A parallelized, automated platform enabling individual or sequential ChIP of histone marks and transcription factors

Significance

Chromatin immunoprecipitation followed by next-generation sequencing (ChIP-seq) has become the most widely used method to infer the genomic binding locations of chromatin-associated proteins such as histones and transcription factors. Since its emergence in 2007, ChIP-seq has been performed through a long sequence of manual steps, with consequent limitations in throughput and sensitivity of the technique. Here, we present a microfluidic implementation of the ChIP procedure, named FloChIP, which greatly streamlines the experimental workflow. We show that FloChIP can perform standard ChIP and sequential ChIP assays in parallel and reproducibly in an automated fashion for histone marks and transcription factors.

Abstract

Despite its popularity, chromatin immunoprecipitation followed by sequencing (ChIP-seq) remains a tedious (>2 d), manually intensive, low-sensitivity and low-throughput approach. Here, we combine principles of microengineering, surface chemistry, and molecular biology to address the major limitations of standard ChIP-seq. The resulting technology, FloChIP, automates and miniaturizes ChIP in a beadless fashion while facilitating the downstream library preparation process through on-chip chromatin tagmentation. FloChIP is fast (<2 h), has a wide dynamic range (from 10⁶ to 500 cells), is scalable and parallelized, and supports antibody- or sample-multiplexed ChIP on both histone marks and transcription factors. In addition, FloChIP’s interconnected design allows for straightforward chromatin reimmunoprecipitation, which allows this technology to also act as a microfluidic sequential ChIP-seq system. Finally, we ran FloChIP for the transcription factor MEF2A in 32 distinct human lymphoblastoid cell lines, providing insights into the main factors driving collaborative DNA binding of MEF2A and into its role in B cell-specific gene regulation. Together, our results validate FloChIP as a flexible and reproducible automated solution for individual or sequential ChIP-seq.

The genome-wide distribution and dynamics of protein−DNA interactions constitute a fundamental aspect of gene regulation. Chromatin immunoprecipitation followed by next generation sequencing (ChIP-seq) (1) has become the most widespread technique for mapping protein−DNA interactions genome-wide. ChIP-seq has been successfully applied to dozens of transcription factors (TFs), histone modifications, chromatin modifying complexes, and other chromatin-associated proteins in humans and other model organisms (2). The ENCODE (Encyclopedia of DNA Elements) and modENCODE consortia alone have already performed more than 8,000 ChIP-seq experiments, which have greatly enhanced our collective understanding of how gene regulatory processes are orchestrated in humans as well as several model organisms (3). In addition, ChIP-seq proved to be essential to acquire new insights into genomic organization (4–6) and into the mechanisms underlying genomic variation-driven phenotypic diversity and disease susceptibility (5, 7, 8). More specifically, this assay proved crucial in determining the DNA binding properties of hundreds of TFs (9). Nevertheless, in comparison to other widespread methods based on next generation sequencing (NGS)—e.g., RNA-seq (10) and assay for transposase-accessible chromatin using sequencing (ATAC-seq) (11)—ChIP-seq lags behind in some key metrics, that is, throughput, sensitivity, and automation, which hinders its wider adoption and reproducibility. For example, while RNA-seq can now be regularly performed on hundreds or thousands of single cells using readily available workflows (12, 13), ChIP-seq has largely remained labor intensive and limited to a few samples per run, each composed of millions of cells. Moreover, while a typical preamplification RNA-seq workflow consists of only three steps—that is, cell lysis, RNA capturing and reverse transcription—ChIP-seq typically involves several preamplification steps (cross-linking, lysis, fragmentation, IP, end repair, and adapter ligation). Finally, any given RNA transcript is present in each cell in numerous copies, which increases the likelihood of its capture and detection, whereas each locus-specific protein−DNA contact occurs a maximum of two times in a diploid cell. The combination of these idiosyncratic differences, together with the lack of enabling solutions, has thus far prevented the ChIP-seq technology, as opposed to other NGS-based methods, from reaching its full potential in terms of adoption, utility, and biomedical relevance.

In addition to the standard ChIP protocol, a modification of its workflow involving sequential ChIP has been developed to infer the genomic cooccurrence of two distinct (protein) targets. In principle, sequential ChIP consists of performing ChIP twice on the same input chromatin, which leads to a multiplication of the inefficiencies mentioned above. Therefore, not only does sequential ChIP show the same limitations as regular ChIP-seq, but these come in an augmented form due to its consecutive nature. As a result, few studies have so far performed sequential ChIP followed by NGS (14, 15) (sequential ChIP-seq), and most of the available studies have so far relied on qPCR to validate putative bivalent regions (16–18) (sequential ChIP-qPCR).

In recent years, several attempts have been made to alleviate some of the limitations of the ChIP-seq and sequential ChIP approaches. Gasper et al. (19) and Aldridge et al. (20) addressed the issue of automation by implementing the manual steps of a conventional ChIP-seq workflow on robotic liquid handling systems. However, in these examples, automation was balanced with sensitivity, since these workflows still require tens of millions of cells per experiment. Van Galen et al. (21) and Chabbert et al. (22) addressed the issue of throughput by barcoding and pooling chromatin samples before IP. Although van Galen et al. proved that their approach led to higher sensitivity (500 cells per ChIP), neither approach is automated, and both are, so far, limited to the detection of histone marks. Ma et al. (23) and Rotem et al. (24) addressed the limitation of sensitivity with two different microfluidic-based strategies. Ma et al. focused on improving the efficiency of the IP step by confining it within microfluidic channels. Although they showed good IP efficiency down to as few as 30 cells, their approach requires impractical antibody−oligo conjugates, is not automated, and was not shown to work for TFs. On the other hand, Rotem et al. achieved the remarkable feat of performing ChIP-seq in a single cell by integrating the concept of chromatin barcoding and pooling into a single droplet-based microfluidic chip. However, even though the barcoding step has, indeed, single-cell resolution, the most critical step—that is, the IP step—is performed manually on 100 cells. As a result, their approach—also shown to work only for histone marks—yielded sparse single-cell data, and thousands of assays are still required to identify specific cell subpopulation signatures. In an effort to simplify the sequential ChIP workflow, Weiner et al. (15) complemented the IP steps with sequential chromatin barcoding, thus achieving a high degree of multiplexing. However, their approach increases the number of experimental steps, which makes it significantly more labor intensive given that the workflow is not automated.

In this work, we aimed to address the major limitations of current ChIP-seq and sequential ChIP solutions (throughput, sensitivity, and automation), by developing a microfluidic strategy that we named FloChIP. We show that high quality and parallel/multiplexed ChIP-seq for histone marks (down to 500 cells) and TFs (100,000 cells) is achieved in less than 2 h through a combination of microvalves, micropillars, flexible surface chemistry, and on-chip chromatin tagmentation. Moreover, by designing an interconnected and modular device, FloChIP enables straightforward re-IP of eluted chromatin, effectively establishing a half-day sequential ChIP pipeline. Finally, we performed FloChIP for the TF MEF2A using chromatin derived from 32 lymphoblastoid cell lines (LCLs). Our data highlight the main drivers of MEF2A collaborative DNA binding and provide insights into MEF2A’s role in the regulatory network underlying lymphoblastoid proliferation.

Results

FloChIP Is Engineered for Automated, Beadless, and Miniaturized ChIP-seq.

Conventional ChIP-seq requires hours of manual work and a wide range of consumables provided by a variety of suppliers and has limited sensitivity. The main rationale behind FloChIP’s design was the development of an efficient solution that would address these drawbacks in a convenient and compact manner. The two core elements of FloChIP’s technology are the assembly of a multilayered stack of biomolecules, enabling versatility in antibody pull-down (Fig. 1A), and an engineered pattern of high surface-to-volume micropillars for efficient chromatin capture and washing (Fig. 1B). FloChIP’s surface chemistry is based on strong although noncovalent molecular interactions, enabling the immobilization of an antibody of choice prior to IP. The first layer is obtained by flowing on-chip a concentrated solution of biotinylated BSA, which passively adsorbs to the hydrophobic walls of the microfluidic device. This layer has both an insulating role, preventing nonspecific adsorption of chromatin to the chip walls, and a docking role for the next layer, which is obtained by flowing on-chip a solution containing neutravidin that strongly binds to the biotin groups of the first layer. The third layer is formed by flowing a solution of biotinylated protein A/G, which becomes firmly immobilized by the unsaturated binding sites of the neutravidin layer. Protein A/G is a recombinant protein used in a variety of immunoassays due to its ability to strongly bind to a large number of different antibodies. This ability is retained by FloChIP’s surface functionalization, which thus constitutes a general substrate for antibody pull-down.

Fig. 1.

FloChIP’s architecture for miniaturized ChIP-seq. (A) FloChIP’s processing phases in descending chronological order. In the first “surface functionalization phase” (∼80 min), the inner walls are functionalized by sequentially introducing chemical species that firmly interact with both the previous and following layers of functionalization. Also, in chronological order, these species are biotin−BSA, neutravidin, biotin−protein A/G, and antibody. Following functionalization, the IP takes place by flowing sonicated chromatin on-chip in a total time of 30 min to 60 min (depending on the chromatin volume that is introduced), followed by low-salt, high-salt, and LiCl buffer washes. Subsequently, the antibody-bound chromatin is tagmented directly on-chip in order to introduce Illumina-compatible adapters. Finally, the tagmented chromatin is eluted off-chip using an SDS-containing buffer and high temperature. (B) Top-view microscopy picture of a portion containing numerous micropillars. Each portion is itself repeated several times along the length of one IP lane. (C) Top-view schematic of one IP lane. Each IP lane can be repeated n times across a FloChIP device. Flow channels are in blue, and control channels are in red. (D) Fluorescence micrographs showing the requirement for protein A/G in the correct formation of FloChIP’s functionalization. Images were taken at 10× magnification. (E) Top-view schematic of the eight-unit FloChIP device. Flow channels are in blue, and control channels are in red.

Another critical feature of FloChIP’s workflow is the microfluidic tagmentation of immunoprecipitated chromatin. In a previous study, tagmentation of bead-bound chromatin (ChIPmentation) was shown to generally increase cost-effectiveness and sensitivity of the ChIP-seq workflow (25). Here, we built upon this concept and adapted it to obtain an example of ChIPmentation performed directly on chromatin bound to the inner walls of a microfluidic device (Fig. 1A, indexing and elution). Briefly, this is achieved by flowing a Tn5 solution into the device while heating the chip surface to 37 °C, allowing direct on-chip indexing of chromatin-bound DNA. Importantly, microfluidic ChIPmentation streamlines the downstream library preparation workflow and reduces hands-on time (see also below).

For the successful initiation of the multilayered surface functionalization, the only substrate requirement is the hydrophobic surface of the device polymer. Therefore, to maximize the surface-to-volume ratio of our devices, we designed an array of micropillars (Fig. 1B), which repeats multiple times across each IP lane (Fig. 1C). With the goal of visually validating the successful assembly of our multilayered on-chip chemistry and to confirm that every layer is essential to this end, we first sought to immunoprecipitate chromatin derived from a HeLa H2B-mCherry cell line using an anti-H2B antibody. The resulting fluorescence micrographs confirmed that each layer of the molecular species is necessary for successful IP of cellular chromatin (Fig. 1D and SI Appendix, Fig. S1A).

The IP lane (Fig. 1C) is the fundamental unit of the FloChIP architecture, and it can itself be repeated n times, where n is the desired throughput of the device. For our initial tests, we used an eight-lane FloChIP device (Fig. 1E and SI Appendix, Fig. S1B). To gain accurate flow control, automation, and multiplexing, a network of soft microvalves was added to the design; different multiplexing modes can be achieved with the same microfluidic architecture by actuating distinct sets of valves. For instance, we used the name “FloChIP mode 1 - sample multiplex,” for the option of coating all IP lanes of the device with one antibody and introducing different samples from dedicated individual inlets (Fig. 2A). In this multiplexing configuration, the number of samples that can be processed is determined by the number of IP lanes of the microfluidic chip, for example, here, eight lanes and thus eight antibodies. Alternatively, “FloChIP mode 2 - antibody multiplex” provides the option of coating each IP lane with a different antibody and of distributing one sample equally across the whole device (thus probing distinct antibodies in parallel using one sample) (Fig. 2B). In this configuration, one can process as many antibodies as the number of IP lanes present on-chip, here, eight lanes and thus eight antibodies.

Fig. 2.

FloChIP-based derivation of chromatin landscapes in LCL GM12878. (A) Schematic depiction of FloChIP’s mode 1: sample multiplex. One antibody solution is introduced through the common inlet and distributed equally across all IP lanes. During IP, each IP lane is loaded separately by introducing different samples through the individual inlets. (B) Schematic depiction of FloChIP’s mode 2: antibody multiplex. Each IP lane is functionalized separately by introducing different antibodies through the individual inlets. During IP, one sample is introduced through the common inlet and distributed equally across all IP lanes. (C) H3K27ac profiles at three different genomic loci obtained by FloChIP with decreasing cell numbers (100,000 to 500 cells). For comparison, ENCODE data generated by conventional ChIP-seq are also shown. (D) Normalized read density metaprofiles at 2 kilobases (kb) around TSS for samples of decreasing cell numbers and ENCODE. (E) Normalized pairwise correlation of samples with decreasing cell numbers and ENCODE (see Methods for details). (F) Signal tracks for H3K27ac, H3K4me1, and H3K4me3 profiles obtained by FloChIP with 100,000 cells are shown at three different genomic loci. (G) H3K9me3 signal track comparison between ENCODE and FloChIP. (H) Normalized read density metaprofiles at 2 kb around TSS for H3K4me3, H3K27ac, and H3K4me1. For comparison, ENCODE data generated by conventional ChIP-seq are also shown. (I) Correlation plots between FloChIP (x axis) and ENCODE (y axis) data for all targets tested, that is, H3K27ac, H3K4me1, H3K9me3, and H3K4me3. (J) Comparison in terms of fraction of reads in peaks (FRiP) between FloChIP and ENCODE for histone mark samples.

FloChIP Reliably Reproduces ENCODE Data across a Wide Range of Input Cells.

To evaluate the performance of our FloChIP strategy, we first set out to empirically estimate FloChIP’s dynamic range. To this end, we performed FloChIP in sample multiplex mode with GM12878 lymphoblastoid cells, by functionalizing the whole chip with an anti-H3K27ac antibody and immunoprecipitating different chromatin dilutions (ranging from 1 million to 500 cells). Despite the observed differences in recovered DNA (SI Appendix, Fig. S2A), we obtained high and stable fold enrichment results across the whole series of dilutions tested (SI Appendix, Fig. S2B). To obtain a genome-wide perspective on its dynamic range, we sequenced FloChIP’s libraries for chromatin samples obtained from 100,000, 50,000, 5,000 and 500 cells. After sequencing, the rate of uniquely mapped reads remained high for all samples (SI Appendix, Fig. S2C), while the fraction of reads falling into peaks (FRiP score) decreased with decreasing input amounts—from over 60% for the largest sample to just above 10% for the smallest (SI Appendix, Fig. S2D). Nevertheless, both locus-specific inspection and genome-wide analysis of the obtained libraries revealed reproducible profiles (Fig. 2C) and characteristic accumulation of reads into regions in proximity of transcription start sites (TSS; Fig. 2D). Moreover, the pairwise correlation of genome-wide read densities demonstrated the high accuracy of our approach, since we uncovered a high correlation between all library pairs (between r = 0.78 and r = 0.96). This included ENCODE−FloChIP pairs, among which the highest correlation was obtained for the 100,000 cell sample, that is, r = 0.88 (Fig. 2E and SI Appendix, Fig. S2E). These values support the robustness of our approach, since they are in line with those from a previous study (3) which assessed ChIP-seq data reproducibility and antibody quality and concluded that correlation values of ∼0.8 between matched conditions are entirely acceptable. In addition, to evaluate individual chip-to-chip variation, we analyzed the correlation between two libraries obtained from 100,000 cells in identical conditions but derived from different FloChIP devices. Again, we obtained high correlation (r = 0.98; SI Appendix, Fig. S2F), indicating that our system is refractory to batch variability.

Next, we set out to evaluate the reproducibility of our approach with other genomic targets. To this end, by using FloChIP’s mode 2 - antibody multiplex, we performed ChIP-seq of four histone marks in parallel (H3K27ac, H3K4me3, H3K4me1, and H3K9me3) using the same sample, going from chromatin to sequencing-ready libraries, in 1 d. Following qPCR enrichment analysis (SI Appendix, Fig. S2G) and sequencing (SI Appendix, Fig. S2H), we found that the obtained signal tracks closely resemble those of ENCODE (Fig. 2F for H3K27ac, H3K4me1, and H3K4me3; Fig. 2I for H3K9me3). In addition, to evaluate FloChIP’s performance with greater resolution, we determined the extent of genome-wide distribution of reads around the TSS (Fig. 2H) and the overall correlation (see Methods) between FloChIP and ENCODE datasets (26–29). Comparison of signal intensities between the respective datasets confirmed an overall high read density correlation in peaks (H3K4me3: r = 0.85; H3K27ac: r = 0.90; H3K4me1: r = 0.89; H3K9me3: r = 0.87; Fig. 2I). Moreover, comparison in terms of the FRiP score showed that, despite the ChIP cell input for ENCODE being two orders of magnitude greater than that of FloChIP, our technology consistently yields highly enriched libraries, with FRiP scores between 1.1× and 4.1× higher for FloChIP compared to ENCODE (Fig. 2J). Finally, to investigate the probability of cross-talk across FloChIP lanes, we performed FloChIP on H3K27ac and H3K4me3 for human (GM12878 cells) and mouse (mESCs) chromatin, each loaded in adjacent IP lanes. Our results were compared to human ENCODE data for each respective mark as a control, revealing that virtually no such cross-talk could be observed between FloChIP lanes (SI Appendix, Fig. S2I and Table S1). These data show that FloChIP can be used to robustly generate chromatin landscapes for histone marks on the same sample in a parallelized manner and over a wide input range.

FloChIP's “Sequential IP” Mode Provides Genome-Wide Information on Bivalent Regulatory Regions.

Conventional ChIP-seq provides information on the genome-wide localization of one specific protein or histone mark at a time. However, DNA regulatory elements tend to be characterized by much more complex chromatin states that involve multiple histone marks and collaborating TFs (5, 16, 30). For instance, it has been shown that promoters showing both repressive (H3K27me3) and active (H3K4me3) marks are a characteristic feature in embryonic stem (ES) cells (16, 31). This class of promoters was originally named “bivalent” (16, 31) and is strongly associated with key spatially regulated developmental genes (32). To obtain direct information on the genomic location of bivalent promoters, a variant of the standard ChIP protocol called sequential ChIP was developed (16). Despite the advantage of sequential ChIP over standard ChIP in discerning true bivalency, its manual involvement and laboriousness have thus far prevented widespread usage. To address the technical limitations of the current sequential ChIP workflow, we exploited FloChIP’s intrinsic modularity, highly efficient IP, and multiplexing features to generate an automated and miniaturized sequential ChIP solution (Fig. 3A). Briefly, FloChIP’s “sequential IP” consists of two consecutive IPs taking place in two adjacent IP lanes. The chromatin immobilized and washed in the first IP lane is resuspended by means of a peptide elution strategy (14) and then transferred on-chip to a neighboring IP lane where the second IP is carried out (Fig. 3A), followed by salt washes and tagmentation.

Fig. 3.

FloChIP’s “sequential IP” mode for the study of factor cooccupancies including bivalent chromatin in mouse ES cells E14TG2a. (A) FloChIP’s sequential IP steps in descending chronological order as applied for H3K4me3−H3K27me3 cooccupancy. Chromatin that is derived from the first IP is collected through a peptide elution strategy (14) into off-chip reservoirs connected to the device. Following collection, the control channels are actuated to isolate the first IP lane from the chromatin, while opening the path to the second IP lane. At this point, the chromatin flows into the second prefunctionalized IP lane. Finally, the bivalent chromatin is eluted again in off-chip reservoirs. (B) Normalized signal tracks for the two individual IP libraries (H3K4me3 and H3K27me3) as well as the corresponding sequential IP samples (H3K27me3/H3K4me3 and H3K4me3/ H3K27me3) at three different genomic loci presented in the original study by Bernstein et al. (16). (C) Normalized signal tracks for the two IP samples (H3K27me3/H3K4me3 and H3K4me3/ H3K27me3) at different genomic loci reported as bivalent in the Co-ChIP study (15). (D) Normalized correlation plots (see Methods) between FloChIP in sequential ChIP mode in the two orders (i.e., H3K27me3 followed by H3K4me3 and vice-versa) and the corresponding Co-ChIP replicates.

We validated this approach by focusing on bivalent chromatin in embryonic development, given its well-studied role in this context. Specifically, we acquired genome-wide direct cooccupancy profiles for H3K27me3 and H3K4me3 in mESCs in both IP directions—that is, H3K27me3 first followed by H3K4me3 (H3K27me3/H3K4me3) and vice versa. As mentioned above, H3K4me3 and H3K27me3 bivalency has been originally attributed to promoters of developmental genes, leading to the hypothesis that a bivalent state maintains genes in a poised state (16). To evaluate the performance of FloChIP’s “sequential IP” mode, we first focused on three distinct regions that have been previously used as proof-of-concept models by Bernstein et al. (16), who illustrated the methylation status difference of these regions by using ChIP-qPCR and sequential ChIP-qPCR. With this approach, they were able to distinguish regions displaying only H3K4me3 (e.g., Tcf4 TSS) and only H3K27me3 (e.g., upstream of Hoxa3) versus those displaying true bivalency (e.g., Irx2 TSS). FloChIP-based genomic profiles (Fig. 3 B and C and SI Appendix, Fig. S3) validated these previous findings (15, 16). We observed that, as in the original study by Bernstein et al., the Tcf4 promoter shows high H3K4me3 but low H3K27me3 enrichment, and thus low bivalency. On the other hand, Hoxa3 was mainly marked by H3K27me3, with low H3K4me3 enrichment and consequently low bivalency signal (Fig. 3B). Finally, the TSS of Irx2 showed high bivalency, with all four genomic tracks, two individual and two sequential FloChIPs, showing high coverage. Moreover, the promoter of genes such as Gata3, Gata4, Tal1, and the Hoxc clusters showed pronounced bivalency, as expected for important developmental genes (15, 33) (Fig. 3C). Next to analyzing specific loci, we also validated our data on a broader scale by measuring the number of reads mapping into the TSS of known bivalent promoters (SI Appendix, Table S2) and achieving high correlation (between r = 0.87 and r = 0.93) with the results obtained by Weiner et al. (15) using their Co-ChIP system (Fig. 3D). Taken together, our data indicate that FloChIP’s “sequential IP” mode constitutes a miniaturized, low-input (100,000 cells) and rapid (between 5 h and 6 h) sequential ChIP-seq workflow for the genome-wide analysis of cooccurring binding events.

FloChIP Robustly Profiles TF Binding.

As mentioned above, previous attempts at improving either the sensitivity or multiplexing ability of ChIP-seq experiments were successful but mostly, so far, in the context of histone marks (21–24). In fact, performing ChIP on TFs poses additional challenges compared to ChIP on histone marks, such as the fact that 1) TF/DNA interactions are less abundant and weaker than histone mark/DNA interactions and 2) antibodies for TFs normally show lower affinity for their epitopes compared to histone mark antibodies. These challenges translate into the need for more-abundant sample inputs and longer incubation times. Indeed, during FloChIP optimization, we also experienced these challenges, rendering FloChIP’s indirect method—that is, with 2- to 4-h antibody/chromatin preincubation in tubes—to be the only robust way to obtain high-quality TF ChIP results. We benchmarked our system on different TF targets that were already assayed in LCLs by the ENCODE consortium, but using only 100,000 cells per ChIP: C/EBPβ, PU.1, and RUNX3 (SI Appendix, Fig. S4). For RUNX3 and PU.1, we obtained good agreement and correlation between our data and ENCODE’s (RUNX3: average [avg] r = 0.75; PU.1: avg r = 0.72) and high correlations between technical replicates (RUNX3: r = 0.95; PU.1: r = 0.9; SI Appendix, Fig. S4), once again highlighting the reproducibility of our system. For C/EBPβ, results were more ambiguous, since the FloChIP and ENCODE libraries did not correlate well (between r = 0.49 and r = 0.52) (SI Appendix, Fig. S4I). However, we found that the ENCODE replicates themselves did not have a high correlation (r = 0.51), which questions the reliability of this dataset (SI Appendix, Fig. S4I). In contrast, FloChIP replicates still yielded a high correlation (r = 0.97), which suggests that our FloChIP data are at least more consistent than ENCODE’s (SI Appendix, Fig. S4I). In addition, genome-wide quality measures such as FRiP score (SI Appendix, Fig. S4E) and motif enrichment (SI Appendix, Fig. S4H) also revealed high quality and reproducibility of FloChIP data. For example, the PU.1, RUNX3, and C/EBPβ motifs were consistently found to be the top hits in their respective libraries (SI Appendix).

In order to demonstrate the applicability of our approach on TFs, we performed MEF2A ChIP-seq on chromatin derived from distinct LCLs derived from 32 unrelated European individuals whose genomes were sequenced and aligned to the hg19 assembly as part of the 1000 Genomes Project (34) (Fig. 4A). We specifically targeted this TF, given its association with variable chromatin modules that were inferred from histone mark and PU.1 ChIP-seq data from LCLs, as presented in one of our previous studies (4). Before sequencing, we verified the IP quality of each library by qPCR (Fig. 4B). Fold change results indicated consistent, high enrichment across the 32 IP lanes (avg(Fold enrichment) = 70). In addition, we noticed that, during library preparation, the number of amplification cycles that was required to obtain sufficient DNA amounts for NGS sequencing was the same for all 32 samples (17 PCR cycles), reflecting homogeneity in DNA yield. This result was obtained without manually adjusting the volume or concentration of the sample prior to IP. We therefore reasoned that FloChIP itself, when used in saturating conditions, provides the additional advantage of equalizing the amount of recovered DNA, which renders the library preparation process straightforward. After sequencing, we confirmed the accumulation of mapped reads in selected genomic loci (based on MEF2A ENCODE data: upstream CCL3, VOPP1, and upstream USP7; Fig. 4C).

Fig. 4.

FloChIP’s MEF2A IP on 32 cell lines. (A) List of the 32 cell lines used in this study. (B) The qPCR enrichment for each library at the VOPP1 locus. The average fold enrichment across all libraries is 70, represented as a dotted line. (C) Signal tracks are reported for each library for three different genomic loci. (D) Number of peaks called for each library (8,600 peaks on average, represented by a dotted line). (E) MEF2A motif enrichment for each library (the median P value is 1e-10, represented as a dotted line). (F) Percent of called peaks containing the MEF2A motif (average is 15%, represented as a dotted line). (G) Results of de novo motif search and enrichment of each detected motif across libraries. (H) Venn diagram showing the percentages of MEF2A-bound promoters and the occurrence of other detected motifs. (I) Gene ontology enrichment analysis of promoters containing MEF2A, BATF, RUNX3, and IRF:PU.1 motif.

Subsequently, we again analyzed the quality of our data using a number of genome-wide measures. First, we called peaks for each sample (Fig. 4D; avg(#peaks) = 8,600) and observed a generally good degree of read density correlation between sample pairs (avg(cor) = 0.84) (SI Appendix, Fig. S5A). Surprisingly, despite detecting strong enrichment of the MEF2-like consensus motif in each individual peak set (Fig. 4E, median(pval) = 1e-10), examining the presence of motifs in peaks revealed that only a small portion of peaks contained the MEF2A motif (Fig. 4F; avg(motif density in peaks) = 15%). Therefore, we set out to explore alternative drivers of the observed MEF2A binding. To this end, we performed de novo motif discovery on each individual peak file and ranked the obtained motifs according to their statistical significance. Interestingly, the MEF2A motif emerged as only the fifth most enriched one, preceded by the motifs for BATF, RUNX3, NFKB, and the IRF:PU.1 dimer (Fig. 4G). This observation is consistent with previous motif-based analyses in LCLs (4, 35) and suggests that MEF2A DNA binding could be largely motif independent and driven by the interaction with other factors (36). Of note is that a previous study aimed at deciphering the Epstein–Barr virus-based mechanism of B cell/lymphoblastoid conversion had also found the same motifs to be strongly enriched in their EBNA3C ChIP-seq dataset (35). Thus, we hypothesized that, in LCLs, MEF2A may play a role in the gene regulatory network underlying B cell proliferation. To test this hypothesis, we first split the observed peaks in unique and overlapping sets based on the presence of one or more of the MEF2A motifs and the three other motifs with the highest enrichment, that is, IRF:PU.1, RUNX3 and BATF (Fig. 4H). We found that the largest set of the resulting Venn diagram consists of peaks containing all four motifs (20% of all peaks) and is enriched for ontology terms linked to immune cell regulation and proliferation (Fig. 4I and SI Appendix, Fig. S5B). In comparison, the second largest set (13%) involved peaks featuring, exclusively, the MEF2A motif and did not show insightful ontology enrichment. These findings indicate that MEF2A may play a role in the gene regulatory network underlying lymphoblastoid cell proliferation and that it does so specifically as part of a larger complex of collaborating TFs.

Next, we examined the impact of genetic variation on MEF2A binding. To achieve this, we exploited our large MEF2A ChIP-seq dataset by considering allele-specific binding (ASB; see Methods) across all 32 samples and found that only a very small fraction (0.7%) of ASBs could be explained by MEF2A motif variation (Fig. 5A and SI Appendix, Fig. S6 A–D). Indeed, only 5 out of 751 ASBs were significantly disrupting or creating an MEF2 motif (considering the entire MEF2 family) at false discovery rate (FDR) < 5% and were concordant in terms of ASB allele effect. Similarly, we found 3 ASBs for BATF, 19 for RUNX, and 2 for IRF (∼4% of all ASBs). In total, we tested 401 mononucleotide human core TF motifs [from HOCOMOCO v11 (37)], revealing only 223 ASBs that were significantly disrupting at least one motif at FDR 5% (∼30%). This result is consistent with our observation that few MEF2A DNA binding events are dependent on MEF2A's own motif and also aligns with the notion that the majority of variable TF DNA binding events are driven by motif-independent mechanisms (5, 36).

Fig. 5.

Genetic variation-based analyses reveal RUNX3 to be an important mediator of MEF2A DNA binding. (A) Proportion of the 751 identified ASBs that can be explained by motif disruption (at FDR 5%). RUNX, IRF, and BATF sum the results of all of the TF members in their family (because the respective motifs are very close). Only ASBs that are concordant with motif disruption or creation are counted. (B) ABC test for inferring motif disruption of other TFs that are modulating MEF2A DNA binding. The x axis represents the motif score/quality difference (for MEF2A in red, and RUNX3 in blue), while the y axis represents the fold change between the number of reads mapping to the ref vs. the alt allele, summed over all heterozygous samples. This was computed for all 5 identified MEF2A ASBs, and 22 RUNX3 ASBs (at FDR 5%). (C) The rs6912511 significantly disrupts the RUNX3 motif (FDR < 5%). (D) Allelic Imbalance highlighted for rs6912511 by summing read counts of each allele over all heterozygous samples. A binomial test yielded a P value of 6.35E-3, which revealed rs6912511 as an ASB. (E) IGV tracks showing MEF2A binding enrichment at the rs6912511 locus. The coverage tracks are from ENCODE (NA12878 MEF2A) and our 32 samples, stratified by rs6912511 genotype, and merged into three tracks (final bigwigs are normalized by reads per kilobase per million mapped reads (RPKM) using deepTools). (F) Boxplots showing the effect of rs6912511 on the three different genotypes (AA, AG, and GG). MEF2A binding enrichment (y axis) was computed from the peak in which rs6912511 was found, and normalized using DESEq2, followed by qqnorm functions in R.

Interestingly, and as indicated above, we found that a large portion of motif-affected ASBs are linked to RUNX-family binding sites (Fig. 5A). In agreement with this, allelic binding cooperativity (ABC) analysis showed that variation in RUNX-like motifs was significantly correlated with MEF2A DNA binding differences (nominal p-val < 5%), pointing to cooperative DNA binding mechanisms between these TFs (Fig. 5B). Indeed, an ABC (see Methods) test revealed that RUNX3 (with 22 single-nucleotide polymorphisms [SNPs] tested) is one of the top significant cobinders with MEF2A (P value = 0.02). MEF2A and MEF2D motifs also appeared in the top seven list, with P values of 0.068 and 0.051, respectively. Of note is that, since the motifs of RUNX factors are very similar to one another, our results cannot specifically distinguish between RUNX1, RUNX2, and RUNX3. Nevertheless, given that only RUNX3 ChIP-seq data exist for LCLs (https://www.encodeproject.org/targets/RUNX3-human/ and this study) and that previous evidence also linked RUNX3 to B cell proliferation pathways (35), we narrowed our analyses to this TF. Interestingly, within the set of RUNX3 ASBs, we detected several cases in which genetic variation from the reference sequence either creates (Fig. 5C) or disrupts a RUNX motif (SI Appendix, Fig. S6E). Accordingly, RUNX motif creation induced MEF2A DNA binding to the alternate allele (Fig. 5 D–F), whereas motif disruption led to MEF2A ChIP signal loss (SI Appendix, Fig. S6 F–H). Taken together, these genetic variation-based results suggest that, within the TF complex governing lymphoblastoid cell proliferation, RUNX3 and MEF2A cooperate, with RUNX3 acting as a key driver of MEF2A DNA binding. In conclusion, our findings demonstrate the value of FloChIP in catalyzing the relatively straightforward acquisition of TF binding data across many genotypes.

Discussion

Profiling the interactions between proteins and DNA has both fundamental and biomedical value (2, 38, 39), but continues to constitute an important technological challenge for genomics research (2, 3). ChIP-seq allows the probing of protein−DNA interactions on a genome-wide scale, thus achieving high throughput in terms of DNA sequence space coverage. However, in terms of experimental output, ChIP-seq largely remains a low-throughput technique. This is mainly due to the long and manually intensive ChIP-seq pipeline which, although widespread, only offers limited automation, reproducibility, and sensitivity. Valuable efforts have already been devoted to addressing ChIP-seq’s main limitations, but these efforts tended to target only specific issues while skipping others, as outlined in the Introduction (14, 15, 18, 20–25). Orthogonal approaches have recently emerged, such as cleavage under targets and release using nuclease (CUT&RUN) (40) and cleavage under targets and tagmentation (CUT&Tag) (41), that are capable of profiling chromatin in a one-tube format and down to single cells. Rather than relying on solid-state separation of protein/DNA complexes, these approaches exploit fusion proteins (protein A-Mnase and protein A-Tn5, respectively) to selectively digest or tagment the genomic DNA in the proximity of chromatin-bound antibodies. Such alternative strategies hold great potential as sensitive and streamlined techniques for genomic profiling of protein/DNA complexes, even down to the single-cell level. But, as opposed to ChIP-seq, they still need to establish themselves as widely implemented tools, which remains uncertain due to the fact that data correlations tend to still be greater among conventional ChIP-seq tools than between these novel and already established approaches (e.g., ENCODE data) (40, 41) or FloChIP, and they cannot, at this stage, be used to perform sequential IP (SI Appendix, Table S3). To more comprehensively address the limitations of standard ChIP-seq, we developed a microfluidic system that we named FloChIP. In this work, we demonstrate that FloChIP enables both rapid single and sequential IP across a wide input range with high target flexibility and experimental scalability.

We first characterized FloChIP’s dynamic range by targeting H3K27ac in samples containing 500 up to 1 million cells. We show high fold enrichment and good genomic coverage for all tested samples, indicating that FloChIP performs robustly across a wide cell number range. In a second step, we tested different antibodies in what we define as the “antibody multiplex mode.” In this configuration, the different IP lanes of our device are individually functionalized with distinct antibodies, with one sample being distributed in parallel to all IP units. The geometric layout of the microchannels thereby ensures uniform distribution of the sample (Fig. 1E and SI Appendix, Fig. S1C). After observing good correlations with the benchmarking data (ENCODE) for all tested histone mark targets, we conclude that FloChIP can generate chromatin state landscapes with reliability and flexibility.

After establishing this proof of concept on histone marks, we set out to expand the applicability of FloChIP in two directions: sequential ChIP-seq and ChIP-seq on TFs. For the former, we took advantage of the use of microvalves to compartmentalize distinct sections of the microfluidic device in a controllable manner, as reported previously for other applications (42, 43). More specifically, the use of microvalves allows easy orchestration of the functionalization of adjacent IP lanes with different antibodies, in this case, H3K4me3 and H3K27me3, and transfer of eluted chromatin from one IP lane to the next. With this approach, we recapitulated the landmark qPCR (16) and sequencing (15) results on the genomic distribution of H3K4me3/H3K27me3 bivalency in ES cells. Here, we note that the key advantage of FloChIP lies in turning sequential ChIP assays from a long (>2 d), intensive, and error-prone protocol into a fast (half-day) and automated procedure. We therefore believe that FloChIP could catalyze renewed interest in histone mark bivalency (15, 17), whose molecular function and relevance remains poorly understood, in part, because of the cumbersome nature of the sequential ChIP-seq technique.

Consistent with the general purpose scope of our microfluidic ChIP-seq system, our final goal was to demonstrate the ability of FloChIP to carry out IP on TF targets. Such capacity would constitute important technical progress, since previous microfluidic or multiplexing implementations of ChIP focused mostly on histone marks (21, 23, 24). To cope with the generally lower affinity of TF antibodies for their targets, we complemented the regular FloChIP procedure with antibody−chromatin preincubation and oscillatory sample loading (44). We first benchmarked FloChIP on three well-studied TFs (PU.1, RUNX3, and C/EBPβ), after which we further validated our platform by completing 32 MEF2A IPs and on-chip tagmentations in less than 1 d. After sequencing, we used the data to investigate the role of MEF2A in the gene regulatory network underlying lymphoblastoid cell proliferation. By integrating genotypic and molecular phenotypic data, we found that MEF2A DNA binding is, in large part, controlled by other TFs such as RUNX3, which, as part of a larger complex of collaborating TFs, seem to coordinate B cell proliferation. Importantly, both sequential ChIP-seq and TF ChIP-seq experiments were carried out on samples of 100,000 cells, therefore constituting a significant improvement in sensitivity for these ChIP-seq variants, which usually require several million cells (3).

Finally, we anticipate that FloChIP can still benefit from further optimization. Despite its ability to automate several steps, such as IP, washes, and tagmentation, FloChIP demands important preparatory hands-on work, for example, wiring the device and interfacing it to the control system. Moreover, even though FloChIP streamlines a significant portion of the ChIP-seq workflow, it remains sensitive to the pre-IP protocol steps and to the choice of specific antigen targets. In other words, similar to standard ChIP-seq, high-affinity antibodies and correctly fragmented chromatin will remain essential requirements for the correct functioning of FloChIP. Finally, while we expect that the implementation of FloChIP will be easily attainable by laboratories with basic microfluidic infrastructure, we aim to also serve the broader scientific community. To do so, we are currently working toward integrating FloChIP in a more compact solution that will allow laboratories without prior microfluidic expertise to adopt it as well. Specifically, we aim to develop a FloChIP benchtop instrument and off-the-shelf FloChIP cartridges, paralleling comparable efforts to, for example, accommodate single-cell transcriptomic analyses (45).

In conclusion, we believe that FloChIP has the potential to empower the community with a practical and reliable IP solution. Its future integration with standardized user-friendly devices will thereby pave the way toward full automation.

Methods

Detailed info can be found in SI Appendix, Supplementary Methods.

Cell Line Information.

For all comparisons with ENCODE data, LCL GM12878 was used. For the sequential ChIP experiments, we used mouse ES cells (E14TG2a) in order to enable the comparison with previous datasets (15, 16, 46). For the large-scale MEF2A experiments, we used the following LCLs: GM06985, GM06986, GM06994, GM07037, GM07048, GM07051, GM07056, GM07346, GM07357, GM10847, GM10851, GM11829, GM11830, GM11831, GM11832, GM11840, GM11881, GM11894, GM11918, GM11920, GM11931, GM11992, GM11993, GM11994, GM12005, GM12043, GM12154, GM12156, GM12249, GM12275, GM12282, GM12283, GM12286, GM12287, GM12383, GM12489, GM12750, GM12760, GM12761, GM12762, GM12763, GM12776, GM12812, GM12813, GM12814, GM12815, and GM12873.

Chromatin Preparation.

Lymphoblastoid and mouse embryonic cells were harvested, washed once with phosphate-buffered saline (PBS), and resuspended in 1 mL of cross-linking buffer. Cross-linking was quenched, and cells were then washed twice with ice-cold PBS, pelleted, deprived of the supernatant, snap frozen, and stored at −80 °C. The frozen cell pellet was resuspended in ice-cold PBS, lysed, and resuspended in sonication buffer. Nuclei were sonicated on a Covaris E220 machine (see SI Appendix; for representative fragment analyzer output, see SI Appendix, Fig. S7A).

FloChIP.

FloChIP devices were fabricated as previously reported (47) (details in SI Appendix, Supplementary Methods).

A FloChIP experiment starts with preloading the control lines with distilled water and activating all valves. Subsequently, all of the reagents required for the surface chemistry are loaded into pipette tips and inserted into the inlets of the microfluidic device. At this stage, all valves are closed, and there is virtually no cross-talk between any of the reagents above. Immediately after completing their insertion into the chip, the tips themselves are connected to solenoid valves for electronic pressure control, and the automated protocol is launched by running the respective script. The protocol entails, in sequential order, the following steps: 20 min of BSA−biotin (100 μL at 2 mg/mL), 30 s of PBS wash, 20 min of Neutravidin (100 μL at 1 mg/mL), 30 s of PBS wash, 20 min of biotin−protein A/G (100 μL at 2 mg/mL), and 30 s of PBS wash. The antibodies used in this study are anti-H3K27ac ab4729, anti-H3K4me3 ab8580, anti-H3K4me1 ab8895, anti-H3K9me3 ab8898, anti-H3K27me3 ab6147, anti-MEF2A sc-17785, anti-RUNX3 sc-101553, anti-PU.1 sc-390405, and anti-C/EBPβ sc-7962. Following antibody loading, chromatin samples are loaded on-chip by opening and closing the respective microvalves. Following IP, salt washes are performed to eliminate nonspecific binding. Subsequently, Tn5 buffer is flowed on-chip to tagment the immunoprecipitated chromatin.

Following Tn5 buffer, sodium dodecyl sulfate (SDS) is loaded on-chip at 65 °C for 10 min in order to elute the antibody-bound chromatin from the device. Of note is that the first elution step in the FloChIP sequential mode is performed through a peptide elution strategy (14) in order to preserve epitopes and the surface chemistry in the adjacent IP lanes. The eluate is independently collected from each IP lane into PCR tubes and de−cross-linked at 65 °C for 4 h. Following de−cross-linking, DNA is purified in Qiagen EB buffer using Qiagen MinElute purification kits (SI Appendix, Fig. S7B).

Species Mixing Experiment.

The species mixing experiment was carried out on two different FloChIP devices, one functionalized with anti-H3K27ac antibody and one with anti-H3K4me3 antibody. Mouse and human chromatin were then loaded in adjacent IP units in order to test for contamination across said units. During analysis, we first created a mixed genome, merging mouse genome mm10 (GRCm38) and human genome hg38 (GRCh38) of Ensembl release 93. All libraries were then aligned, using STAR (Spliced Transcripts Alignment to a Reference, v. 2.6.1b) (48), to this genome, using default parameters and “–outFilterMultimapNmax 1.” This parameter allows discarding of any multiply mapping reads and thus recovery of reads that map uniquely to the human or the mouse genome. Then, we counted the number of reads per chromosomes (here chromosome + species) using “samtools idxstats” and summed the reads belonging to the same species.

FloChIP Read Mapping and Processing for Histone Marks and TFs.

Sequencing reads were mapped to the human (hg38 for all histone mark and TF comparisons with ENCODE data; hg19 for the 32 samples of the 1000 Genomes Project) and mouse (mm10) genomes using STAR (48) with default parameters. We counted the number of reads that are falling into different features (introns, intergenic, CDS, etc.) using RSeQC read_distribution.py script (v.2.6.4) (49). This revealed no major anomalies, as most reads for all libraries mapped to either intergenic or intronic regions, with ENCODE libraries (especially for RUNX3, PU.1, and H3K9me3) tending to have a slightly greater proportion of intronically mapped reads (SI Appendix, Fig. S8). Uniquely mapped reads were used to call peaks using the HOMER (Hypergeometric Optimization of Motif EnRichment) (50) command “findPeaks.pl” with the appropriate flag, that is, “–histone” for histone marks and “–factor” for TFs with the parameter size set to 1,000 in the case of H3K27ac for decreasing cell numbers (SI Appendix, Table S4). FRiP scores were calculated using HOMER’s command “annotatePeaks.pl” on individual peak files, dividing the total number of reads that fall within peaks by the total number of mapped reads. Correlation scatter plots were generated using the “multiBamSummary” command of the deepTools package in “bins” mode, that is, by binning the entire genome and measuring read densities in each bin. Metaprofiles were generated using HOMER’s command “annotatePeaks.pl” in “hist” mode, which automatically normalizes each directory by the total number of mapped reads. Genomic coverage profiles were generated using the IGV (Integrative Genomics Viewer) browser; all profile groups of the same histone mark or TF were normalized by the same factor to allow a reliable comparison. Motif enrichment analysis of known motifs was performed with HOMER’s findMotifsGenome.pl for the following motifs: 1) Mef2a(MADS)/HL1-Mef2a.biotin-ChIP-Seq(GSE21529)/Homer, 2) BATF(bZIP)/Th17-BATF-ChIP-Seq(GSE39756)/Homer, 3) PU.1-IRF(ETS:IRF)/Bcell-PU.1-ChIP-Seq(GSE21512)/Homer, 4) PU.1(ETS)/ThioMac-PU.1-ChIP-Seq(GSE21512)/Homer, 5) RUNX(Runt)/HPC7-Runx1-ChIP-Seq(GSE22178)/Homer, and 6) CEBP(bZIP)/ThioMac-CEBPb-ChIP-Seq(GSE21512)/Homer.

Allele-Specific Binding.

For identifying variants subject to Allele-Specific Binding of MEF2A, we downloaded the 1000G genotyping data from the European Bioinformatics Institute (EBI) server (hg19) and removed variants with a minor allele frequency lower than 5%, while restricting our analysis to the 32 samples of interest, which yielded a total of 1,268,985 SNPs. Then, we applied the “ASEReadCounter” tool from GATK v.4.0.4.0, on each of the 32 samples. These results were then merged by summing read counts for every SNP across heterozygous samples. We filtered out all SNPs with a coverage lower than 10 reads, which yielded 6,330 SNPs. Then we opted to keep only SNPs falling into called peaks, which led to 4,554 SNPs that were further analyzed. On these, we performed a binomial test to assess significant allelic imbalance (nominal P value of 5%), yielding 751 potential ASBs (37 at FDR 5%).

TF Motif Disruption Analysis.

All downstream analyses were performed using R v. 3.5.0. We analyzed all 751 ASBs to check whether they were significantly impacting TF motifs using atSNP v.1.0 (51) with the BSgenome.Hsapiens.UCSC.hg19 package as genome library and SNPlocs.Hsapiens.dbSNP144.GRCh37 package as SNP library. We tested 401 mononucleotide human core TF motifs that were downloaded from HOCOMOCO v11 (37).

ABC.

ABC was assessed for all 751 ASBs, using a linear regression between the motif disruption/creation log likelihood ratio computed by atSNP (see Methods) and the fold change between the ASB ref and alt counts. This analysis was performed for all SNPs significantly disrupting any of the 401 motifs analyzed by atSNP at FDR 5%, and allowed us to find which ASBs were concordant with motif disruption across several tested SNPs.

Data Availability.

All data have been deposited on ArrayExpress (E-MTAB-7179, E-MTAB-7150, E-MTAB-7186, E-MTAB-8553, E-MTAB-8551).