High-resolution mapping of transcription factor binding sites on native chromatin

Robust ORGANIC profiles of Reb1 and Abf1 binding sites

We performed MNase digestion of uncrosslinked intact nuclei from S. cerevisiae strains expressing Reb1-FLAG and Abf1-FLAG, solubilized chromatin by needle extraction, and immunoprecipitated tagged transcription factors at 80, 150, or 600 mM NaCl to obtain different levels of stringency (Fig. 1a and Supplementary Fig. 1). We then prepared TF-bound and input DNAs for paired-end sequencing using a modified library preparation protocol¹⁹ (Fig. 1a and Supplementary Fig. 1). Consistent with immunoprecipitation of proteins with small footprints, we found that small fragments were enriched in Reb1 ChIP relative to input (Supplementary Fig. 2a), and we therefore profiled the <100 bp (len50) size class.

The Reb1 immunoprecipitated (IP) fractions showed sharp peaks over a negligible background relative to the corresponding input chromatin (Fig. 1b). Similar peaks were identified when fragments were not filtered by size (Supplementary Fig. 2). Interestingly, the len50 size class inputs showed strong peaks corresponding to Reb1 binding sites seen in the IP samples, though at a lower level of occupancy. In the input, we observed highly occupied peaks not corresponding to Reb1 binding sites in intergenic regions (len50 tracks, Fig. 1b). With increasing salt concentration, there was a dramatic reduction in both total number and dynamic range of ORGANIC peaks (Fig. 1b), consistent with disruption of relatively weak electrostatic TF-DNA interactions at low affinity sites. Some but not all ORGANIC peaks corresponded to Reb1 binding sites previously identified by ChIP-chip and ChIP-exo (Fig. 1b)^13,20. Similar results were obtained for Abf1 when compared to ChIP-chip data (Supplementary Fig. 3). For both Abf1 and Reb1, we observed a high degree of overlap between sites detected at different extents of MNase digestion (Supplementary Figs. 2–4). We conclude that ORGANIC profiling robustly detects both previously published and new Reb1 and Abf1 binding sites.

ORGANIC TF sites have characteristic sequence motifs

In order to characterize putative Reb1 and Abf1 binding sites, we applied a peak-calling algorithm with a conservative threshold to the len50 ChIP data and asked whether detected peaks were associated with characteristic consensus motifs using the MEME algorithm²¹. We identified 1,992 ORGANIC peaks in the Reb1 len50 size class 80 mM (low-salt) experiment (Fig. 2a). The low-salt ORGANIC Reb1 sites included 204 (83.3%) ChIP-chip and 935 (52.6%) ChIP-exo sites (Fig. 2b and Supplementary Fig. 5). Low-salt Abf1 ORGANIC peaks included 162 of 278 (58.3%) ChIP-chip peaks, whereas 600 mM (high-salt) Abf1 ORGANIC peaks identified more total sites (1,258), including 214 (of 278) sites also identified by ChIP-chip (Fig. 2b, d). The ORGANIC Reb1 and Abf1 motifs matched those reported in previous studies^13,20 (Fig. 2a, b).

We characterized the reproducibility of our method by performing pairwise comparisons of positions and occupancies of peaks called using independent biological replicates and from peak sets using varying salt concentrations, and found that datasets were well correlated (R² = 0.80 – 0.95, Supplementary Fig. 4). Occupancies at Reb1 sites called by both ChIP-exo and ORGANIC profiling were poorly correlated (R² < 0.05, Supplementary Fig. 5b). We conclude that ORGANIC profiling reproducibly captures a large fraction of previously published Abf1 and Reb1 binding sites, while identifying 2–8 fold more motif-associated sites than other methods.

ORGANIC profiles are highly sensitive and specific

Given the strong sequence specificities of Abf1²² and Reb1²³, we evaluated the accuracy of ORGANIC profiling by using the presence of a MEME-derived motif within a peak region as the ‘gold-standard’ for classifying a peak as a true positive. Strikingly, 99.3% of low-salt Reb1 sites contained the TTACCCG motif (Fig. 2a). The percentage of peaks containing the motif decreased to 61.5% at high salt (Fig. 2a). Whereas virtually all ORGANIC peaks had Reb1 motifs, only 59.6% of all ChIP-exo sites were found to be associated with a Reb1 motif¹³. In contrast to Reb1 ORGANIC sites, the 1,066 low-salt Abf1 ORGANIC sites contained a smaller percentage of peaks with motifs (63.3%) than peaks called using high-salt extraction (93.7%, Fig. 2b). We estimate a false negative rate of ~0.5% for ORGANIC profiling at Reb1 motifs (see Online Methods).

In order to evaluate the specificity of ORGANIC profiling, we determined how well peak sequences matched consensus motifs by scoring peaks using MEME-derived position-specific scoring matrices (PSSMs). Using the Reb1 ORGANIC PSSM, we found a distribution of high motif scores (true positives) with no strongly negative scores at low salt (Fig. 3a). When the salt concentration was increased to 150 mM and 600 mM, we observed a graded reduction in the fraction of true positive Reb1 sites and the appearance of strongly negative scores, giving a bimodal distribution (Fig. 3a). In comparison, ChIP-exo Reb1 calls included a high number of negative calls and showed a motif score distribution similar to the high-salt ORGANIC Reb1 dataset (Fig. 3a bottom panel). A similar trend was obtained using the ChIP-exo-derived PSSM (Supplementary Fig. 6).

Abf1 motif scores were also narrowly distributed (Fig. 3b) and, as expected from the increase in Abf1 motif-containing peaks at high salt, we observed a reduction in false-positive calls and an increase in true-positive calls at 600 mM salt (Fig. 3b). Given the structural differences in the Reb1 and Abf1 DNA-binding domains^24,25, it is likely that optimal extraction and ChIP conditions for the proteins differ, explaining the differential specificity of ORGANIC profiles across varying salt concentrations and different DNA binding proteins. The increase in apparent specificity to >90% at higher salt concentrations for Abf1 despite reduction in dynamic range (Supplementary Fig. 3) suggests that experimental parameters can be tailored to the DNA-binding protein of interest.

ORGANIC sites display DNaseI footprints and are conserved

In order to confirm that ORGANIC sites are bound in vivo, we used published S. cerevisiae DNaseI-Seq data²⁶ to ask whether sites detected by ORGANIC profiling are associated with classical footprints indicative of in vivo occupancy^26–28. For both Reb1 and Abf1, average DNaseI-Seq profiles at ORGANIC sites showed characteristic footprints (Fig. 4a, b). In contrast, average DNaseI-Seq tag counts at ChIP-exo sites did not show a footprint (Fig. 4c), except at the subset of sites that are also ORGANIC sites (Supplementary Fig. 5). We found that DNaseI footprint depth, which is correlated with in vivo occupancy^26,28, corresponds well with Reb1 and Abf1 ORGANIC site occupancies (Supplementary Figs. 7 and 8). These results suggest that ORGANIC sites are qualitatively occupied in vivo and also that relative occupancies determined by ORGANIC profiling are quantitatively correlated with in vivo binding.

In order to further exclude the possibility of TF redistribution during chromatin preparation or ChIP, we mixed equal numbers of isolated Drosophila S2 cell nuclei with Reb1-FLAG budding yeast nuclei and performed ORGANIC profiling at 150 mM salt. We expected that, if redistribution occurred, the ~300-fold excess of Drosophila sequences with Reb1 binding sites would be enriched in the ChIP fraction. However, we detected only a background level of Drosophila DNA in the ChIP fraction relative to the input (Fig. 4d). We detected good correlation between Reb1 sites detected in replicates of the mixing experiment and, consistent with stable binding under conditions used in ORGANIC profiling, we found a high level of correlation (R² = 0.995) between occupancies detected in experiments with mixed and unmixed nuclei (Supplementary Fig. 9a, b). The motif score distribution of Drosophila Reb1 peaks was dominated by negative scores (Supplementary Fig. 9c). These analyses suggest that Reb1 does not shift detectably from yeast to Drosophila chromatin during chromatin preparation and ChIP.

We also considered specificity of identified binding sites by analyzing the correlation between motif score and occupancy of Reb1 and Abf1. At low salt, we observed poor correlation between occupancy and motif score (R² = 0.02, Fig. 4e). Consistent with a bimodal distribution of motif scores between true positives and false positives (Fig. 3), we observed that low-scoring sites are poorly occupied by the TF, while high-scoring sites show an expected broad distribution of occupancies (Fig. 4e). A similar relationship between occupancy and motif score was observed at high salt and with ChIP-exo Reb1 sites (Supplementary Fig. 10). Because high motif scores are correlated with favorable binding energies²⁹, this analysis also suggests that TFs do not redistribute to thermodynamically favored binding sites during ORGANIC profiling.

We expected that the new Abf1 and Reb1 sites would show evolutionary conservation above background levels because conservation of TF binding sites implies purifying selection³⁰. We plotted phastCons scores³¹, which represent the probability that a given base is in a conserved region, in windows centered at Reb1 or Abf1 sites (Fig. 4f, g). Interestingly, we observed increased conservation of new sites at motif positions relative to background (Fig. 4f, g and Supplementary Fig. 5). In general, new sites had either higher or comparable conservation scores when compared to sites detected by ChIP-chip or ChIP-exo. Consistent with a role for Abf1 and Reb1 in positioning flanking nucleosomes at a subset of promoters^23,32, we detected well-positioned flanking nucleosomes around Reb1 and Abf1 sites (Fig. 5a, b). Since virtually all of the ORGANIC Reb1 sites have a Reb1 binding motif, we asked whether there is a difference in motif strength or TF occupancy that could explain differential phasing of nucleosomes. We ranked ORGANIC sites by nucleosome occupancy and considered the top and bottom 200 sites, which corresponded to sites with relative nucleosome occlusion and depletion, respectively (Fig. 5a, b and Supplementary Fig. 11). We detected no substantial difference in MEME-derived motifs between the two groups (Fig. 5a, b), suggesting that the degree of nucleosome phasing is not associated with motif strength.

ORGANIC profiles do not require input normalization

Sequencing formaldehyde cross-linked and sonicated chromatin (Sono-Seq) is known to preferentially recover regions of accessible chromatin⁹. In order to characterize the observed differences in MNase protection, we performed Sono-Seq and generated profiles of average normalized Sono-Seq counts in 2-kb windows centered at Reb1 binding sites determined by different methods. Strikingly, ChIP-exo Reb1 sites showed enrichment for Sono-Seq reads, whereas there was virtually no enrichment at ORGANIC or ChIP-chip Reb1 sites (Fig. 5c and Supplementary Fig. 5f). Similarly, we detected no enrichment of Sono-Seq reads at ORGANIC or ChIP-chip Abf1 sites (Fig. 5d). Sono-Seq enrichment is consistent with the observation of increased DNase cleavage²⁶ at ChIP-exo sites and comparatively lower cleavage at ORGANIC sites (Fig. 4a–c). We obtained similar results using previously published Sono-Seq data (Supplementary Fig. 12) and using sensitivity to MNase digestion as an independent measure of chromatin accessibility (Supplementary Results). The Reb1 ChIP-chip study was performed using a two-color spotted microarray on which ChIP DNA was co-hybridized with DNA from an un-enriched sample (input)³³; this input normalization procedure likely corrected for bias from preference for accessible chromatin, as the input sample was obtained from cross-linked and fragmented DNA. In contrast, input normalization is not performed with ChIP-exo, which likely explains the observed strong preference for binding sites in accessible chromatin. Unlike the Reb1 ChIP-exo map, the ORGANIC accessibility profile is similar to input-normalized ChIP-chip. As the cross-linking and sonication steps of ChIP-exo are essentially the same as those used to obtain accessible chromatin by Sono-Seq^9,13, the absence of these steps could account for the insensitivity of ORGANIC profiling to the degree of chromatin accessibility. We conclude that ORGANIC maps do not show chromatin accessibility preferences and therefore do not require input normalization.

Highly specific ORGANIC profiles of Drosophila TFs

We assessed the applicability of ORGANIC profiling to eukaryotes with larger genomes by mapping GAF and Psq from Drosophila S2 cell active chromatin extracted with 80 mM salt³⁴ (Supplementary Fig. 1). In order to determine whether GAF is lost from the nucleus under native conditions, we summed losses incurred during processing of nuclei during the ORGANIC and modENCODE X-ChIP protocols. In both cases, ~15–20% of total GAF was lost (Supplementary Fig. 13). We observed enrichment for peaks in the len25 (1–50 bp) and len50 size class ChIP fractions relative to input for both GAF and Psq (Supplementary Fig. 14). Enriched peaks were associated with DNaseI hypersensitive sites and were evolutionarily conserved, suggesting they represent bona fide in vivo sites (Supplementary Fig. 15). Consistent with previous work demonstrating that GAF and Psq heterodimerize and act in concert at many loci³⁵, we observed similar genome-wide profiles for GAF and Psq. Using the same peak-calling method and de novo motif analysis approach to characterize yeast TF binding sites, we called 3,300 GAF and 957 Psq sites and recovered expected GAG-repeat containing motifs in 76.5% of GAF and 40% of Psq ORGANIC peaks (Fig. 6a). In contrast, ChIP-chip identified 4,567 GAF sites³⁶, of which only ~5% had characteristic motifs. Therefore, ORGANIC profiling is greater than an order of magnitude more specific than ChIP-chip for factor binding in the Drosophila genome.

In addition to sites with the characteristic GAF motif, GAF is known from X-ChIP-chip data to bind TF ‘hotspots’^36,37, which are thought to reflect dynamic, low-affinity binding of multiple transcription factors³⁷. Remarkably, TF hotspots were absent from GAF ORGANIC peak calls (Fig. 6b and Supplementary Fig. 14). When we searched for the GAF consensus motif among TF hotspot regions, only 17.5% displayed such a stringent motif. This lack of robust GAF motifs at TF hotspots can account for the absence of ORGANIC signals at these sites, despite ready detection by X-ChIP. Consistent with the cross-linking of highly transient interactions at commonly used formaldehyde crosslinking times¹², we suggest that the dynamic binding of GAF at TF hotspots resulted in trapping of transiently bound GAF by formaldehyde cross-linking to these sites when using X-ChIP. In contrast, ORGANIC profiling detects only sites that are stably bound under native extraction conditions, and so the lack of robust GAF motifs at TF hotspots is further evidence that ORGANIC maps are highly specific for stable, direct TF-DNA interactions.