Calculate phenotypic activity

A common first analysis for morphological datasets is the activity of the cells’ phenotypes. We will use the copairs package, which makes use of mean average precision to obtain a metric of replicability for any set of morphological profiles. In other words, it indicates how similar a given set of compounds are, relative to their negative controls, which is usually cells that have experienced no perturbation.

import polars as pl
import polars.selectors as cs
import seaborn as sns
from broad_babel.query import get_mapper
from copairs.map import average_precision

Downloading data from 'https://zenodo.org/records/12211976/files/babel.db' to file '/home/runner/.cache/pooch/2eaa6a2f4915f72d7100683f53982ed8-babel.db'.

We will be using the CRISPR dataset specificed in our index csv, but we will select a subset of perturbations and the controls present.

INDEX_FILE = "https://raw.githubusercontent.com/jump-cellpainting/datasets/50cd2ab93749ccbdb0919d3adf9277c14b6343dd/manifests/profile_index.csv"
CRISPR_URL = pl.read_csv(INDEX_FILE).filter(pl.col("subset") == "crispr").item(0, "url")
profiles = pl.scan_parquet(CRISPR_URL)

Sample perturbations and add known negative control.

jcp_ids = (
    profiles.select(pl.col("Metadata_JCP2022")).unique().collect().to_series().sort()
)
subsample = jcp_ids.sample(10, seed=42)
subsample = (*subsample, "JCP2022_800002")  # Add the only control in CRISPR data
profiles_subset = profiles.filter(pl.col("Metadata_JCP2022").is_in(subsample)).collect()
unique_plates = profiles_subset.filter(pl.col("Metadata_JCP2022") != subsample[-1])[
    "Metadata_Plate"
].unique()
perts_controls = profiles_subset.filter(pl.col("Metadata_Plate").is_in(unique_plates))
with pl.Config() as cfg:
    cfg.set_tbl_cols(7)  # Limit the number of columns printed
    print(perts_controls.head())

shape: (5, 263)
┌──────────────┬──────────────┬──────────────┬─────────────┬───┬───────────┬───────────┬───────────┐
│ Metadata_Sou ┆ Metadata_Pla ┆ Metadata_Wel ┆ Metadata_JC ┆ … ┆ X_257     ┆ X_258     ┆ X_259     │
│ rce          ┆ te           ┆ l            ┆ P2022       ┆   ┆ ---       ┆ ---       ┆ ---       │
│ ---          ┆ ---          ┆ ---          ┆ ---         ┆   ┆ f32       ┆ f32       ┆ f32       │
│ str          ┆ str          ┆ str          ┆ str         ┆   ┆           ┆           ┆           │
╞══════════════╪══════════════╪══════════════╪═════════════╪═══╪═══════════╪═══════════╪═══════════╡
│ source_13    ┆ CP-CC9-R1-05 ┆ I23          ┆ JCP2022_800 ┆ … ┆ -0.024703 ┆ 0.148245  ┆ 0.069796  │
│              ┆              ┆              ┆ 002         ┆   ┆           ┆           ┆           │
│ source_13    ┆ CP-CC9-R1-05 ┆ J02          ┆ JCP2022_800 ┆ … ┆ -0.119861 ┆ 0.047197  ┆ -0.146323 │
│              ┆              ┆              ┆ 002         ┆   ┆           ┆           ┆           │
│ source_13    ┆ CP-CC9-R1-05 ┆ L23          ┆ JCP2022_800 ┆ … ┆ -0.123536 ┆ 0.038496  ┆ 0.156573  │
│              ┆              ┆              ┆ 002         ┆   ┆           ┆           ┆           │
│ source_13    ┆ CP-CC9-R1-05 ┆ O23          ┆ JCP2022_800 ┆ … ┆ 0.07883   ┆ 0.080706  ┆ -0.003756 │
│              ┆              ┆              ┆ 002         ┆   ┆           ┆           ┆           │
│ source_13    ┆ CP-CC9-R1-05 ┆ M02          ┆ JCP2022_800 ┆ … ┆ -0.018471 ┆ -0.074819 ┆ 0.057594  │
│              ┆              ┆              ┆ 002         ┆   ┆           ┆           ┆           │
└──────────────┴──────────────┴──────────────┴─────────────┴───┴───────────┴───────────┴───────────┘

Now we create a mapper to label treatments and controls. See the previous tutorial for details on fetching metadata.

pert_mapper = get_mapper(
    subsample, input_column="JCP2022", output_columns="JCP2022,pert_type"
)
perts_controls_annotated = perts_controls.with_columns(
    pl.col("Metadata_JCP2022").replace(pert_mapper).alias("pert_type")
)

Finally we use the parameters from . See the copairs wiki for more details on the parameters that copairs requires.

pos_sameby = ["Metadata_JCP2022"]  # We want to match perturbations
pos_diffby = []
neg_sameby = []
neg_diffby = ["pert_type"]
batch_size = 20000

metadata_selector = cs.starts_with(("Metadata", "pert_type"))
meta = perts_controls_annotated.select(metadata_selector)
features = perts_controls_annotated.select(~metadata_selector)

result = average_precision(
    meta.to_pandas(),
    features.to_numpy(),
    pos_sameby,
    pos_diffby,
    neg_sameby,
    neg_diffby,
    batch_size,
)
result = pl.DataFrame(
    result
)  # We convert back to polars because we prefer how it prints dataframes
result.head()

shape: (5, 8)

Metadata_Source	Metadata_Plate	Metadata_Well	Metadata_JCP2022	pert_type	n_pos_pairs	n_total_pairs	average_precision
str	str	str	str	str	i64	i64	f64
"source_13"	"CP-CC9-R1-05"	"I23"	"JCP2022_800002"	"negcon"	419	471	0.920554
"source_13"	"CP-CC9-R1-05"	"J02"	"JCP2022_800002"	"negcon"	419	471	0.920515
"source_13"	"CP-CC9-R1-05"	"L23"	"JCP2022_800002"	"negcon"	419	471	0.931227
"source_13"	"CP-CC9-R1-05"	"O23"	"JCP2022_800002"	"negcon"	419	471	0.92076
"source_13"	"CP-CC9-R1-05"	"M02"	"JCP2022_800002"	"negcon"	419	471	0.951237

The result of copairs is a dataframe containing, in addition to the original metadata, the average precision with which perturbations were retrieved. Perturbations that look more similar to each other than to the negative controls in the plates present in the same plates will be higher. Perturbations that do not differentiate themselves against negative controls will be closer to zero.

To wrap up we pull the standard gene symbol and plot the distribution of average precision.

name_mapper = get_mapper(
    subsample, input_column="JCP2022", output_columns="JCP2022,standard_key"
)

to_plot = result.filter(pl.col("pert_type") == "trt").with_columns(
    pl.col("Metadata_JCP2022").replace(name_mapper).alias("Perturbed gene")
)

# Plotting
sns.stripplot(data=to_plot.to_pandas(), x="average_precision", y="Perturbed gene")

We can see that only some perturbations can be easily retrieved when compared to negative controls, in this case KIF16B and CDK20. For a deeper dive into how mean Average Precision (mAP) works, you can explore this notebook.

Other Formats