High-throughput single-cell epigenomic profiling by targeted insertion of promoters (TIP-seq)

Abstract

Chromatin profiling in single cells has been extremely challenging and almost exclusively limited to histone proteins. In cases where single-cell methods have shown promise, many require highly specialized equipment or cell type-specific protocols and are relatively low throughput. Here, we combine the advantages of tagmentation, linear amplification, and combinatorial indexing to produce a high-throughput single-cell DNA binding site mapping method that is simple, inexpensive, and capable of multiplexing several independent samples per experiment. Targeted insertion of promoters sequencing (TIP-seq) uses Tn5 fused to proteinA to insert a T7 RNA polymerase promoter adjacent to a chromatin protein of interest. Linear amplification of flanking DNA with T7 polymerase before sequencing library preparation provides ∼10-fold higher unique reads per single cell compared with other methods. We applied TIP-seq to map histone modifications, RNA polymerase II (RNAPII), and transcription factor CTCF binding sites in single human and mouse cells.

Figure 1.

Overview of TIP-seq, a robust low-cell mapping method that combines CUT&Tag with RNA-mediated linear amplification. (A) See text for details. (B) Custom ME-T7p transposon used in TIP-seq to insert T7 promoters near antibody-bound targets. Standard CUT&Tag ME-A shown for comparison. *Upstream AT-rich stabilizer sequence increases affinity of T7 RNAP for promoter and the efficiency of promoter clearance (Tang et al., 2005). Ab., antibody; pol., polymerase; tspn, transposon.

Figure 2.

Bulk TIP-seq substantially increases library complexity and sensitivity over PCR-based library preparation. (A) Histone modifications. IGV track view across a 3-Mb segment of the human genome for H3K27me3 TIP-seq on 10,000, 1,000, 100, and 50 HCT116 cells. Tracks show CUT&Tag data in 10,000 cells for comparison. TIP-seq for normal IgG in 1,000 cells shown as negative (neg.) control. (B) TFs. IGV track view across a 3-Mb segment of the human genome showing TIP-seq targeting CTCF in 5,000, 1,000, 100, and 50 HCT116 cells and two replicates of CUT&Tag data in 5,000 cells performed in parallel with TIP-seq. Top track shows ENCODE ChIP-seq for comparison. CUT&Tag samples from 1,000, 100, and 50 cells failed to yield sufficient library DNA and/or sequencing reads. (C) Single cells. IGV track view showing TIP-seq data collected from HCT116 single cells. Cells were processed and tagmented in bulk until cells were FACS sorted into individual PCR tubes for DNA purification and subsequent IVT and library preparation. (D) Library complexity of CTCF TIP-seq vs. CUT&Tag as a fraction of unique reads (red) or T7-duplicate reads (gray) over total reads. Samples were processed in parallel, with CUT&Tag (pA-Tn5 loaded with ME-A/B adapters), or TIP-seq (pA-Tn5 loaded with ME-T7 transposons), and after tagmentation and DNA purification, CUT&Tag DNA was PCR amplified using 15 cycles, while TIP-seq DNA was processed as described in Fig. 1 and indexed with nine PCR cycles. Libraries were pooled to equimolar ratios and paired-end sequenced. (E) Peak enrichment heatmaps of CTCF TIP-seq vs. CUT&Tag surrounding ±2 kb ENCODE CTCF peaks. Samples were normalized to the sum of per-base read coverage and scaled to 1× genome coverage before plotting heatmaps with deepTools. (F) Pearson correlations among bulk TIP-seq, single-cell TIP-seq, and ENCODE ChIP-seq using 50-kb bins.

Figure S1.

Bulk TIP-seq quality metrics. (A) H3K27me3 Pearson correlations among TIP-seq, CUT&Tag, and ENCODE ChIP-seq targeting H3K27me3 in an HCT116 cell using 50-kb bins. (B) H3K27me3 Venn diagram displaying peak overlap among samples. (C) CTCF Pearson correlations among TIP-seq, CUT&Tag, and ENCODE ChIP-seq targeting TF CTCF in an HCT116 cell using 50-kb bins. (D) CTCF Venn diagram displaying peak overlap among samples. (E) MEME logo representations for CTCF bulk, single-cell aggregate (Agg.), and five single cells chosen at random showing the most prevalent motif identified for each respective library and the P value associated with it.

Figure 3.

Overview of sciTIP-seq. (A) Cells are harvested, permeabilized, and treated with primary and secondary antibodies in bulk. Cells are counted and distributed to a 96-well plate (∼2,000 cells/well) where they are incubated with custom, uniquely indexed pA-Tn5. Cells are washed to remove unbound pA-Tn5 before activating tagmentation. Tagmentation is terminated by addition of EDTA, and cells are pooled together and redistributed at random to a new 96-well plate (∼15–100 cells/well, depending on number of barcode combinations used during index 1). DNA undergoes a gap-fill reaction via Taq polymerase and IVT via T7 RNAP. RNA is purified and reverse transcribed using a random hexamer primer, then primed for second-strand synthesis using primer complementary to the ME-T7 transposon. ME-B–only Tn5 was used to simultaneously fragment and adapter-tag 3′ end of cDNA to prepare for PCR indexing. (B) Resulting library fragments contain an r5 index added during targeted pA-Tn5 tagmentation and an i5 and i7 index added during PCR to enable retroactive demultiplexing of single cell reads. Ab., antibody; exp., experimental; pol., polymerase; tspn, transposon.

Figure S2.

sciTIP-seq quality metrics. (A) Violin plots showing the number of unique reads per cell for each sample before filtering out low-read cells. (B) Violin plots showing the percent unique reads for each sample before filtering out low-read cells. (C) Violin plots showing the percent unique reads for each sample after filtering out low-read cells. (D) PCR duplication rate vs. read count for each single cell (samples grouped by color). Mean and median duplication rates (after filtering out cells with <1,000 reads). (E) T7 duplication rate vs. read count for each single cell (samples grouped by color). (F) Overall (PCR and T7) duplication rate vs. read count for each single cell (samples grouped by color). Mean and median duplication rates (after filtering out cells with <1,000 reads) are shown in bottom right corner of duplication scatterplots D–F. (G) Table displaying the mean PCR, T7, and overall duplication rates before and after filtering of low-reads cells. (H) Barnyard analysis scatterplot displaying cross mapping of mouse and human cells to mm10 and hg38 reference genomes. 5.6% of cells had >10% reads mapping to both reference genomes after removal of cells with <200 reads. (I) FRiP violin plot with cell read counts (x axis) showing that the cohort of cells with fewer reads retains a high signal-to-noise ratio. dup, duplication; frag, fragment.

Figure 4.

Multiplexed sciTIP-seq in F121-9 mESCs. (A) Distribution of unique reads of 3,557 single cells in F121-9 mESCs and human HCT116 cells after removal of PCR/T7 duplicates and filtering out cells with <1,000 reads. Red line indicates the mean (38,054) unique reads per cell. (B) Violin plots showing the number of unique reads per cell for each sample after filtering. (C) Violin plots showing the FRiP for each respective sample. (D) IGV track view across an 11-Mb segment of the mouse genome for sciTIP-seq. Tracks show 100 single cells together with both pseudo-bulk (aggregate) and bulk TIP-seq/CUT&Tag (CnT)/ChIP data (when available) at the representative loci for each sample (RNAPII, CTCF, H3K9me3 H3K27me3, H3K27ac). Bulk ChIP data are from ENCODE and the 4D Nucleome consortium.

Figure S3.

Comparison to Paired-Tag and scRNA-seq. Library comparisons between downloaded Paired-Tag from mouse hippocampus cells (red) and TIP-seq from F121-9 mESCs (blue) using identical filters and peak calling parameters between both library types. (A and B) Violin plots showing unique read counts and FRiP for H3K27me3. (C and D) Violin plots showing unique read counts and FRiP for H3K9me3. Means are marked with a black dot. Peak numbers provided below FRiP violins. (E) UMAPs of 10x Genomics scRNA-seq data (GEO accession no. GSM3160745) and RNAPII sciTIP-seq data. scRNA-seq cell numbers were randomly downsampled to match that of RNAPII sciTIP-seq (480 cells).

Figure 5.

Single-chromosome mapping of RNAPII to parsed alleles in hybrid F121-9 mESCs. (A) IGV track view across a 9-Mb segment of the mouse chromosome 13 comparing parsed vs. unparsed bulk CUT&Tag, sciTIP pseudo-bulk, and single cells. Tracks show unparsed bulk ChIP (4D nucleome consortium), unparsed bulk CUT&Tag, parsed bulk CUT&Tag, parsed sciTIP pseudo-bulk (aggregate), and 100 parsed single cells. Red asterisk marks loci with differential RNAPII binding. (B) IGV track view across a 7-Mb segment of the mouse chromosome 2 with all else the same as in A. agg., aggregate; CnT, CUT&Tag.