CFU Plot Studio

Version 0.4.0. Statistics from 0.1.x are not comparable with 0.2.0 and later — see CHANGELOG.md and AUDIT-2026-09.md before re-using an older number.

CFU Plot Studio is an R Shiny app for turning replicate-level colony forming unit data into publication-focused bar plots, summary tables, QC checks, and statistics.

The app is designed for microbiology lab workflows where data are usually collected as rows of replicate measurements across samples, treatments, timepoints, and CFU counts.

What It Does

• Imports replicate-level CSV files without requiring a fixed column order.
• Maps sample/vector, treatment/dose/condition, timepoint, replicate, and CFU columns inside the app.
• Plots CFU summaries as bar plots or point plots with SD, SEM, 95% CI, IQR, or min-max variation intervals.
• Shows individual replicate points on top of bars.
• Draws the figure preview at the exact export geometry, so the font sizes on screen are the font sizes in the exported file.
• Runs statistics on log10(CFU) at the level of the biological replicate, reporting the confidence interval, fold-change interval, and Hedges' g alongside each p value.
• Reports advisory spread and normality diagnostics on the same units the tests use.
• Censors counts at or below zero at a limit of detection, or excludes them, and records which.
• Supports sample/vector comparisons, timepoint comparisons, treatment versus control, and all treatment-pair comparisons.
• Exports cleaned data, summary tables, QC tables, statistics, and ANOVA tables.
• Exports figures as high-resolution PNG, PDF, SVG, animated GIF, and PowerPoint.
• Can create editable PowerPoint vector figures when the rvg package is installed.
• Saves and reloads plot-style presets as JSON.
• Keeps one style for a whole paper: colours belong to sample and timepoint names, one font family for every text, and a paper style that is applied every time the app starts.
• Exports an analysis manifest and a reproducible R script for the current figure.
• Includes a Figure QA checklist for publication-readiness checks.

Data Format

Your CSV should contain one row per replicate measurement. Column names can vary because the app lets you map them.

Required information:

Field	Example
Sample/vector	Control strain, Test strain
Treatment/dose/condition	Baseline, Treatment A, Treatment B
Timepoint	Early, Late
Replicate	1, 2, 3
CFU	1200000

The included dummy_cfu_example.csv is synthetic example data and can be downloaded from the app as a template.

Biological versus technical replicates

This is the single most consequential thing the app cannot work out from your file, so it asks. Biological replicate column in the sidebar says which column identifies independent cultures. Rows that share a value in that column are technical repeats — the same culture plated more than once — and are averaged on the log10 scale (the geometric mean of the plates) before any test runs. n then counts cultures, not plates.

By default the app uses the Replicate column for both, which is right when one row is one culture. If your file has a separate plate or well column, point the biological replicate selector at the culture column instead. If the file genuinely has no independent replication, choose None in this file — the app will not invent one. It labels the banner, the figure caption, the QC table and a unit_of_replication column on every result "technical replicates only", and those numbers must not be reported as replicate-level statistics.

Before version 0.2.0 every unpaired test counted every row, so three cultures plated four times each entered a t-test as n = 12. See AUDIT-2026-09.md.

Decimal-comma files (1,5E+08 with ; separators) and tab-separated files are read too; the CSV dialect control sniffs the delimiter by default.

Running The App

Windows: double-click

Double-click Run CFU Plot Studio.bat. It finds R, installs anything missing into a private .Rlibrary folder beside the app, picks a free port and opens your browser. Close the console window to stop the app. Nothing is installed system-wide and nothing leaves your machine.

The first run installs packages and can take several minutes. Later runs start in seconds.

If R is installed somewhere the launcher does not look, point it at your Rscript.exe:

setx CFU_RSCRIPT "C:\Program Files\R\R-4.5.0\bin\Rscript.exe"

Don't have R? Install it from https://cran.r-project.org/bin/windows/base/ first.

Any platform: from a shell

Rscript run_app.R

run_app.R does the same dependency check and port selection, and works from any working directory. Overrides:

Variable	Default	Purpose
CFU_APP_HOST	127.0.0.1	Bind address
CFU_APP_PORT	first free from 4267	Fixed port; errors if that port is busy
CFU_APP_LIB	<app>/.Rlibrary	Where missing packages are installed
CFU_NO_INSTALL	unset	Set to 1 to fail rather than install anything

From an R session

shiny::runApp(".")

This path assumes the packages below are already installed.

Installation

The launchers handle this for you. To install by hand:

install.packages(c(
  "shiny", "ggplot2", "dplyr", "readr", "tibble", "tidyr", "scales",
  "emmeans", "broom", "DT", "colourpicker", "jsonlite"
))

Optional export packages. Each one only affects the export format named, and the app reports which are absent at startup:

install.packages(c(
  "officer",    # PowerPoint export
  "rvg",        # editable vector art inside PowerPoint
  "gganimate",  # animated GIF export
  "gifski"      # GIF encoding
))

Figure Size

Width and height are entered in inches or millimetres; switching the unit converts the numbers so the physical figure stays the same size. A readout under the size boxes always states the geometry both ways plus the resulting pixel dimensions, for example 3.50 x 2.76 in (89 x 70 mm) at 600 DPI = 2100 x 1656 px.

The on-screen preview is rendered at that exact geometry rather than stretched to fill the browser pane. Because the preview device resolution scales with it, an 8 pt tick label occupies the same fraction of the figure on screen as it will in the exported file — so a figure tuned against the preview does not need re-tuning after export.

Presets: single column (3.35 x 2.65 in), double column (7.0 x 4.2 in), square (4.5 x 4.5 in), Nature single column (89 mm), Nature double column (183 mm). All set 600 DPI.

Download Everything

Download everything (.zip) bundles the current figure and everything needed to defend or rebuild it:

File	What it is
figure.png / .pdf / .svg / .pptx	The figure in every static format
legend_key.png / .pdf / .svg	The figure's legend key on its own (see Legend key)
legend_text.txt	The legend text starter: each group with its colour and marker in words
figure_reveal.gif	Optional, off by default because it is slow
data_cleaned.csv	Every row that survived import
data_plotted.csv	The rows actually plotted — in survival mode the paired ratios, each with both source counts
summary.csv	Exactly what the Summary tab shows
statistics.csv	Every comparison with intervals, effect sizes and adjusted q values
anova.csv	The ANOVA table, naming its sum-of-squares type
assumption_diagnostics.csv	Per-cell spread and normality checks on the tested units (advisory)
lod_wells.csv	Present only when wells were censored: every affected well and what was done with it
qc_replicates.csv, figure_qa.csv	Source checks, replicate structure, publication checklist
recreate_figure.R	Standalone script that rebuilds the figure from embedded data
plot_preset.json	Every setting, reloadable from the sidebar
analysis_manifest.json	What was loaded, what was filtered, how the figure was made
README.txt	Contents, the figure geometry, the readout, and what could not be produced

Each item is attempted independently, so a missing optional package costs that one file rather than the whole archive, and the README says which are absent and why.

The bundle names the code that produced it. analysis_manifest.json, README.txt and the header of recreate_figure.R all carry the app version, plus the short git SHA when the app is run from a checkout. A bundle that cannot identify its own version is not reproducible, whatever else it contains.

A figure file is verified after it is written. ggsave reports a graphics device that failed to open as a warning, not an error, so a download could previously hand back an empty file and the bundle could lose figure.pdf in silence — capabilities("cairo") returns TRUE on exactly the machines where this happens (macOS without XQuartz). Every export now checks that the file exists, is non-empty, and begins with the magic bytes of the format requested, and falls back from cairo_pdf to the base pdf device with a visible note about font embedding. A format that cannot be written at all is reported as missing rather than shipped as a zero-byte file.

Legend Key

Every figure comes with a legend key: a small image listing each plotted sample or group with its exact colour, marker and outline, next to its name, in the figure's font. It sits in the Legend key box directly under the PNG / PDF / SVG buttons:

• Key PNG / Key PDF / Key SVG download the key in the same formats as the figure. In PDF and SVG the names stay editable text.

• Legend text starter is a read-only box with a sentence ready to paste into the figure legend, and a Copy button. Its format is NAME (colour, marker); NAME (colour, marker); ..., followed by the outline colour, e.g.

  MG1655 (black, filled circles); MG1655 + WT (orange, filled circles). Circles are outlined in black.
  

  Colour words are the nearest plain name (by CIE Lab distance); two groups that would get the same word are told apart with "light" / "dark". Bar figures say bars, or bars, replicates as filled circles when replicate points are shown.

• Place key on figure draws the same key on the figure itself, at the Legend position you chose (the right-hand margin if that is none). It also gives the one-sample figure and per-bar colouring a key, which they otherwise do not show.

The key is the figure's own legend cut out of the figure, and the text is read from the same colour mapping the figure is drawn from, so both always follow the current sample names, colours and selection, and the key on the figure and the downloaded key are the same. It covers every plot mode, the PowerPoint and reveal slides and the GIF (through Place key on figure), Download everything (legend_key.* and legend_text.txt) and recreate_figure.R, which rebuilds the key and the text alongside the figure.

One Style For The Whole Paper

A paper's figures come from several CSVs and several sessions; they should still look like one set.

• Colours belong to names. Under Apply Okabe-Ito colorblind palette there is one colour picker per sample and one per timepoint in the loaded data, labelled with its name. The colour is stored against the NAME, not the position: load another CSV and every name the app has seen keeps its colour, wherever it falls in the level order, and a new name takes the first default colour no other name is using. Samples default to Okabe-Ito in order (blue, vermillion, bluish green, ...); timepoints to the two blues the app has always used, then Okabe-Ito. Presets, the manifest and recreate_figure.R carry the colours as sample_palette / time_palette (name -> hex). Loading a preset merges them by name, keeping names the current CSV lacks, so one preset can name every strain of the paper. A preset saved before 0.4.0, which numbered its colours (sample_color_1, sample_color_2, time_color_1, time_color_2), gives them to the first and second name loaded, and says so.
• Font. Font (above Base font size) sets the family of every text on the figure: titles, axes, legend, caption, statistic labels, the n row, the legend key and every export. Custom… takes any installed family by name. A family this machine does not have is drawn in the default sans; the app says so once, and Figure QA carries a Font row with the font requested and the font used. The manifest records both.
• Paper style. In Reproducibility, Save as paper style writes today's settings and colours to paper_style.json beside app.R (or wherever the CFU_PAPER_STYLE environment variable points), and the app applies that file every time it starts; the line under the buttons names the file and when it was saved. A style is a style, so it leaves the settings that belong to one figure alone: the plot title, plot mode, y minimum, y maximum, survival baseline, survival readout and the statistics shown. Forget paper style deletes the file after asking; nothing else changes. The file has the shape of a plot preset, so it can also be loaded through Load plot preset JSON, which, as for any preset, applies everything.

Keeping The Data Area Fixed

ggplot sizes the panel last: it gets whatever space is left after the legend, the titles, the axis labels and the caption have taken theirs. Two figures exported at the same width therefore have different-sized data areas the moment their labelling differs. Measured on a 7 x 4.4 in figure from this app, moving the legend from the top to the right side shrinks the panel from 6.39 in wide to 3.83 in — a 40% loss of data area with the data untouched. Panels meant to be compared side by side cannot be built that way.

Set Width and height describe to The plot panel only and the priority inverts: the data area is pinned to exactly the size you asked for and the figure grows around it. The readout then shows both, for example

Panel pinned at 4.20 x 2.60 in (107 x 66 mm)  ->  figure 6.18 x 4.27 in ...

Two survival figures from the same data, one plain and one with a side legend, a long title and angled axis labels, both come out with a panel of exactly 4.200 x 2.600 in while their figures differ (4.80 x 4.68 in versus 6.18 x 4.27 in). That is the point: the comparable thing stays constant and the rest gives way.

Notes:

• With facets, each panel gets the size, so one facet of a multi-panel figure is directly comparable with a single-panel figure at the same setting.
• The journal size presets are figure widths. In panel mode they set the panel that produces that figure, rather than silently reinterpreting an 89 mm figure as an 89 mm panel.
• PNG, PDF, SVG, the preview and PowerPoint all honour the pinned panel. Animated GIF does not — gganimate builds its frames from the ggplot rather than the laid-out table — and the app says so when you export one.
• The default is still The whole figure, and that path is unchanged.

Bars Or Points

A bar encodes its value as a length measured from zero. On a log10(CFU) axis that baseline is 1 CFU/mL, which is arbitrary, and viable counts spanning 10^8 to 10^11 leave most of the panel empty. Setting Mean shown as to Points (no bars) draws the group mean as a marker instead, which allows the axis to be framed on the data. Auto y-axis respects the distinction: it keeps the zero baseline for bars and fits the range to the data for points.

Colouring Bars Individually

Bar colouring switches between one colour per group (the default) and one colour per bar. In per-bar mode you get a colour picker for every drawn bar, plus a button that lays Okabe-Ito across them and one that returns to the group colours.

A caveat the app enforces rather than hides: when bars are dodged by sample or timepoint, colouring them individually means colour no longer identifies the group. The fill legend is hidden in that case (it would just list every bar) unless Place key on figure is on, and Figure QA raises it. Per-bar colour is unambiguous in the one-sample-one-timepoint mode, and elsewhere it is best used to highlight rather than to classify.

Moving Things On The Canvas

The Canvas bar above the figure has three modes:

• Off — clicks do nothing.
• Place legend — click anywhere on the figure and the legend centre moves there.
• Move stat labels — click a significance label to pick it up, click again to drop it.

Reset placement returns the legend and every label to its automatic position. Nudges are stored per label in axis units, so they survive re-rendering, filtering, resizing, preset save/load, and the reproducible-script export.

Title, subtitle, methods caption and both axis titles have alignment sliders under Text placement (0 left, 0.5 centred, 1 right).

Paired Survival

An induction time-course is not really a question about absolute counts, it is a question about survival: how much of each culture is left at the readout timepoint relative to where that same culture started. Plot mode > Paired survival computes, for every (sample, treatment, replicate) that has BOTH timepoints:

log10 survival ratio = log10(CFU at readout) - log10(CFU at baseline)

then summarises those per-replicate ratios. Pick the baseline and readout timepoints in the sidebar; a dashed reference line marks no change.

Why pair. When every replicate has both timepoints the paired point estimate is algebraically identical to the unpaired one — mean(a - b) is mean(a) - mean(b). What pairing buys is the uncertainty: the standard error becomes the spread of within-culture differences rather than the pooled spread across cultures. On replicates whose starting titres differ by orders of magnitude that is the difference between p = 0.002 and p = 0.5 on the same estimate.

And when a replicate is missing one timepoint, the two stop agreeing entirely. The unpaired calculation subtracts a baseline mean containing a replicate the readout mean cannot contain, charging that replicate's whole titre to the treatment. On the bundled gp75 dummy file that flips the sign of the biology at one dose: paired says 48% survival, unpaired says 119% growth.

Axis. The three display scales — log10 ratio, fold change, percent of baseline — are the same numbers with different tick labels. Switching never rescales or distorts anything. Percent is offered but is intrinsically asymmetric (a 10x drop is 10%, a 10x rise is 1000%), so log10 is the default. The raw-CFU log axis is not available in this mode: the plotted quantity is already a log ratio.

Statistics. The default test asks whether survival differs from no change, as a one-sample t-test of the per-replicate log ratios against zero — which is the paired test. The reported effect size is d_z (mean of differences over their SD), reported under its own column name because it is a different quantity from the two-sample Hedges' g and the two must not be compared. Between-construct and dose-versus-control comparisons run on the same ratios. The timepoint comparison is refused, because both timepoints have already been consumed to form the ratio.

Pairing between constructs and between doses. The within-culture pairing above is guaranteed by the design: the two timepoints came from one flask. Whether replicate 1 of one construct and replicate 1 of another are also the same experiment — one split culture, one day — is a claim the CSV cannot settle, so Replicate labels match across samples and treatments is off by default.

Tick it and the between-construct and dose-versus-control comparisons become paired t-tests on the matched differences. When the claim is true this is markedly more sensitive: on three days whose baselines span four logs, a consistent half-log construct effect gives p = 0.002 paired and p = 0.37 unpaired, from the same point estimate. When the claim is false, ticking it invents a pairing and the p values are wrong.

Pairing is not free. A replicate with no counterpart on the other side is excluded, and the Statistics tab reports n_matched per row plus a note naming how many were dropped. On the bundled gp75 dummy file pairing actually reduces the number of testable doses from three to two, because different replicate numbers survived on each side — which is exactly the sort of thing worth seeing before trusting the result. The rank test has no paired form here, so selecting both leaves the comparison unpaired rather than silently substituting a signed-rank test.

What pairing costs, stated. Wells whose partner did not survive the CFU > 0 filter contribute nothing, and a cell can end up with no pairs at all and vanish from the figure. The banner above the plot and the Figure QA "Pairing completeness" check report the complete pairs, the orphans, any unlabelled wells, and name any cell that disappeared. The plotted n counts pairs, not wells.

Publication Figure Controls

The app includes controls for:

• Exact export width, height, and DPI, in inches or millimetres.
• Reproducible size presets for single-column, double-column, square, and Nature-width figures.
• A custom Y-axis quantity (CFU/mL, CFU/plate, CFU/OD600), wrapped in log10 automatically.
• A replicate-n row under the axis when n differs between groups, or a single n statement in the methods caption when it does not.
• A jitter seed, so replicate points land in the same place on every re-export.
• Log10 CFU, or raw CFU on a log axis. Both summarise on the log10 scale — the scale every test uses — so the raw-axis bar is the geometric mean with a multiplicative error interval, and the caption says so.
• Manual y-axis minimum and maximum.
• Major and minor y-axis tick spacing.
• Major and minor y-axis guide lines.
• Plot box, axis line width, tick length, and minor tick length.
• Bar width, dodge width, outline width, error-bar width, point size, point alpha, and jitter.
• Capped error bars, uncapped whiskers, mean point plus whiskers, mean crossbar intervals, or replicate-points-only variation display.
• Font sizes for base text, title, subtitle, and statistic labels.
• Legend position, including inside-plot positioning.
• Custom colors for samples, timepoints, bars, outlines, axes, grids, and statistic labels.
• Treatment units, time units, and optional unit suffixes.

Data That Cannot Be Plotted, And The Limit Of Detection

A non-numeric or blank CFU value cannot be plotted or tested and is excluded from the figure, the summary, and every test. The app states how many rows were dropped and why in a banner above the figure, and the plotted n is always the surviving n — not the number of wells plated.

A count of zero is different: log10 is undefined there, but the reading is a real measurement — below the limit of detection — and it happens exactly where a treatment worked. Excluding those wells biases a log-reduction towards the null. On a control at 1e8 against a treated arm of 1e2, 1e2 and 0, dropping the zero reports a 6.0-log reduction where censoring it at 1 CFU/mL gives 6.67. Background-subtracted counts can legitimately land below zero as well, and are treated the same way.

Counts at or below zero in the sidebar offers the two defensible treatments:

• Exclude the well — the default, and what every version before 0.2.0 did.
• Substitute at the limit of detection — set your LOD (default 1 CFU/mL, the smallest countable titre on a neat plate) and substitute at the LOD or at LOD/2.

Whichever you choose, the rule, the count and every affected well are written into the figure caption, the Figure QA tab, analysis_manifest.json, and lod_wells.csv inside the export bundle. The choice does not survive only as a banner on screen.

Statistics

The default statistics use Welch t-tests on log10(CFU) values. The app also supports Student t-tests, Wilcoxon rank-sum tests, and model-based marginal means through emmeans, with optional Tukey HSD or Dunnett-versus-control post-hoc procedures.

Every test runs on one value per biological replicate. Technical repeats are collapsed first, on the log10 scale, and each result row names its unit_of_replication and the n_a / n_b it actually compared.

Each comparison reports the log10 difference with its confidence interval, the same interval back-transformed to a fold-change range, and Hedges' g (Cohen's d with the small-sample correction, which matters at n = 3).

The rank test is offered because log10(CFU) normality is an assumption, not a fact — but note that with three versus three replicates the smallest attainable two-sided p is 0.1, so no comparison can reach 0.05. Rows produced under that condition carry a note saying so. It uses the exact null distribution unless ties force the normal approximation, which is also reported per row.

Every pair is tested. With three or more samples or timepoints the figure annotates only the first pair — stacked stars at one x position are unreadable — but all pairs are computed, corrected together, named in the caption, and exported in the Statistics tab.

The ANOVA uses Type II sums of squares via car::Anova when car is installed, because the CFU > 0 filter usually leaves the design unbalanced and Type I sequential SS would make each main effect depend on the order the terms appear in the formula. The table names the type it used.

How the ANOVA relates to the pairwise table

It depends on which test you picked, so the ANOVA tab states it in words rather than leaving the layout to imply it:

• Welch or Student t-tests — the pairwise table is an independent family, not a post-hoc of the ANOVA. Its correction is applied across that family alone, and a non-significant omnibus does not gate it. This is a legitimate standalone procedure, but it is not protected post-hoc testing and should not be described as such.
• emmeans with Tukey or Dunnett — the pairwise table is a protected post-hoc family on the model the ANOVA reports. Those procedures control the family-wise error themselves, so the app's own correction setting is disabled and the table records p_adjust_method = none rather than correcting twice.
• Wilcoxon — the omnibus becomes a Kruskal-Wallis rank-sum test, so the two tables describe the same kind of procedure. (Dunn's post-hoc is not implemented; the pairwise family remains Wilcoxon tests with the chosen correction.)

Assumption diagnostics

The QC tab reports per-cell n, mean and SD on log10(CFU), a Shapiro-Wilk p where one is defined, and each cell's variance as a ratio to the narrowest cell — all computed on the same aggregated units the tests use. A note appears when you have selected Student, or a pooled-variance emmeans model, and the widest cell has more than 4x the variance of the narrowest.

These are advisory. Nothing switches a test and nothing is gated on a p-value: at n = 3 Shapiro-Wilk has almost no power, so a passing p is not evidence of normality.

Which mean the summary table reports

summary.csv and the Summary tab report both centres, named so they cannot be confused:

• geometric_mean_cfu with geometric_sd_factor — the quantity the figure draws and the tests analyse. The interval is multiplicative: geometric mean ×/÷ the factor, given as geometric_mean_low and geometric_mean_high.
• arithmetic_mean_cfu, arithmetic_sd_cfu, arithmetic_sem_cfu — kept because some readers want them, but they are not what was tested. On three wells at 1e6, 1e7 and 1e8 the arithmetic mean is 3.7e7 against a geometric mean of 1e7.

Before 0.2.0 these were mean_cfu / sd_cfu / sem_cfu sitting beside geometric_mean_cfu, with nothing but the word to tell them apart and no dispersion at all on the geometric side.

Testing

Rscript tests/test_bundle.R                # optionally: ... path/to/your.csv
Rscript tests/test_panel_size.R
Rscript tests/test_survival.R
Rscript tests/test_statistics.R
Rscript tests/test_column_matching.R
Rscript tests/test_end_to_end.R            # optionally: ... path/to/your.csv
Rscript tests/test_legend_key.R
Rscript tests/test_paper_style.R

test_end_to_end.R drives the real Shiny server with a real uploaded CSV, so it covers the reactives rather than the helper functions. Point it at your own file to check the whole pipeline against your data.

test_statistics.R re-derives each expected value from first principles or a hand-worked example rather than from the app's own helpers, so a bug that is self-consistent across the app still fails the test. Both scripts exit non-zero on failure.

test_legend_key.R checks the legend key from the files the app writes: every group in the figure's colour mapping appears in the key with the same colour, marker and name; renaming, recolouring and deselecting a sample are followed; the text starter names every group with a colour word; the key placed on the figure matches the downloaded one; names stay text in SVG and PDF; and the bundled recreate_figure.R rebuilds the key in a fresh R session.

test_paper_style.R renders the default figure in every plot mode with the 0.3.0 code (git show 0762ca1:app.R, the 0.3.0 commit) and with the current code, in separate R processes, and checks the SVG files are byte-identical. It then checks that colours follow sample names across CSVs and through presets (old numbered presets included), that the font reaches every text node, the key and a pinned panel, that the paper style is saved, auto-loaded without its figure-specific settings and forgotten, and that the bundled recreate_figure.R reproduces the palette and the font in a fresh R session.

Multiple-comparison correction options include:

• BH
• Holm
• Bonferroni
• None

The scope of that correction — across every comparison in the figure, or within each panel or group — is honoured by all of the t-test, rank-test and emmeans paths, and the scope actually used is recorded in the adjustment_scope column of statistics.csv. (Before 0.2.0 the emmeans path silently corrected globally whatever the control said.) Tukey and Dunnett are the exception, and say so: they control the family-wise error themselves, so the app's correction is disabled rather than applied on top.

Statistic labels can be shown as significance stars or exact adjusted values.

Notes For Manuscripts

For manuscript figures, a good starting workflow is:

1. Upload or map the data.
2. Check the QC tab for replicate issues.
3. Choose the plot mode and statistics comparison.
4. Set y-axis boundaries and tick spacing.
5. Use a reproducible figure size preset.
6. Export SVG/PDF for vector editing or editable PowerPoint if rvg is installed.
7. Export the statistics table alongside the figure for record keeping.
8. Save the plot preset, analysis manifest, and reproducible R script with the project folder.

Files

• app.R: main Shiny application.
• Run CFU Plot Studio.bat: double-click launcher for Windows.
• run_app.R: cross-platform launcher (dependency check, free-port selection, browser).
• tests/test_column_matching.R: regression tests for CSV header auto-detection and CSV dialects.
• dummy_cfu_example.csv: synthetic example/template data.
• CHANGELOG.md: release history, including which versions produce comparable statistics.
• AUDIT-2026-09.md: the September 2026 statistical audit and what was done about each finding.
• outputs/: local validation outputs and screenshots.

License

This project is released under the MIT License.

Citation And DOI

Citation metadata is provided in CITATION.cff. Zenodo metadata is provided in .zenodo.json.

After the repository is public on GitHub, connect it to Zenodo and create a GitHub release. Zenodo will archive that release and mint a DOI for it.

The initial release metadata lists Michael Baffour Awuah as maintainer. Update the affiliation field in CITATION.cff and .zenodo.json later if a formal institutional affiliation should be displayed.