Skip to main content
NAR Genomics and Bioinformatics logoLink to NAR Genomics and Bioinformatics
. 2026 Nov 10;8(4):lqag138. doi: 10.1093/nargab/lqag138

CLIPittyClip: a single-command framework for CLIP sequencing data analysis

Soon Yi 1,2,3,✉, Joseph M Luna 4,5,✉
PMCID: PMC13748876  PMID: 43011589

Abstract

Crosslinking immunoprecipitation (CLIP) sequencing is the standard method for mapping RNA-binding protein (RBP) binding sites transcriptome wide, yet most analysis pipelines are either modular toolkits requiring manual assembly or large-scale workflows built for consortium infrastructure. Here, we present CLIPittyClip, a single-command pipeline that processes CLIP data from raw FASTQ files to a peak matrix and crosslink site outputs, supports five major CLIP read architectures and runs entirely locally on standard laptop hardware. The pipeline introduces CLink, a Python reimplementation of CLIP Tool Kit crosslink-induced truncation and mutation site analyses that is up to 27-fold faster while recovering equivalent binding motifs and crosslink site resolution. Using CLIP data for a prototypical RBP, we performed systematic parameter sweeps for two peak callers to establish empirically calibrated defaults for the pipeline. We validated CLIPittyClip against published ELAVL1, HNRNPC, and RBFOX2 datasets spanning BrdU-CLIP, iCLIP, eCLIP, and PAR-CLIP architectures, and confirmed compatibility with downstream analyses for differential binding and motif analysis. CLIPittyClip provides a fast, local, and practical processing layer for CLIP data analysis in a single-lab setting and is freely available at https://github.com/LunaRNALab/CLIPittyClip.

Introduction

RNA-binding proteins (RBPs) govern virtually every stage of post-transcriptional gene expression, and crosslinking immunoprecipitation (CLIP) sequencing technologies have become a standard method for defining RBP-binding sites transcriptome wide [1]. Since its introduction and initial implementation [2], to its early adaptations for high-throughput sequencing [3–5], CLIP has diversified considerably, with many variants improving yield, reducing background, and sharpening crosslink-site resolution through targeted technical innovations [6–11]. Computational tools for CLIP analysis have multiplied in parallel, with each addressing a particular use case or scale [12, 13]. The CLIP Tool Kit (CTK) enables processing of HITS-CLIP, PAR-CLIP, and iCLIP data with single-nucleotide crosslink resolution [14], but functions as a toolkit rather than a pipeline, requiring users to chain steps manually and manage dependencies. Early pipeline attempts such as FAST-iCLIP [15] addressed this for HITS-CLIP and iCLIP but are no longer actively maintained. The ENCODE eCLIP and nf-core/clipseq pipelines implement rigorous, standardized workflows built for large-scale consortium work; however, their infrastructure requirements make them less well suited to single-lab use [8, 16, 17]. Galaxy-based workflows, such as PIPE-CLIP, lower the barrier to entry but trade command-line flexibility for a graphical interface that is often externally hosted, limiting speed, reproducibility, and customization [18]. More recently, racoon_clip implemented an accessible Snakemake-based pipeline for iCLIP and eCLIP centered on single-nucleotide crosslink site identification [19]. A comparable lightweight solution that handles multiple CLIP read architectures and produces both peak and crosslink outputs for downstream analysis has not been available.

Recently, PCLIPtools addressed a version of this problem for PAR-CLIP data, providing a lightweight end-to-end solution from FASTQ to peaks with a Poisson-based peak caller that outperforms PARalyzer in speed, cluster yield, and reproducibility [20]. The analogous need for 254 nm UV CLIP data prompted us to develop CLIPittyClip to fulfill a similar role. This pipeline was first introduced as the analysis component of coCLIP, a method combining APEX proximity labeling with CLIP-seq [21], but was not described or benchmarked as a standalone tool. Here, we introduce version 3.5, which adds full crosslink site analysis through both the established CTK toolkit and a new Python reimplementation (CLink), as well as native support for five major CLIP read architectures, completing a local end-to-end workflow from FASTQ to peak matrix and crosslink sites in a single command with modest hardware requirements (Table 1). Unlike toolkits that require manual step assembly or consortium pipelines built for cluster infrastructure, CLIPittyClip combines local single-command operation, coverage across five read architectures, and both peak and crosslink outputs in one tool. CLIPittyClip is freely available on GitHub at https://github.com/LunaRNALab/CLIPittyClip.

Table 1.

Comparison of hardware requirements, installation, and analysis features across CLIP pipelines

Feature CLIPittyClip CTK racoon_clip nf-core/clipseq PCLIPtools eCLIP
Minimum RAM 32 GB (16 GB for CLink) 32 GB 120 GB 32 GB (varies up to 84 GB) 32 GB 32 GB
Infrastructure Laptop/workstation Laptop/workstation HPC/cluster HPC/cluster HPC/cluster HPC/cluster
Installation Conda Manual (Perl) Conda Nextflow + container None (clone and run) Docker or manual
Single-command operation Yes No Yes Yes Yes Yes
Input architectures 5 3 2 2 1 (PAR-CLIP) 1 (eCLIP)
Peak calling Yes Yes Yes Yes Yes Yes
Crosslink site analysis Yes Yes Yes Yes T→C only No

A persistent practical problem in CLIP analysis is that peak calling parameters are rarely benchmarked systematically and are instead set by convention or inherited from prior implementations. While parameters governing trimming, unique molecular identifier (UMI) handling, and alignment are handled by well-established tools with sensible defaults, peak calling is where arbitrary choices compound. We address this directly with parameter sweeps across prototypical CLIP datasets to identify settings that perform well for representative experiments, providing an empirical basis for the defaults used in CLIPittyClip. We then validate the pipeline against published datasets for HuR [21–24], HNRNPC [7, 24, 25], and RBFOX2 [24, 26], demonstrating comparable binding site annotation, motif recovery, and single-nucleotide crosslink site resolution across multiple CLIP modalities. Together, these results establish CLIPittyClip as a reliable and practical framework for CLIP data analysis in a single-lab setting.

Materials and methods

Pipeline overview

CLIPittyClip (v3.5.0) is a modular pipeline written in Bash for complete CLIP-seq data analysis. The pipeline is designed to minimize external dependencies and supports a broad range of CLIP-seq library chemistries. All key parameters are set as command-line flags, so users do not need to edit configuration files between runs. The workflow includes FASTQ deduplication, sample demultiplexing, adapter trimming and UMI extraction, optional repeat RNA prefiltering, genome alignment, BAM processing with PCR duplicate collapse and coverage generation, peak calling, and optional crosslink-induced mutation site (CIMS) or crosslink-induced truncation site (CITS) analysis using the CTK or CLink, introduced here. CLIPittyClip is freely available on GitHub at https://github.com/LunaRNALab/CLIPittyClip. A stable release v3.5.0 is also available on Zenodo at https://doi.org/10.5281/zenodo.20875662. All benchmarking, downstream analysis scripts, and a full list of CLIPittyClip commands performed for the data presented in this work can be found at https://github.com/LunaRNALab/BiTS_CLIPittyClip.

Datasets and accession numbers

Published and in-house datasets used in this study are presented in Supplementary Table S1 from the following sources: HuR BrdU-CLIP [21], HNRNPC BrdU-CLIP [25], HNRNPC iCLIP [7], RBFOX2 BrdU-CLIP [26], HuR eCLIP [27] (ENCODE accession: ENCSR090LNQ), HNRNPC eCLIP and RBFOX2 eCLIP [24] (ENCODE accession: ENCSR550DVK and ENCSR249ROI for HNRNPC HepG2 and K562 eCLIP, respectively; ENCSR987FTF and ENCSR756CKJ for RBFOX2 HepG2 and K562 eCLIP, respectively), and HuR PAR-CLIP [22, 23].

Reference genomes and annotation files

Human samples were aligned to the GRCh38 primary assembly and mouse samples to the GRCm39 primary assembly (Ensembl), with STAR genome indices built using the corresponding Ensembl release 115 Gene Transfer Format (GTF) file for splice-junction guidance (--sjdbOverhang 100). The same release 115 GTF was used downstream for annotation, motif-density sequence extraction, and BindingSiteFinder (BSF) analysis. Repeat-element annotation used the University of California Santa Cruz (UCSC) genome browser RepeatMasker track (hg38, extracted in August 2023); the optional repeat prefilter index additionally drew on NCBI 45S ribosomal DNA (U13369.1), GtRNAdb tRNA sequences, and Dfam/RepeatMasker transposable-element consensus sequences restricted to mammalian clades. Genomic sequence retrieval for motif analyses used the Bioconductor BSgenome.Hsapiens.UCSC.hg38 and BSgenome.Mmusculus.UCSC.mm39 packages. Because Ensembl GTFs use bare chromosome names (e.g. 1, MT) while UCSC-style tools and BSgenome packages expect chr-prefixed names (e.g. chr1, chrM), chromosome naming was harmonized at each tool boundary requiring it, rather than maintaining a single converted annotation file throughout.

FASTQ-level deduplication

Prior to alignment, reads were collapsed by exact sequence identity. The CTK tool fastq2collapse.pl performs this via a Unix sort | uniq approach, sorting the full read set such that identical sequences become adjacent. Sorting cost grows faster than the number of reads, meaning run-time increases disproportionately as library size increases, despite modest memory requirements. To speed this up, we implemented two deduplication engines. The default is a hash-based approach that uses a single pass to tally sequence occurrences in an in-memory hash table, and a second pass writes the first occurrence of each unique sequence with its observed count appended to the read ID, discarding subsequent duplicates. Run-time scales linearly with read count, and memory scales with the number of unique sequences rather than total reads. A sort-based engine is available via a --low-memory flag. It replicates the sort-then-collapse logic of fastq2collapse.pl as a standalone implementation, holding a fixed, flat memory footprint regardless of library size by spilling to disk rather than an in-memory table. A console warning recommends --low-memory above 30 million reads. Both engines write output identical in format to fastq2collapse.pl (@readID#count header followed by sequence, +, and quality lines), preserving compatibility with all downstream tools without modification. An inline sort | uniq pipeline is retained as a fallback for environments where neither engine’s dependencies are available.

Demultiplexing, adapter trimming, and UMI extraction

In-house CLIP libraries are frequently pooled with custom barcoding schemes; our demultiplexing step was designed to accommodate variable read architectures rather than a fixed layout. Given a barcode (BC) file, pooled reads are split with cutadapt using an unanchored 5′ search (-g, --no-indels, --action = none), with the mismatch tolerance expressed as an error rate scaled to barcode length (default one mismatch) and a barcode-collision check run beforehand to abort if any two barcodes fall within that mismatch distance of each other. Because the search is unanchored, it tolerates a UMI of arbitrary length preceding the barcode, and because --action = none leaves the barcode in place rather than removing it, the read is split into per-sample files without committing to a UMI/barcode boundary at this stage. The default read layout is [UMI] [barcode] [spacer] [READ]; a [barcode] [UMI] [spacer] [READ] layout is also supported (--bc-first) for protocols where the barcode precedes the UMI. A trailing 3′ UMI, as in iCLIP3, is removed with --trim3, which trims a fixed number of nucleotides after adapter removal, extending the grammar to protocols that place a second UMI between the read and the 3′ linker. Because read architecture is specified through these flags rather than hardcoded layouts, new protocols following this grammar can be accommodated and are catalogued in Supplementary Data S1. After splitting, each per-sample file is passed through fastp, which extracts the UMI into the read ID and trims the 3′ adapter; a default in-house adapter (L32, GTGTCAGTCACTTCCAGCGG) is used but may be overridden with any sequence via -a. More details on CLIP read architectures supported by CLIPittyClip can be found in Supplementary Data S1.

A dedicated GEO mode (--geo) reuses the same barcode-based splitting logic but performs no further modification; the resulting per-sample files are byte-identical to what was sequenced aside from the barcode split itself, accompanied by MD5 checksums for repository deposit.

Two additional modes handle external protocols with fixed, noncustomizable read architectures that fall outside the barcode-file paradigm. For ENCODE deposited eCLIP data, paired-end mode accepts Read 2 as output by eclipdemux, with the UMI already present as a colon-delimited token in the read header; this is moved into the read sequence prior to deduplication (so that deduplication is UMI-aware) and returned to the header afterward in CTK-compatible format (READ#count#UMI). A single-end variant supports seCLIP data [28], in which the UMI is fixed at the first 10 nt of Read 1 and the adapter is hardcoded to the TruSeq Read 1 sequence, reflecting the fixed nature of that protocol rather than a user-configurable parameter. For fPAR-CLIP data, reads follow a fixed [UMI] [READ] [2 nt spacer] [6mer barcode] [adapter] layout; preprocessing uses two fastp passes, the first extracting the UMI and trimming the barcode and adapter using a bundled 43-barcode adapter FASTA (lib/parclip_adapters.fa), and the second removing the residual 2 nt random spacer. A custom adapter FASTA may be supplied in place of the bundled set via --parclip-adapters.

Repeat prefiltering

Reads were optionally prefiltered against a dedicated repeat reference prior to genome alignment, comprising 45S ribosomal DNA, tRNA sequences (GtRNAdb), and transposable element consensus sequences (Dfam/RepeatMasker, restricted to mammal-relevant clades). This step is opt-in and off by default, since repeat-element binding can itself be biologically meaningful for some RNA-binding proteins. Where enabled, the rationale is that ribosomal RNA, tRNA, and transposable elements are typically dominant, highly repetitive background in CLIP libraries that STAR cannot place uniquely; removing them before genome alignment reduces multimapping noise and spurious peak calls while preserving the reads for separate quantification rather than discarding them outright. Other noncoding RNA classes, including miRNA, Y RNA, lncRNA, snRNA, and snoRNA, were deliberately excluded from the repeat reference and left to align normally against the genome, as these are frequently the analyte of interest rather than background. Filtering was performed with Bowtie2 in local alignment mode (--local --very-sensitive-local) rather than end-to-end, prioritizing sensitive detection of any repeat-derived subsequence within a read over full-length placement. Reads mapping to the repeat reference are quantified separately at per-element and per-family resolution, RPM-normalized to total input reads; unmapped reads proceeded to genome alignment unchanged.

Alignment

Adapter-trimmed reads were aligned to the genome with STAR by default; Bowtie2 is offered as an alternative but is not validated for crosslink-site analysis and is flagged accordingly at run-time, lacking both empirical tuning against CIMS/CITS output and splice awareness. Mismatch filtering with STAR used a fractional rather than absolute cutoff (--outFilterMismatchNoverReadLmax 0.1), scaling the permitted mismatch count with the post-trim read length akin to fractional edit distance (-n 0.06) in BWA, as used by CTK. We observed that an absolute cutoff, in contrast, discards reads carrying crosslink-induced deletion alongside ordinary sequencing error, suppressing CIMS signal on short reads; a loose absolute backstop (--outFilterMismatchNmax 5) was retained to exclude only extreme outliers. Gap-open and gap-extend penalties for insertions and deletions were lowered to parity with the mismatch penalty (--scoreDelOpen -1 --scoreDelBase -1 --scoreInsOpen -1 --scoreInsBase -1), since standard RNA-seq defaults in STAR penalize indels more heavily and would otherwise bias alignment toward resolving a true deletion as a run of substitutions. Alignment was performed end-to-end, without soft-clipping, as CLIP reads are short and fully adapter-trimmed and clipping risks discarding the crosslink-proximal end of the read. Reads aligning ambiguously to more than ten loci were discarded (--outFilterMultimapNmax 10). By default, only uniquely mapped reads (NH:i:1) proceed to deduplication and crosslink calling; an optional multimapper rescue mode, inspired by CLAM [29], hard-assigns each multimapped read to the candidate locus with the greatest unique-read pileup support within a ± 50 bp window as a single-pass approximation of the iterative expectation-maximization procedure used in CLAM. This mode is off by default, and all analyses presented here used uniquely mapped reads only.

BAM processing, PCR collapse, and coverage generation

Following alignment, reads are deduplicated at the BAM level to remove PCR duplicates, distinct from the global deduplication applied to raw FASTQ reads prior to alignment. As with alignment, only uniquely mapped reads (NH:i:1) are retained before deduplication by default. Deduplication is performed with umi_tools dedup, in one of two modes depending on whether a UMI is present. When a UMI is available, reads are first grouped by their strand-aware 5′ mapping position, then within each position, UMIs that likely represent the same original molecule are merged: two UMIs are merged when they differ by only one edit and one is supported by at least twice as many reads as the other, with the more abundant UMI kept as the representative (--method directional). This treats a low-count, near-identical UMI as a probable PCR or sequencing error of a more abundant one at the same position, rather than a distinct molecule. When no UMI is available, all reads sharing the same position and strand are collapsed outright (--method unique, position only). Reads are also partitioned by splice status (--spliced-is-unique), such that two reads sharing a 5′ position but differing in whether they span an intron are not collapsed together.

The deduplicated BAM is used to generate two coverage products. RPM-normalized, strand-specific bedGraphs are computed directly from the BAM with bedtools genomecov; the scale factor for each sample is calculated as 1 000 000 divided by its total uniquely mapped read count and recorded, alongside that read count, in a shared scale_factors.tsv for traceability. Separately, when single-nucleotide crosslink resolution is requested, the 5′-most position of each read (read start on the + strand, read end on the − strand, matching the truncation-site convention used throughout) is extracted into a BED file and converted to strand-specific bigWigs via bedtools genomecov in BED-interval mode followed by bedGraphToBigWig; this crosslink-resolution signal, rather than the read-extent bedGraph, can be used for downstream per-nucleotide analyses such as BindingSiteFinder [30]. Both coverage paths draw chromosome sizes from a chrom.sizes file extracted once from the genome index at start-up.

Abundant RNA read-level quantification

CLIPittyClip optionally reports the fraction of reads assigned to abundant RNA classes (rRNA, tRNA, snRNA, snoRNA, miRNA, and others), enabled with --report-abundant-rna. This check runs on the deduplicated collapsed BED files already produced by the pipeline, requiring no separate alignment or processing step. Reads are counted against the same annotation used downstream, as a percentage of both unique reads and raw reads before deduplication, independent of peak calling, since short abundant elements can absorb substantial read depth while yielding few called peaks. Custom BED or GTF coordinates may be supplied to quantify any additional class of interest. When a grouping file is provided, quantification is pooled per group; a CLIP sample and its size-matched input (SMI) can be compared directly at the read level.

Crosslink site identification: CLink CITS/CIMS implementation and CTK comparison

Both truncation (CITS) and mutation (CIMS) site calling share a common signal-extraction stage. A single forward pass over a sorted, deduplicated BAM accumulates four per-position signals independently for each strand: coverage, 5′ read-end (truncation) counts, 1-nt CIGAR deletions, and substitutions by type. The truncation site is defined to match CTK convention as the last reference position synthesized before RT stalling at the crosslink. Crosslink positions are recorded as one position upstream of a forward read’s alignment start, or at a reverse read’s alignment end. Truncation counts are tracked separately for reads with and without a co-occurring deletion, since a read carrying a crosslink-induced deletion would otherwise inflate both signals from the same underlying event. This scan is cached to a single compressed array file so that CITS and CIMS calling both draw on one BAM pass rather than rescanning.

Site calling uses a shared statistical model. A background rate per signal type is estimated as the ratio of total signal to total coverage across sufficiently covered positions, with the top 1% of positions by signal fraction excluded so that crosslink sites do not inflate the background estimate. Each qualifying position, meeting minimum coverage, minimum raw signal count, and minimum signal fraction thresholds, is then tested under a binomial model: the background rate is treated as the null probability of signal at a given position, and the probability of observing at least as many signal-bearing reads as were seen is computed as P(X ≥ observed | Binomial(coverage, λ)) via the binomial survival function. This accounts for the fact that high-coverage positions need proportionally more signal reads to be statistically surprising than low-coverage ones. Raw p values are then corrected for multiple testing across all positions tested genome-wide using the Benjamini-Hochberg procedure, which controls the expected proportion of false positives among called sites rather than the probability of any single false positive. Sites with a corrected q-value below a user-specified false discovery rate (FDR) threshold are retained as significant.

CITS departs from this shared model in how the background rate itself is computed. Rather than testing truncation enrichment against a single genome-wide rate, positions are first grouped into clusters by proximity (gap ≤ 25 nt, matching CTK), and each position is tested against the rate computed locally within its own cluster, rather than the rate computed across the whole genome. Testing against a local cluster background here asks whether a position stands out relative to its immediate neighborhood. CIMS, in contrast, tests deletions and substitutions against the pooled genome-wide background directly, without local clustering, since these signals reflect a discrete crosslink-induced lesion at a specific base rather than a stalling probability that scales with local transcript abundance. Substitution testing covers all twelve mismatch types. Throughout, forward- and reverse-strand signals are tested independently, while background rate estimation pools both strands for a more stable estimate.

CTK (tag2collapse.pl, parseAlignment.pl, CITS.pl, CIMS.pl) [14] was run alongside CLink on the same deduplicated reads as an independent, permutation-based reference for truncation and mutation site calling, rather than as the primary caller. Where both were enabled, CTK CIMS permutation count was raised from its default of 5 to 10 or 20 iterations to stabilize the permutation-based FDR estimate at the depths used here, with CIMS FDR and CITS p-value thresholds left at their defaults of 0.05; CTK CITS clustering gap (25 nt) matches the cluster gap used in CLink local background calculation, so the two callers operate over comparable neighborhoods despite differing in how the background itself is estimated. CLink was run with its default FDR threshold of 0.05 and minimum coverage of 5. For datasets with multiple replicates, both callers were run in pooled mode (--group-xlsite for CLink, --ctk-group for CTK), merging per-sample deduplicated alignments by group before site calling to increase effective depth, rather than calling sites per replicate and combining afterward.

Peak calling with HOMER findPeaks and CTK tag2peak

Two peak callers are supported on the collapsed BED file. Hypergeometric Optimization of Motif EnRichment (HOMER) peaks are called with findPeaks-style factor, the punctate transcription-factor-style mode. Peaks are called independently per strand (-strand separate). Local enrichment is required at twofold over a 1 kb local background window (-L 2 -localSize 1000), with a default minimum peak width of 20 bp, a default merge distance of 50 bp between adjacent peaks, and a default fragment length of 25 bp used for tag-shift estimation; these three values (-size, -minDist, -fragLength) are exposed as pipeline-level parameters and were the dimensions varied in the parameter sweeps described below.

CTK peak calling (tag2peak.pl) takes a different approach to peak boundaries. Rather than scanning fixed-width windows, we used its --valley-seeking argument to bound each peak at the nearest local minimum in the read-density profile on either side. As with HOMER, calling is strand-specific (-ss), and nearby peaks within the same merge distance (-gap, defaulting to 50 bp, matching HOMER’s -minDist) are combined. A minimum peak height of 2 reads (-minPH 2) was used as a default; -big was used to enable the CTK memory-efficient mode for large libraries.

Parameter sweeps and performance benchmarking

A consistent design principle underlies all sweeps described here: Each comparison began with fixed input data and varied only the single parameter or tool under test so that output differences can be attributed to that variable alone. These tests used ELAVL1 BrdU-CLIP data generated in our laboratory [21]. All benchmarking and secondary analysis scripts presented downstream of CLIPittyClip are accessible at https://github.com/LunaRNALab/BiTS_CLIPittyClip.

Read-depth saturation was assessed by subsampling a fixed collapsed-read pool to a range of depths (100 000 to several million reads) and calling peaks at each depth with fixed parameters, separately for HOMER, CTK, and CLIPper, to characterize each caller independently. CLIPper required a separate sweep from the same depth-matched BAM subsamples, since it required a sorted, indexed BAM directly (clipper -b <bam> -s hg38 -o <out.bed> --processors 8) rather than the collapsed BED used by HOMER and CTK, and was run in a dedicated conda environment. All three callers nonetheless received the same depth-matched pool of uniquely mapped (NH:i:1) reads, only reencoded into whichever format each tool required. A complementary sweep ran HOMER and CTK on identical subsampled reads at each depth, rather than independently subsampled replicates, so peak counts and pairwise overlap could be compared at matched depth without depth itself confounding the comparison.

Sensitivity to each caller’s internal parameters was swept separately from depth. The fold-enrichment threshold (-L) and local background window size (-localSize) in HOMER were each varied in turn while holding the other fixed, and the minimum peak height (-minPH) in CTK was varied independently, all on the same fixed collapsed-read input. Crosslink-site significance thresholds (FDR or p value, depending on caller) were swept separately for CLink and CTK CITS/CIMS output, measuring at each threshold what fraction of called sites fell within an independently called peak. Based on the motif enrichment plots presented in Figs 3C and 4C, we observed a positional offset between peaks and CITS/CIMS; therefore, all overlap calculations allowed small positional tolerance of 25 nucleotides

Figure 3.

Comparison of CLIP Tool Kit and CLink peaks and crosslink-induced truncation sites in four datasets spanning HuR, HNRNPC, and RBFOX2 RNA binding proteins. Stacked bars show genomic annotation, with sequence logos, motif density plots, and Venn diagrams of overlap.

Comparison of Peak and CITS analysis between CTK and CLink. (A) Distribution of peaks and CITS generated from CTK and CLink across all datasets: HuR BrdU-CLIP [21], HNRNPC BrdU-CLIP [25], HNRNPC iCLIP [7], and RBFOX2 BrdU-CLIP in mouse [26]. Genomic features are color coded with the total number of sites depicted. (B) Sequence logos of top enriched motifs from peaks, CTK CITS, and CLink CITS for each dataset. The enriched motifs were calculated using the findMotifsGenome.pl program in HOMER. (C) Density plots for the enriched motifs near peaks, CTK CITS, and CLink CITS. (D) Venn diagrams showing overlaps between peaks and CTK or CLink CITS, with the number of sites per condition shown. In each panel, columns correspond to the four datasets.

Figure 4.

Comparison of peaks, truncation sites, and deletion sites across eCLIP datasets for RBFOX2, HNRNPC, and HuR, plus PAR-CLIP HuR with thymidine-to-cytidine conversions. Panels show genomic annotation, sequence logos, motif density, and Venn diagrams of overlap.

Comparison of Peak, CITS, and CIMS analysis across eCLIP and PAR-CLIP datasets with CLink. (A) Distribution of peaks, CLink CITS, and CLink deletion (Del) across all datasets: RBFOX2 HepG2 eCLIP, RBFOX2 K562 eCLIP, HNRNPC eCLIP, and HuR eCLIP. For HuR PAR-CLIP, peaks, CLink Del, and CLink T-to-C mutation sites are shown instead. Genomic features are color coded with the total number of sites noted. (B) Sequence logos of enriched motifs from the site types shown in panel (A) for each dataset. The enriched motifs were calculated using the findMotifsGenome.pl program in HOMER. (C) Density plots for the enriched motifs near the corresponding site types. (D) Venn diagrams showing overlaps among the corresponding site types: peaks, CLink CITS, and CLink Del for the eCLIP datasets, and peaks, CLink Del, and CLink T-to-C mutation sites for HuR PAR-CLIP. The number of sites within each overlap is shown inside the diagram. In each panel, columns correspond to the datasets.

Run-time and memory were benchmarked separately from these peak-count sweeps. Eight threads (-t 8) were used wherever a tool exposed a thread or processor option, namely STAR and Bowtie2 alignment, fastp, CLIPper (--processors), and CLink chromosome-parallel pileup scan; several of the steps directly compared, including HOMER findPeaks and CTK Perl tools (tag2collapse.pl, parseAlignment.pl, CITS.pl, CIMS.pl), are single-threaded. FASTQ-level deduplication was timed head-to-head between fastq2collapse.pl and the hash-based Python implementation on matched input. BAM-level PCR deduplication was benchmarked between CLink and CTK from an identical pool of mapped reads, with the unique-mapper filter (NH/MAPQ) and the deduplication method varied independently to isolate which step drives any difference in run-time or output. Peak-caller wall time and peak memory were benchmarked for HOMER, CTK, and CLIPper given an identical pool of uniquely mapped reads (NH:i:1) from the same BAM; HOMER and CTK additionally received identical input via shared parseAlignment.pl/tag2collapse.pl preprocessing into one collapsed BED, so timing differences reflect the peak-calling step itself, while CLIPper ran on the BAM directly per its native workflow. Finally, the full crosslink-site-calling pipelines were benchmarked end to end, CTK against CLink, from a matched N-read BAM subsample through deduplication, signal extraction, and CIMS/CITS calling, reporting wall time and peak RAM for the complete pipeline.

Annotation, motif, and overlap analysis

Peaks and crosslink sites from each caller (HOMER/CTK peaks, CTK CITS/CIMS, CLink truncations/deletions/T→C substitutions) were annotated using a single 25-level, first-wins hierarchical scheme. Each site was assigned to the first category it overlaps, with the following priority order: 3′UTR > 5′UTR > coding DNA sequence (CDS) > discrete ncRNA biotypes (miRNA, lncRNA, Y RNA) > structural RNAs (rRNA, snRNA, snoRNA, tRNA) > RepeatMasker-derived repeat classes > distal 3′UTR > first intron > proximal intronic > deep intronic. Unmatched sites outside of these categories were considered intergenic.

Annotated sites were filtered to confident calls using protocol-specific thresholds. CLIP peaks were retained above the per-group median normalized read count; eCLIP peaks required above-mean enrichment and presence in both replicates; PAR-CLIP peaks required presence in every replicate and above-mean enrichment, reflecting the lack of an input control. Crosslink sites were filtered on signal strength and each caller’s own significance metric (CTK p value/FDR; CLink q value), with a stricter score cutoff (mean + 2 SD rather than median) for eCLIP and PAR-CLIP.

Motif density around each filtered site type was assessed by extracting a fixed window (default ± 100 nt, strand-aware) and plotting the fraction of sites containing a user-specified motif by distance from the site center, testing whether a binding protein’s known motif is sharply enriched at the centered crosslink or peak position. De novo motif discovery was run separately with HOMER (findMotifsGenome.pl -rna -oligo) across multiple k-mer lengths per site type.

For each pair of site types expected to mark the same event (peak versus CTK, peak versus CLink, CTK versus CLink, separately for truncations and deletions, plus PAR-CLIP T→C), reciprocal overlap and three-way intersection counts were computed with bedtools intersect across a range of positional tolerances (0, 10, 25 nt slop), and decomposed by genomic context (genic versus nongenic). Results were visualized as three-way Euler diagrams and per-set recovery bar charts in R.

To compare sites called by only one output with sites shared between outputs, the genic Peak, CITS, Del, and TtoC sites from Fig. 4 were each extended by 25 nt on both sides before intersection, such that sites within 50 nt were counted as overlapping. Each site was labeled by which of the other outputs were called an overlapping site (none, one specific partner, or both). For each class, we calculated the fraction of sites with an exact match to the cognate motif (UGCAUG for RBFOX2, UUUUUU for HNRNPC and HuR) within 50 nt of the site center, read on the site strand, with 95% binomial (Wilson) confidence intervals treating each site as an independent observation. Background was estimated from the same sites shifted 1 kb upstream and downstream, and enrichment is the ratio of the observed to the background fraction.

RBPBench and BindingSiteFinder analysis

RBPBench (rbpbench search) was run on combined peaks from ELAVL1 (HuR) BrdU-CLIP CLIPittyClip output as an orthogonal motif check, scanning its curated database, querying ELAVL1, HNRNPC, and RBFOX2 (specificity controls) with peaks extended 25 nt, plus a custom regex (TTTTTT) to cross-check HOMER’s motif:

  • rbpbench search \

  • --in /path/to/04_PEAKS/FINAL_PEAKS.bed \

  • --genome /path/to/Homo_sapiens.GRCh38.dna.primary_assembly.fa \

  • --gtf /path/to/Homo_sapiens.GRCh38.115.chr.gtf.gz \

  • --out /rbpbench_test/ELAVL1 \

  • --rbps ELAVL1 HNRNPC RBFOX2\

  • --ext 25 \

  • --regex TTTTTT \

  • --plot-motifs

BSF was applied on ELAVL1 BrdU-CLIP CLIPittyClip output, using per-replicate crosslink BED and BigWig files directly, to perform differential binding analysis (Mock versus Arsenite), which the peak-matrix workflow does not support. Sites were defined at a fixed 15 nt width and required reproducibility in ≥ 2 replicates at a 5% cutoff (relaxed from BSF’s stricter default). Differential binding was tested using calculateBsFoldChange, a DESeq2 likelihood-ratio test with poscounts size-factor estimation.

CLIPittyClip testing environments, AI-agent usage, and agent-facing documentation

The CLIPittyClip pipeline was tested on machines with the following hardware specifications: a Linux workstation running Ubuntu 22.04.5 LTS equipped with an AMD Ryzen 7 7800 × 3D (8 Cores/16 Threads) processor and 64 GB of RAM; a Linux workstation running Ubuntu 24.04.4 LTS equipped with an Intel Core i7-12700K (12 Cores/20 Threads) processor and 32 GB of RAM; an Apple MacBook Air running macOS 15.5.1 equipped with an Apple M2 processor and 16 GB of RAM; and an Apple MacBook Pro running macOS 15.7.4 equipped with an Apple M1 Max processor and 64 GB of RAM. Anthropic Claude Code Sonnet 4.6 and Google Antigravity (accessed from December 2025 to June 2026) agents were used to assist in code organization, implementation, documentation drafting, and error troubleshooting. Neither AI tool was used to generate or interpret data. All computational tests, analysis decisions, scientific interpretations, and figures were devised and made by the authors. All AI-assisted code outputs were reviewed, edited, and tested by the authors who independently confirmed the reproducibility of results across the above hardware configurations. Context and glossary markdown files (CONTEXT.md and docs/GLOSSARY.md in the CLIPittyClip GitHub repository) document all CLIPittyClip domain-specific terminology, pipeline architecture, and algorithmic conventions; AI agents assisting users with this codebase should be directed to these files first to ground their understanding of CLIPittyClip usage.

Results

End-to-end CLIP data processing with CLIPittyClip

CLIPittyClip was designed to provide end-to-end processing of CLIP data, from a variety of FASTQ inputs to a complete peak matrix for downstream analysis. The pipeline accepts FASTQ files from pooled experiments requiring demultiplexing, directories of predemultiplexed FASTQ files (as downloaded from the Gene Expression Omnibus, for instance), or single files. It is organized into three modules covering QC and alignment prep (PREPittyPrep), alignment with crosslink-induced truncation and mutation analysis (CITS/CIMS) (MAPittyMap), and peak calling with optional cross-sample aggregation (PEAKittyPeak) (Fig. 1A). CLIPittyClip supports a range of CLIP read architectures (Fig. 1B), which typically encode a per-read UMI, a per-library BC, and spacer sequences upstream of the CLIP read, followed by a 3′ adapter (see Supplementary Data S1 on supported CLIP architectures).

Figure 1.

Metro-style workflow map of CLIPittyClip from sequencing reads to peaks and crosslink sites, with diagrams of supported read layouts and plots comparing CLink processor time and memory use with CLIP Tool Kit (CTK).

CLIPittyClip for end-to-end CLIP sequencing data processing. (A) A “metro map” summary of the pipeline, from input fastq files through genome alignment, CTK and CLink tracks for CITS\CIMS analysis, peak calling, and peak matrix aggregation. Numbered files indicate key outputs within the CLIPittyClip output directory; primary tools for each step are in light gray. Vertical blocks demarcate the three pipeline modules: PREPittyPrep (preprocessing), MAPittyMap (mapping and crosslink site identification), and PEAKittyPeak (peak calling). (B) Major CLIP read architectures compatible with CLIPittyClip and their source protocols. (C) CPU time and peak RAM for global fastq deduplication, comparing the hash-based and sort-based implementations to CTK fastq2collapse across 1 to 100M input reads. (D) CPU time and peak RAM for PCR deduplication, comparing CLink to CTK tag2collapse across 1 to 25M uniquely mapped reads. CLink may use either the UMI sequence or the stranded 5′ mapping position for deduplication. (E) CPU time and peak RAM for CITS/CIMS analysis, comparing CLink to CTK across 1 to 25M uniquely mapped reads. CTK was run with 10 CIMS iterations.

Processing begins with global deduplication of identical reads. The hash-based implementation introduced here produces output identical to CTK fastq2collapse.pl and runs approximately 10-fold faster, from 36 s against 3 s at 1 million (M) reads to 67 min against 6.5 min at 100M reads (Fig. 1C). This speed comes at a memory cost, reaching 9.8 GB at 100M reads against roughly 120 MB for the sort-based approach; a sort-based fallback is retained for memory-limited machines. Demultiplexing, adapter trimming, and UMI extraction follow with fastp [31]. An optional repeat prefiltering step aligns reads to rRNA, tRNA, and transposable element sequences with Bowtie2 [32]; these are quantified separately and excluded from genome alignment. The remaining reads are aligned with STAR (default) [33] or Bowtie2 [32].

The resulting alignment feeds one of two processing tracks. The CTK track wraps CTK tools for mutation parsing, PCR deduplication, and CITS and CIMS analysis. The CLink track introduced with CLIPittyClip reimplements these steps in Python at the BAM level. CLink deduplicates roughly two-fold faster than CTK while using less than half the memory, with position-based deduplication holding at 1.2 GB across the full depth range tested against 44 GB for CTK at 25M reads (Fig. 1D). For CITS and CIMS analysis, CLink is 10-fold faster at 1M mapped reads and 27-fold faster at 25M mapped reads (Fig. 1E). Both tracks feed a common read-level peak-calling step using HOMER findPeaks or CTK tag2peak, with an optional aggregation stage that consolidates peaks across samples into a combined matrix for downstream analysis.

Peak calling parameter sweeps establish sensible CLIPittyClip defaults

A primary goal of most CLIP analyses is to identify regions of locally elevated read density, typically referred to as peaks, which mark sites where an RBP is bound to its target at steady state. CLIPittyClip implements read-level peak callers rather than crosslink site-based callers, reflecting the distinction between crosslink sites and binding sites, which do not necessarily coincide [12]. We consider peaks to mark regions of RBP occupancy, while CITS and CIMS modules operate on the same reads to provide crosslink site resolution within or near those peaks as a separate, complementary output.

We initially tested three callers representing distinct approaches: CLIPper [34], the reference eCLIP caller applying a per-gene negative binomial model; CTK tag2peak [26], which uses valley-seeking approach to filter peak heights; and HOMER findPeaks [35], a local background enrichment approach developed for ChIP-seq that has been applied to CLIP [36–38]. As a test dataset for a representative CLIP experiment, we used HuR CLIP data generated in our laboratory [21] (Supplementary Table S1). HuR (ELAVL1) is a prototypical RBP with well-characterized binding to AU-rich elements in 3′ UTRs and introns [39, 40], making it a useful benchmark for evaluating peak caller behavior under typical conditions.

CLIPper produced comparable annotation profiles and peak overlap with HOMER and CTK at matched depth, but its run-time of approximately 3 h against 9 s for HOMER and 40 s for CTK on identical input rendered it impractical for laptop use (Supplementary Fig. S1). HOMER and CTK were therefore carried forward.

Sweeping the HOMER -L parameter, which sets the minimum fold change over a local background window, and the CTK -minPH parameter, which sets the minimum number of overlapping reads qualifying as a peak, showed that HOMER annotation composition was stable across stringency thresholds while CTK shifted toward intronic calls at lower -minPH (Fig. 2A and B). Peak counts behaved differently as well, declining gradually for HOMER from 6.3k at -L 2 to 145 at -L 500, but dropping steeply for CTK from 33k at -minPH 2 to 8 at -minPH 200. Noncoding peaks reinforced this pattern: across all -L values, HOMER composition remained stable (Supplementary Fig. S2A). At low -minPH, CTK called substantial numbers of peaks in repeat element categories, while at high -minPH, we noticed an increase in snRNA/snoRNA proportion (Supplementary Fig. S2B). Sweeping -localSize across nearly three orders of magnitude left peak count and annotation composition largely unchanged (Supplementary Fig. S2C and  D).

Figure 2.

Line plots and bar charts showing how peak numbers and overlap with crosslink-induced truncation sites change across HOMER fold-enrichment and CLIP Tool Kit peak-height thresholds and read depths, with a genome browser track of the DDX17 gene.

Peak calling parameter sweeps for CLIPittyClip defaults. (A) Genic peak distributions and counts called using HOMER across different -L values (min fold enrichment over local background) starting from 1M mapped unique reads. (B) Same as in panel (A) but for CTK across different -minPH values (minimum peak height by read count). (C) Peaks with overlapping CITS across different -L values for HOMER and -minPH values for CTK. (D) Impact of input read depths on genic peaks for HOMER and CTK under their respective -L = 2 and -minPH = 2 default values. Green line marks shared peaks between callers. (E) Gene track example of DDX17 showing HuR CLIP reads (Bedgraph), CTK peaks, and HOMER peaks under respective default settings.

The presence of crosslink sites provides an orthogonal measure of peak quality. We next asked, at each threshold, what fraction of called peaks contained an independently called CITS, as measured by our CLink crosslink caller (see next section). We observed that the presence of a crosslink site varies sharply with parameter value and in opposite directions for the two callers (Fig. 2C). CTK peaks were crosslink-supported in only 6% of cases at -minPH 2, rising through 19% at -minPH 3% to 87% at -minPH 10 and 100% at the most stringent settings, when few peaks are called. HOMER peaks held near 26% to 28% across most of the -L range, rising to 82% only at -L 500, where fewer than 300 peaks remain. Crosslink support is therefore threshold-dependent for CTK and relatively threshold-independent for HOMER, with CTK approaching full support only where few peaks remain and HOMER holding a stable fraction across the -L range through its local background model. Read-depth sweeps showed HOMER peak counts scaling modestly with depth while CTK-unique peaks accumulated rapidly above 1M reads, with HOMER peaks remaining largely a high-confidence subset of CTK calls at all depths (Fig. 2D). CLIPittyClip therefore defaults to -L 2 and -localSize 1000 for HOMER and -minPH 3 for CTK; these represent sensible starting defaults that balance analysis stringency and exhaustiveness. Users with deeply sequenced libraries should note that CTK is substantially more sensitive to read depth and may require additional parameter adjustment to maintain consistent stringency across samples. The DDX17 locus, as an example, illustrates this tradeoff: HOMER peaks coincide with the highest-signal HuR binding sites while CTK additionally recovers lower-signal peaks distributed across the gene body (Fig. 2E).

Crosslink site determination and analysis with CLIPittyClip

CLink reimplements CTK CITS/CIMS calling but with three architectural differences. First, the PCR deduplication step uses the directional UMI-graph algorithm in umi_tools [41] in place of CTK tag2collapse.pl (Fig. 1D). Second, crosslink-position evidence is extracted in a single pass over the deduplicated BAM into a compressed array (pileup.npz) that is shared between CITS and CIMS, eliminating the duplicate BED and mutation-file scans required by CITS.pl and CIMS.pl steps in CTK. Third, statistical significance is assessed with Benjamini-Hochberg FDR correction rather than a permutation-based approach, substantially reducing run-time (Fig. 1E) while preserving control of the false discovery rate.

Turning to crosslink sites, we next compared CLink CITS and CIMS against CTK directly. To test this, we processed three datasets through both implementations: HuR and HNRNPC BrdU-CLIP data generated in our laboratory [21, 25], the original HNRNPC iCLIP datasets [7], and RBFOX2 BrdU-CLIP data from mouse brain [26], used originally for CTK CITS benchmarking [14]. Annotation profiles were consistent with known binding preferences in all cases, with peaks and CITS calls distributed similarly across genomic features (Fig. 3A and Supplementary Fig. S3A). Both implementations recovered the expected motifs, U-rich sequences for HuR and HNRNPC and the canonical UGCAUG for RBFOX2 (Fig. 3B), with motif density sharply enriched at the crosslink position (Fig. 3C). The top motif for each dataset is most enriched approximately 25 nt upstream of the peak center, while CITS calls coincide with the sharpest motif density at position 0. This offset is expected, since CITS marks the 5′ truncation position of each read while peaks reflect the distribution of read lengths traversing crosslink sites, whose center falls downstream of the read start by roughly half a typical CLIP tag length (30–60 nt). CITS calls from the two implementations overlapped extensively and fell predominantly within peaks (Fig. 3D). CIMS deletion analysis showed the same pattern at lower total counts, as expected given the lower frequency of crosslink-induced mutations relative to truncations, while retaining crosslink site resolution for the top motif of each RBP (Supplementary Fig. S3B and  C).

To assess how these results generalize to additional CLIP protocols beyond iCLIP-style datasets, we applied CLink to analogous eCLIP and PAR-CLIP datasets: HuR eCLIP from K562 cells [27], HNRNPC and RBFOX2 eCLIPs from both K562 and HepG2 cell lines [24], and HuR PAR-CLIP [22, 23]. In all cases, CITS and CIMS showed annotation profiles consistent with known binding preferences (Fig. 4A), recovered the expected top binding motifs (Fig. 4B), and yielded enriched motif densities near CITS or offset peaks (Fig. 4C). In the case of PAR-CLIP HuR data, where crosslink-induced truncations are not generated, we used T-to-C single-nucleotide mutation outputs from CLIPittyClip and recovered the expected U-rich motif with broad enrichment around peak centers (Fig. 4D). Across the eCLIP datasets, CITS consistently yielded more sites than CIMS, consistent with truncation being the dominant crosslink signature for UV-based crosslinking. Sites called by only one CLink or peak output remained enriched for the cognate motif relative to background, while agreement between two outputs, most often between Peaks and CITS, further increased motif frequency in most datasets (Supplementary Fig. S4). Taken together, these results establish CLink as a functionally equivalent and faster replacement for CTK that generalizes across CLIP protocol types and input sources.

CLIPittyClip outputs support biologically interpretable downstream analysis

To demonstrate that CLIPittyClip outputs support biological interpretation beyond motif recovery, we reprocessed previously published HuR fractionation-CLIP data from nuclear and cytoplasmic compartments [21] and compared binding between the two. Annotation of the resulting peaks and crosslink sites recovered the expected compartmental pattern for HuR with nuclear peaks dominated by intronic binding, while cytoplasmic peaks bound mostly to 3′UTRs (Fig. 5A), consistent with the established role of HuR in pre-mRNA processing in the nucleus and mRNA stability in the cytoplasm. CLink CITS and CIMS calls tracked the same compartmental shift within each fraction. To test differential binding directly, we passed the CLIPittyClip peak matrix to a DESeq2-based comparison of cytoplasmic versus nuclear peaks. Differentially bound sites separated cleanly by annotation, with 3′UTR peaks enriched in the cytoplasm and intronic peaks enriched in the nucleus (Fig. 5B). This compartment-specific pattern was recovered end to end from raw FASTQ using CLIPittyClip outputs and standard downstream tools.

Figure 5.

Application of CLIPittyClip outputs: stacked bars of HuR binding site locations in nuclear and cytoplasmic fractions, a volcano plot and an MA plot of differential binding, a diagram of the output folders, and a screenshot of the interactive wizard.

CLIPittyClip outputs are organized to be compatible with downstream tools. (A) Genic peaks, CLink CITS, and CLink deletion sites distribution between HuR CLIP from nuclear and cytoplasmic fractions. Genomic features are color coded with the total number of peaks/sites noted. (B) Differential binding analysis between nuclear and cytoplasmic HuR CLIP using DESeq2. Dotted vertical line denotes fold change threshold at two-fold. (C) CLIPittyClip output structure with key outputs as inputs for downstream analysis tools. (D) A command-line “Wizard” mode helps users interactively construct their CLIPittyClip commands.

These downstream analyses draw on CLIPittyClip outputs directly, without reformatting (Fig. 5C). Normalized bedgraph or BigWig files make visualization in browsers such as the UCSC Genome Browser [42] easy to implement. Bed file outputs, such as peaks.bed or CITS.bed, can be directly used for motif enrichment analysis, for which we used HOMERs findMotifs.pl [35] (Fig. 3B and 4B). Peak matrices from the aggregation module are compatible with DESeq [43] style analyses that require a count matrix. We also confirmed compatibility with two CLIP-seq-specific analysis packages: RBPBench [44], which scores peak and crosslink site sets against a curated database of known RBP-binding signatures; and BindingSiteFinder [30], which calls consensus binding sites across replicates using a DESeq2-based statistical framework. In both cases, CLIPittyClip peak and crosslink site outputs could be passed directly to these tools and produced results consistent with known RBP-binding preferences.

Finally, command-line pipelines present a meaningful barrier for new and less computationally experienced users. To lower this barrier, CLIPittyClip includes an interactive wizard mode that guides users through each parameter decision in sequence, generating a complete, ready-to-run command at the end (Fig. 5D). The wizard requires no prior knowledge of CLIPittyClip flag structure and handles the most common experimental configurations out of the box. The same flag-based design lets users configure architectures not yet in the index, and we maintain the index in the repository as new protocols appear.

Discussion

CLIPittyClip was built to fill a specific gap with a locally installable, single-command pipeline for CLIP-seq data that runs on hardware most labs already have. CLIPittyClip installs via conda in a single step and requires no cluster access, containers, or cloud resources. The pipeline is not designed to replace consortium-scale workflows such as the ENCODE eCLIP pipeline [8], its more recent Skipper implementation that replaces peak calling with transcriptome-wide window testing [17], nf-core/clipseq [16], or racoon_clip [19], which assume cluster or workstation-scale resources, and in the case of Skipper, a size-matched input library. CLIPittyClip is instead designed to get a new lab member from a FASTQ file to peaks and crosslink sites in a single command from their laptops, with outputs compatible with standard genome browsers and downstream analysis tools. CLIPittyClip does not perform gene-aware peak calling or differential binding analysis; these belong downstream, with established tools for each [30, 45].

We opted to keep peak calling and crosslink site analysis as separate, complementary outputs rather than collapsing them into a single step, a choice that reflects how differently the two outputs describe a protein-RNA interaction [46]. A peak marks a region of locally elevated read density that reflects RBP occupancy at steady state, but the reads contributing to it may span multiple contact positions or accumulate nonspecifically through mechanisms unrelated to crosslinking [47]. A crosslink site is a different kind of evidence, marking the position where reverse transcriptase stalled or introduced a mutation at the point of protein-RNA contact [7, 48]. The two often overlap at high-confidence sites, but what discordance means for any given site remains an open question. Callers such as Clippy [49] and iCount [50] that fold crosslink information into peak calling were therefore not tested, as this would collapse two biologically distinct outputs into one. The variable overlap between peaks, CITS, and CIMS in our hands is consistent with patterns reported broadly across CLIP datasets [12], and reflects the distinct information content of each output type.

This separation also informed the choice of peak caller. CLIPittyClip uses HOMER findPeaks [35] or CTK tag2peak [14], both selected because they are annotation-free and handle background without reference to gene models or transcript boundaries. This preserves noncoding and unannotated binding sites, generalizes across organisms and cell types without a matched GTF, and makes no assumptions about the transcriptome that is eligible for binding. Gene-aware callers such as CLIPper [34] can improve specificity for well-annotated exonic binding but apply per-gene models that are computationally prohibitive at our scale and break down for intronic or intergenic targets. HOMER, though initially developed for ChIP-seq [35] is actively maintained, straightforward to configure (e.g. -rna), and its local enrichment model has produced consistent results across a range of CLIP datasets [36–38]. We also evaluated Piranha [51] and omniCLIP [52] but excluded them, the former for build failures on modern toolchains, the latter for requiring a matched RNA-seq background unavailable for most archived data.

An important question on peaks is whether to normalize against an SMI library. SMI, introduced with eCLIP, uses input RNA at matched molecular weight as a background for crosslink-specific signal, and is informative for proteins with narrow target repertoires [8]. Its value for any given protein, however, is difficult to know in advance, since the normalization itself can introduce bias in a protein-dependent manner. Kuret et al. found, for example, that input normalization improves motif enrichment for A-rich binders while reducing it for U-rich binders [53]. For this reason, CLIPittyClip does not apply SMI normalization at the processing layer. Instead, SMI libraries are processed through the same layer as CLIP samples and assembled into the same combined peak matrix making differential enrichment against inputs with DESeq [43] or DEWSeq [45] easy to implement when a suitable input library exists. Most CLIP protocols, including iCLIP, irCLIP, and PAR-CLIP, do not generate SMI, and the eCLIP datasets reanalyzed here were processed without their SMI controls. CLIPittyClip nonetheless recovered the expected motifs and annotation profiles for all three RBPs, indicating that internal background estimation alone is sufficient for the recovery tasks shown here. Nevertheless, as internal estimation does not flag abundant structural RNAs, CLIPittyClip reports the read-level fraction mapping to rRNA, tRNA, snRNA, and snoRNA (--report-abundant-rna; Methods). A user can then determine whether a dataset is dominated by these classes before interpreting it. Where an SMI library is available, the same quantification applies to both, and comparing target-class enrichment in the CLIP against its input gives a direct read on whether SMI normalization is warranted for a given RBP.

This preference for internal background estimation extends to crosslink-site calling. CLink estimates crosslink-site significance internally, from the read-depth background at each position rather than against a separate control library, as in the original CTK implementation. Both implementations recover consensus binding motifs with similar positional precision, and CLink-unique sites show the same crosslink-position enrichment. For RBFOX2, where the canonical UGCAUG motif provides a stringent benchmark, CLink and CTK performed equivalently on the same dataset used for the original CTK implementation [14]. The advantage of CLink is both memory and speed. At 25M mapped reads, CTK requires 44 GB for deduplication and 41 GB for CITS/CIMS, exceeding a typical laptop, while CLink performs the same analyses with 1.2 and 2 GB. The same holds for run-time with full CITS and CIMS analysis of 25M mapped reads completed in 36 min with CLink against roughly 16 h with CTK. Because no single set of parameters is likely optimal across RBPs, new proteins typically require testing several peak-calling and crosslink thresholds to find settings that hold up across depth and replicates. CLIPittyClip enables fast, low-memory iteration to make this process feasible on ordinary hardware. Overall, these choices reflect the design of CLIPittyClip as a lightweight platform for processing CLIP data, whether for a new lab member analyzing their first experiment or an experienced user characterizing a new CLIP dataset.

Supplementary Material

lqag138_Supplemental_Files

Acknowledgements

We thank Chaolin Zhang for thoughtful discussion and advice. We further thank Vidusha Kothwela and Rohan Krishna for testing additional Luna lab computers for CLIPittyClip usage. The name was inspired by an offhand comment on CLIP analysis from Troels K. H. Scheel (U. Copenhagen); we thank him.

Author contributions: Soon Yi (Conceptualization [equal], Formal analysis [equal], Investigation [equal], Methodology [equal], Software [equal], Validation [equal], Visualization [equal], Writing—original draft [equal], Writing—review & editing [equal]) and Joseph M. Luna (Conceptualization [equal], Formal analysis [equal], Funding acquisition [lead], Investigation [equal], Methodology [equal], Project administration [lead], Software [equal], Supervision [lead], Validation [equal], Visualization [equal], Writing—original draft [equal], Writing—review & editing [equal]).

Contributor Information

Soon Yi, Center for RNA Science and Therapeutics, School of Medicine, Case Western Reserve University, Cleveland, OH 44106, United States; Department of Biochemistry, School of Medicine, Case Western Reserve University, Cleveland, OH 44106, United States; Department of Psychiatry, University of Pennsylvania, Philadelphia, PA 19104, United States.

Joseph M Luna, Center for RNA Science and Therapeutics, School of Medicine, Case Western Reserve University, Cleveland, OH 44106, United States; Department of Biochemistry, School of Medicine, Case Western Reserve University, Cleveland, OH 44106, United States.

Supplementary data

Supplementary data is available at NAR Genomics & Bioinformatics online.

Conflict of interest

None declared.

Funding

This work was supported in part by the NIH (R35 GM154771 to J.M.L.) as well as start-up funds (to J.M.L.) from the Department of Biochemistry at the Case Western Reserve University School of Medicine.

Data availability

CLIPittyClip is freely available on GitHub at https://github.com/LunaRNALab/CLIPittyClip. A stable release v3.5.0 is also available on Zenodo at https://doi.org/10.5281/zenodo.20875662. All benchmarking, downstream analysis scripts, and a full list of CLIPittyClip commands performed for the data presented in this work can be found at https://github.com/LunaRNALab/BiTS_CLIPittyClip. Published and in-house datasets used in this study are listed in Supplementary Table S1 and are as follows: HuR BrdU-CLIP [21], HNRNPC BrdU-CLIP [25], HNRNPC iCLIP [7], RBFOX2 BrdU-CLIP [26], HuR eCLIP [27] (ENCODE accession: ENCSR090LNQ), HNRNPC eCLIP and RBFOX2 eCLIP [24] (ENCODE accession: ENCSR550DVK and ENCSR249ROI for HNRNPC HepG2 and K562 eCLIP, respectively; ENCSR987FTF and ENCSR756CKJ for RBFOX2 HepG2 and K562 eCLIP, respectively), HuR PAR-CLIP [22, 23].

References

  • 1. Hafner  M, Katsantoni  M, Köster  T  et al.  CLIP and complementary methods. Nat Rev Methods Primers. 2021;1:20. 10.1038/s43586-021-00018-1 [DOI] [Google Scholar]
  • 2. Ule  J, Jensen  KB, Ruggiu  M  et al.  CLIP identifies Nova-regulated RNA networks in the brain. Science. 2003;302:1212–5. 10.1126/science.1090095 [DOI] [PubMed] [Google Scholar]
  • 3. Licatalosi  DD, Mele  A, Fak  JJ  et al.  HITS-CLIP yields genome-wide insights into brain alternative RNA processing. Nature. 2008;456:464–9. 10.1038/nature07488 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 4. Yeo  GW, Coufal  NG, Liang  TY  et al.  An RNA code for the FOX2 splicing regulator revealed by mapping RNA-protein interactions in stem cells. Nat Struct Mol Biol. 2009;16:130–7. 10.1038/nsmb.1545 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 5. Chi  SW, Zang  JB, Mele  A  et al.  Argonaute HITS-CLIP decodes microRNA–mRNA interaction maps. Nature. 2009;460:479–86. 10.1038/nature08170 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 6. Hafner  M, Landthaler  M, Burger  L  et al.  Transcriptome-wide identification of RNA-binding protein and microRNA target sites by PAR-CLIP. Cell. 2010;141:129–41. 10.1016/j.cell.2010.03.009 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 7. König  J, Zarnack  K, Rot  G  et al.  iCLIP reveals the function of hnRNP particles in splicing at individual nucleotide resolution. Nat Struct Mol Biol. 2010;17:909–15. 10.1038/nsmb.1838 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 8. Van Nostrand  EL, Pratt  GA, Shishkin  AA  et al.  Robust transcriptome-wide discovery of RNA-binding protein binding sites with enhanced CLIP (eCLIP). Nat Methods. 2016;13:508–14. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 9. Zarnegar  BJ, Flynn  RA, Shen  Y  et al.  irCLIP platform for efficient characterization of protein–RNA interactions. Nat Methods. 2016;13:489–92. 10.1038/nmeth.3840 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 10. Buchbender  A, Mutter  H, Sutandy  FXR  et al.  Improved library preparation with the new iCLIP2 protocol. Methods. 2020;178:33–48. 10.1016/j.ymeth.2019.10.003 [DOI] [PubMed] [Google Scholar]
  • 11. Ducoli  L, Zarnegar  BJ, Porter  DF  et al.  irCLIP-RNP and Re-CLIP reveal patterns of dynamic protein assemblies on RNA. Nature. 2025;641:769–78. 10.1038/s41586-025-08787-5 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 12. Sahadevan  S, Sekaran  T, Hentze  MW  et al.  Signal or noise? RNA-binding proteins and the challenges of binding site assignments. Nucleic Acids Res. 2026;54:gkag452. 10.1093/nar/gkag452 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 13. Uhl  M, Houwaart  T, Corrado  G  et al.  Computational analysis of CLIP-seq data. Methods. 2017;118-119:60–72. 10.1016/j.ymeth.2017.02.006 [DOI] [PubMed] [Google Scholar]
  • 14. Shah  A, Qian  Y, Weyn-Vanhentenryck  SM  et al.  CLIP Tool Kit (CTK): A flexible and robust pipeline to analyze CLIP sequencing data. Bioinformatics. 2017;33:566–7. 10.1093/bioinformatics/btw653 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 15. Flynn  RA, Martin  L, Spitale  RC  et al.  Dissecting noncoding and pathogen RNA–protein interactomes. RNA. 2015;21:135–43. 10.1261/rna.047803.114 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 16. West  C, Capitanchik  C, Cheshire  C  et al.  nf-core/clipseq—a robust Nextflow pipeline for comprehensive CLIP data analysis. Wellcome Open Res. 2023;8:286. 10.12688/wellcomeopenres.19453.1 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17. Boyle  EA, Her  H-L, Mueller  JR  et al.  Skipper analysis of eCLIP datasets enables sensitive detection of constrained translation factor binding sites. Cell Genom. 2023;3:100317. 10.1016/j.xgen.2023.100317 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18. Chen  B, Yun  J, Kim  MS  et al.  PIPE-CLIP: a comprehensive online tool for CLIP-seq data analysis. Genome Biol. 2014;15:R18. 10.1186/gb-2014-15-1-r18 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 19. Klostermann  M, Zarnack  K. racoon_clip—a complete pipeline for single-nucleotide analyses of iCLIP and eCLIP data. Bioinform Adv. 2024;4:vbae084. 10.1093/bioadv/vbae084 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 20. Polash  AH, Hafner  M. PCLIPtools: A robust framework for identifying RNA-protein interaction sites from PAR-CLIP experiments. Nucleic Acids Res. 2026;54:gkag062. 10.1093/nar/gkag062 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 21. Yi  S, Singh  SS, Rozen-Gagnon  K  et al.  Mapping RNA–protein interactions with subcellular resolution using colocalization CLIP. RNA. 2024;30:920–37. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 22. Kishore  S, Jaskiewicz  L, Burger  L  et al.  A quantitative analysis of CLIP methods for identifying binding sites of RNA-binding proteins. Nat Methods. 2011;8:559–64. 10.1038/nmeth.1608 [DOI] [PubMed] [Google Scholar]
  • 23. Mukherjee  N, Corcoran  DL, Nusbaum  JD  et al.  Integrative regulatory mapping indicates that the RNA-binding protein HuR couples pre-mRNA processing and mRNA stability. Mol Cell. 2011;43:327–39. 10.1016/j.molcel.2011.06.007 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 24. Van Nostrand  EL, Freese  P, Pratt  GA  et al.  A large-scale binding and functional map of human RNA-binding proteins. Nature. 2020;583:711–9. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 25. Yi  S, Singh  SS, Ye  X  et al.  Inherent Specificity and Variation Sensitivity as Quantitative Metrics for RBP Binding. bioRxiv, 10.1101/2025.03.28.646018, 1 April 2025, preprint: not peer reviewed. [DOI] [Google Scholar]
  • 26. Weyn-Vanhentenryck  SM, Mele  A, Yan  Q  et al.  HITS-CLIP and integrative modeling define the Rbfox splicing-regulatory network linked to brain development and autism. Cell Rep. 2014;6:1139–52. 10.1016/j.celrep.2014.02.005 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 27. Kagda  MS, Lam  B, Litton  C  et al.  Data navigation on the ENCODE portal. Nat Commun. 2025;16:9592. 10.1038/s41467-025-64343-9 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 28. Blue  SM, Yee  BA, Pratt  GA  et al.  Transcriptome-wide identification of RNA-binding protein binding sites using seCLIP-seq. Nat Protoc. 2022;17:1223–65. 10.1038/s41596-022-00680-z [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 29. Zhang  Z, Xing  Y. CLIP-seq analysis of multi-mapped reads discovers novel functional RNA regulatory sites in the human transcriptome. Nucleic Acids Res. 2017;45:9260–71. 10.1093/nar/gkx646 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 30. Lewinski  M, Brüggemann  M, Köster  T  et al.  Mapping protein-RNA binding in plants with individual-nucleotide-resolution UV cross-linking and immunoprecipitation (plant iCLIP2). Nat Protoc. 2024;19:1183–234. 10.1038/s41596-023-00935-3 [DOI] [PubMed] [Google Scholar]
  • 31. Chen  S, Zhou  Y, Chen  Y  et al.  fastp: an ultra-fast all-in-one FASTQ preprocessor. Bioinformatics. 2018;34:i884–90. 10.1093/bioinformatics/bty560 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 32. Langmead  B, Salzberg  SL. Fast gapped-read alignment with Bowtie 2. Nat Methods. 2012;9:357–9. 10.1038/nmeth.1923 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 33. Dobin  A, Davis  CA, Schlesinger  F  et al.  STAR: ultrafast universal RNA-seq aligner. Bioinformatics. 2013;29:15–21. 10.1093/bioinformatics/bts635 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 34. Lovci  MT, Ghanem  D, Marr  H  et al.  Rbfox proteins regulate alternative mRNA splicing through evolutionarily conserved RNA bridges. Nat Struct Mol Biol. 2013;20:1434–42. 10.1038/nsmb.2699 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 35. Heinz  S, Benner  C, Spann  N  et al.  Simple combinations of lineage-determining transcription factors prime cis-regulatory elements required for macrophage and B cell identities. Mol Cell. 2010;38:576–89. 10.1016/j.molcel.2010.05.004 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 36. Rozen-Gagnon  K, Gu  M, Luna  JM  et al.  Argonaute-CLIP delineates versatile, functional RNAi networks in Aedes aegypti, a major vector of human viruses. Cell Host Microbe. 2021;29:834–848.e13. 10.1016/j.chom.2021.03.004 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 37. Scheel  TKH, Moore  MJ, Luna  JM  et al.  Global mapping of miRNA-target interactions in cattle (Bos taurus). Sci Rep. 2017;7:8190. 10.1038/s41598-017-07880-8 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 38. Luna  JM, Barajas  JM, Teng  K-Y  et al.  Argonaute CLIP defines a deregulated miR-122-bound transcriptome that correlates with patient survival in human liver cancer. Mol Cell. 2017;67:400. 10.1016/j.molcel.2017.06.025 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 39. Abdelmohsen  K, Gorospe  M. Posttranscriptional regulation of cancer traits by HuR. Wiley Interdiscip Rev RNA. 2010;1:214–29. 10.1002/wrna.4 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 40. Lebedeva  S, Jens  M, Theil  K  et al.  Transcriptome-wide analysis of regulatory interactions of the RNA-binding protein HuR. Mol Cell. 2011;43:340–52. 10.1016/j.molcel.2011.06.008 [DOI] [PubMed] [Google Scholar]
  • 41. Smith  T, Heger  A, Sudbery  I. UMI-tools: modeling sequencing errors in Unique Molecular Identifiers to improve quantification accuracy. Genome Res. 2017;27:491–9. 10.1101/gr.209601.116 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 42. Casper  J, Speir  ML, Raney  BJ  et al.  The UCSC Genome Browser database: 2026 update. Nucleic Acids Res. 2026;54:D1331–5. 10.1093/nar/gkaf1250 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 43. Love  MI, Huber  W, Anders  S. Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol. 2014;15:550. 10.1186/s13059-014-0550-8 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 44. michauhl. RBPBench: Evaluate CLIP-seq and other genomic region data using a comprehensive collection of known RBP binding motifs. GitHub. https://github.com/michauhl/RBPBench (May 2026, date last accessed). [Google Scholar]
  • 45. Schwarzl  T, Sahadevan  S, Lang  B  et al.  Improved discovery of RNA-binding protein binding sites in eCLIP data using DEWSeq. Nucleic Acids Res. 2024;52:e1. 10.1093/nar/gkad998 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 46. Feng  H, Bao  S, Rahman  MA  et al.  Modeling RNA-binding protein specificity in vivo by precisely registering protein-RNA crosslink sites. Mol Cell. 2019;74:1189–204.e6. 10.1016/j.molcel.2019.02.002 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 47. Kristofich  J, Nicchitta  CV. High-throughput quantitation of protein-RNA UV-crosslinking efficiencies as a predictive tool for high-confidence identification of RNA-binding proteins. RNA. 2024;30:644–61. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 48. Zhang  C, Darnell  RB. Mapping in vivo protein-RNA interactions at single-nucleotide resolution from HITS-CLIP data. Nat Biotechnol. 2011;29:607–14. 10.1038/nbt.1873 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 49. Ule Lab . Clippy: a wrapper around SciPy's find_peaks function to enable peak calling of CLIP data. [Computer software]. GitHub. https://github.com/ulelab/clippy  (May 2026, date last accessed). [Google Scholar]
  • 50. Curk  T. iCount: Protein–RNA interaction analytics. [Computer software]. GitHub. https://github.com/tomazc/iCount  (May 2026, date last accessed).
  • 51. Uren  PJ, Bahrami-Samani  E, Burns  SC  et al.  Site identification in high-throughput RNA–protein interaction data. Bioinformatics. 2012;28:3013–20. 10.1093/bioinformatics/bts569 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 52. Drewe-Boss  P, Wessels  H-H, Ohler  U. omniCLIP: Probabilistic identification of proteinRNA interactions from CLIP-seq data. Genome Biol. 2018;19:183. 10.1186/s13059-018-1521-2 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 53. Kuret  K, Amalietti  AG, Jones  DM  et al.  Positional motif analysis reveals the extent of specificity of protein-RNA interactions observed by CLIP. Genome Biol. 2022;23:191. 10.1186/s13059-022-02755-2 [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

lqag138_Supplemental_Files

Data Availability Statement

CLIPittyClip is freely available on GitHub at https://github.com/LunaRNALab/CLIPittyClip. A stable release v3.5.0 is also available on Zenodo at https://doi.org/10.5281/zenodo.20875662. All benchmarking, downstream analysis scripts, and a full list of CLIPittyClip commands performed for the data presented in this work can be found at https://github.com/LunaRNALab/BiTS_CLIPittyClip. Published and in-house datasets used in this study are listed in Supplementary Table S1 and are as follows: HuR BrdU-CLIP [21], HNRNPC BrdU-CLIP [25], HNRNPC iCLIP [7], RBFOX2 BrdU-CLIP [26], HuR eCLIP [27] (ENCODE accession: ENCSR090LNQ), HNRNPC eCLIP and RBFOX2 eCLIP [24] (ENCODE accession: ENCSR550DVK and ENCSR249ROI for HNRNPC HepG2 and K562 eCLIP, respectively; ENCSR987FTF and ENCSR756CKJ for RBFOX2 HepG2 and K562 eCLIP, respectively), HuR PAR-CLIP [22, 23].


Articles from NAR Genomics and Bioinformatics are provided here courtesy of Oxford University Press

RESOURCES