Overview

mantispy.ds ships example screens that span several assays and resolutions. This page places them against each other, so you can find the one closest to your own data, and then shows where the per-well cell counts come from.

Which dataset is like yours

dataset	assay and perturbation	unit of analysis	size	the example of
bbbc021	compound screen, 38 compounds and DMSO, MCF-7	well	632 wells, 103 treatments	the mechanism-of-action benchmark
rohban	ORF overexpression (gain of function), U2OS	well	1,918 wells, 5 plates	pathway hits by overexpression
pki	kinase inhibitors, few doses and many replicates, U2OS	well	3,072 wells, 8 plates	hit calling and normalization
oasis_pilot	36 compounds over a 10-point dose range, U2OS and HepaRG	well	12 plates, two cell lines	concentration-response curves
jump_target2	one JUMP compound plate map run at every site	well	up to 141 plates, 11 sources	cross-laboratory batch effects
jump_crispr	CRISPR knockout (loss of function), U2OS	well (guides as replicates)	51,185 wells	an arrayed genetic screen
jump_lite	one JUMP plate map under six feature sets	well (one field each)	4 plates, 1,536 wells	embeddings against CellProfiler
chroma	alternative-dye pilot, 8 channels, 91 compounds	well	8 acquisitions	a non-standard stain panel
neuropainting	iPSC neurons and astrocytes, no perturbation	well	1,691 wells, 20x and 63x	a design-factor screen (genotype, donor)
pooled_rare	pooled coding-variant screen, pre-aggregated	barcode	290 barcodes, 132 genes	a pre-aggregated allele series
scallops_arv471	optical pooled CRISPR under ARV-471, targeted readout	single cell	2M+ cells, 9 features	a single-cell pooled screen
cp_posh	optical pooled CRISPR knockout, broad morphology	single cell	124 genes, ~1,278 features	broad single-cell pooled morphology
jump_cells	single cells of one JUMP plate	single cell	13,578 cells, 24 wells	per-cell CellProfiler measurements
jump_plate	images and segmentations of one well	image (field of view)	2 fields, 8 channels	raw images as SpatialData
jump_export	one raw CellProfiler export	field of view	1 field	the input to read_profiles

Well counts

Every well-level screen, with how many wells, features and controls it has, and how many cells a well holds.

import matplotlib.pyplot as plt
import numpy as np
import pandas as pd

import mantispy as mt

WELLS = [
    "bbbc021",
    "rohban",
    "pki",
    "jump_target2",
    "jump_crispr",
    "neuropainting",
    "chroma",
    "oasis_pilot",
]
datasets = {name: getattr(mt.ds, name)() for name in WELLS}

rows = []
for name, adata in datasets.items():
    obs = adata.obs
    rows.append(
        {
            "dataset": name,
            "accession": adata.uns["mantispy"]["dataset"],
            "wells": adata.n_obs,
            "features": adata.n_vars,
            "plates": obs["Metadata_Plate"].nunique(),
            "controls": int(obs["Metadata_Control"].sum()) if "Metadata_Control" in obs else pd.NA,
            "cells per well": obs["Metadata_CellCount"].median(),
            "fields per well": obs["Metadata_SiteCount"].median() if "Metadata_SiteCount" in obs else pd.NA,
        }
    )
pd.DataFrame(rows).set_index("dataset")

	accession	wells	features	plates	controls	cells per well	fields per well
dataset
bbbc021	BBBC021	632	467	55	330	592.0	4.0
rohban	cpg0017-rohban-pathways	1918	3616	5	120	633.0	9.0
pki	cpg0008-pki	3072	5839	8	832	1669.0	9.0
jump_target2	cpg0016-jump	5374	3616	11	898	1427.5	9.0
jump_crispr	cpg0016-jump-assembled	51185	595	148	7478	3093.0	<NA>
neuropainting	cpg0038-tegtmeyer-neuropainting	1691	204	4	<NA>	92.0	4.0
chroma	cpg0029-chroma-pilot	3455	64	8	216	625.0	9.0
oasis_pilot	cpg0033-oasis-pilot	4604	99	12	511	4443.5	9.0

controls is Metadata_Control, which the loaders of bbbc021, rohban, pki, jump_target2, jump_crispr and oasis_pilot set. chroma marks its untreated wells negcon in Metadata_control_type instead. Five datasets are left out of the well-count table:

• pooled_rare holds one profile per barcode of a pooled screen, so there is no well to count cells in;

• jump_cells holds 13,578 single cells of one JUMP plate, see its page;

• jump_plate holds the images and segmentations of one well of the same plate;

• jump_export is the CellProfiler export directory of one field of view of that plate;

• jump_lite holds the same 1,536 JUMP wells under six feature sets, see its page.

Where the cell counts come from

Metadata_CellCount is the number of cells behind a row, and Metadata_SiteCount the number of fields of view that contributed cells. Every count is exact, taken from what the upstream pipeline measured:

dataset	cell count
bbbc021	summed from the Image.csv of each well, see its page
rohban, pki	Metadata_Count_Cells in the augmented profiles
oasis_pilot, chroma	Metadata_Count_Cells in the profiles
neuropainting	Metadata_Object_Count in the profiles
jump_target2	each plate’s backend table, see its page
jump_crispr	crispr_cell_counts.csv.gz of jump-profiling-recipe, the counts its cell-count correction regresses out

Metadata_Object_Count is the number of cells pycytominer aggregated into a well. It equals Metadata_Count_Cells in every well of every file here that publishes both, so it is as exact. read_profiles() copies either one to Metadata_CellCount, and Metadata_Site_Count to Metadata_SiteCount, so a Gallery file read directly carries them too.

Cells per field

Datasets image different numbers of fields per well, and so do plates within some of them. Cells per field compares them on one scale.

per_field = {
    name: (adata.obs["Metadata_CellCount"] / adata.obs["Metadata_SiteCount"]).dropna().to_numpy()
    for name, adata in datasets.items()
    if "Metadata_SiteCount" in adata.obs
}
order = sorted(per_field, key=lambda name: np.median(per_field[name]))
fig, ax = plt.subplots(figsize=(8, 4))
ax.boxplot([per_field[name] for name in order], tick_labels=order, showfliers=False)
ax.set_yscale("log")
ax.set_ylabel("cells per field of view")
ax.tick_params(axis="x", rotation=45)

Metadata_SiteCount counts the fields of view that contributed cells, which need not be all those imaged: an empty field contributes none, and so does one left out of the analysis. JUMP-Target-2 has examples of both. jump_crispr has no field count, because the recipe’s table publishes none.