Pairwise Compare

Compare the rows of a tidy pandas dataframe in pairs, for example the correlation between the profiles of wells that differ in their cell line. You choose which columns must be different and which must be the same between the two rows of a pair, and get one result per pair.

The problem it solves: as an analyst of image-based profiles, I need the similarity of every pair of samples that differ in one way and match in another (for example, wells of different cell lines on the same plate), so that I can see how replicates agree and how groups differ, because the average of a group hides how its samples relate.

flowchart LR
    subgraph U2OS["U2OS (plate P1)"]
        A1["well A1"]
        A2["well A2"]
    end
    subgraph HeLa["HeLa (plate P1)"]
        B1["well B1"]
        B2["well B2"]
    end
    A1 -.-|"r = 1.00"| A2
    B1 -.-|"r = 1.00"| B2
    A1 ---|"r = 0.90"| B1
    A2 ---|"r = 0.90"| B1
    A1 ---|"r = 0.92"| B2
    A2 ---|"r = 0.92"| B2

Each line is one result: the Pearson correlation r of two wells. Solid lines join wells of different cell lines, and dashed lines join the replicate wells of one cell line.

The example notebook uses a dataset from the nf1_schwann_cell_painting_data. Most of the development is in the src folder.

Quick start

Take four wells of one plate, with the metadata that identify them and four measured features:

import pandas as pd
from pairwise_compare import compare_pairs

wells = pd.DataFrame(
    {
        "plate": ["P1", "P1", "P1", "P1"],
        "cell_line": ["U2OS", "U2OS", "HeLa", "HeLa"],
        "well": ["A1", "A2", "B1", "B2"],
        "area": [10.0, 11.0, 20.0, 19.0],
        "intensity": [5.0, 5.5, 2.0, 2.5],
        "texture": [0.1, 0.3, 0.9, 0.7],
        "eccentricity": [0.5, 0.4, 0.8, 0.9],
    }
)

compare_pairs(
    wells,
    different=["cell_line", "well"],  # the two wells of a pair are of different cell lines
    features=["area", "intensity", "texture", "eccentricity"],
    same=["plate"],  # and are on the same plate
)

The result has one row per pair of wells, with their Pearson correlation across the features:

This call compares wells of different cell lines (the solid lines in the picture). Wells of the same cell line are not compared with each other.

To compare the replicate wells of each cell line instead (the dashed lines), keep the cell line the same and ask for different wells:

compare_pairs(
    wells,
    different=["well"],  # the two wells of a pair are different wells
    features=["area", "intensity", "texture", "eccentricity"],
    same=["plate", "cell_line"],  # of the same cell line on the same plate
)

Each column must either match (same) or differ (different), so one call gives one kind of pair. To get every pair of different wells in one call, use different=["well"] and same=["plate"]. The results only hold the columns you name, so add the cell line to each pair afterwards by looking it up from its well.

Installation

You can install the package with either of the following commands:
pip install pairwise-compare
uv add pairwise-compare

Data format

The dataframe is tidy, as described by Wickham (2014): one row per sample (for example, one well) and one column per variable.

• Group columns hold the metadata that define which rows are compared, such as the plate, the cell line, or the well. They are the columns listed in different, same and drop.
• Feature columns hold the numbers that are compared, such as measurements of the cell morphology. They are listed in features and must be numeric.

All other columns are ignored.

Which rows are compared

• Two rows are compared when they have a different value in every different column and, if same is given, the same value in every same column.
• different needs at least two columns, or one column and at least one same column. A column cannot be in both.
• Each pair is compared once: the pair (A, B) is not repeated as (B, A), and a row is not compared with itself.
• group0 and group1 in the output are the two rows of a pair. For a same column both have the same value.
• drop leaves columns of same or different out of the results. They still decide which rows are compared. For example, drop=["Metadata_Well"] removes the well columns and keeps the others.

Missing values

• A row with a missing value in a same or different column is left out of the comparisons, with a warning that says how many rows and which columns.
• A row with a missing value in a feature column is kept, and the results of its comparisons are missing (NaN), with a warning.
• Missing values in other columns are ignored.
• A row with the same value in every feature has no Pearson correlation (NaN), with a warning.

Order of the results

• The first row of a pair (group0) is the one that comes first by its values in the group columns, and the second row (group1) comes after it. With same columns, rows are first compared by their same values, and then by their different values, column by column. Without same columns, rows are compared by the first different column, and then by the other ones.
• The pairs are sorted by these keys, in this order:
  1. the same columns (both rows of a pair have the same values in them), or, without same columns, the first different column of group0 and then of group1;
  2. the different columns of group0 (the other ones without same columns) and then those of group1;
  3. the position in the dataframe of group0 and then of group1.

Other comparisons

compare_pairs uses the Pearson correlation by default. To compare the rows another way, pass a comparator, which is a class that inherits from Comparator (see Comparator.py and the examples in test_comparator.py). A comparator has a name, which is the name of its column in the results, and two methods:

• prepare(values) is called once with the feature values of all the compared rows (a numpy array with one row per sample), and returns whatever compare needs. By default, it returns the values as they are.
• compare(prepared, rows0, rows1) returns one value for each pair of rows (rows0[k], rows1[k]), as a numpy array.

A comparator keeps no results, so the same comparator can be used for any number of calls to compare_pairs.

Speed

All the pairs are found and compared with array operations, so the time grows with the number of pairs and the number of features, not with a loop over the groups. The speed test notebook times compare_pairs for plates of several sizes, so you can see how long a comparison of your size takes.

The memory use depends on the number of pairs. The results table needs memory for every pair, and finding the pairs inside one group of same values needs memory for every combination of two rows of the group. Together they peaked at about 80 bytes for every combination of two rows in a test with one group (100 features, a group of 6,000 rows: 18 million pairs, 2.9 GB, 3 seconds), so a group of 20,000 rows would need about 32 GB. Add a same column that splits very large groups into smaller ones.

Citation

If you use Pairwise Compare, please cite it with the metadata in CITATION.cff (the "Cite this repository" button of GitHub uses it). The tidy data format is described in:

Wickham, H. (2014). Tidy Data. Journal of Statistical Software, 59(10), 1-23.

Development

See CONTRIBUTING.md for how to run the tests and make a release.