bartab 0.0.2

Analysis of barcoded, pooled functional genomics assays.

🍹 bartab

bartab estimates relative fitness from sequencing-based pooled competition experiments.

It is designed for experiments where barcoded strains, guides, mutants, or constructs are grown together, sampled over time, and quantified by NGS read counts or UMI counts. bartab estimates the fitness of each barcode relative to a reference barcode, usually WT, using either spike-in normalisation or measured culture growth.

bartab provides:

• a command-line interface for fitting, plotting, and simulating pooled fitness experiments
• a Python API built around AnnData
• weighted least-squares fitness estimation
• dose-response fitting with a Hill/log-logistic model
• diagnostic plots
• synthetic data simulation

---

Contents

• Installation
• Conceptual model
• Input tables
• Command-line interface
  • bartab fit
  • bartab plot
  • bartab sim
• Python API
  • Load data
  • Compute log-ratios
  • Fit fitness models
  • Fit dose-response models
  • Plot results
  • Simulate data
• Output structure
• Interpretation
• Limitations
• Issues, problems, suggestions
• Documentation

---

Installation

From PyPI

pip install bartab

From source

git clone https://github.com/scbirlab/bartab.git
cd bartab
pip install -e .

For development, install test and documentation dependencies as appropriate for your local setup.

---

Conceptual model

Interactive tutorial on analysis principles.

---

Input tables

bartab expects three tables:

1. count table
2. sample sheet
3. barcode sheet

Files may be CSV, TSV, TXT, or other formats supported by carabiner.pd.read_table.

---

Count table

One row per barcode per sample.

Required columns:

Meaning	Default CLI option	Typical name
barcode / strain identifier	--barcode-column	strain_id
sample identifier	--sample-column	sample_id
count value	--count-column	count

Example:

strain_id	sample_id	count
wt	sample_0	12034
mutant_A	sample_0	8312
spike	sample_0	5021
wt	sample_1	18420
mutant_A	sample_1	6420
spike	sample_1	2100

Counts must be non-negative.

---

Sample sheet

One row per sequencing sample.

Required columns:

Meaning	Default CLI option	Typical name
sample identifier	--sample-column	sample_id
culture / biological replicate	--culture-column	culture_id
timepoint	--timepoint-column	timepoint

Optional columns:

Meaning	CLI option	Used for
concentration	--concentration-column	dose-response analysis
growth measurement	--growth-column	growth-based normalisation
sampled volume	--volume-column	adaptive-volume correction with spike-in

Example:

sample_id	culture_id	timepoint	dose	growth	volume
sample_0	rep1	0	0	0.05	1.0
sample_1	rep1	1	0	0.20	1.0
sample_2	rep1	2	0	0.80	1.0

If --t0 is not supplied, the minimum value in the timepoint column is used as the initial timepoint.

---

Barcode sheet

One row per barcode.

Required columns:

Meaning	Default CLI option	Typical name
barcode / strain identifier	--barcode-column	strain_id

Optional metadata columns are preserved in the output AnnData.obs.

Example:

strain_id	gene	annotation
wt	WT	reference
mutant_A	geneA	deletion mutant
spike	spike	non-growing spike-in

---

Command-line interface

Run:

bartab --help

bartab has three main commands:

bartab fit   # estimate fitness
bartab plot  # plot fitted results
bartab sim   # simulate pooled competition data

---

bartab fit

Estimate relative fitness per strain or barcode.

Minimal spike-in example

bartab fit counts.csv \
  --sample-sheet sample_meta.csv \
  --barcode-sheet strain_meta.csv \
  --output results.h5ad \
  --reference wt \
  --spike-name spike \
  --use-spike \
  --barcode-column strain_id \
  --sample-column sample_id \
  --culture-column culture_id \
  --count-column count \
  --timepoint-column timepoint

This writes:

results.h5ad
results.csv

when the output path ends in .h5ad.

---

Minimal growth-based example

bartab fit counts.csv \
  --sample-sheet sample_meta.csv \
  --barcode-sheet strain_meta.csv \
  --output results.h5ad \
  --reference wt \
  --barcode-column strain_id \
  --sample-column sample_id \
  --culture-column culture_id \
  --count-column count \
  --timepoint-column timepoint \
  --growth-column growth \
  --growth-type density

---

Dose-response example

bartab fit counts.csv \
  --sample-sheet sample_meta.csv \
  --barcode-sheet strain_meta.csv \
  --output dose_response.h5ad \
  --reference wt \
  --spike-name spike \
  --use-spike \
  --barcode-column strain_id \
  --sample-column sample_id \
  --culture-column culture_id \
  --count-column count \
  --timepoint-column timepoint \
  --concentration-column dose

If --concentration-column is supplied, bartab first fits per-concentration WLS fitness estimates and then fits a Hill/log-logistic dose-response model.

---

Output format

If --output ends in:

Output suffix	Behaviour
.h5ad	writes full annotated AnnData object and a companion .csv table
.h5ad.gz	writes compressed AnnData and companion .csv table
.csv	writes fitted parameter table only
.tsv / .txt	writes fitted parameter table only, tab-separated

---

Main fit options

Option	Default	Description
input	required	count table
--output, -o	required	output file
--sample-sheet, -c	required	sample metadata table
--barcode-sheet, -b	required	barcode metadata table
--reference, -r	required	reference barcode/strain
--barcode-column, -u	strain_name	barcode column in count and barcode tables
--sample-column, -s	sample_id	sample ID column
--culture-column	culture_id	biological replicate/culture column
--count-column, -a	count	count column
--timepoint-column, -t	timepoint	timepoint column
--t0, -0	minimum timepoint	initial timepoint value
--spike-name, -x	None	spike-in barcode name
--use-spike	False	use spike-in to estimate culture expansion
--growth-column, -d	OD600	growth measurement column
--growth-type	density	either density or generations
--volume-column	None	sampled volume column
--concentration-column, -k	None	concentration column for dose-response
--pseudocount	1.0	value added to counts before log transform
--model-type	WLS	currently accepted: WLS, OLS, HillFitnessModel

---

bartab plot

Generate diagnostic plots from a fitted .h5ad file.

bartab plot results.h5ad \
  --output results_plots \
  --model-type WLS \
  --plot-format png

This writes plots using the supplied prefix:

results_plots_time-count.png
results_plots_time-ratio.png
results_plots_expansion-ratio.png
results_plots_pred-obs.png
results_plots_volcano.png

For dose-response/Hill results:

bartab plot dose_response.h5ad \
  --output dose_response_plots \
  --model-type HillFitnessModel \
  --plot-format png

This additionally writes:

dose_response_plots_dr.png

---

Plot options

Option	Default	Description
input	required	fitted .h5ad file
--output, -o	required	filename prefix
--highlight, -X	None	barcode names to highlight
--model-type	WLS	model to plot; WLS, OLS, or HillFitnessModel
--plot-format	png	png, PNG, pdf, or PDF

---

bartab sim

Simulate count data from a pooled competition experiment.

The simulator takes a JSON file describing true strain fitness values and writes three CSV files:

<output>_count.csv
<output>_sample_meta.csv
<output>_strain_meta.csv

---

Single-concentration simulation

Input JSON:

{
  "wt": 1.0,
  "spike": 0.0,
  "mutant_fast": 1.25,
  "mutant_slow": 0.5
}

Run:

bartab sim fitness.json \
  --output synthetic/single \
  --timepoints 8 \
  --n-cultures 3 \
  --reads-per-barcode 1000 \
  --seed 42

Then fit:

bartab fit synthetic/single_count.csv \
  --sample-sheet synthetic/single_sample_meta.csv \
  --barcode-sheet synthetic/single_strain_meta.csv \
  --output synthetic/single_results.h5ad \
  --reference wt \
  --spike-name spike \
  --use-spike \
  --barcode-column strain_id \
  --sample-column sample_id \
  --culture-column replicate \
  --count-column count \
  --timepoint-column timepoint

---

Dose-response simulation

For dose-response simulation, the JSON values are interpreted as [IC50, bottom].

Input JSON:

{
  "wt": [10000.0, 1.0],
  "spike": [10000.0, 0.0],
  "sensitive": [10.0, 0.0],
  "resistant": [1000.0, 0.0]
}

Run:

bartab sim dose_fitness.json \
  --output synthetic/dose \
  --n-dose 6 \
  --dose-max 1000 \
  --dose-fold 2 \
  --timepoints 8 \
  --n-cultures 3 \
  --reads-per-barcode 1000 \
  --seed 42

Then fit:

bartab fit synthetic/dose_count.csv \
  --sample-sheet synthetic/dose_sample_meta.csv \
  --barcode-sheet synthetic/dose_strain_meta.csv \
  --output synthetic/dose_results.h5ad \
  --reference wt \
  --spike-name spike \
  --use-spike \
  --barcode-column strain_id \
  --sample-column sample_id \
  --culture-column replicate \
  --count-column count \
  --timepoint-column timepoint \
  --concentration-column dose

---

Simulation options

Option	Default	Description
input	required	JSON file of strain fitness values
--output, -o	required	output prefix
--generate-controls, -z	0	number of neutral controls to generate
--generate-more, -m	0	number of additional random-fitness strains
--inoculum, -n	1000	cells per strain in inoculum
--carrying-capacity, -K	10.0	maximum fold-change supported
--timepoints, -s	10	number of sampled timepoints
--max-time, -t	10	final simulated timepoint
--n-cultures, -r	3	biological replicate cultures
--reads-per-barcode, -b	1000	mean sequencing depth per barcode per sample
--seed, -e	42	random seed
--n-dose, -d	None	number of dose-response concentrations
--dose-max	1000	maximum simulated dose
--dose-fold	2	dilution spacing parameter

---

Python API

The Python API is built around AnnData.

A typical workflow is:

1. load count, sample, and barcode tables into AnnData
2. compute barcode/reference log-ratios and expansion axis
3. fit a model
4. inspect adata.obs
5. optionally plot or write output

---

Load data

from bartab.io import load_anndata

adata = load_anndata(
    counts="counts.csv",
    sample_meta="sample_meta.csv",
    strain_meta="strain_meta.csv",
    reference="wt",
    spike="spike",
    count_column="count",
    strain_id="strain_id",
    sample_id="sample_id",
    culture_id="culture_id",
    timepoint_column="timepoint",
)

The same function also accepts pandas DataFrame objects:

adata = load_anndata(
    counts=counts_df,
    sample_meta=sample_df,
    strain_meta=strain_df,
    reference="wt",
    spike="spike",
    count_column="count",
    strain_id="strain_id",
    sample_id="sample_id",
    culture_id="culture_id",
    timepoint_column="timepoint",
)

For dose-response data:

adata = load_anndata(
    counts=counts_df,
    sample_meta=sample_df,
    strain_meta=strain_df,
    reference="wt",
    spike="spike",
    count_column="count",
    strain_id="strain_id",
    sample_id="sample_id",
    culture_id="culture_id",
    timepoint_column="timepoint",
    concentration_column="dose",
)

---

Compute log-ratios

Spike-in normalisation

from bartab.transforms import compute_log_ratios

adata = compute_log_ratios(
    adata,
    pseudocount=1.0,
    use_spike=True,
)

With adaptive-volume correction:

adata = compute_log_ratios(
    adata,
    pseudocount=1.0,
    use_spike=True,
    volume_column="volume",
)

Growth-based normalisation

adata = compute_log_ratios(
    adata,
    pseudocount=1.0,
    growth_column="growth",
    growth_type="density",
    use_spike=False,
)

If the growth column already contains generations:

adata = compute_log_ratios(
    adata,
    pseudocount=1.0,
    growth_column="generations",
    growth_type="generations",
    use_spike=False,
)

This adds:

adata.layers["__log_ratio__"]
adata.var["__log_expansion__"]

---

Fit fitness models

Weighted least squares

from bartab.models.anndata import AnnDataWLSModel

model = AnnDataWLSModel()
adata = model.fit(adata)

Results are written into adata.obs with names prefixed by the model name.

For WLS, common output columns include:

WLS:slope
WLS:slope_p
WLS:slope_se
WLS:slope_ci_low
WLS:slope_ci_high
WLS:fitness
WLS:fitness_low
WLS:fitness_high
WLS:nobs
WLS:rsq
WLS:fit_status

Predictions are written to:

adata.layers["WLS:predicted"]

The model name is also appended to:

adata.uns["models_fitted"]

---

Ordinary least squares

from bartab.models.anndata import AnnDataOLSModel

model = AnnDataOLSModel()
adata = model.fit(adata)

OLS is mainly useful as an unweighted baseline or diagnostic comparison.

---

Fit dose-response models

For dose-response analysis, first load data with concentration_column set, then compute log-ratios, fit WLS, and fit the Hill model.

from bartab.io import load_anndata
from bartab.transforms import compute_log_ratios
from bartab.models.anndata import AnnDataWLSModel, AnnDataHillModel

adata = load_anndata(
    counts="dose_count.csv",
    sample_meta="dose_sample_meta.csv",
    strain_meta="dose_strain_meta.csv",
    reference="wt",
    spike="spike",
    count_column="count",
    strain_id="strain_id",
    sample_id="sample_id",
    culture_id="replicate",
    timepoint_column="timepoint",
    concentration_column="dose",
)

adata = compute_log_ratios(
    adata,
    pseudocount=1.0,
    use_spike=True,
)

adata = AnnDataWLSModel().fit(adata)

adata = AnnDataHillModel().fit(
    adata,
    concentration="dose",
)

Hill model outputs include:

HillFitnessModel:log_ic50
HillFitnessModel:log_ic50_p
HillFitnessModel:log_ic50_se
HillFitnessModel:log_ic50_ci_low
HillFitnessModel:log_ic50_ci_high
HillFitnessModel:ic50
HillFitnessModel:ic50_low
HillFitnessModel:ic50_high
HillFitnessModel:h
HillFitnessModel:h_p
HillFitnessModel:nobs
HillFitnessModel:dof
HillFitnessModel:fit_status

The Hill model uses a log-logistic form fit on log concentration. Concentration-zero samples are omitted from the non-linear fit.

---

Plot results

Plotting functions live in bartab.plotting.

from bartab.plotting import (
    time_vs_count,
    time_vs_ratio,
    expansion_vs_count,
    expansion_vs_ratio,
    pred_vs_true,
    volcano,
    dose_response,
)

Examples:

fig, ax = time_vs_count(
    adata,
    highlight_barcodes=["mutant_A"],
    filename="plots/time_count.png",
)

fig, ax = expansion_vs_ratio(
    adata,
    highlight_barcodes=["mutant_A"],
    filename="plots/expansion_ratio.png",
)

fig, ax = pred_vs_true(
    adata,
    model_name="WLS",
    highlight_barcodes=["mutant_A"],
    filename="plots/pred_obs.png",
)

fig, ax = volcano(
    adata,
    model_name="WLS",
    param="fitness",
    p="slope_p",
    vline=1.0,
    highlight_barcodes=["mutant_A"],
    filename="plots/volcano.png",
)

For dose-response:

fig, ax = dose_response(
    adata,
    model_name="WLS",
    highlight_barcodes=["sensitive", "resistant"],
    filename="plots/dose_response.png",
)

Plot files are saved at high resolution. When a filename is provided, bartab also writes the plotted data alongside the figure through the internal figsaver utility.

---

Simulate data

The simulator can be used from Python.

from bartab.simulation import calculate_growth_curves, reads_sampler

fitness = {
    "wt": 1.0,
    "mutant_slow": 0.5,
    "mutant_fast": 1.25,
}

t, ref_expansion, growths = calculate_growth_curves(
    inoculum=1_000,
    fitness=fitness,
    inoculum_var=0.1,
    carrying_capacity=10,
    n_timepoints=8,
    max_time=10.0,
    ref_key="wt",
    seed=42,
)

counts = reads_sampler(
    growths,
    seq_depth=1_000,
    sample_frac=0.1,
    reps=3,
    variance=0.001,
    seed=42,
)

counts has shape:

n_strains × n_replicates × n_timepoints

---

Low-level model API

The lower-level model classes operate directly on arrays.

import numpy as np
from bartab.models.linear import OLSModel, WLSModel

Y = np.array([
    [0.0, -0.5, -1.0],
    [0.0,  0.0,  0.0],
])

x = np.array([0.0, 1.0, 2.0])

model = OLSModel()
results, preds = model.fit(Y, x)

results is a list of dictionaries, one per row of Y.

For most users, the AnnData model API is preferable.

---

Output structure

bartab stores results in an AnnData object.

Main matrix

adata.X

Raw count matrix, with:

rows    = barcodes / strains
columns = samples

---

Barcode metadata

adata.obs

Contains barcode metadata and model outputs.

Important internal columns include:

Column	Meaning
__is_reference__	whether the barcode is the reference
__is_spike__	whether the barcode is the spike-in
model-prefixed columns	fitted parameters and statistics

Example model columns:

WLS:fitness
WLS:fitness_low
WLS:fitness_high
WLS:slope_p
HillFitnessModel:ic50
HillFitnessModel:log_ic50_p

---

Sample metadata

adata.var

Contains sample metadata.

Important internal columns include:

Column	Meaning
__is_t0__	whether the sample is an initial timepoint
__culture_index__	culture/replicate identifier
__inducer__	concentration or "single dose"
__log_expansion__	fitted x-axis: culture expansion

---

Layers

adata.layers

Important layers include:

Layer	Meaning
__log_ratio__	observed barcode/reference log-ratio change
WLS:predicted	WLS fitted values
OLS:predicted	OLS fitted values
HillFitnessModel:predicted	Hill model fitted values

---

Unstructured metadata

adata.uns

Includes:

Key	Meaning
reference	reference barcode name
spike	spike-in barcode name
timepoint_column	timepoint column used
count_column	count column used
concentration_column	concentration column, if any
strain_id	barcode ID column(s)
sample_id	sample ID column(s)
culture_id	culture/replicate column(s)
models_fitted	list of fitted models

---

Interpretation

For single-concentration WLS output:

Output	Interpretation
WLS:fitness = 1	barcode grows like the reference
WLS:fitness < 1	growth disadvantage
WLS:fitness > 1	growth advantage
WLS:slope_p	evidence that the barcode deviates from neutral fitness
WLS:fitness_low, WLS:fitness_high	confidence interval for relative fitness

For dose-response output:

Output	Interpretation
HillFitnessModel:ic50	concentration giving 50% inhibition
lower IC50	greater sensitivity
higher IC50	greater resistance
HillFitnessModel:h	fitted Hill/logistic slope
HillFitnessModel:log_ic50_p	evidence for non-zero IC50 parameter estimate

Dose-response estimates are most reliable when the tested concentrations bracket the fitness transition.

---

Practical guidance

• Use the same barcode identifiers across the count table and barcode sheet.
• Use the same sample identifiers across the count table and sample sheet.
• If using spike-in normalisation, include exactly one spike-in barcode.
• If using growth-based normalisation, growth values must be positive.
• For dose-response analysis, concentrations must be numeric.
• Low-count barcodes can produce unstable estimates.
• Inspect diagnostic plots before interpreting individual hits.
• Check reference and spike-in behaviour before trusting global results.

---

Limitations

Results may be unreliable when:

• the reference barcode is depleted or poorly counted
• the spike-in grows, dies, degrades, or is sampled inconsistently
• bottlenecks dominate the experiment
• counts are extremely low
• barcode identities are mismatched between tables
• dose-response concentrations do not span the active range

The WLS weights are estimated using an approximate delta-method variance on log count ratios with a method-of-moments dispersion estimate.

---

Issues, problems, suggestions

Please add bug reports, feature requests, or suggestions to the issue tracker:

https://www.github.com/scbirlab/bartab/issues

---

Related resources

• Interactive Hugging Face Space: https://huggingface.co/spaces/scbirlab/bartab
• Tutorial Space: https://huggingface.co/spaces/scbirlab/tutorial-seq-fitness
• Source code: https://github.com/scbirlab/bartab