Overview

Every dataset mantispy.ds serves, 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",
    "luad",
    "agnp",
    "neuropainting",
    "amish",
    "chroma",
    "oasis_pilot",
    "miami",
]
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	473	55	330	592.0	4.0
rohban	cpg0017-rohban-pathways	1918	3634	5	120	633.0	9.0
pki	cpg0008-pki	3072	5857	8	832	1669.0	9.0
jump_target2	cpg0016-jump	5374	3634	11	898	1427.5	9.0
jump_crispr	cpg0016-jump-assembled	51185	597	148	<NA>	3093.0	<NA>
luad	cpg0031-caicedo-cmvip	6144	432	16	<NA>	1757.5	9.0
agnp	cpg0040-garcia-fossa-AgNP	180	112	4	<NA>	317.5	28.5
neuropainting	cpg0038-tegtmeyer-neuropainting	1691	204	4	<NA>	92.0	4.0
amish	cpg0047-amish	468	575	3	<NA>	373.0	9.0
chroma	cpg0029-chroma-pilot	3455	64	8	<NA>	625.0	9.0
oasis_pilot	cpg0033-oasis-pilot	4604	99	12	<NA>	4443.5	9.0
miami	cpg0006-miami	2775	66	14	<NA>	1586.0	9.0

controls is Metadata_Control, which the loaders of bbbc021, rohban, pki and jump_target2 set. agnp and chroma mark their untreated wells negcon in Metadata_control_type instead. Three datasets are not wells and are left out of the 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.

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
luad, miami, oasis_pilot, amish, chroma	Metadata_Count_Cells in the profiles
agnp, 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.