mismap-qc

Missing-data validation for proteomics and RNA-Seq experiments. Detects outlier samples, classifies dropout mechanism (MNAR vs MAR), tests for batch effects, and gates pipelines on configurable QC rules. Every check has a matching plot for when you want to see the problem rather than just check it.

Documentation: foertsch.github.io/mismap-qc

Install

pip install mismap-qc

Optional extras: [interactive] for HTML matrices (plotly), [anndata] for AnnData input.

Quick start

import pandas as pd
from mismap_qc import qc, assert_qc

df = pd.read_csv("proteomics.tsv", sep="\t", index_col=0)

# 1. Inspect: full QC report in one call
report = qc(df, group_level="condition")
print(report)
# MismapReport(n=8412x96, 3 outliers, 412 MNAR features, passed=True)

# 2. Drill in on anything flagged (the report holds pandas DataFrames)
report.sample_outliers.query("flagged")
report.feature_mechanism.query("mechanism == 'MNAR'")

# 3. Gate a pipeline (raises MismapQCFailure on rule violation)
assert_qc(df, thresholds={
    "min_sample_completeness": 0.60,
    "max_mnar_fraction": 0.30,
    "max_sample_outliers": 3,
})

For a human-readable multi-section summary instead of the one-line repr:

print(report.summary())

Starting from an AnnData object:

from mismap_qc import from_anndata, qc
df = from_anndata(adata, obs_levels=["batch", "condition"])
report = qc(df, group_level="condition")

Six of the nine plot functions pair with their underlying numbers through return_data=True. missing_mechanism() always returns (Figure, DataFrame); missing_abundance_density() and missing_matrix_html() return a figure and an HTML string respectively.

from mismap_qc import detection_waterfall
fig, table = detection_waterfall(df, return_data=True)
# table: feature, detection_rate, rank

What this validates

Check	What it catches
Sample completeness	Samples with too few detected features
Outlier detection	Samples with anomalous missingness vs group peers
Missingness mechanism	Dropouts driven by low abundance (MNAR) vs random (MAR)
Batch effects	Features whose detection differs between conditions
Run order drift	Instrument degradation over a long acquisition

Why use this

mismap-qc fills a specific gap in the Python omics ecosystem:

• missingno does general missing-data visualization but has no omics awareness (groups, MultiIndex sample annotations, MNAR mechanism).
• protti (R) classifies missingness mechanism but has no Python equivalent.
• great-expectations validates tabular data but does not understand missingness mechanism or omics-specific patterns.

mismap-qc handles all three with a single API and reads AnnData natively.

Examples

• CPTAC Lung Adenocarcinoma proteomics -- real-world tutorial using public CPTAC LUAD data (~100 tumour/normal samples). Shows how missingness clusters by tumour/normal status.

Running the demo

No virtual environment needed -- the demo uses PEP 723 inline script dependencies with uv:

uv run demo.py

Feature types

The feature_type parameter controls labels in axes and tooltips:

Value	Labels
"PROT"	Protein / Proteins (default)
"GENE"	Gene / Genes
"PEPTIDE"	Peptide / Peptides

Input format

A pandas DataFrame with:

• Rows = features (proteins, genes, peptides)
• Columns = samples, optionally as a MultiIndex for annotation strips
• NaN = missing / not detected

When columns are a MultiIndex, level names automatically become annotation strip labels.

missing_matrix() -- static plot

fig = missing_matrix(
    df,
    title="Gene Detection Matrix",
    subtitle="80 genes x 30 samples | 23% missing",
    save="output.png",
)

Layout (top to bottom)

Component	Description
Title + subtitle	Bold title, italic subtitle for metadata
Dendrogram	Hierarchical clustering of samples by nullity pattern
Annotation strips	One colour bar per MultiIndex column level
Nullity matrix	Dark = detected, light = missing
Completeness sparkline	Per-sample or per-feature detection rate

Parameters

Data & labels

Parameter	Type	Default	Description
df	DataFrame	required	Features (rows) x samples (columns). NaN = missing.
title	str	""	Bold figure title
subtitle	str	""	Italic line below title (e.g. dataset metadata)
feature_type	str	"PROT"	Feature type: "PROT", "GENE", or "PEPTIDE"
label_level	int	-1	Which column level to use for x-axis tick labels

Clustering & sorting

Parameter	Type	Default	Description
cluster_samples	bool	True	Cluster samples by binary nullity pattern
cluster_method	str	"average"	scipy linkage method
show_dendrogram	bool	True	Show dendrogram above the matrix
sort_features	str | None	"descending"	Sort features by completeness ("ascending", "descending", or None)

Annotations

Parameter	Type	Default	Description
annotation_levels	list[int] | None	None	Column levels to show as colour bars (default: all except innermost)
annotation_colors	dict | None	None	Custom colours per level (see below)

Custom annotation colours accept level indices or names as keys:

missing_matrix(
    df,
    annotation_colors={
        "Medium_Type": {"Fresh": "#88CCEE", "Conditioned": "#CC6677"},
        "Medium_Condition": {"SF": "#44AA99", "FBS": "#DDCC77", "AS": "#AA4499"},
    },
)

Unspecified factor levels fall back to built-in palettes.

Completeness sparkline

Parameter	Type	Default	Description
completeness	str	"below"	"below" = per-sample (horizontal), "side" = per-feature (vertical)
completeness_threshold	float | None	None	Draws a dashed red line at this value (0--1)

Legends & layout

Parameter	Type	Default	Description
legend_loc	str	"upper right"	Corner for legends: "upper right", "upper left", "lower right", "lower left"
figsize	tuple | None	None	Figure size (auto-calculated if None)
color_present	str	"#2d2d2d"	Colour for detected cells
color_missing	str	"#f0f0f0"	Colour for missing cells

Font sizes

Parameter	Type	Default	Description
fontsize	int	10	Base font size (fallback)
fontsize_legend	int | None	None	Legend entries
fontsize_rows	int | None	None	Gene/row labels
fontsize_cols	int | None	None	Sample/column labels
fontsize_annotations	int | None	None	Annotation strip labels

Group summary

Parameter	Type	Default	Description
group_summary	int | str | None	None	Column level to group by; prints per-group completeness to console

fig = missing_matrix(df, group_summary="Medium_Condition")

Output:

Group Completeness (Medium_Condition)
--------------------------------
  SF               63%  (n=10)
  AS               80%  (n=10)
  FBS              88%  (n=10)

Only prints when the level has more than one group.

Split by factor

Parameter	Type	Default	Description
split_by	int | str | None	None	Split into side-by-side panels by this column level

fig = missing_matrix(df, split_by="Medium_Condition", annotation_levels=[0])

Each panel is independently clustered. The split level is automatically removed from annotation strips.

Output

Parameter	Type	Default	Description
save	str | None	None	Save figure to this path
dpi	int	150	Save resolution

missing_matrix_html() -- interactive HTML

Plotly-based interactive version with hover tooltips showing feature name, sample ID, all annotation levels, and detection status.

from mismap_qc import missing_matrix_html

missing_matrix_html(
    df,
    title="Gene Detection Matrix (Interactive)",
    subtitle="80 genes x 30 samples",
    completeness_threshold=0.5,
    save="output/interactive.html",
)

Supports the same clustering, sorting, annotation, and completeness options as the static version. Additional parameters:

Parameter	Type	Default	Description
width	int | None	None	Plot width in pixels (auto-calculated if None)
height	int | None	None	Plot height in pixels (auto-calculated if None)

Requires plotly (pip install plotly or included via PEP 723 in demo.py).

Generating toy data

uv run make_toy_data.py

Creates data/toy_rnaseq.csv: 80 genes x 30 samples with structured missingness patterns across 6 groups (Fresh/Conditioned x SF/FBS/AS).

Dependencies

• numpy
• matplotlib
• scipy
• pandas
• plotly (optional, for HTML export only)

missing_upset() -- co-missingness intersections

Which combinations of samples share missing features. For each intersection, how many features are missing in exactly that combination and no others. At small n this is what separates technical dropout from biology: if two replicates always lose the same proteins together, that is not biology.

Requires upsetplot (pip install mismap-qc[upset]).

from mismap_qc import missing_upset

fig, table = missing_upset(df, max_intersections=12, return_data=True)

The example above recovers two real patterns: Fresh3 alone accounts for 60 features (one bad sample), and Cond2|Cond3 shares 34 (a co-dropping pair).

Parameter	Type	Default	Description
by	str | int	"sample"	"sample" for one set per sample, or a column level name/index for one set per group
group_min_frac	float	0.5	Group mode only. A feature counts as missing in a group once it is missing in at least this fraction of the group's samples
min_size	int	1	Intersections smaller than this are not drawn
max_intersections	int	50	Draw at most this many intersections, largest first
feature_type	str	"PROT"	"PROT", "GENE", or "PEPTIDE"
return_data	bool	False	Return (Figure, DataFrame)

Intersection count grows quickly with sample count, so the plot caps at the 50 largest by default. Truncation is stated on the figure ("showing the 12 largest of 33 intersections") and return_data=True still returns every intersection, with a plotted column marking what made the cut. Nothing is silently dropped.

Every feature with at least one missing value belongs to exactly one intersection, so the returned table has one row per feature:

Column	Description
feature	Feature ID
members	The samples or groups it is missing in, pipe-joined ("Cond2|Cond3")
n_features	Size of that intersection
rank	Intersection rank by size, 1 = largest
plotted	Whether it survived min_size and max_intersections

Which makes the follow-up question a one-liner:

table.query("members == 'Cond2|Cond3'").feature   # proteins lost in exactly those two

Fully detected features carry no intersection information and are excluded.

Contributing

Bug reports and pull requests are welcome. See CONTRIBUTING.md for the development install, how to run the tests and linter, and the conventions for adding a check or a plot. Participation is covered by the Code of Conduct.

Development install:

git clone https://github.com/foertsch/mismap-qc.git
cd mismap-qc
uv sync --extra dev      # or: pip install -e ".[dev]"

Citation

If you use mismap-qc in published work, please cite it. Machine-readable metadata is in CITATION.cff; GitHub renders it as "Cite this repository" in the sidebar.

Förtsch, A. (2026). mismap-qc: missing-data validation for proteomics and RNA-Seq (version 0.4.1). https://github.com/foertsch/mismap-qc

Use of generative AI

Most of this package's code, tests, and prose was written by an AI coding agent working from my specifications and under my review. The design decisions and the responsibility are mine. The full disclosure, including what was and was not generated, is at Use of generative AI.

License

MIT. See LICENSE.