RBPeek

RBPeek quantifies the co-occupancy of RNA-binding proteins (RBPs) at a defined set of genomic loci. Given an inference BED (here, reproducible THRAP3 binding sites) and a panel of CLIP peak files, it ranks the panel samples by their depth-normalised binding at those loci and plots the binding profiles of the highest-ranked samples.

Requirements

intersect_inference_bed.py requires Python ≥ 3.9, bedtools, numpy, scipy, matplotlib, seaborn and scikit-learn; rbpeek.yml defines a conda environment that provides them. The other scripts require only Python ≥ 3.6 and bedtools.

Usage

1. Partition the inference loci into exonic and intronic sets and generate the normalisation regions:

   python3 scripts/split_inference_bed_by_region.py -b THRAP3/THRAP3_merged_min2rep_anchors.bed -g <annotation.gtf> -o THRAP3 --drop-chrM

2. Run the analysis for one region:

   python3 scripts/intersect_inference_bed.py \
     -x <panel directory> \
     -s THRAP3/RBPeekSamplesheet_eCLIP.tsv \
     -b THRAP3/THRAP3_merged_min2rep_anchors_exonic.bed \
     --norm-bed THRAP3/regions_exonic.bed \
     --genome <chromosome sizes> \
     -o results/thrap3_exonic \
     --support-pct 20 --tsne

   scripts/run_thrap3_region.sbatch <exonic|intronic> runs this step under SLURM with the settings used for the THRAP3 analysis.

Options

option	default	description
-x/--xldir	required	directory against which the samplesheet's file paths are resolved
-s/--samplesheet	required	TSV with columns file and group (sample label)
-b/--bed	required	inference BED6+, strand in column 6; each locus is anchored at (start + end) // 2
--norm-bed	primary chromosomes of --genome, both strands	BED6 of the regions over which each sample's normalising cDNA is summed (regions_exonic.bed for exonic loci, regions_intronic.bed for intronic loci)
--genome	required	chromosome sizes file for bedtools slop
-o/--outdir	results	output directory
--window	100	half-width (nt) of the window around each locus
--central-window	10	half-width (nt) of the central window used for ranking
--support-pct	30	percentage of top-ranked samples shown in the heatmap and tSNE
--highlight	none	sample labels, or substrings of them, drawn in black on the metaprofile in addition to the ten highest-ranked samples
--heatmap-scale-percentile	99	percentile of non-zero heatmap values mapped to the top of the colour scale
--tsne, --tsne-perplexity	off, 30	tSNE embedding of the heatmap loci

Chromosome naming (chr1 or 1) is harmonised automatically between the inference BED and the panel files.

Method

1. Signal. Each locus is extended by --window nt on either side and intersected with each panel file on the same strand. The cDNA count of a peak (BED score) is distributed uniformly across its width, so a peak that partly overlaps a window contributes only the overlapping fraction, and a 1 nt crosslink site retains its full score at its own position. Offsets are strand-aligned, with positive values downstream of the locus. This yields, for each sample, a locus × offset support matrix c(l, o).
2. Normalisation. region_cdna is the cDNA of the sample's peaks within --norm-bed on the same strand (by default, the primary chromosomes of --genome on both strands). Each peak is weighted by the fraction of its width inside the regions, overlapping regions are merged beforehand, and mitochondrial peaks are excluded because mitochondrial rRNA is a major source of background in eCLIP libraries. Normalised support is x(l, o) = c(l, o) × 10⁶ / region_cdna.
3. Metaprofile. For each sample, m(o) is the mean over all loci of log(1 + x(l, o)). The logarithm is applied per locus before averaging, so that the curve reflects the breadth of binding across loci and is not dominated by a small number of strongly bound loci. Loci without signal contribute 0.
4. Ranking. central_binding is the area under m(o) within ±--central-window nt. Samples are ranked by central_binding, and the top --support-pct percent are selected. Because the ranking score and the plotted curve are the same statistic, curve height reflects rank. Samples without region cDNA, or without a peak in the central window of any locus, are not selected.
5. Figures.
   • metaprofile.pdf shows m(o) for the ten highest-ranked samples, without smoothing, so the area between the red dotted lines equals central_binding. These lines mark the central window, and the legend gives the rank and central_binding of each sample. Samples named with --highlight are added to the plot and drawn in black.
   • binf_support_heatmap.pdf shows loci × selected samples. Each cell is the strongest single peak of the sample within the central window of the locus (the cDNA of that peak inside the window, per million region cDNA), log(1 + x)-transformed and scaled to the --heatmap-scale-percentile of the non-zero cells. Samples are ordered by hierarchical clustering (cosine distance, average linkage) and labelled with their rank; loci are ordered by summed cell value, and loci without a central peak in any selected sample are omitted.
   • binf_summary_tsne.png is a tSNE embedding of the heatmap loci (fixed random seed).

The width of the central peak in the metaprofile is determined by the panel peak calls and should not be interpreted as a binding footprint: the panel peaks are approximately 11 nt wide, and the cDNA of each peak is distributed uniformly across its width.

Outputs

file	contents
metaprofile.pdf	m(o) of the ten highest-ranked selected samples
sample_summary.tsv	one row per panel sample, sorted by rank
binf_support_heatmap.pdf	loci × selected samples
binf_summary_tsne.png	tSNE embedding of the heatmap loci (--tsne)

Columns of sample_summary.tsv:

column	description
sample	sample label from the samplesheet
rank, selected	rank by central_binding; whether the sample was selected
central_binding	area under m(o) within the central window
region_cdna	normalising cDNA (Method 2)
locus_cdna, proportional_binding	cDNA of the distinct peaks within any locus window, each weighted by the fraction of its width inside; and its ratio to region_cdna
total_peak_support, mean_peak_support	support summed over all locus windows (a peak within the windows of two loci is counted in both), and that sum divided by the number of loci
loci_with_signal, frac_loci_with_signal	number and fraction of loci with any support within the window
mean_binding_offset	mean, over loci with signal, of the centre of the 5 nt window with the highest support
mean_variance, mean_pearson_skew, mean_kurtosis	means, over loci with signal, of the variance (normalised scale), Pearson median skewness and excess kurtosis of the support vector

THRAP3 data

The inference loci derive from four HEK293 HA-THRAP3 iCLIP replicates (GRCh38). Crosslink sites were derived from read 1 of each pair, and the Clippy peak calls are stored in THRAP3/raw/. build_thrap3_inference_bed.py converts chromosome names to UCSC style, merges overlapping peaks across replicates on each strand, retains regions supported by at least two of the four replicates, and reduces each region to its 1 nt midpoint. This yields 26,030 loci (THRAP3_merged_min2rep_anchors.bed; the score column records replicate support), of which 216 are mitochondrial and are removed before the loci are partitioned into exonic and intronic sets.

The panel (THRAP3/RBPeekSamplesheet_eCLIP.tsv) comprises 224 HepG2 and K562 eCLIP samples; RBPeekSamplesheet.tsv additionally lists the iCLIP and PAR-CLIP samples. Panel entries under peaks/merged/ were produced with merge_replicate_peaks.py.

Limitations. THRAP3 was profiled in HEK293 cells, whereas the panel consists of HepG2 and K562 data, so cell line is confounded with RBP identity; ranks should be compared, not absolute values. central_binding increases with the fraction of loci at which a sample has signal, and the number of peaks called depends on library depth, so normalisation by region_cdna does not remove the influence of depth entirely.

Centrosome data

build_centrosome_inference_bed.py builds inference loci for proximity-CLIP (APEX) data, in which peak cDNA reflects the proportion of a transcript in the labelled compartment rather than a discrete binding site. Peak cDNA is assigned to gene bodies (each peak weighted by the fraction of its width inside the gene, same strand) and expressed per million of the sample's non-mitochondrial peak cDNA. Genes are retained if they pass the differential-expression filter (log2FoldChange > 1, pvalue <= 0.05) and the ratio of the experimental to the control group mean, with a pseudocount of 1 per million added to both, is at least 2. The anchors are the midpoints of the pooled experimental peaks that lie within retained genes.

python3 scripts/build_centrosome_inference_bed.py --deseq <DESeq2 table.xlsx> -g <annotation.gtf>

Because the loci are not partitioned by region, intersect_inference_bed.py is run without --norm-bed, so that each sample is normalised by its peak cDNA on the primary chromosomes.

Scripts

script	purpose
intersect_inference_bed.py	the analysis described above
split_inference_bed_by_region.py	exonic and intronic locus sets (strand-aware, exons given priority) and the corresponding normalisation regions; --drop-chrM removes mitochondrial loci
build_thrap3_inference_bed.py	replicate THRAP3 peak calls → inference BED
build_centrosome_inference_bed.py	proximity-CLIP peak calls and differential-expression table → inference BED of enriched genes
merge_replicate_peaks.py	pools replicate peak BEDs by summing scores at identical intervals
run_thrap3_region.sbatch	SLURM job for one region of the THRAP3 analysis