for generating barcoded DNA fragments. Left: schematics and false-colored microscope images of microuidic devices. Right: schematic of molecular biology reactions occurring inside droplets. (a) A ow focus drop maker is used to encapsulate single templates into droplets. Inside droplets, PCR or MDA is used to clonally amplify the single template. (b) The splitmerger device is used to add transposases into template drops while achieving a 10 dilution of the templates. The template droplets are injected on the left side, split at junction (1) so that 1/10th of the droplet

continues on to pair with a reagent droplet generated on chip at (2) and the pair merges at the channel widening (3). The transposase reaction inside droplets fragments templates while adding adaptors to each fragment. (c) The microuidic device used for attaching barcodes to DNA fragments. Templates droplets (1) and barcode droplets (2) are injected into the device where they pair with each other and a large PCR reagent droplet generated on chip (3). The three droplets merge at the electrode (4) and are split into smaller droplets for thermal cycling (5). Barcodes are spliced onto fragments by overlap-extension PCR. Scale bars, 100 mm.

NATURE COMMUNICATIONS | 7:11784 | DOI: 10.1038/ncomms11784 | http://www.nature.com/naturecommunications

Web End =www.nature.com/naturecommunications 3

ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/ncomms11784

fragmentation must be performed while maintaining compartmentalization, to prevent pieces of different templates from mixing before barcodes have been attached. To fragment in the droplets, we use a microuidic device to add Tn5 transposase into each droplet, which randomly fragments and attaches short sequences to the amplied templates25 (Fig. 2b). Because transposases are single-turnover enzymes, an optimal stoichiometric ratio of transposase to templates must be maintained with a 10-fold dilution of the template droplet into the fragmentation droplet. To address this need, we develop a module combining droplet splitting and merging (Fig. 2b and Supplementary Fig. 1). The incoming droplets pass through a junction sampling B1/10th of their volume, which is then merged with a new droplet approximately equal to the size of the original droplet. This device accomplishes the necessary tasks of diluting the starting droplet and adding the new reagent, while maintaining the droplet size constant throughout the process. After the transposase is added, the droplets are collected into a syringe and incubated in a water bath at 55 C for the transposase reaction.

After the templates have been fragmented, the barcodes used to tag fragments belonging to the same template are attached by overlap-extension PCR in the droplets (Fig. 2c). In this reaction, barcode sequences attach to the fragments through regions of sequence homology on the adaptor sequences added by the transposase. This step thus requires merging three droplets: template, barcode and PCR reagent. We design a triple merger device for merging three droplets at once. Improving on the designs of conventional mergers26, we concatenate multiple merging junctions, which act independently to achieve robust merging of all three droplets (Fig. 2c and Supplementary Fig. 2). The volumes and reagent concentrations of the droplets are controlled to ensure correct stoichiometry for PCR barcoding. In addition, the channels enable one of each type of droplet to combine in the electro-coalescence junction, shown to the right in Fig. 2c. The resultant droplets are 90 mm spherical diameter and can coalesce during thermal cycling (see Supplementary Note 1 for details on coalescing droplets). To make them more robust, we split the merged droplets into four portions using a splitter27. The split droplets are collected into PCR tubes and thermally cycled to attach the barcodes. Even with the small size, B1050%

of droplets coalesce (Supplementary Fig. 3a), which is undesirable since it can lead to multiple templates or barcodes in a single droplet, and hence improper barcoding. We therefore remove these droplets using a combination of gravity-induced and pinched-ow fractionation28 (Supplementary Fig. 3b and Supplementary Methods). The remaining droplets are chemically ruptured and the DNA contents are puried over a spin column, then size selected to remove free barcodes, resulting in a sequence-ready library.

Generation of barcode droplets. Uniquely barcoding millions of DNA templates requires tens of millions of barcode droplets, each containing a clonal population of one barcode sequence. To generate these barcode droplets, we individually encapsulate and amplify random barcode molecules using the same technique shown in Fig. 2a (also see Supplementary Fig. 4a). Barcode molecules consisting of random N-mers anked by constant sequences are chemically synthesized and encapsulated with PCR reagents for amplication. The molecules are loaded at a limiting dilution of B1 in 10 droplets. The droplets are thermally cycled, generating within each loaded droplet a clonal population of amplied product; these droplets can then be merged with the template droplets for the barcoding step shown in Fig. 2c. Using this approach, we generate B10 million barcode droplets

in o1 h for B$10 of PCR reagent, which is sufcient to barcode B1 million templates in the SMDB workow.

Because barcode sequences are random, it is possible for two barcodes of the same sequence to label different templates. In in silico simulations, we nd that the likelihood of this undesirable event is extremely low for barcodes of sufcient length (Supplementary Fig. 4b). During PCR amplication and sequencing of the barcodes, errors and mutations generate a cloud of related sequences around the original barcode sequence. By sequencing our barcode library, we nd that the original barcode sequences are on average three Hamming distances from their nearest neighbour, while the sequences within the cloud of mutated barcodes around each original barcode are, on average, only 1 Hamming distance from their nearest neighbour (Supplementary Fig. 4c). However, the mutated barcodes typically comprise o5% of all reads and do not represent a signicant source of inefciency. To address this issue, we develop an algorithm to cluster mutated barcodes and their parent sequences into a single barcode cluster (Supplementary Note 2). These barcode clusters represent all fragments that originate from the same template, and thus, are used for template analysis, SNP identication and reassembly.

Validation of single-molecule barcoding. A key property of SMDB is its ability to barcode single molecules, which greatly simplies bioinformatic analysis since all reads in a given cluster are known to originate from only one template. To validate that SMDB indeed barcodes single molecules, we apply it to a library of eight templates from 3 to 5 kb long (for details on known template library, see Supplementary Methods). Because only one-tenth of barcode droplets contain barcodes, we expect only one-tenth of encapsulated templates to be barcoded. Starting with B1 M template droplets encapsulated at one in ten droplets containing templates, we expect a theoretical yield of B10,000 barcoded templates. Practically, the yield of sequenced templates would be lower due to the sample losses incurred during the start-up of microuidic devices and during the removal of coalesced droplets. Sequencing the library, we obtain B10 million reads using a MiSeq 2 250 run, yielding 3,563 clusters, which represents B35% of

theoretical yield. For perfect barcoding of single molecules, all reads in all clusters should map to only one template. Aligning reads from each cluster to the eight reference sequences, we calculate for each barcode cluster the fraction of reads mapping to the dominant template, dened as the single (out of eight possible) template to which the majority of reads in a cluster map (Fig. 3a). We nd that 490% of clusters contain 490% reads mapping to the dominant template. Nevertheless, we observe a low background of o2% of reads mapping to the non-dominant template in less than half of the barcode clusters, which we attribute to mis-tagging, a phenomenon often observed in barcoded sequence libraries prepared in well plates, and thought to originate from chimeric PCR products generated during library amplication and sequencing29. Since many barcode clusters contain some degree of non-dominant template reads, we dene clusters containing 490% dominant template as single-template clusters. The overwhelming majority (B90%) of clusters are single-template clusters (Fig. 3a, inset). Instances of multiple templates in the same barcode cluster are infrequent, and consistent with the rate of co-encapsulation expected by Poisson statistics (see Supplementary Note 3 for details). Multiple-encapsulations can be reduced by lowering template concentration, which reduces the instances of multiple templates in the same barcode clusters at the expense of barcoding throughput.

The ideal sequencing data provides full-length, high-accuracy coverage of all templates in the sample. However, bias in

4 NATURE COMMUNICATIONS | 7:11784 | DOI: 10.1038/ncomms11784 | http://www.nature.com/naturecommunications

Web End =www.nature.com/naturecommunications

NATURE COMMUNICATIONS | DOI: 10.1038/ncomms11784 ARTICLE

100

80

60

40

20

a b

c

0.7

1.00.80.60.40.2

0.6

Per cent reads as dominant

template

0.5

Fraction of clusters

1.0

0.60.40.2 0

0.8

1 2 3 4 5 6

Templates in cluster

Normalized coverage

Local GC content

Coverage entropy

0.3

9

8

7

6

5

0.4

0

0.7

1.00.80.60.40.2

0.6

0.5

0.4

0

0.3

0

0.25

0.50

0.75

1.0

0 0 1.0 2.0 3.0 4.0

100 101 102 103 104 105 106

Cumulative fraction of clusters Number of reads in cluster

Base position (kb)

5.0

Figure 3 | Mapping reads from barcode clusters to the known template references. (a) Cumulative distribution of barcode clusters based on per cent of reads that map to the dominant template. The majority of clusters contain reads mostly from a single template. Inset: the number of templates with 410%

mapping reads in each barcode cluster is counted and plotted as a histogram. (b) Aggregate coverage of two randomly chosen templates for all barcode clusters with corresponding local GC content (dashed line). See Supplementary Fig. 5 for the corresponding plots for all templates. (c) All barcode clusters plotted based on coverage entropy and number of reads in each barcode cluster. Each point represents one barcode cluster. Insets: coverage distribution for the individual barcode clusters denoted in corresponding colour on the main plot; Y axis: normalized coverage, X axis: base position.

sequencing can yield excessive coverage in certain regions and insufcient coverage in others. To investigate whether our approach is susceptible to such bias, we plot the coverage distribution for each template (Fig. 3b and Supplementary Fig. 5). We observe systematic coverage bias for all templates, much of which correlates with local GC content, and hence, is likely the result of the PCR amplication of the libraries for sequencing30. We also observe decreased coverage at the ends of templates, a known bias of transposase fragmentation25. Thus, the primary forms of bias in our data are the same as those observed in standard NGS, and result from the same sources.

To quantify how bias affects coverage, we dene the coverage entropy as the informational entropy of the coverage distribution for each barcode cluster (see Supplementary Note 4 for discussion on coverage entropy). Clusters with high-coverage entropy exhibit at distributions with uniform coverage, while the clusters with low-coverage entropy exhibit peaky distributions with non-uniform coverage. Consequently, coverage entropy is a good predictor of whether a cluster contains sufcient information to reassemble a template, and is thus an overall good metric for coverage uniformity (Supplementary Fig. 6a). Plotting the coverage entropy of each barcode cluster against the number of reads contained within it, we observe two populations, one in which entropy saturates rapidly with coverage (upper left) and another in which entropy rises more slowly (Fig. 3c). The clusters where entropy rises slowly with number of reads are more biased, and therefore require more sequencing to obtain the requisite information for assembly. On the basis of our results, an entropy 47 is required for successful assembly (Supplementary

Fig. 6a). This corresponds 4100 reads in the barcode cluster (Fig. 3c). Therefore, one measure for the efcient utilization of sequencing reads is the number of barcode clusters with 4100 reads obtained for a xed amount of total sequencing reads used (Supplementary Fig. 7). While more sequencing produces more viable barcode clusters, exhaustively sequencing the library results in inefcient utilization of reads.

SMDB detects rare SNPs and captures haplotypes. An important application of NGS is to detect rare single-nucleotide polymorphisms (SNPs) in heterogeneous populations, such as viruses, cells or human beings8,10,31,32. Characterizing that SNPs are physically linked on the same template, called haplotyping, is important for understanding how multiple variants at distant loci can contribute to a given phenotype. However, performing these

tasks with conventional NGS is often extremely challenging or impossible due to the inability of the short reads to span multiple SNPs. Moreover, standard NGS is error-prone, generating one error in every B1,000 bases; this prevents condent detection of rare variants without accepting a large proportion of false-positives8,9,33. To enhance sensitivity, known patterns of error production can be modelled and used to correct data, providing modest improvements8. Molecular techniques can greatly increase sensitivity to detect rare SNPs but reduce read length even further34.

SMDB is able to condently detect rare SNPs because each molecule is sequenced to great depth, allowing reads to be averaged together to obtain an accurate consensus for every base. To demonstrate this, we generate a population of DNA templates via 35 cycle PCR of a bacterial plasmid extracted from a culture grown from a single colony. In this population, every sequence shares signicant homology, but rare variants exist. Variants like these can have important biological consequences, such as allowing HIV to evolve drug resistance or the development of rare alleles that increase risk for disease in human populations11,33. We sequence the population using SMDB on a MiSeq 2 150 run, obtaining 4.6 million reads in B6,000

barcode clusters. Because each barcode cluster represents fragments amplied from a single molecule, we expect a fraction of the fragmentsand therefore readsto contain amplication errors. In the worst case scenario where an error is made in the rst round of amplication, we expect B50% of the reads to be erroneous for any one position in the sequence. Since these cases are reported as di-allelic SNPs by the SNP-caller, we keep only the mono-allelic SNP calls to ensure the highest accuracy of our mutation calls. We identify 457 high-condence SNPs in B10% of templates, whereas B90% of the templates contain no SNPs compared to the reference (Fig. 4a and Supplementary Fig. 8). With the exception of SNP C1067G existing in B5.5% of templates, all others are present in o0.1%

of the templates, far below the limit of detection for standard NGS. To compare our results to standard SNP calling methods, which do not use barcode information, we call SNPs while disregarding the barcode grouping of reads and detect only the C1067G variant. Hence, SMBD amplies the sensitivity of sequencing and allows capture of biological information invisible to standard methods. Unlike conventional NGS, the limit of detection of SMDB scales with the number of molecules sequenced and can be easily orders of magnitude more sensitive than conventional NGS (Fig. 4b).

NATURE COMMUNICATIONS | 7:11784 | DOI: 10.1038/ncomms11784 | http://www.nature.com/naturecommunications

Web End =www.nature.com/naturecommunications 5

ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/ncomms11784

a c

b

103

55

*

Conventional NGS

Frequency of variant

102

103

104

105

[afii9826] = 0.01

[afii9826] = 0.05

[afii9826] = 0.10

[afii9826] = 0.15

[afii9826] = 0.20

6

*

SNP frequency

C1067G

4

2

0 0 1.6

0.8

0.2 0.6

0.4 1.4

1.2

1.0

0 8

6

4

104

2

10

Base position (kb)

Molecules sequenced

Figure 4 | Calling SNPs and haplotypes of single templates from barcode clusters. (a) Frequency of SNPs detected at each position in the template. *Indicates the C1067G SNP that can also be detected without barcodes. (b) Limit of detection for SMDB for a given number of molecules sequenced.

b denotes the expected type II error (probability of not detecting the variant). Dashed line represents limit of detection for conventional NGS29.(c) Phylogenetic tree constructed using consensus sequences generated by the SNP calls. The C1067G mutation and all its derivatives are highlighted in blue. Each node represents a new mutant.

In addition to detecting rare SNPs, SMDB naturally generates haplotypes, which are important for characterizing mutations that have synergistic effects and are broadly relevant from virus evolution to human genetics35,36. SMDB provides haplotyping information because SNPs that occur on the same template are grouped into the same barcode cluster, allowing haplotypes to be condently identied for each template. To demonstrate SMDB haplotyping, we plot the haplotypes determined by SMDB in a phylogenetic tree, allowing us to determine the order of mutations that occurred during replication (Fig. 4c). The mutations in the population are generated by replication, and thus, in the absence of selection, ones that occur early in replication exist in a large subset of the progeny. The phylogenetic tree shows that C1067G was the rst mutation that arose in the population, consistent with the fact that C1067G mutation is the most abundant SNP.

SMDB facilitates de novo assembly. De novo assembly, the process of piecing together short reads into long contigs, is necessary to extract useful information from short reads when a reference sequence is not available, such as when sequencing new genomes or metagenomes37,38. Despite years of improvement, de novo assemblers continue to struggle with datasets comprising multiple sets of highly homologous sequences18,37,38. In some cases, de novo assembly is practically impossible because the information needed to uniquely generate a contig spans a length beyond the accessible read length of short-read sequencing. SMDB simplies de novo assembly by ensuring that all reads in a cluster originate from one template, allowing unambiguous assembly of a contig that was previously impossible when all reads from all templates must be considered concurrently.

To demonstrate de novo assembly with SMDB, we sequence a test library of known templates 35 kb long with a MiSeq 2 250,

obtaining B9 million reads clustering into 2,043 groups. We perform de novo assembly on each barcode cluster independently, yielding 245 contigs 42 kb long. The contigs span a range of lengths, and a signicant portion of the assembled contigs cover the full length of the templates (Fig. 5a). To account for low-read coverage at the ends of the templates due to biased transposase insertion, we trim the rst and last 250 bp of the contigs. The resultant sequences are accurate when compared to the known reference sequences, having an overall error rate of 4.3 10 4

per base and no detectable structural variations or chimeras. If the errors in the contigs are artifacts of assembly or sequencing, we expect them to be negatively correlated with the coverage entropy

of the barcode groups used to assemble them. However, we nd contig accuracy is independent of coverage entropy, and rather, depends slightly on position in the contig (Fig. 5b and inset). This is reminiscent of the pattern of SNPs seen in the previous experiment (Fig. 4a), indicating that these are likely rare SNPs rather than errors in the assembled contigs.

Theoretically, any DNA template can be barcoded by SMDB if it can be encapsulated and amplied. However, PCR amplication becomes inefcient for templates longer than 5 kb. To sequence molecules longer than this, we implement multiple displacement amplication (MDA), a non-specic, isothermal method that can amplify whole genomes39. We generate fragments of the E. coli genome 710 kb in length and sequence the resulting library on a MiSeq 2 300 run from which we

obtain B13 million reads clustering into B1,000 groups after quality ltering. As expected, de novo assembly with barcodes yields signicantly longer and more accurate contigs than assembly without barcodes (Fig. 5c and inset). Interestingly, B26% of these contigs do not map to the E. coli genome, but to other bacterial genomes in the NCBI refseq database, and thus represent contaminating DNA in the library rather than sequencing errors (Supplementary Fig. 6b,c). Thus, SMDB enables sequencing of long templates with arbitrary sequence, but care must be taken to limit contamination.

DiscussionA challenge when performing molecular biology reactions in droplets is that, often, multiple reagents must be added to the droplets at different times. Since reagent addition always increases the size of the droplets, adding multiple reagents can produce nal droplets that are too large to be robustly handled. To perform reagent addition while maintaining droplets at a reasonable size, we have developed a splitmerge device that combines droplet splitting with droplet merger26,40. This device has the unique and valuable property of producing nal droplets that are equal in size to the initial droplets; hence, this same device can be used to perform multiple additions on an emulsion while maintaining constant droplet size. The degree of dilution can be adjusted by varying the amount sampled from the split droplet, which is adjusted by controlling the ow rate of the splitting outlet. This obviates the need to construct a unique device with increasing dimensions for each round of reagent addition, and maintains the droplets in the size range that is optimized for handling and incubation. The splitmerge device should be valuable when multiple reagent additions must be

6 NATURE COMMUNICATIONS | 7:11784 | DOI: 10.1038/ncomms11784 | http://www.nature.com/naturecommunications

Web End =www.nature.com/naturecommunications

NATURE COMMUNICATIONS | DOI: 10.1038/ncomms11784 ARTICLE

# Contigs

a 35

25

15

5

15

5

35

25

2.0 3.0 4.0 5.0

0.4

0.6

0.8

1.0

Length of contigs (kb)

Fraction of full-length

b

100.0

99.5

Per cent identity

Accuracy (Phred)

45

99.0

25

98.5

98.0

0 5.0

7.0 7.5 8.0 8.5

Base position (kb)

Entropy of cluster

c

1.4

1.0

103

Unbarcoded Barcoded

1.2

50

40

30

0.8

0.4

Frequency

Accuracy (Phred)

0.6

0.2

20

10

0 2.0 4.0 6.0 8.0

Base position (kb)

Length of contig (kb)

0 2.0 14.0

4.0 6.0 8.0 10.0 12.0

Figure 5 | De novo assembly of single templates from barcode clusters. (a) Distribution of assembled contigs by length (left panel) and fraction of the template covered (right panel). (b) Per cent sequence match of each assembled contig to the reference is plotted against the entropy of the barcode cluster that produced the contig. Inset: the accuracy of assemblies for each base position, on a Phred scale, binned by every 50 bp or until the rst mismatch to reference if no mismatches are found within 50 bp.(c) Distribution of read-lengths of contigs obtained from SMDB of 710 kb fragments of the E. coli genome. Inset: per-base accuracy of the contigs on a Phred scale.

performed on an emulsiona task that has thus far been a signicant challenge for droplet microuidic workows.

The random Poisson encapsulation of templates and barcodes is a source of inefciency in SMDB, but one that is overcome by leveraging the ultrahigh-throughput nature of droplet

microuidics. To ensure that most templates are paired with a single barcode, barcodes and templates are loaded at B1 in 10 per droplet, yielding a single pairing event for B1 in 100 droplets. Even with this inefciency, the throughput of our device enables barcoding of B3,500 molecules in B15 min. Assuming a modest template length of 5 kb, this is sufcient to cover an E. coli genome at B5 coverage. With higher-throughput droplet

generation and manipulation, such as emulsication under jetting conditions41 and parallelization of channel networks42,43, it should be possible to increase throughput by an order of magnitude. In addition, the template and barcode emulsions can be sorted to discard empty droplets, which should increase efciency B10-fold by ensuring that every pairing event comprises one of each component with no wasted droplets.

Encapsulation of templates into small volumes reduces amplication bias during PCR but also limits the amount of DNA generated for each barcoded template. Therefore, the number of starting templates is directly correlated with the amount of DNA obtained at the end of the workow. We have empirically determined that 410,000 productive droplets are required to provide the minimum B20 nanomoles for sequencing after accounting for sample loss through the workow. Although it is possible to additionally PCR amplify lower yield libraries, this results in more bias, yielding uneven coverage of templates, and uneven distribution of reads into barcode groups.

Droplet microuidic workows have been successfully adapted into non-microuidic labs through collaboration with labs with microuidic expertise21,22. For labs interested in adopting SMDB, we suggest collaborating with a droplet microuidics lab, because although the fabrication and operation of the microuidic devices is straight forward, the handling of droplets outside of devices is quite nuanced. Dolomite, a company dedicated to providing offthe-shelf and custom designed droplet microuidic devices for research, is also an excellent resource for implementing droplet microuidics workows into the lab.

New technologies for sequencing DNA while retaining long-range information are becoming available20,44. While these technologies share some similarity to ours, there are critical differences that make each approach better or worse for different applications. For example, recent methods that encapsulate many template molecules in each droplet provide very high throughput and are an inexpensive solution for barcoding large amounts of DNA, but the resulting sequence data cannot be deconvoluted back to single molecules since within each barcode cluster (droplet) many templates of different sequences exist. This may be acceptable for applications in which the templates are highly dissimilar or in which single-molecule resolution is not required, but in others it may prove problematic. In particular, for samples in which the molecules share signicant homology but small sequence differences are biologically relevant, such as when studying viral diversity and evolution, these technologies are ineffective and the SMDB approach is better suited. A similar technology specically targeted to sequence human genomes is available and therefore applications of SMDB to human genome sequencing are not investigated45.

We have applied SMDB to the barcoding of single DNA molecules from virus and microbial genomes, but the principle of encapsulating and barcoding nucleic acids in microuidic droplets is broadly applicable. For example, droplet microuidics has been used to encapsulate, lyse, and amplify single viruses and cells46,47. The SMDB workow we describe here could be combined with these methods to barcode the genomes of these organisms, to perform whole-genome single virus or cell sequencing. This could make the barcoding workow valuable for characterizing genetic reassortment in seasonal inuenza. Indeed, while barcoding up to B10,000 single entities is immediately

NATURE COMMUNICATIONS | 7:11784 | DOI: 10.1038/ncomms11784 | http://www.nature.com/naturecommunications

Web End =www.nature.com/naturecommunications 7

ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/ncomms11784

practical with the methods we describe, if single cells rather than long templates were to be barcoded, the number of individual genomes that can be sequenced is limited by the sequencing throughput of NGS. Even with the massive capacity available with present-day instruments, it is not enough to fully leverage the throughput of our droplet method. However, as sequencing instruments continue to decrease in cost and increase in throughput, sequencing large barcoded populations of cells and viruses should become practical, impacting applications in which genetic diversity is important, such as in microbial communities.

Methods

Microuidic devices. Photoresist masters are created by spinning on a layer of photoresist SU-8 3025 (Microchem) onto a 3 inch silicon wafer (University Wafer) at 3,000 rpm, then baking at 95 C for 5 min. Then, the photoresist is subjected to 3 min ultravoilet exposure over photolithography masks (CAD/Art Services) printed at 12,000 DPI. After ultravoilet exposure, the wafers are baked at 95 C for 10 min then developed for 10 min in fresh propylene glycol monomethyl ether acetate (Sigma Aldrich) then rinsed with fresh propylene glycol monomethyl ether acetate and baked at 95 C for 5 min to remove solvent. To fabricate the triple merger device, a second layer of photoresist was patterned on top of the rst layer after the rst ultravoilet exposure to generate a two-layered master. The micro-uidic devices are fabricated by curing poly(dimethylsiloxane) (10.5:1 polymer-tocrosslinker ratio) over the photoresist master48. The devices are cured in an 80 C oven for 1 h, extracted with a scalpel, and inlet ports added using a 0.75 mm biopsy core (World Precision Instruments, catalogue no. 504529). The device is bonded to a glass slide using O2 plasma treatment and channels are treated with Aquapel (PPG Industries) to render them hydrophobic. Finally, the devices are baked at 80 C for 10 min to dry the Aquapel before they are ready for use.

Barcode emulsion. Chemically synthesized barcode oligonucleotides (GCAGCTGGCGTAATAGCGAGTACAATCTGCTCTGATGCCGCATAGNNN NNNNNNNNNNNNTAAGCCAGCCCCGACACT) (IDT) are added at 0.01 pM concentration into a PCR reaction mix containing 1 NEB Hotstart Phusion

polymerase (NEB, catalogue no. M0536L), 2% w/v Tween 20, 2% w/v PEG 6000, 400 nM forward and reverse primers (FL128 CTGTCTCTTATACACATC TCCGAGCCCACGAGACGTGTCGGGGCTGGCTTA) (FL129 CAAGCAGA AGACGGCATACGAGATCAGCTGGCGTAATAGCG). The reaction mixture and HFE 7500 uorinated oil (3 M) with 2% (w/w) PEG-PFPE amphiphilic block copolymer surfactant (Ran Biotechnologies) are loaded into separate 1 ml syringes and injected at 300 and 500 ml h 1, respectively, into a ow-focusing droplet maker using syringe pumps (New Era, catalogue no. NE-501) controlled with a custom Python script (https://github.com/AbateLab/Pump-Control-Program

Web End =https://github.com/AbateLab/Pump-Control-Program). After collecting the emulsion in PCR tubes, the oil underneath the emulsion is removed using a pipette and replaced with FC-40 uorinated oil (Sigma Aldrich, catalogue no. 51142-49-5) with 5% (w/w) PEG-PFPE amphiphilic block copolymer surfactant for improved thermal stability (see Supplementary Note 1 for details on thermostability). The emulsion is transferred to a T100 thermocycler (BioRad) and thermally cycled with the following program: 98 C for 3 min, followed by 40 cycles with 2 C per second ramp rates of 98 C for 10 s, 62 C for 20 s and 72 C for 20 s, followed by a nal hold at 12 C. SYBR staining using10 SYBR GREEN I in HFE 7500 oil is used to quantify encapsulation rate under

a uorescent microscope.

Generating template droplets. For SMDB using PCR, DNA template molecules are encapsulated and amplied in the same manner as described above, except the primers used are FL178 (CCACTACGCCTCCGCTTTC) and FL179 (CCATC TCATCCCTGCGTGT), and input DNA is a library of long molecules with universal adaptors on either side. Input DNA concentration is adjusted until one in ten droplets are uorescent under SYBR staining. To construct the library of seven known templates, eight DNA templates are amplied from 5 ng of phage lambda genomic DNA (NEB: N3013S) using 500 nM of primer sets (see oligonucleotides listed in Supplementary Table 1) using 1 NEB phusion hotstart ex mix (NEB:

M0536S) with the following cycling conditions: 98 C 3 min, 35 cycles of: 98 C 15 s, 62 C 30 s, 72 C 3 min, followed by 72 C 5 min and optional holding at 12 C overnight. The PCR products are gel-extracted using 1% agarose gel and Zymo gel extraction kit. To attach constant sequence adaptors to all the fragments, 100 ng of gel-extracted amplicons are added to an adaptor ligation mix of: 1 mM adaptors,0.20 mM dNTPs, 0.5 ml (60 units) of Bst 2.0 polymerase warmstart (NEB: M0538M), 2.5 ml T4 DNA ligase from the quick ligation kit (NEB: M2200S),1 ligase buffer from the quick ligase kit. The reaction is incubated at 25 C for

15 min then 65 C for 10 min for heat inactivation, then DNA is puried using the Zymo DNA concentrator kit. The concentration of resulting DNA is quantied using the bioanalyzer high sensitivity kit and pooled together at equal molar concentration to generate the eight templates library.

For SMDB using MDA, reactions are performed using REPLI-g single cell kit (Qiagen, catalogue no. 150343). E. coli genomic fragments are from E. coli K12(DH10B) cells purchased from New England BioLabs (catalogue no. C3019H), lysed and puried using PureLink Genomic DNA Mini Kit (Life Technologies, catalogue no. K1820-00). Ten kilobase fragments are gel-extracted following a 10-min digestion with NEBNext dsDNA Fragmentase (NEB, catalogue no. M0348S) of 800 ng DNA and quantied using a NanoDrop (Thermo Scientic). The fragmented input DNA is incubated with 3 ml Buffer D2 and 3 ml H2O for 10 min at 65 C. After stopping by adding 3 ml stop solution, a master mix comprising nuclease-free H2O, REPLI-g reaction buffer, and REPLI-g DNA polymerase is added. The MDA reactions are then emulsied in the manner described above and incubated at 30 C for 3 h then 70 C for 20 min for heat inactivation.

Fragmentation of templates in droplets. Droplets containing amplied templates, a Nextera Transposase reaction mixture composed of 1 TD buffer,

2% w/v Tween 20, 2% w/v PEG 6000, and 1/10 volume of TDE from Nextera Kit (Illumina, catalogue no. FC-121-1031, or puried in lab as described49), deionized water, HFE 7500 with 2% w/v EA surfactant and 2 M NaCl are loaded into 1 ml syringes (BD scientic) and connected to the split-merge microuidic device (Supplementary Fig. 1). The electrode is connected by clipping the output of a cold cathode uorescent inverter connected to a DC power supply (Mastech) to the needle of the electrode syringe using an alligator clip. Setting a voltage of2.0 V at the power supply results in a B2 kV AC at the electrode, which causes droplets close to the electrode to merge. The resulting emulsion is collected in a 1 ml syringe and incubated at 55 C for 10 min and then 70 C for 20 min in large water baths.

Barcoding of fragmented templates. Fragmented template droplets, barcode droplets and a PCR mixture composed of 1 Invitrogen Platinum Multiplex mix

(ThermoFisher, catalogue no. 4464268), 400 nM Primers FL127 (AATGATAC GGCGACCACCGAGATCTACACTCGTCGGCAGCGTC) and FL129 (CAAGC AGAAGACGGCATACGAGATCAGCTGGCGTAATAGCG), 1 in 50 dilution of the NT buffer from the Nextera XT Kit (0.2% SDS) (Illumina, catalogue no. FC-131-1024), 1% Tween 20 w/v, 1% PEG 6000 w/v, 2.5 U ml 1 Bst Polymerase 2.0

Warmstart (NEB catalogue no. M0538S) are loaded into a syringe and injected into the double merger device as shown in Supplementary Fig. 2. The emulsion is collected in a 0.5 ml thin-walled PCR tube, and the oil is replaced with FC-40 with 5% w/v EA surfactant before thermal cycling at: 65 C for 5 mins, 95 C for 2 mins, then 25 cycles at 2 C/s ramp rates of 95 C for 15 s, 60 C for 1 min, 72 C for1 min, and then 72 C for 5 min followed by optional 12 C hold overnight. After thermal cycling, the oil is replaced with HFE 7500 with 2% w/v EA surfactant, then loaded into a syringe injected into a pinched-ow fractionation device to remove large droplets as shown in Supplementary Fig. 3b. After removal oflarge droplets, the emulsion is broken by adding 20 ml of 1H,1H,2H,2H-Peruoro-1-octanol (Sigma Aldrich, catalogue no. 370533) and brief centrifugation in a micro-centrifuge. The aqueous top phase is collected and DNA is puried using a Zymo DNA concentrator kit.

Final library amplication. Overall, 2 ng of the barcoded library is added to a PCR mixture containing 1 Phusion master mix and 400 nM of primer FL127/129 and

thermal cycled as follows: 98 C 3 min, and 10 cycles of 98 C 10 s, 62 C 20 s, 72 C 1 min, 72 C 5 min. The resulting DNA is loaded into a Blue Pippin (Sage Biosciences) 100600 bp cassette to extract DNA from 300700 bp range to remove free untagged barcodes. The resultant DNA is concentrated using a Zymo DNA concentrator kit, quantied using the Bioanalyzer high sensitivity DNA chip (Agilent) and sequenced on the MiSeq using a custom index primer (FL166).

Bioinformatics analysis. Barcodes are clustered using a python program dfs clustering (Supplementary Note 2), which uses raw Miseq fastq les and outputs barcode clusters and the IDs of their associated reads along with quality control metrics. Barcode clusters containing o500 reads are removed because they contain too few reads for analysis. For SNP calling, reads from each barcode cluster are mapped onto the template sequence using Bowtie 2 (ref. 50)very-sensitive-local settings and outputted as a SAM le then converted into a BAM le using Samtools. To call SNPs, Samtools v1.2 mpileup is used with options-d 9999-u-V-I and Bcftools call is used with options -c -v to lter only positions that contain SNPs. The SNPs are then ltered for homozygous calls as there should be only one template per barcode cluster. The phylogenetic tree was constructed using consensus sequences generated by the replacing each position of the reference with the SNP called for each barcode cluster. Duplicate sequences are removed, and then the list of non-redundant consensus sequences are used to generate a phylogenetic tree using the maximum likelihood method in Phylip v3.696. For de novo assembly, reads from each barcode cluster are written into one le in fasta format with bases lower than Q20 replaced with N. Each fasta le is fed into the IDBA-UD assembler v1.1.1 (ref. 51) with parametersmink 20maxk 120step 20min_contig 2000min_count 1max_mismatch 3.

8 NATURE COMMUNICATIONS | 7:11784 | DOI: 10.1038/ncomms11784 | http://www.nature.com/naturecommunications

Web End =www.nature.com/naturecommunications

NATURE COMMUNICATIONS | DOI: 10.1038/ncomms11784 ARTICLE

Data availability. Sequencing data generated from SMDB of the seven templates control and E. coli genomic fragments libraries are available at the Sequence Read Archive (SRA) under accession code SRP072529.

References

1. Metzker, M. L. Sequencing technologiesthe next generation. Nat. Rev. Genet. 11, 3146 (2010).

2. Shendure, J. & Ji, H. Next-generation DNA sequencing. Nat. Biotechnol. 26, 11351145 (2008).

3. Li, R. et al. De novo assembly of human genomes with massively parallel short read sequencing. Genome Res. 20, 265272 (2010).

4. Nagarajan, N. & Pop, M. Sequence assembly demystied. Nat. Rev. Genet. 14, 157167 (2013).

5. Saeed, I., Tang, S.-L. & Halgamuge, S. K. Unsupervised discovery of microbial population structure within metagenomes using nucleotide base composition. Nucleic Acids Res. 40, e34 (2011).

6. Wommack, K. E., Bhavsar, J. & Ravel, J. Metagenomics: read length matters. Appl. Environ. Microbiol. 74, 14531463 (2008).

7. Wooley, J. C. & Ye, Y. Metagenomics: facts and artifacts, and computational challenges. J. Comput. Sci. Technol. 25, 7181 (2009).

8. Bansal, V. A statistical method for the detection of variants from next-generation resequencing of DNA pools. Bioinformatics 26, i318i324 (2010).

9. Nielsen, R., Paul, J. S., Albrechtsen, A. & Song, Y. S. Genotype and SNP calling from next-generation sequencing data. Nat. Rev. Genet. 12, 443451 (2011).

10. Acevedo, A., Brodsky, L. & Andino, R. Mutational and tness landscapes of an RNA virus revealed through population sequencing. Nature 505, 686690 (2014).

11. Tennessen, J. A. et al. Evolution and functional impact of rare coding variation from deep sequencing of human exomes. Science 337, 6469 (2012).12. Scholz, M. B., Lo, C.-C. & Chain, P. S. Next generation sequencing and bioinformatic bottlenecks: the current state of metagenomic data analysis. Curr. Opin. Biotechnol. 23, 915 (2012).

13. Chin, C.-S. et al. Nonhybrid, nished microbial genome assemblies from long-read SMRT sequencing data. Nat. Methods 10, 563569 (2013).

14. Laszlo, A. H. et al. Decoding long nanopore sequencing reads of natural DNA. Nat. Biotechnol. 32, 829833 (2014).

15. Kinde, I., Wu, J., Papadopoulos, N., Kinzler, K. W. & Vogelstein, B. Detection and quantication of rare mutations with massively parallel sequencing. Proc. Natl Acad. Sci. USA 108, 95309535 (2011).

16. Casbon, J. A., Osborne, R. J., Brenner, S. & Lichtenstein, C. P. A method for counting PCR template molecules with application to next-generation sequencing. Nucleic Acids Res. 39, e81 (2011).

17. Lundin, S. et al. Hierarchical molecular tagging to resolve long continuous sequences by massively parallel sequencing. Sci. Rep. 3, 1186 (2013).

18. Hiatt, J. B., Patwardhan, R. P., Turner, E. H., Lee, C. & Shendure, J. Parallel, tag-directed assembly of locally derived short sequence reads. Nat. Methods 7, 119122 (2010).

19. Kuleshov, V. et al. Whole-genome haplotyping using long reads and statistical methods. Nat. Biotechnol. 32, 261266 (2014).

20. Amini, S. et al. Haplotype-resolved whole-genome sequencing by contiguity-preserving transposition and combinatorial indexing. Nat. Genet. 46, 13431349 (2014).

21. Macosko, E. Z. et al. Highly parallel genome-wide expression proling of individual cells using nanoliter droplets. Cell 161, 12021214 (2015).

22. Klein, A. M. et al. Droplet barcoding for single-cell transcriptomics applied to embryonic stem cells. Cell 161, 11871201 (2015).

23. Rotem, A. et al. High-throughput single-cell labeling (Hi-SCL) for RNA-seq using drop-based microuidics. PLoS ONE 10, e0116328 (2015).

24. Hindson, B. J. et al. High-throughput droplet digital PCR system for absolute quantitation of DNA copy number. Anal. Chem. 83, 86048610 (2011).25. Adey, A. et al. Rapid, low-input, low-bias construction of shotgun fragment libraries by high-density in vitro transposition. Genome Biol. 11, R119 (2010).

26. Jin, B.-J., Kim, Y. W., Lee, Y. & Yoo, J. Y. Droplet merging in a straight microchannel using droplet size or viscosity difference. J. Micromech. Microeng. 20, 035003 (2010).

27. Abate, A. R. & Weitz, D. A. Faster multiple emulsication with drop splitting. Lab. Chip 11, 19111915 (2011).

28. Yamada, M., Nakashima, M. & Seki, M. Pinched ow fractionation: continuous size separation of particles utilizing a laminar ow prole in a pinched microchannel. Anal. Chem. 76, 54655471 (2004).

29. Carlsen, T. et al. Dont make a mista(g)ke: is tag switching an overlooked source of error in amplicon pyrosequencing studies? Fungal Ecol. 5, 747749 (2012).

30. Aird, D. et al. Analyzing and minimizing PCR amplication bias in Illumina sequencing libraries. Genome Biol. 12, R18 (2011).

31. Meyerson, M., Gabriel, S. & Getz, G. Advances in understanding cancer genomes through second-generation sequencing. Nat. Rev. Genet. 11, 685696 (2010).

32. Out, A. A. et al. Deep sequencing to reveal new variants in pooled DNA samples. Hum. Mutat. 30, 17031712 (2009).

33. Simen, B. B. et al. Low-abundance drug-resistant viral variants in chronically HIV-infected, antiretroviral treatment-naive patients signicantly impact treatment outcomes. J. Infect. Dis. 199, 693701 (2009).

34. Lou, D. I. et al. High-throughput DNA sequencing errors are reduced by orders of magnitude using circle sequencing. Proc. Natl Acad. Sci. USA 110, 1987219877 (2013).

35. Sabeti, P. C. et al. Detecting recent positive selection in the human genome from haplotype structure. Nature 419, 832837 (2002).

36. Giallonardo, F. D. et al. Full-length haplotype reconstruction to infer the structure of heterogeneous virus populations. Nucleic Acids Res. 42, e115e115 (2014).

37. Baker, M. De novo genome assembly: what every biologist should know. Nat. Methods 9, 333337 (2012).

38. Wences, A. H. & Schatz, M. C. Metassembler: merging and optimizing de novo genome assemblies. Genome Biol. 16, 207 (2015).

39. Dean, F. B. et al. Comprehensive human genome amplication using multiple displacement amplication. Proc. Natl Acad. Sci. USA 99, 52615266 (2002).

40. Tan, Y.-C., Collins, J. & Lee, A. P. in Transducers, Solid-State Sensors, Actuators and Microsystems, 12th International Conference on 2003 Vol. 1, 2831 (Boston, Massachusetts, 2003).

41. Abate, A. R. & Weitz, D. A. Air-bubble-triggered drop formation in microuidics. Lab. Chip 11, 1713 (2011).

42. Nisisako, T., Torii, T., Takahashi, T. & Takizawa, Y. Synthesis of monodisperse bicolored janus particles with electrical anisotropy using a microuidic co-ow system. Adv. Mater. 18, 11521156 (2006).

43. Romanowsky, M. B., Abate, A. R., Rotem, A., Holtze, C. & Weitz, D. A. High throughput production of single core double emulsions in a parallelized microuidic device. Lab. Chip 12, 802 (2012).

44. Eisenstein, M. Startups use short-read data to expand long-read sequencing market. Nat. Biotechnol. 33, 433435 (2015).

45. Zheng, G. X. Y. et al. Haplotyping germline and cancer genomes with high-throughput linked-read sequencing. Nat. Biotechnol. 34, 303311 (2016).

46. Tao, Y. et al. Rapid, targeted and culture-free viral infectivity assay in drop-based microuidics. Lab. Chip 15, 39343940 (2015).

47. Tao, Y. et al. Artifact-free quantication and sequencing of rare recombinant viruses using drop-based microuidics. Chembiochem. Eur. J. Chem. Biol. 16, 21672171 (2015).

48. Duffy, D. C., McDonald, J. C., Schueller, O. J. A. & Whitesides, G. M. Rapid prototyping of microuidic systems in poly(dimethylsiloxane). Anal. Chem. 70, 49744984 (1998).

49. Picelli, S. et al. Tn5 transposase and tagmentation procedures for massively scaled sequencing projects. Genome Res. 24, 20332040 (2014).

50. Langmead, B. & Salzberg, S. L. Fast gapped-read alignment with Bowtie 2. Nat. Methods 9, 357359 (2012).

51. Peng, Y., Leung, H. C. M., Yiu, S. M. & Chin, F. Y. L. IDBA-UD: a de novo assembler for single-cell and metagenomic sequencing data with highly uneven depth. Bioinformatics 28, 14201428 (2012).

Acknowledgements

We thank R. Hernandez and R. Andino for helpful scientic discussions. We thank Eric Chow and the Center for Advanced Technologies at UCSF for technical expertise with sequencing. We thank B. Demaree, S. Poust and C.Q. Lan for helpful comments on the manuscript. This work was supported by the National Science Foundation through a CAREER Award (Grant Number DBI-1253293); the National Institutes of Health (NIH) (Grant Numbers HG007233-01, R01-EB019453-01 and DP2-AR068129-01); and the Defense Advanced Research Projects Agency Living Foundries Program (Contract Numbers HR0011-12-C-0065, N66001-12-C-4211 and HR0011-12-C-0066). Funding for open access charge: (NIH grant number DP2-AR068129-01).

Author contributions

F.L. and A.R.A. proposed the concept and prepared the manuscript. J.H. contributed to the conceptualization and design of the droplet barcodes. F.L. performed the experiments and analysis of data. A.Y. designed and implemented the barcode clustering algorithm.

Additional information

Accession codes: Sequencing data generated from SMDB of the 7 templates control andE. coli genomic fragments libraries are available at the Sequence Read Archive (SRA) under accession code SRP072529.

Supplementary Information accompanies this paper at http://www.nature.com/naturecommunications

Web End =http://www.nature.com/ http://www.nature.com/naturecommunications

Web End =naturecommunications

Competing nancial interests: The authors declare no competing nancial interests.

NATURE COMMUNICATIONS | 7:11784 | DOI: 10.1038/ncomms11784 | http://www.nature.com/naturecommunications

Web End =www.nature.com/naturecommunications 9

ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/ncomms11784

Reprints and permission information is available online at http://npg.nature.com/reprintsandpermissions/

Web End =http://npg.nature.com/ http://npg.nature.com/reprintsandpermissions/

Web End =reprintsandpermissions/

How to cite this article: Lan, F. et al. Droplet barcoding for massively parallel single-molecule deep sequencing. Nat. Commun. 7:11784 doi: 10.1038/ncomms11784 (2016).

This work is licensed under a Creative Commons Attribution 4.0 International License. The images or other third party material in this article are included in the articles Creative Commons license, unless indicated otherwise in the credit line; if the material is not included under the Creative Commons license, users will need to obtain permission from the license holder to reproduce the material. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/

Web End =http://creativecommons.org/licenses/by/4.0/

10 NATURE COMMUNICATIONS | 7:11784 | DOI: 10.1038/ncomms11784 | http://www.nature.com/naturecommunications

Web End =www.nature.com/naturecommunications