pairwise.py

from typing import Mapping, Optional, Sequence
import warnings
import seaborn as sns
import matplotlib.pyplot as plt
import scipy.stats
import pandas as pd
from math import ceil
import numpy as np
import altair as alt
from itertools import zip_longest
from .util import _choose_mtx_rep
import statsmodels.stats.multitest


def plot_paired_fc(
    adata,
    groupby,
    paired_by,
    *,
    metric="log2_fc",
    var_names=None,
    layer=None,
    n_top_vars=30,
    fold_changes=None,
    metric_name=None,
    pvalues=None,
):
    """Plot fold changes as a bar chart with overlayed scatterplot to represent the variability between samples.

    Essentiallly this is a more compact representation of the same results as in :func:`plot_paired`.
    """
    groups = adata.obs[groupby].unique()
    if len(groups) != 2:
        raise ValueError(
            "The number of groups in the group_by column must be exactely 2"
        )

    if var_names is None:
        var_names = adata.var_names
        if len(var_names) > 20:
            warnings.warn(
                "You are plotting more than 20 variables which may be slow. "
                "Explicitly set the `var_names` parameter to turn this off. "
            )

    adata = adata[:, var_names]
    X = _choose_mtx_rep(adata, False, layer)
    try:
        X = X.toarray()
    except AttributeError:
        pass

    groupby_cols = [groupby]
    if paired_by is not None:
        groupby_cols.insert(0, paired_by)

    df = adata.obs.loc[:, groupby_cols].join(
        pd.DataFrame(X, index=adata.obs_names, columns=var_names)
    )

    # remove unpaired samples
    df[paired_by] = df[paired_by].astype(str)
    df.set_index(paired_by, inplace=True)
    has_matching_samples = df.groupby(paired_by).apply(
        lambda x: sorted(x[groupby]) == sorted(groups)
    )
    has_matching_samples = has_matching_samples.index[has_matching_samples].values
    removed_samples = adata.obs[paired_by].nunique() - len(has_matching_samples)
    if removed_samples:
        warnings.warn(f"{removed_samples} unpaired samples removed")

    df = df.loc[has_matching_samples, :]
    df.reset_index(drop=False, inplace=True)

    if metric == "diff":
        if metric_name is None:
            metric_name = "mean score difference"

        def metric(a, b):
            return b - a

    elif metric == "log2_fc":
        if metric_name is None:
            metric_name = "log2(FC)"

        def metric(a, b):
            return np.log2(b + 1) - np.log2(a + 1)

    df_fc = (
        df.sort_values([groupby, paired_by])
        .groupby(paired_by)
        .apply(
            lambda x: (
                pd.DataFrame(
                    metric(
                        x.loc[x[groupby] == groups[0], x.columns != groupby].values,
                        x.loc[x[groupby] == groups[1], x.columns != groupby].values,
                    ),
                    columns=x.columns[x.columns != groupby],
                )
            )
        )
        .reset_index(drop=False)
        .drop(columns=["level_1"])
    )

    df_fc_melt = (
        df_fc.melt(id_vars=paired_by)
        .groupby(["variable"])
        .apply(lambda x: x.assign(mean=np.mean(x["value"])))
    ).reset_index(drop=True)

    order = (
        df_fc_melt.loc[:, ["variable", "mean"]]
        .drop_duplicates()
        .sort_values("mean", ascending=False)["variable"]
        .tolist()
    )
    domain = np.max(np.abs([np.min(df_fc_melt["mean"]), np.max(df_fc_melt["mean"])]))

    return alt.Chart(
        df_fc_melt.loc[:, ["variable", "mean"]].drop_duplicates()
    ).mark_bar().encode(
        x=alt.X("variable", sort=order),
        y="mean",
        color=alt.Color(
            "mean",
            scale=alt.Scale(
                scheme="redblue",
                reverse=True,
                domain=[-domain, domain],
            ),
        ),
    ) + alt.Chart(
        df_fc_melt
    ).mark_circle(
        color="black"
    ).encode(
        x=alt.X("variable", sort=order), y=alt.Y("value", title=metric_name)
    )


def plot_paired(
    adata,
    groupby,
    *,
    paired_by=None,
    var_names=None,
    show=True,
    return_fig=False,
    n_cols=4,
    panel_size=(3, 4),
    show_legend=True,
    hue=None,
    size=10,
    ylabel="expression",
    pvalues: Sequence[float] = None,
    pvalue_template=lambda x: f"unadj. p={x:.2f}, t-test",
    adjust_fdr=False,
    boxplot_properties=None,
    palette=None,
):
    """
    Pairwise expression plot.

    Makes on panel with a paired scatterplot for each variable.

    Parameters
    ----------
    adata
        adata matrix (usually pseudobulk).
    group_by
        Column containing the grouping. Must contain exactely two different values.
    paired_by
        Column indicating the pairing (e.g. "patient")
    var_names
        Only plot these variables. Default is to plot all
    """
    if boxplot_properties is None:
        boxplot_properties = {}
    groups = adata.obs[groupby].unique()
    if len(groups) != 2:
        raise ValueError(
            "The number of groups in the group_by column must be exactely 2"
        )

    if var_names is None:
        var_names = adata.var_names
        if len(var_names) > 20:
            warnings.warn(
                "You are plotting more than 20 variables which may be slow. "
                "Explicitly set the `var_names` parameter to turn this off. "
            )

    X = adata[:, var_names].X
    try:
        X = X.toarray()
    except AttributeError:
        pass

    groupby_cols = [groupby]
    if paired_by is not None:
        groupby_cols.insert(0, paired_by)
    if hue is not None:
        groupby_cols.insert(0, hue)

    df = adata.obs.loc[:, groupby_cols].join(
        pd.DataFrame(X, index=adata.obs_names, columns=var_names)
    )

    if paired_by is not None:
        # remove unpaired samples
        df[paired_by] = df[paired_by].astype(str)
        df.set_index(paired_by, inplace=True)
        has_matching_samples = df.groupby(paired_by).apply(
            lambda x: sorted(x[groupby]) == sorted(groups)
        )
        has_matching_samples = has_matching_samples.index[has_matching_samples].values
        removed_samples = adata.obs[paired_by].nunique() - len(has_matching_samples)
        if removed_samples:
            warnings.warn(f"{removed_samples} unpaired samples removed")

        # perform statistics (paired ttest)
        if pvalues is None:
            _, pvalues = scipy.stats.ttest_rel(
                df.loc[
                    df[groupby] == groups[0],
                    var_names,
                ].loc[has_matching_samples, :],
                df.loc[
                    df[groupby] == groups[1],
                    var_names,
                ].loc[has_matching_samples],
            )

        df = df.loc[has_matching_samples, :]
        df.reset_index(drop=False, inplace=True)

    else:
        if pvalues is None:
            _, pvalues = scipy.stats.ttest_ind(
                df.loc[
                    df[groupby] == groups[0],
                    var_names,
                ],
                df.loc[
                    df[groupby] == groups[1],
                    var_names,
                ],
            )

    if adjust_fdr:
        pvalues = statsmodels.stats.multitest.fdrcorrection(pvalues)[1]

    # transform data for seaborn
    df_melt = df.melt(
        id_vars=groupby_cols,
        var_name="var",
        value_name="val",
    )

    # start plotting
    n_panels = len(var_names)
    nrows = ceil(n_panels / n_cols)
    ncols = min(n_cols, n_panels)

    fig, axes = plt.subplots(
        nrows,
        ncols,
        figsize=(ncols * panel_size[0], nrows * panel_size[1]),
        tight_layout=True,
        squeeze=False,
    )
    axes = axes.flatten()
    if hue is None:
        hue = paired_by
    for i, (var, ax) in enumerate(zip_longest(var_names, axes)):
        if var is not None:
            sns.stripplot(
                x=groupby,
                data=df_melt.loc[lambda x: x["var"] == var],
                y="val",
                ax=ax,
                hue=hue,
                size=size,
                linewidth=1,
                palette=palette,
            )
            if paired_by is not None:
                sns.lineplot(
                    x=groupby,
                    data=df_melt.loc[lambda x: x["var"] == var],
                    hue=hue,
                    y="val",
                    ax=ax,
                    legend=False,
                    errorbar=None,
                    palette=palette,
                )
            sns.boxplot(
                x=groupby,
                data=df_melt.loc[lambda x: x["var"] == var],
                y="val",
                ax=ax,
                color="white",
                fliersize=0,
                **boxplot_properties,
            )

            ax.set_xlabel("")
            ax.tick_params(
                axis="x",
                # rotation=0,
                labelsize=15,
            )
            ax.legend().set_visible(False)
            ax.set_ylabel(ylabel)
            ax.set_title(var + "\n" + pvalue_template(pvalues[i]))
        else:
            ax.set_visible(False)
    fig.tight_layout()

    if show_legend == True:
        axes[n_panels - 1].legend().set_visible(True)
        axes[n_panels - 1].legend(bbox_to_anchor=(1.1, 1.05))

    if show:
        plt.show()

    if return_fig:
        return fig