Quantify a western blot from a scanned image

Hand Claude Code a scanned blot image and get back per-lane band intensities, a target-over-loading-control normalized value for every lane, replicate-level statistics, and a QC overlay — produced by the Western Blot Quantification skill and saved as a re-runnable script.

   
Problem class Data analysis
Subject areas Molecular and Cellular Biology
Evidence level Reported
Complexity One skill or MCP
Availability Fully open
Compute Laptop

Problem

Densitometry of western blots is one of the most common — and most quietly irreproducible — quantitative steps in cell biology. A postdoc scores a knockdown or a phospho-signal by drawing rectangles in ImageJ, subtracting a background band she picks by eye, dividing by a GAPDH lane, and pasting the numbers into a spreadsheet. Every one of those choices (ROI placement, background model, which loading control, how replicates are combined) is a degree of freedom that changes the fold-change, and none of it is captured anywhere the reviewer can see. Housekeeping-protein normalization is itself a known bias: β-actin, GAPDH, and tubulin are not constant across injury, disease, or hormone state.

“Solved” looks like: point at the blot image, get back one tidy table with a row per lane (lane, target raw intensity, loading-control raw intensity, normalized ratio), replicate-aggregated fold-changes with SD/SEM, and a verification image showing exactly which pixels each band ROI covered — all produced by a committed script that re-runs identically on the next blot, so the number in the figure is auditable rather than eyeballed once.

  1. Install the Western Blot Quantification skill. It ships in the SciAgent-Skills collection — clone once and load as a plugin:

    git clone https://github.com/jaechang-hits/SciAgent-Skills

    Then inside Claude Code run /plugin install sciagent-skills and confirm it appears under /plugin → Installed. The skill runs its Python locally — no upload.

  2. Stage the image and record the lane map. Save the raw scan (uncropped, unadjusted TIFF/PNG — never a JPEG a journal compressed) in one folder. Write down, per lane: the sample/condition, which bands are the target, and which are the loading control. Prefer a total-protein loading control (stain-free or REVERT) over a single housekeeping protein where you have it — it is materially more accurate and reproducible (see Evidence).

  3. Generate a committed quantification script. Have Claude write the analysis to a file rather than eyeballing ROIs interactively, so the same detection and normalization apply to every blot:

    Use the Western Blot Quantification skill on ./blots/blot01.tif.
Write a script quantify_blot.py that:
  - uses analyze_pixel_distribution / find_roi_from_image to detect
    the target and loading-control band ROIs per lane,
  - subtracts local background and integrates band intensity,
  - normalizes each lane's target to its loading control, then to
    the control lane, emitting one row per lane to blot01_quant.csv
    (lane, condition, target_raw, loading_raw, normalized_ratio),
  - aggregates biological replicates (mean, SD, SEM, fold-change vs
    control) into replicate_summary.csv,
  - saves a verification overlay (roi_overlay.png) showing every ROI,
  - records skill commit, package versions, the image sha256, the
    normalization method, and background model in provenance.json.
Then summarize replicate_summary.csv.

    Pin the environment (requirements.txt / environment.yml) with the exact package versions, and commit quantify_blot.py, the pinned env, the raw image, and provenance.json. The recorded ROI method, background model, and normalization choice are what make the fold-change reproducible — see the reproducibility guide.

  4. QC the ROI overlay before you trust the numbers. Open roi_overlay.png and check the two failure modes that silently corrupt densitometry: an ROI that clips a saturated band (integrated intensity plateaus — the signal is out of linear range, re-scan at lower exposure) and a background window that overlaps a neighboring band. If either appears, adjust the detection threshold or background window in the script and re-run — never hand-edit blot01_quant.csv.

  5. Hand off downstream. replicate_summary.csv is the artifact: feed it to your statistics (report the test and n) and cite the committed script and provenance.json in the methods. The per-lane table is the audit trail for any reviewer who asks how a band became a number.

Why this assembly

Rung 2 of the simplicity ladder. Plain Claude Code (rung 1) can write ad-hoc densitometry code, but it has no encoded convention for band ROI detection, background subtraction, or the two-step target-to-loading-control normalization — so every run drifts exactly where reproducibility matters. The skill encodes that workflow (analyze_pixel_distribution/find_roi_from_image for detection, loading-control-then-target normalization, replicate aggregation) and emits a verification grid, which is a real gain over rung 1. There is nothing to escalate to at rung 3/4: this is a single, well-bounded image-to-table task. The one judgment call the recipe surfaces explicitly is the loading control (prefer total-protein), because that is the dominant bias in the published literature.

Availability

Fully open. The Western Blot Quantification skill is part of the community SciAgent-Skills collection (CC BY 4.0); it installs its own Python dependencies and runs locally with no account, API key, or image upload. TIFF/PNG and CSV are open formats.

Compute requirements

Laptop-sufficient. Band detection and integration on a single scanned blot (a few megapixels) run in seconds on a laptop CPU — no GPU. A folder of dozens of blots is laptop-scale as a loop. The only “heavy” resource is disk for keeping uncompressed raw scans under version control.

Evidence

Reported. The quantitative core — integrated-density band quantification normalized to a loading control, with total-protein normalization preferred over a single housekeeping protein — is field-standard and directly evidenced. A stain-free total-protein control reduced variability enough to cut the samples needed for significance by >50% versus actin/tubulin (Maloy et al., Anal. Biochem. 2022); REVERT total-protein stain gave a wider linear range and better gel-to-gel consistency than GAPDH/actin/tubulin in brain homogenates (Kirshner & Gibbs, Mol. Cell. Endocrinol. 2018); and titration-based WB further removes housekeeping-normalization bias using regression-curve quantification (Maestri et al., PLOS ONE 2025). The skill’s two-step normalization and replicate aggregation follow this practice.

No head-to-head benchmark of the agent-driven assembly versus a hand-run ImageJ densitometry exists — the skill buys a local, committed, reproducible run with a verification overlay, not a new measurement method. That gap is why this recipe is Reported, not Validated.

Alternatives considered

  • ImageJ/Fiji gel-analysis GUI (rung 0–1). The classic path: draw lanes, plot profiles, measure peaks by hand. Reach for it for a one-off blot where interactive curation is faster than scripting. The skill is worth it when you want the ROIs, background model, and normalization pinned and re-runnable across every blot in a paper.
  • scikit-image by hand (rung 1–2). scikit-image can do the thresholding/regionprops that underlie band detection, but you then re-implement ROI logic and the two-step normalization yourself. The skill encodes that; drop to scikit-image only for a non-standard blot geometry the skill can’t detect.
  • Segment and quantify cells in a microscopy image. Different problem — instance segmentation of cells, not lane/band densitometry. Cross-referenced because both are committed image-to-table cell-biology quantifications.

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.