Profile ChIP-seq or ATAC-seq signal around genomic features

Hand Claude Code a set of aligned ChIP-seq or ATAC-seq BAM files; get back normalized bigWig coverage tracks, a sample-correlation QC matrix, and profile/heatmap figures of signal centered on TSSs or a peak set.

   
Problem class Data analysis
Subject areas Molecular and Cellular Biology, Immunology and Microbiology
Evidence level Proposed
Complexity One skill or MCP
Availability Fully open
Compute Laptop (no GPU)

Problem

After aligning a ChIP-seq, CUT&RUN, or ATAC-seq experiment, the next step is almost always the same: convert BAMs to normalized coverage tracks, check that replicates agree, and visualize signal aggregated around a feature set (transcription start sites, called peaks, enhancers). deepTools is the field-standard toolkit for this, but it is a chain of CLI tools (bamCoverage, multiBamSummary, plotCorrelation, computeMatrix, plotHeatmap, plotProfile) with normalization flags (RPKM, CPM, BPM, --effectiveGenomeSize) that are easy to mis-set, and the matrix/plotting steps need their arguments to line up. Solved looks like: BAMs in, a QC correlation figure plus a publication-ready TSS or peak heatmap out, with every normalization and binning choice explicit and reproducible.

  1. Install the deepTools Claude Skill from the K-Dense collection:

    npx skills add K-Dense-AI/scientific-agent-skills

    Enable the deeptools skill when prompted. deepTools’ own dependencies install on first use (the skill uses uv / pip).

  2. Prepare inputs. Provide coordinate-sorted, indexed BAMs (*.bam + *.bam.bai) for each sample/replicate, a genome build (so the effective genome size is correct), and a feature file: a BED/GTF of TSSs or a peak BED (e.g., MACS2 narrowPeak). For ATAC-seq, decide up front whether to shift reads for Tn5 (--Offset) — state it in the prompt.

  3. Make normalized tracks and a replicate-QC matrix. A minimal prompt:

    Use the deeptools skill. For each BAM in bams/, run bamCoverage with
--normalizeUsing BPM --binSize 25 --effectiveGenomeSize 2913022398
(GRCh38) and write bigWigs to tracks/. Then run multiBamSummary in
bins mode over all BAMs and plotCorrelation (Spearman, heatmap) to
qc/replicate_correlation.png so I can confirm replicates cluster.

  4. Center the signal on features and plot. Continue:

    Now run computeMatrix reference-point --referencePoint TSS -b 3000
-a 3000 --binSize 25 over the bigWigs in tracks/ against
features/tss.bed, then plotHeatmap (sorted by mean signal) to
figures/tss_heatmap.png and plotProfile to figures/tss_profile.png.
Use scale-regions instead of reference-point if I give you gene
bodies rather than TSSs.

  5. Inspect the correlation matrix first. If replicates don’t cluster above the rest, stop and check the alignment/QC upstream before trusting the heatmaps.

Why this assembly

Rung 2 of the simplicity ladder. Plain Claude Code could shell out to deepTools from documentation, but the skill encodes the correct tool ordering (coverage → summary → matrix → plot), the normalization vocabulary (BPM vs RPKM vs CPM), and the effective-genome-size and bin-size conventions that make the tracks comparable across samples — the exact places ad-hoc scripts go wrong. This is a single well-bounded analysis served by a single toolkit, so there is no reason to escalate to a multi-tool harness or an autonomous system. Peak calling (MACS2) and differential binning (DiffBind/csaw) are separate steps not in this skill; this recipe stops at normalized tracks and feature-centered visualization.

Availability

Fully open. The deepTools skill is OSS (BSD) in K-Dense-AI/scientific-agent-skills; deepTools itself is BSD-licensed. No subscription or institutional access required.

Compute requirements

Laptop-sufficient for a small experiment; no GPU. bamCoverage and computeMatrix are I/O- and CPU-bound, and runtime scales with BAM size and read depth. A typical mammalian ChIP-seq sample (20–40 M reads) converts to a 25-bp-bin bigWig in a few minutes per sample on 4–8 cores, and whole-experiment computeMatrix over thousands of TSSs is minutes more; expect a few hundred MB of bigWig per sample. Move to a workstation (16–32 GB RAM, more cores via --numberOfProcessors) only when you have many samples or run whole-genome bins.

Evidence

Proposed. No documented attempt at an LLM-driven (Claude + deepTools skill) ChIP-seq/ATAC-seq profiling workflow is known. The grounding is component-level: deepTools is a peer-reviewed, widely cited toolkit (Ramírez et al., Nucleic Acids Research 44:W160 (2016)) whose bamCoverage/computeMatrix/plotHeatmap chain is the established convention for exactly this task, and the deepTools skill catalog entry documents that the skill drives these tools locally. The closest analogous documented LLM workflow is the cataloged Biomni agent, which composes genomics CLI tools (including coverage/visualization steps) inside an autonomous loop; this recipe pulls that capability down to the simplest rung that solves the stated problem.

Alternatives considered

  • Plain Claude Code, no skill. Fine for a one-off where you want to audit every flag, but you lose the encoded normalization conventions and re-derive the tool ordering each time.
  • A Nextflow/Snakemake pipeline (e.g., nf-core/chipseq). The right choice for production cohorts run repeatedly on a cluster — see the Nextflow development skill. This recipe is for the interactive, exploratory “I have BAMs, show me the signal” case, not a hardened batch pipeline.
  • An autonomous-science system (Biomni). Overkill for a fixed visualization task; the autonomous loop only earns its overhead when track generation is one node in a larger generated analysis.

See also

Sources

Tried this recipe?

Share feedback — what worked, what didn’t, what you’d change. The form opens with this recipe pre-selected and a link back to this page.