Single Molecule and Single Cell Epigenomics

Abstract

Dynamically regulated changes in chromatin states are vital for normal development and can produce disease when they go awry. Accordingly, much effort has been devoted to characterizing these states under normal and pathological conditions. Chromatin immunoprecipitation followed by sequencing (ChIP-seq) is the most widely used method to characterize where in the genome transcription factors, modified histones, modified nucleotides and chromatin binding proteins are found; bisulfite sequencing (BS-seq) and its variants are commonly used to characterize the locations of DNA modifications. Though very powerful, these methods are not without limitations. Notably, they are best at characterizing one chromatin feature at a time, yet chromatin features arise and function in combination. Investigators commonly superimpose separate ChIP-seq or BS-seq datasets, and then infer where chromatin features are found together. While these inferences might be correct, they can be misleading when the chromatin source has distinct cell types, or when a given cell type exhibits any cell to cell variation in chromatin state. These ambiguities can be eliminated by robust methods that directly characterize the existence and genomic locations of combinations of chromatin features in very small inputs of cells or ideally, single cells. Here we review single molecule epigenomic methods under development to overcome these limitations, the technical challenges associated with single molecule methods and their potential application to single cells.

1. Epigenetics and Epigenomics

The term “epigenetics” was originally coined by Conrad Waddington [1], and can be defined as the heritable changes in traits that occur without direct alteration of the underlying primary DNA sequence. At the molecular level, epigenetic phenomena are regulated by so-called epigenetic marks that include covalent changes to DNA and chromatin proteins, and remodeling of nucleosome position on the DNA. These are often controlled by non-coding RNAs and can respond to environmental influences. The epigenome refers to the constellation of epigenetic marks that exist in a cell and their chromosomal locations, which varies by cell type and developmental stage. Normal epigenomic states are important for cell-type specific gene expression and a variety of cellular and developmental processes. In humans, the aberrant placement of epigenetic marks has been linked to many diseases, including cancer (reviewed in [2]). The best studied DNA modification in mammals is 5-methylcytosine (5mC), which plays an important role in genomic imprinting, and the suppression of transposable elements [3]. Recently, it has been shown that 5mC can be converted to 5-hydroxymethylcytosine (5hmC) by the TET (Ten-Eleven-Translocation) family of proteins [4]. In addition to converting 5mC to 5hmC, TET proteins can catalyze further hydroxylation of 5hmC to 5-formylcytosine (5fC) and 5-carboxylcytosine (5caC) in an enzymatic activity-dependent manner [5, 6]. Numerous histone modifications exist primarily on the protruding amino-terminal tails of histones [7]; these include methylation (mono-, di-, trimethylation), acylation, phosphorylation, ubiquitylation, sumoylation, ADP ribosylation, carboxylation, glutathionylation, O-GlcNAcylation, hydroxylation, formylation and citrullination. The covalent modifications are placed by “writer” activities (reviewed in [8–11]), removed by “eraser” activities (reviewed in [12–14]). They exert their cellular effects by altering chromatin structure and by recruiting “reader” factors (reviewed in [15, 16]), which in turn recruit additional factors necessary for epigenetically regulated processes.

2. Epigenomic methods

Characterization of epigenomes has been greatly enhanced by use of various high-throughput sequencing technologies, which have been used to construct detailed epigenomic landscapes in many systems. The currently used, highest resolution methods include bisulfite sequencing (BS-seq) to assess 5mC [17–20], variants of BS-seq to assess 5hmC [21, 22], and chromatin immunoprecipitation followed by sequencing (ChIP-seq) for the analysis of histone modifications and chromatin binding proteins [23–33]. Additional methods have been developed to assess epigenetically influenced chromatin states diagnostic for transcriptionally regulatory domains including DNAse-seq [34], FAIRE (Formaldehyde-Assisted Isolation of Regulatory Elements) [35] and Sono-seq [36]. Among the technical challenges associated with these methods are biased amplification of bisulfite treated DNAs [37, 38] and the problem of antibody quality for ChIP studies. Antibody specificity can limit ChIP-seq accuracy and reproducibility. Standards have been established to validate antibody quality [32, 39] that include blotting approaches. Semi-synthetic DNA-barcoded nucleosome libraries may also prove useful [40] for antibody validation, and emerging microfluidic-based platforms [41] and the Nanostring nCounter platform [42] have the potential for streamlining this process. Synthetic affinity reagents [43] may ultimately replace antibodies that have batch-to-batch variation. Two issues that are more fundamental than these technical challenges are (1) that the cited epigenomic methods generally provide information about a single epigenetic feature at a time, rather than the combinations of epigenomic features that exist in the genome; and (2) that they work best with an abundance of materials, and not as well with inputs from few or even single cells. The remainder of the discussion focuses on why these are limitations, and approaches under development to overcome them.

3. Combinatorial Chromatin States

In describing the histone code hypothesis, Strahl and Allis suggested “that multiple histone modifications, acting in a combinatorial or sequential fashion on one or multiple histone tails, specify unique downstream functions” [44]. The strongest support for this at the time was provided by observations that the N-terminal tails of histones H3 and H4 contain multiple sites where covalent modifications may be placed. Since then, many lines of evidence lent support for the histone code hypothesis, and in particular, that combinations of chromatin modifications, and not just single modifications alone, are critical for genomic regulation. First, mass spectrometric analysis revealed that individual histones do in fact carry multiple modifications on their N-terminal tails [45, 46]. Second, plant homeodomain (PHD) and bromodomain (BRD) motifs within so-called “reader” proteins that bind to specific modified histone amino acids, bind their targets in a manner that is sensitive to the modification states of nearby amino acids [47]. The histone acetyltransferase p300 in particular was shown to display this property [40]. Third, the reader protein Bromodomain PHD-finger Transcription Factor (BPTF), which has a methyl lysine-binding PHD motif adjacent to an acetyl lysine-binding BRD motif, was shown to bind H3K4me2/3 containing nucleosomes preferentially when the nucleosome also carried H4K16ac [48]. The observation that modifications on two independent histone molecules together affected BPTF binding highlights the fact that biologically relevant combinatorial states are not limited to those in cis on a single histone molecule. Fourth, combinations of histone modifications influence the biochemical activities of factors that bind and further modify histones. For example, the demethylase KDM7A that targets methylated forms of H3K9 and H3K27 for demethylation [49] contains a PHD motif that binds H3K4me3, suggesting that KDM7A is directed to its H3K9me and H3K27me targets in chromatin by adjacent H3K4me3 [50].

The histone code hypothesis can be extended to include effects coordinated with DNA modifications, as the combined importance of DNA and histone modifications to gene expression has been documented. The NuRD complex contains methyl binding domain (MBD) proteins, which bind 5mC and 5hmC, histone deacetylases (HDAC) and chromatin remodeling activity [51]. Gene silencing by HDAC activity in these complexes is enabled by MBD recruitment of the complex to modified DNA [52]. Given the cross talk among chromatin modifications, it should come as no surprise that their effects are coordinated by mechanisms that sense the modifications in combination. As the number of known reader proteins [53] and chromatin modifications [6, 54] increases, so does the potential complexity of the histone code, or more broadly, the chromatin code. These trends elevate the importance of identifying and mapping the genomic locations of combinations of chromatin features in order to understand how those features regulate genomic information in normal and disease states.

4. Technologies that overcome some limitations of ChIP-seq and BS-seq

4.1. Re-ChIP and ChIP-BS-seq

The most widely used ChIP protocols query chromatin sources for chromatin features one at a time. Several sets of efforts have characterized where in the genome combinations of chromatin features can be found. One of these used sequential- or re-ChIP experiments, whereby chromatin immunoprecipitated with a first antibody was subjected to re-precipitation with a second antibody before analyzing the DNA [55–62]. In one application of re-ChIP, a bivalent state comprising H3K4me3 and H3K27me3 modifications at genes important for lineage specification was found in pluripotent stem cells [63]; in another application, histone variants H3.3 and H2AZ were found together on active promoters, enhancers and insulator regions [64]. Re-ChIP methods require large inputs of chromatin, given the inefficiencies with which each antibody precipitates the chromatin, and in some cases, the low abundance of the chromatin feature. There are few examples of whole genome re-ChIP studies. Studies with more than two sequential ChIP reactions will likely require antibodies or other affinity reagents with dissociation constants well below those of existing reagents in order to have high enough modification capture efficiencies; small reaction volumes that enable use of high concentrations of chromatin and capture reagents; and improvements in library preparation or sequencing methods that make most efficient use of the DNA isolated by ChIP.

In other efforts to define coincidence between 5mC and H3K27me3, DNA isolated by anti H3K27me3 ChIP was subjected to bisulfite sequencing [65]. In principle, this strategy can be applied to any DNAs isolated from a single or re-ChIP experiment, if sufficient amounts of DNA are recovered. Single molecule sequencing technologies have enabled the identification of DNA modifications, including 5mC and 5hmC without bisulfite and other chemistries (see below [66–68]). These provide an alternate means of characterizing the sequences and DNA modification states of chromatin isolated by ChIP.

4.2. Statistical methods

With high quality, high coverage sequencing results in hand, the workflow for ChIP-seq data analysis consists of read mapping and peak calling, including peak modeling and identification (reviewed in [69]). Statistical strategies have been combined with single ChIP experiments to identify sites where combinations of chromatin features are likely to be found [42, 70–74]. These include use of hierarchical or k-means cluster analyses, heat map analysis, and Venn and Euler diagram analyses, (reviewed in [69]), however, direct observations using methods specifically designed to report combinations of chromatin features will be needed to validate these approaches. A distinct set of statistical considerations exists for single cell and single molecule analyses. These are discussed separately below.

4.3. Mass spectrometry

Mass spectrometry has been responsible for the dramatic expansion of the histone modifications known (reviewed in [75]). When applied to histones that were not previously fractionated, this can reliably identify modifications that arise together in pairwise and higher order combinations [45, 46]. This is true if the modifications are close enough on the peptide chain to remain on informative peptide fragments (reviewed in [76]). When amino acids of interest are localized on different histones, or far apart on a single histone, it is difficult or impossible to identify combinations of modifications. However, if nucleosomes are previously isolated by ChIP, quantitative mass spectrometry can reveal modifications on histones that associate with the modification recognized by the antibody [77]. A question still unaddressed by mass spectrometry is where on the DNA combinations of histone modifications reside, and if the histone modifications are associated with DNA modifications. In principle, it should be possible to perform ChIP and reserve the DNA for sequencing or bisulfite sequencing, and subject the isolated histones to mass spectrometry, but it is unlikely that this will enable unambiguous characterization of the range of histone modifications at a given locus.

4.4. Single molecule methods

Methods that interrogate individual molecules for a range of chromatin features may provide a means for determining which chromatin features are coincident, mutually exclusive, the frequency with which combinations actually appear in the genome and their location. Ensemble methods based on averaged measurements of typical single feature ChIP experiments cannot provide this information. Several distinct single molecule methods have been developed (reviewed in [78–82]), and can be applied to chromatin isolated from a population of cells, or from single cells (discussed below). Several of these methods rely on fluorescent antibodies to detect epigenetic marks. As is true with ChIP, data quality will depend on antibody binding specificity and affinity [39].

4.4.1. DNA curtains, ordered chromatin arrays, nanochannel squeezing and ChIP-string

DNA curtains are prepared by tethering the biotinylated ends of DNA to streptavidin localized in lipids anchored to a solid substrate in a flow cell (reviewed in [83]). When an aqueous medium is flowed through the cell, DNA fibers elongate (reviewed in [84]). These methods have been used to study variables influencing nucleosome deposition [85] and the effects of nucleosomes on DNA scanning by repair proteins [86]. Future embodiments of DNA curtains might make it possible to characterize where modifications exist on naturally occurring chromatin using fluorescent probes recognizing those modifications. This approach could identify patterns of coincidence and mutual exclusion of modifications on single chromatin fragments; however, further enhancements still would be required to reveal what DNA sequences are affected.

Ordered arrays of chromatin fragments stretched across devices fabricated with polydimethylsiloxane (PDMS) provide an alternative to molecular combing [87]. In this procedure the PDMS device contains micron scale pillars protruding from a surface, and spaced at micron scale intervals. Chromatin is placed between the device and a glass coverslip. Individual molecules bind to the pillars and are elongated across them by drawing the coverslip across the device. Once bound, histone H3 was detected using fluorescent antibodies and fluorescence microscopy. It should be possible to extend this method to report patterns of coincidence and mutual exclusion of chromatin features. Methods have not been developed to identify the underlying DNA sequences on ordered arrays.

DNA curtains and ordered arrays entail tethering chromatin in some manner prior to observation. In a complementary method, chromatin can be flowed through an elastic channel, which can be narrowed to a cross section of approximately 200nm × 200nm, thus confining linear fragments. Confinement can also be achieved in narrow rigid channels [88, 89]. When HeLa cell chromatin confined in elastic channels was stained with fluorescent antibodies detecting histone modifications, patterns of H4ac and H3K9me3 were observed [90]. Consistent with expectations, these modifications, that respectively report active and silent chromatin, were not coincident. This method was also use to characterize two histone modifications simultaneously in chromatin reconstituted from natural histones on T7 phage DNA [91]. Additionally, DNA methylation in experimentally methylated lambda DNA was detected using a fluorescently tagged MBD portion of MeCP2 [89].

The resolution of ordered arrays (at least 320 nm [87]) and elongation through confinement (6–10 kbp [88, 89]) correspond to as few as 25 nucleosomes, if chromatin was decompacted from the 30 nm fiber [92]. Further development of these methods is needed to determine the DNA sequences associated with fluorescent signals reporting chromatin features.

ChIP-string applied DNA isolated by ChIP to the NanoString nCounter platform [42]. This single molecule counting method reported the presence of chromatin features at 487 loci through use of fluorescent bar coded probes capable of uniquely identifying the loci. As originally developed, chromatin features were assayed individually. It is not clear if the method is sensitive enough for use with re-ChIP studies for dual mark assessment.

4.4.2. SMRT and nanopore sequencing

Recently developed and emerging single molecule sequencing technologies have led to the observation that modified bases can be detected during sequencing. These methods rely on two essential features of the sequencing platforms: First, sequenced DNA templates are those naturally existing in the cells of interest and not generated by amplification methods; and second, during sequencing, modified nucleotides can exhibit measurably distinct behaviors from unmodified nucleotides, thereby reporting their modification state. With the Pacific Biosciences SMRT platform, sequence data are determined by detecting incorporation of spectrally distinct fluorescent nucleotides during DNA synthesis guided by the naturally occurring template. The latency of incorporation of nucleotides is influenced by adjacent 5mC [66] and 5hmC [67] residues. With nanopore based methods, MBD1 protein bound to 5mC on a DNA template alters the current flowing through the sequencing nanopore [68]. MBD1 or similar probes might not be necessary to distinguish cytosines that are unmodified, methylated and hydroxymethylated on nanopore platforms [93–95]. In principle, pending further advancements with nanopore sequencing [96, 97], either of these single molecule sequencing approaches can enable detection of coincident chromatin modifications, if DNA isolated by ChIP is applied to the platforms.

4.4.3. SCAN

Single chromatin molecule analysis at the nanoscale (SCAN), is analogous in principle to flow cytometry (Figure 1). Flow cytometry entails binding cells to antibodies that recognize cell surface antigens and carry distinct fluorophores, then measuring the fluorescent properties of single cells as they flow through an inspection or focal volume. The results report the abundance of cells that carry different combinations of cell surface antigens. Similarly, chromatin fragments can be bound to fluorescent antibodies or other molecular probes that recognize different chromatin modifications, molecules can be driven by a voltage gradient through a fluidic channel, and defined combinations of chromatin modifications can be detected and their abundances measured [98, 99]. In this embodiment, SCAN is an analytical tool, capable of enumerating the abundance of chromatin modifications or their combinations. Further modifications enable SCAN to be used as a preparative tool to isolate chromatin molecules that carry those different states. This preparative mode is analogous to fluorescence activated cell sorting (FACS) and can provide molecules defined chromatin modifications for applications such as sequencing [100].

Figure 1. Experimental platform for single molecule epigenomics analyses.

A, SCAN device. Nanoscale channels were fabricated on 10 cm silica wafers and bonded to a second top wafer containing holes to access the channels (top). Cylinders attached over holes allow channels to be filled with chromatin or DNA solutions and fitted with electrodes to drive molecules through the channels electrokinetically. Red box outlined in top image contains 16 nanofluidic channels (middle), and wafers contain 27 such elements. Micrograph of one channel showing the 500 nm wide by 250 nm deep constriction where fluorescence detection occurred (bottom). B, Schematic of a wafer mounted on a confocal fluorescence microscope. Overlapping lasers (blue and orange cones) illuminate a 1.3 μm length of the channel forming a 160 aL (160×10⁻¹⁸ L) inspection volume where single molecule detection takes place. Chromatin is shown in a channel with black, yellow, red and blue circles depicting different epigenetic features. Antibodies recognizing those epigenetic features and carrying distinct fluorophores are bound to chromatin. After laser excitation, fluorescent emissions collected with a confocal aperture are detected with avalanche photodiodes (APDs) capable of single photon detection. Devices with this architecture enabled analyses of epigenomic marks individually and combinations [98, 99].

4.4.4. Computational and statistical considerations for single molecule methods

There are at least three major computational and statistical challenges with single molecule methods. These include distinguishing signal from noise for each detection event; managing the large data sets that come from each experiment; difficulties performing data analysis in real time, which is preferable but less practical than performing analysis after data collection.

With the exception of nanopore-based approaches, each of the single molecule methods involves identification of fluorescence bursts from many molecules, and then extracting the characteristics from the data integrated over each burst. This is challenging, especially for platforms that assess molecular properties that change dynamically (SMRT sequencing and SCAN) vs. platforms that collect data from immobilized molecules whose properties do not change over the course of the experiment (DNA curtains, ordered chromatin arrays, nanochannel squeezing and ChIP-string; however, some embodiments of these methods could require dynamic analyses). The challenges arise because the event rate of particle detection is low and fluorescence background contributes significantly to the measured signal. Burst identification has been traditionally achieved by filtering and thresholding the detected photon flux based on the signal time-trace data through the use of statistical analysis and an iterative optimization of an intensity threshold [101–103]. However, development of more accurate statistical tools and faster data processing techniques are among the most important efforts in single molecule detection.

Fluorescent bursts are collected in one of two ways. With single-point detectors (photodiodes or photomultiplier tubes), the raw data from a photo-detector consists of digital pulses that can be time-tagged, binned or counted, processed to detect bursts and then different related quantities extracted. With two-dimensional detectors (CCD and CMOS arrays), data collection and analysis includes acquiring tens of thousands of images of dynamic samples, localizing and identifying individual molecules, and extracting the position of individual molecules. The computational cost of these tasks grows linearly with the number of detectors (or pixels in the arrays) and can easily exceed the capabilities of a typical desktop computer. For example, when a 512 × 512 pixel CCD is used for single molecule detection in high-throughput and multicolor imaging applications, one image generates about 0.5 Megabytes of data collected every 1 ms. Each hour of analysis entails nearly 30 Gigabytes of data.

Until now, single molecule experiments rely on processing data for analysis after collection instead of in real time, which can prove to be limiting. To overcome this limit will require development of new algorithms for molecule identification and tracking, and the use of high performance hardware such as graphics processor units and field programmable gate arrays. The latter has been incorporated into SCAN for single molecule sorting [100].

5. Analyzing Low Cell Inputs

Because single molecule methods provide a high level of sensitivity for chromatin analyses, they also hold promise for enabling analyses of very low cell inputs, including single cells. Such analyses are important for several kinds of studies, three examples of which are outlined below. First, embryonic development, from the single cell zygote to the morula and blastocyst stage when differentiation first begins, is known to involve dramatic changes in chromatin states that are vital to differentiation and development. However, little is known about how those states differ among cells in the preimplantation embryo, if they are affected by location in the embryo, cell-cell or cell-extracellular matrix contacts, and the sequence of changes that occur in the cells at different genomic loci. Mapping these states, spatially, temporally and genomically would provide an unprecedented view of molecular changes during development, which can only be achieved by single cell methods. Second, differentiation has been modeled using embryonic stem cells and culture conditions that drive differentiation into specific lineages. Chromatin states before and differentiation have been characterized in populations of cultured cells using ChIP [63] and bisulfite sequencing [104]. It is not known how individual cells responding to differentiation promoting signals behave, the course of those behaviors throughout differentiation in culture, and if they parallel what occurs in individual cells within the embryo. Third, beyond questions of differentiation and development, it is known that disease states can arise when individual cells exhibit deviations from normal behavior, including at the level of chromatin regulation. Ensemble analyses of tissues or groups of cells reveal population-averaged behaviors, but not the degree of variance among cells within the population, which can be revealed only by single cell analyses. Whether high variance in cellular behavior correlates with disease likelihood will require the application of single cell analytical methods. In addition to these three examples, single cell analyses can apply molecular criteria to assess the number of distinct cell types that exist in organisms and the heterogeneity that arises in clonally derived cells.

Single cell analytical methods have been developed to sequence bacterial genomes [105], mammalian genomes [106, 107], and transcriptomes by microarray [108, 109] and RNA sequencing [110–116] (reviewed in [117]). The C1 system from Fluidigm facilitates isolation of single cells, as well as on-device extraction and processing of RNA and DNA for such studies; however, this platform has not been adapted for chromatin isolation. In an effort to fill this gap, microscale devices have been developed capable of trapping low inputs of cells for standard ChIP [41, 118] and restriction enzyme based 5mC analysis [119], from single cells that enable the isolation and manipulation of single chromatin molecules using antibody conjugated microspheres and optical tweezers [120], or from single cells using a design that enables their integration with the SCAN platform [121]. The latter has been used for extracting chromatin from low cell inputs by sequentially flowing into the devices the reagents used for native chromatin extractions, followed by fluorescent antibodies needed for detecting chromatin features by SCAN (unpublished). Common to these devices is their nanoliter scale volume, which can ensure that reagent concentrations, including antibodies, are kept high, allowing efficient binding and high sensitivity of detection. Picoliter volumes are achievable with micron scale droplets, which hold promise for epigenomic analyses, either of single cells or single molecules. They are readily loaded with materials and reagents for analysis, can be subjected to high throughput fluorescent detection in flow channels [122] and droplets with defined properties can be selectively sorted for subsequent analyses [123].

Other efforts for single cell analyses of DNA and histone modification states have been developed that do not rely on custom-built microscale devices [124–126]. Hi-C methods were reported that characterize chromatin interactions in the genome of single cells [127]. Proximity ligation methods have detected coincidence of chromatin features by microscopy [128] in single cells. Other techniques have characterized nuclear lamina interactions with chromatin in single cells [129]. Analyses of 5mC have been reported for single cells using bisulfite methods [130, 131]. Further development of these and other single cell techniques will be necessary to address the kinds of questions raised above.

Aside from the technical issues associated with collecting epigenomic data from single molecules and single cells (discussed further below), are the statistical issues associated with extracting meaningful biological information from the data. ChIP-seq and BS-seq, when applied to materials isolated from tens of millions of cells, can reliably identify loci harboring a given epigenetic mark when sequencing is done at high enough coverage. However, when querying chromatin from few cells for combinations of epigenetic marks using single molecule approaches, coverage is fundamentally limited: Autosomal loci in a single cell can never be sequenced to more than 2X coverage even with completely efficient library preparation. Such limitations can be overcome by performing multiple single cell and single molecule experiments on a population of cells, with the sampling requirements dictated by the efficiency of each step in the workflow and the magnitudes of the effects being measured.

6. Technical challenges of single molecule SCAN and single cell chromatin analyses

Single molecule and single cell analytical methods bring new opportunities for discovery, and challenges for their implementation. Among the challenges of single molecule methods are achieving signal to noise ratios that enable reliable detection of chromatin features on individual molecules when using fluorescent detection reagents, and achieving high enough throughput rates in analytical and preparative modes to accomplish experimental objectives. Among the challenges of single cell analyses are availability of reagents with favorable binding constants to detect chromatin features with high efficiency, even at low analyte concentrations. The use of nanoliter volumes in microfluidic devices [118, 132, 133] or subnanoliter volumes in micrometer sized droplets [123, 134, 135] can help maximize those concentrations. Challenges common to single molecule and single cell analyses include development and use of high-efficiency library preparation and/or novel sequencing techniques that reliably gather sequence information from DNAs isolated using single molecule methods.

6.1. Signal-to-Noise ratio (SNR) for single-molecule detection (SMD)

Most of the single-molecule analytical methods for epigenomic measurements require measurement of emissions from fluorescent intercalators and antibodies bound to single chromatin molecules [87, 98–100]. The fluorescent signals from single molecules must be distinguished from hardware noise and background, which is complicated by the vast excess of solvent. The number of photons collected in detector (S) of an optical setup is

$$\mathrm{S}=W_{\mathit{abs}}\times QY\times D_{\mathit{eff}}=P/h{v}_A\times {\sigma }_{\mathit{abs}}/A\times QY\times D_{\mathit{eff}},$$

where W_abs is the molecular absorption rate, P (Watts) excitation beam power, σ_abs an absorption cross-section of the single fluorophore for excitation light, A (cm²) the cross-section of the excitation beam in the plane of the molecule, QY fluorescence quantum yield of a fluorophore, and D_eff the global collection efficiency of the optics setup. The environment of the single molecule can add a background contribution modeled by a rate b per unit-volume and unit-excitation power. From these definitions, the signal-to-background ratio (SBR) can be expressed as

$$\text{SBR}=\frac{S\times t_0}{bV\times P\times t_0}=\frac{D_{\mathit{eff}}\times QY\times {\sigma }_{\mathit{abs}}}{bV\times A\times h{v}_A},$$

where t₀ is the integration time of the detector and V is the intersection between the excitation and the detection volume [136]. SBR can be maximized by use of small focal volumes, large absorption cross sections (σ_abs) and fluorophores with high fluorescence quantum yields (QY). Accordingly, microfluidic, nanofluidic and microdroplet platforms are preferred for single molecule and single cell analyses.

Some detectors amplify the incoming signal (S × τ) before readout, introducing a gain (G) in the measured signal. In practice, the shot noise component of the incoming signal is itself multiplied by the gain G, and a correction factor F, called the excess noise factor, σ_amp = F × G × σ_in, where σ_in represents the standard deviation of the input signal and σ_amp that of the amplified signal. For a detector without gain, or for photon-counting detectors, G = 1 and F = 1. For negligible dark count rates and negligible readout noise component, the resulting SNR can thus be expressed generally as

$$\text{SNR}=\frac{\sqrt{S\times t_0}}{F\sqrt{1+1/\mathit{SBR}}}.$$

SBR does not depend on the measurement duration, whereas SNR increases with the measurement duration and SBR. Achieving the required SNR requires close attention to maximizing signal while minimizing background. Higher excitation power produces higher SNR values, but the power cannot be increased arbitrarily because saturation accompanied by excess power can actually lead to reduced fluorescent emission due to photobleaching. Background can have different sources, including detector dark counts and residual laser scattered light or Raman scattering, some of which can in principle be minimized by sample or setup optimization.

To estimate actual SBR and SNR values from a typical experiment, consider the single molecule fluorescence detection for Alexa Fluor dyes in water, using a single-photon avalanche photodiode detector (SPAD) to collect the signal. QY and σ_abs of Alexa Fluor 488 are 0.92 and 2.5 × 10⁻¹⁶ cm² at 488-nm excitation (from the manufacturer), respectively. Using 100 μW of laser power, 1 ms integration time, 0.4 μm of channel width, and a 2 μm of length of laser spot along channel (L_focal), we expect 7063 photons/ms emitted from a single AlexaFluor 488 molecule. With a high-performance optical microscope fitted with a high numerical aperture lens having 5% collection efficiency, there will be an estimated 353 photons/ms reaching the SPAD, leading to a SNR ~ 18.5 and SBR ~ 35.3. More favorable SNR and SBR values would be obtained by replacing organic dyes with quantum dots (QDs), because of their exceptional brightness and stability. QDs have the added benefit of excitation at low wavelengths and narrow emission spectra, which makes more efficient use of the spectrum for higher order multiplexing.

To improve SNR or SBR in SMD, several microscopic and spectroscopic techniques have been suggested and tested. These overcome a limitation of conventional widefield epi-fluorescence, which creates high background fluorescence from out-of-focus light. Each method modifies the excitation laser beam to further confine light on the nanoscale and improve the spatial resolution. This can be achieved by evanescent illumination using total internal reflection fluorescence microscopy [137], and by two photon excitation using two-photon fluorescence microscopy [138]. Confocal microscopy can also reduce background signals originating outside the focal plane using spatial filters [139]. Laser scanning and spinning disk confocal fluorescence microscopy have the added benefit of increasing spatial resolution in single molecule imaging [140]. Photo-activated localization microscopy (PALM) [141] and stochastic optical reconstruction microscopy [142] have been performed to obtain images with a resolution beyond the diffraction limit; these may be of particular value for ordered chromatin arrays or chromatin confinement methods where localization of chromatin features on large fragments is limiting. Other optical methods have been demonstrated to enhance single molecule detection techniques beyond the limits set by optical diffraction (reviewed in [143]).

6.2. Throughput

SCAN presents challenges distinct from flow cytometry and FACS. Micron scale fluidic channels suitable for single cell analyses are not suitable for single molecule analyses, because using such dimensions would mandate use of highly diluted samples to ensure single molecule analyses and result in low sample throughput. Therefore, submicron fluidic structures are used. When flowing molecules through a channel with a 500nm × 500nm cross section, and an inspection laser with a 1.3 μm beam diameter, the focal volume is 325 aL and Poisson statistics reveal the probability that the inspection volume will be occupied at a given concentration, by

$$\mathrm{P}(\mathrm{k},\mathrm{\lambda })=e^{-\lambda }\frac{{\lambda }^k}{k!},$$

where P(k) represents for the probability of finding exactly k number of molecules in the focal volume at a given time. λ is the average number of molecules in the focal volume. Thus, analyte concentrations must be kept low so that the probability of more than one molecule in the inspection volume P(≥ 2), is kept low. P(≥ 2) can be estimated by the expression (1 – probability of zero molecules in the focal volume – probability of one molecule in the focal volume). Requiring that two or more molecules are present in the detection volume no more than 1% of the time gives 0.01 ≥ P(≥ 2) = 1 − e^−λ − λe^−λ = 1 − e^−λ(1 + λ), resulting in λ = 0.149. Then the maximum concentration of molecules (C_max) that ensures this low likelihood of two molecules in the focal volume is given by

$$C_{\mathit{max}}=\frac{\lambda }{A_s\times L_{\mathit{focal}}}=760pM.$$

The throughput (N_T) in number of molecules per unit time is then given by N_T = λ/t₀. Thus, the throughput of the single channel is estimated to 0.149 molecule/ms, using a 1ms value for t₀ from the SNR discussion above. If the size of a human genome size is 3.2 × 10⁶ kbp, and the size of prepared chromatin fragments is 500 bp, then the throughput in number of genomes per hour can be obtained as:

$$\frac{0.149\text{molecules}}{ms}=\frac{4.47\times {10}^6\text{bp}}{\mathit{minute}}\approx \frac{0.0838\text{genome}}{hr}.$$

It will take about 12 hours to analyze one human genome equivalent in a single nanofluidic channel or just over 7 minutes on a 96 channel device like that shown in figure 2, if done with the same optical sensitivity that is possible with single channels. However, it is not likely that throughput will scale linearly with channel number for two reasons. First, SPADs that are ideal for single channels are not appropriate for detecting fluorescence from many channels, given their cost and engineering challenges to integrate 96 SPADs on a device; instead a 2D photon collection array is needed. These are available in the form of EM-CCD and sCMOS cameras; however, both camera formats have lower sensitivity than SPADs, requiring longer integration times. Parameters of high importance are fast frame rates, low dark current and readout noise, and high quantum efficiency and dynamic range. Pixel size and number should be appropriate to the geometry of the fluidic device, and data collection should be restricted to the informative pixels with relevant fluorescent emissions. Second, as channel number increases, so will the areas of excitation illumination and emission detection. These larger areas require lower magnification lenses with lower numerical apertures, which negatively affect photon collection efficiency [144]. Several hardware and data processing advances are needed for single molecule methods to achieve the genomic coverage currently enjoyed by ensemble methods like ChIP.

Figure 2. SCAN devices for increased throughput.

A, Optical micrograph of a 4-inch diameter fused silica wafer containing an array of 27 fluidic devices, each containing 96 parallel channels. B, Differential interference contrast optical micrograph of tens of the parallel fluidic channels in the center of a device. The square structures on each side of the central array are support pillars to prevent collapse when bonding to a cover wafer. C, Scanning electron micrograph showing approximately 8 parallel channels etched between microfluidic reservoirs on a single device. The channels are 20 microns in length (horizontal) and 500 nm deep. D, SEM close-up of an individual fluidic channel that is 750 nm in width. Devices with this parallel architecture can enable analyses like those previously reported [98, 99] with higher throughput, however, fluorescence detection will require replacing APDs with other detectors such as CMOS arrays, which will somewhat compromise sensitivity and hence throughput.

6.3. Reagents for detecting chromatin features and their dissociation constants (K_d)

Antibodies, DNA-binding proteins, and chromatin-binding proteins that bind specifically to their targets with high affinity are essential for single molecule chromatin analyses. Tests for specificity of antibodies recognizing histone modifications often use peptides corresponding to modified histone tails [145]. K_d measurements are rarely taken, but when reported, spanned four orders of magnitude, from 0.2nM to 2μM with high lot-to-lot variation [43, 146]. K_d values for proteins binding to DNA modifications ranged from 10nM to 50μM [147, 148]. Synthetic reagents to detect chromatin features hold promise for future studies [43].

Traditional ChIP-seq, which can use abundant sources of chromatin, does not require the highest affinity reagents as long as enough DNA can be recovered for high coverage sequence. However, single molecule and single cell analyses do rely on high affinity reagents or very small reaction volumes to enable high reagent concentrations. If we assume the K_d of an antibody or binding protein specific for a chromatin feature is 1 nM, and this concentration of detection reagent were used, then the concentration of the chromatin should be at least 10 nM for efficient binding. If a single cell were used for the chromatin source, and the chromatin were fragmented to 500 bp, then to achieve these concentrations requires a 2.2 nL reaction volume. For detection reagents with a 1μM K_d, we would want a chromatin concentration of 10μM, which for a single cell would require a reaction volume of 2.2 pL. These volumes can be achieved using microscale fluidic devices or micron scale droplets [123, 134, 135].

6.4. Sorting single molecules

Whereas analytical SCAN can provide information about coincidence and mutual exclusion of chromatin features, and their relative abundance in samples [99], it does not report the genomic locations of those modifications. Methods that sort individual molecules have been developed [149], including one that is based on DNA modification status that was adapted to SCAN [100]. In proof of principle experiments, methylated DNA was reliably separated from unmethylated DNA by single molecule sorting. In those studies, DNAs were labeled with a fluorescent intercalator and MBD1 protein carrying a spectrally distinct fluorophore. In analytical SCAN, molecules are driven by voltage through a channel with a single input and output through a focal volume where fluorescence information is collected, but in preparative SCAN, the nanoscale fluidic device for sorting includes a bifurcation downstream of the focal volume (Figure 3). As single molecules are driven through a focal volume by the voltage potential, the fluorescent properties of the molecules are measured and the signals directed to a field programmable gate array (FPGA), which places the voltage potential along one outflow track or the other. In this manner, molecules with fluorescent properties diagnostic for the 5mC can be isolated. The false positive rate, whereby unmethylated molecules were inadvertently identified as methylated and isolated was under 2%; the false negative rate, whereby methylated molecules were inadvertently identified as unmethylated and excluded, was under 10% [100]. Ongoing efforts are focused on applying this approach to complex chromatin samples. The technical issues relevant to analytical SCAN that were described above also apply to preparative SCAN. Approaches using single molecules enclosed in droplets offer an alternative to SCAN sorting as droplets can be sorted using acoustic waves [122].

Figure 3. Potential integrated micro-/nano-device architecture for multiplexed epigenomic measurements and amplification-free sequencing from a single cell.

A schematic diagram depicting an integrated device for profiling combinations of epigenomic states from single cells using single molecule methods. Examples of each module of the integrated device were described in the text. From left, isolation of cells for analysis followed by extraction of chromatin from cells [41, 118, 121]; binding of antibodies or other reagents at high concentration in a low volume microreactor to detect epigenomic features [123, 133–135]; removal of unbound antibody can improve throughput of informative chromatin containing complexes; sorting of chromatin carrying various user defined combinations of epigenomic marks [100]; extraction of DNA from sorted chromatin; library preparation after low bias WGA [105, 155–157] followed by sequencing, or amplification free sequencing.

Once single molecules with user specified chromatin modifications are isolated, sequencing can reveal where combinations of modifications reside. DNA that is recovered will be present in picogram quantities or lower, which will require high-efficiency methods for sequencing. Library-free sequencing approaches will be ideal. Alternatively, highly efficient, low bias amplification and library preparation techniques are required.

6.4.1. Sequencing platforms

Each of the next generation sequencing technologies require distinct workflows of varying complexity to prepare templates for sequencing. DNA losses and inefficiencies at each stage present challenges for applications with very low DNA inputs. These can be minimized by performing library preparation reactions in nL volume scale devices [133], which in theory can be integrated with previously mentioned single molecule and single cell analytical platforms. Library preparation for the widely used Illumina HiSeq platforms requires repairing the DNA ends, single nucleotide A-tailing, adaptor ligation, and library amplification prior to final DNA purification and loading onto the sequencer.

Competing sequencers, notably those based on single molecule methods, can have simpler workflows. The formerly available Helicos sequencing platform required no prior ligation and amplification steps. Instead, single molecules, as they existed in cells, were sequenced directly after a 3′ tailing reaction and hybridization to tail-complementary sequences on the platform [150]. Accordingly, with this highly simplified workflow, the Helicos platform proved useful for ChIP-seq studies using as few as 50pg of input DNA [151]. While this platform was a success technically, it was short-lived as a commercial product.

The Pacific Biosciences SMRT system, as previously mentioned, sequences native single-molecule templates without any amplification, although templates must undergo repair, ligation, primer annealing, and DNA polymerase loading steps prior to deposition in zero mode wave guides, where single DNA molecules are optically monitored during DNA polymerase-directed incorporation of fluorescent nucleotides [152]. Given that library preparation is comparable in complexity to the Illumina system, it is not clear if this platform will be suitable for picogram and subpicogram quantities of DNA. As cited above, the Pacific Biosciences platform can report 5mC [66] and 5hmC [67] residues during sequencing based on the reaction kinetics of base incorporation – important features for combinatorial analysis.

Often called the fourth-generation of sequencing platforms, nanopore-based sensing devices provide another method for directly sequencing single DNA molecules (reviewed in [153]). The technology consists of a nanopore within a dielectric membrane that separates two salt solutions. Upon application of a current, the system drives ions through the nanopore where charged molecules, such as DNA, will create resistance, thereby partially blocking the pore as the DNA passes through. Using the translocation time and current blockade, each of the four nucleotides can be distinguished from one another and sequence data collected. A key advantage of this platform is that templates require no preparation beyond their purification, which will eliminate concerns of inefficiencies and losses associated with library preparation. However, it is not clear what will be the lower limit of template abundance required for sequencing on nanopore platforms, or their throughput. Published data reported the use of 1μM DNA solutions in a 60μL volume [97], but these values may drop as the technology matures. Like the SMRT-seq platform, 5mC and 5hmC can be distinguished from one another and unmodified cytosines in nanopores [68, 93, 95]. Nucleosome complexes can also be identified by nanopores [154], raising the possibility that histone modification states can be revealed as well.

6.4.2. Amplification

When single molecule methods are used to isolate chromatin for sequencing, the yields are likely to be lower than when ensemble methods like ChIP are used. Yields will be smaller still when low cell inputs are used. Whole-genome amplification is necessary to accrue enough material for sequencing. The most widely used methods for WGA using low DNA inputs include multiple displacement amplification (MDA) [105, 155], ligation-mediated PCR [156] and multiple annealing and looping-based amplification cycles (MALBAC) [157]. All methods suffer from some amplification bias, with PCR introducing more than MDA [155], and MDA potentially introducing more than MALBAC [157]. However, biases can be minimized with small reaction volumes [158]. These approaches have worked well in traditional single-cell analyses; however, they have yet to be tested thoroughly in epigenetic applications involving whole genome analysis using chromatin inputs from fewer than 5,000–10,000 cells [29, 159].

6.4.3. Tagmentation

Recently, an alternative to the standard methods of fragmentation and adapter ligation has been developed for preparing libraries for whole-genome sequencing. While traditional methods require multiple steps to generate adaptor-ligated fragments, the hyperactive derivative of the Tn5 transposase can be used to simultaneously fragment, end-repair, and tag the DNA with specific adaptors to both ends, all within a single-tube reaction. This method termed tagmentation, enables library preparation from subnanogram amounts of DNA [160]. Tagmentation has been adapted to bisulfite analysis of 5mC using low tens of nanograms of DNA [161, 162], and to analysis of open chromatin using as few as 500 cells [163].

7. Closing remarks

Many technical advances have been vital to modern molecular genetics research – DNA sequencing being prominent among them. It is clear that beyond the information carried by genomic sequences, combinations of chromatin features play vital roles in genomic control. Furthermore, variations in chromatin states that exist in different tissues and cells, and even within a given cell type, have great biological consequence. Accordingly, technical advances are required to characterize these chromatin states at high resolution. Just as the single nucleotide resolution provided by whole genome sequencing has been vital to understand the structure and function of genomes, we speculate that single molecule and single cell methods will be vital to understand phenotypes and mechanisms of inheritance not revealed by primary genomic sequences. Many independent methods that were described here hold promise for filling these needs. Ongoing development and integration of these methods are needed to fulfill this promise.

Table 1.

Comparison of methods

Technique	Feature(s) measured	Information reported	Limitations	References
Ensemble methods
ChIP-seq	Histone modifications Chromatin binding factors	Genomic locations of individual chromatin features	Coincidence of multiple marks inferred; methods to improve inferences exist Requires high quality affinity reagents	[42, 72, 73, 164]
re-ChIP or sequential ChIP	Histone modifications Chromatin binding factors	Genomic locations of combinations of chromatin features	Full genome coverage not readily achieved Requires high quality affinity reagents	[165]
BS-seq	5mC or 5hmC	Genomic locations of either feature Can use DNA from ChIP for dual feature analysis	Does not distinguish 5mC from 5hmC or report locations of 5fC or 5caC	[166] [65]
oxBS-seq	5mC	Genomic locations of 5hmC Can be combined with BS-seq to distinguish locations of 5mC and 5hmC Can use DNA from ChIP for dual feature analysis	Does not report locations of 5fC or 5caC	[22, 167]
TAB-seq	5hmC	Genomic locations of 5hmC Can use DNA from ChIP for dual feature analysis	Does not report locations of 5fC or 5caC	[21]
Mass spectrometry	Histone modifications	Abundance, identity and coincidence of histone modifications	Requires modifications reside on nearby amino acids on the same histone, unless combined with ChIP for prior purification of analyte Does not report genomic locations of histone modifications	[46, 77, 168–171]
Single molecule methods
DNA curtains	DNA binding proteins	Binding properties and dynamics of interactions between DNA binding proteins defined DNA substrates	Not yet adapted to naturally occurring chromatin	[83–86]
Ordered chromatin arrays	Histones	Presence of histones on chromatin; readily adaptable to detect histone modifications	Does not report sequence of DNA Requires high quality affinity reagents Limited resolution	[87]
Chromatin elongated by confinement	Histone modifications and 5mC	Presence of histone modifications and 5mC	Does not report sequence of DNA Requires high quality affinity reagents Limited resolution	[88, 89]
ChIP-string	Chromatin binding factors	Which of up to 800 genomic locations bind feature of interest	Does not provide full genome coverage Requires high quality affinity reagents	[42]
SCAN	Histone modifications and 5mC, individually or in combination	Abundance of chromatin features Can be run in analytical or preparative modes	Requires high quality affinity reagents Limited molecule per minute throughput in preparative mode	[98–100]
SMRT sequencing	5mC and 5hmC	Genomic locations of indicated features	Identification of features is based on latency of nucleotide incorporation, which is also influenced by sequence context	[66, 67]
Nanopore methods	5mC and 5hmC	Presence of indicated features on a DNA template	May require high quality affinity reagents Further development needed for genome wide sequencing	[68, 93–95]