lotsofcells 0.0.0__py3-none-any.whl

lotsofcells/__init__.py

@@ -0,0 +1,39 @@
+"""lotsofcells: proportion-test statistics and visualization on single-cell metadata.
+
+Python port of the R package `lotsOfCells`, designed for the scanpy / AnnData
+framework. Compatible with single-cell (`AnnData`) and spatial transcriptomics
+(`SpatialData` / `MuData`) objects, since metadata is read from `.obs`.
+
+References
+----------
+Óscar González-Velasco; lotsOfCells: data visualization and statistics of
+single cell metadata. bioRxiv 2024.05.23.595582;
+https://doi.org/10.1101/2024.05.23.595582
+"""
+
+from ._utils import get_metadata, get_palette
+from .lotsofcells import lots_of_cells
+from .entropy import entropy_score
+from .plots import (
+    bar_chart,
+    waffle_chart,
+    polar_chart,
+    density_chart,
+    dynamics_chart,
+    plot_abundance_test,
+)
+
+__all__ = [
+    "get_metadata",
+    "get_palette",
+    "lots_of_cells",
+    "entropy_score",
+    "bar_chart",
+    "waffle_chart",
+    "polar_chart",
+    "density_chart",
+    "dynamics_chart",
+    "plot_abundance_test",
+]
+
+__version__ = "0.3.0"

lotsofcells/_stats.py

@@ -0,0 +1,279 @@
+"""Internal statistical primitives.
+
+Direct ports of the R helpers `cellToGamma`, `cellToGammaOriginal` and
+`cellToMontecarlo`. Implementation choices (pseudocounts, transforms) match
+the R version exactly so results are comparable.
+"""
+from __future__ import annotations
+
+from typing import Dict, List, Sequence, Tuple, Union
+
+import numpy as np
+import pandas as pd
+
+
+# --- Transformations used everywhere ---------------------------------------------------
+
+def pseudo_count(counts: np.ndarray) -> np.ndarray:
+    """`counts + 0.5` — matches the R pseudocount in lotsOfCells.R."""
+    return counts + 0.5
+
+
+def pseudo_count_arcsin(counts: np.ndarray) -> np.ndarray:
+    """`counts + sqrt(counts^2 + 1)` — matches the R pseudocount in entropyScore.R."""
+    return counts + np.sqrt(counts * counts + 1)
+
+
+def asrt(p: np.ndarray) -> np.ndarray:
+    """Arcsin square-root transform (Anscombe-style)."""
+    return np.arcsin(np.sqrt(np.clip(p, 0, 1)))
+
+
+def logit(f: np.ndarray) -> np.ndarray:
+    return np.log(f / (1 - f))
+
+
+def geom_mean(x: np.ndarray) -> float:
+    """Geometric mean over the strictly positive entries of ``x``.
+
+    Note: this intentionally diverges from R's literal ``exp(mean(log(x)))``,
+    which collapses to 0 whenever **any** entry is 0. In the symmetric
+    divergence formula used by `entropyScore`, a zero in
+    ``|p * log2(p/q)|`` means ``p[i] == q[i]`` (the two distributions agree
+    on cell type ``i``); such a term should contribute *nothing* to the
+    divergence — not zero out the entire score.
+
+    The 1-class test makes this critical: random partitions inside a single
+    condition often share integer totals after the ``int(sqrt(count_s))``
+    crowd sizing, which forces ``p[i] == q[i]`` for any cell type missing
+    from both subsamples. With strict R semantics every iteration collapses
+    to 0; with this version the geom_mean is taken over the cell types
+    that actually disagree.
+
+    If every entry is zero, the divergence really is 0.
+    """
+    x = np.asarray(x, dtype=float)
+    nonzero = x[x > 0]
+    if nonzero.size == 0:
+        return 0.0
+    return float(np.exp(np.mean(np.log(nonzero))))
+
+
+# --- Contingency tables --------------------------------------------------------------
+
+def _table(groups: Sequence[str], covariable: Sequence[str]) -> pd.DataFrame:
+    """Equivalent of R `table(data.frame(groups, covariable))`."""
+    return (
+        pd.crosstab(pd.Series(groups, name="groups"),
+                    pd.Series(covariable, name="covariable"))
+    )
+
+
+def _ensure_rows(tab: pd.DataFrame, label_order: Sequence[str]) -> pd.DataFrame:
+    """Add zero rows for any missing labels and reindex."""
+    missing = [l for l in label_order if l not in tab.index]
+    if missing:
+        z = pd.DataFrame(0, index=missing, columns=tab.columns)
+        tab = pd.concat([tab, z])
+    return tab.reindex(label_order)
+
+
+def _ensure_cols(tab: pd.DataFrame, indexes: Sequence[str]) -> pd.DataFrame:
+    missing = [c for c in indexes if c not in tab.columns]
+    if missing:
+        for m in missing:
+            tab[m] = 0
+    return tab[list(indexes)]
+
+
+# --- Goodman & Kruskal gamma rank correlation ----------------------------------------
+
+def _ranked_proportions(
+    tab: pd.DataFrame,
+    label_order: Sequence[str],
+    indexes: Sequence[str],
+) -> np.ndarray:
+    """Rows=label_order, cols=covariables.
+
+    Computes per-covariable proportions then ranks across labels.
+    Mirrors `t(apply(dftmp,2,function(row){row/(sum(row)+0.1)}))[labelOrder, indexes]`
+    followed by `t(apply(.,1,rank))`.
+    """
+    tab = _ensure_rows(tab, label_order)
+    tab = _ensure_cols(tab, indexes)
+    # column-wise proportions: row/(sum(row)+0.1) per column => divide each column by (col_sum+0.1)
+    col_sums = tab.values.sum(axis=0) + 0.1  # shape (n_cov,)
+    contig = tab.values / col_sums[np.newaxis, :]  # rows = labels in label_order
+    # rank within each row across covariables (R: apply(contig_tab,1,rank))
+    # 'average' ties to mirror base::rank's default
+    ranks = np.apply_along_axis(_rank_avg, 1, contig)
+    return ranks  # shape (n_labels, n_cov)
+
+
+def _rank_avg(x: np.ndarray) -> np.ndarray:
+    """Equivalent of R base::rank(x, ties.method='average')."""
+    order = np.argsort(x, kind="mergesort")
+    ranks = np.empty_like(order, dtype=float)
+    ranks[order] = np.arange(1, len(x) + 1, dtype=float)
+    # average over ties
+    _, inv, counts = np.unique(x, return_inverse=True, return_counts=True)
+    sums = np.zeros_like(counts, dtype=float)
+    np.add.at(sums, inv, ranks)
+    avg = sums / counts
+    return avg[inv]
+
+
+def _concordant_discordant(
+    ranks: np.ndarray, rank_index: np.ndarray, original: bool
+) -> Tuple[np.ndarray, np.ndarray]:
+    """For each covariable column, count concordant and discordant pairs.
+
+    If `original=False` (random/null): concordant means
+    sign(ranks[i]-ranks[i+1:]) == -1 (matches the R cellToGamma which assumes
+    monotonic 1..N and sign always = -1). Discordant counts where
+    `ranks[i] != ranks[k]` and sign != -1.
+
+    If `original=True`: compare against the actual rank_index sign pattern.
+    """
+    n_labels, n_cov = ranks.shape
+    nconc = np.zeros(n_cov, dtype=int)
+    ndisc = np.zeros(n_cov, dtype=int)
+    for i in range(n_labels - 1):
+        ri = ranks[i]
+        rj = ranks[i + 1 :]  # (rest, n_cov)
+        diff_r = ri[np.newaxis, :] - rj  # (rest, n_cov)
+        if original:
+            idx_diff = rank_index[i] - rank_index[i + 1 :]
+            target_sign = np.sign(idx_diff)[:, np.newaxis]  # (rest, 1)
+            nconc += np.sum(np.sign(diff_r) == target_sign, axis=0)
+            mask_neq = diff_r != 0
+            ndisc += np.sum((np.sign(diff_r) != target_sign) & mask_neq, axis=0)
+        else:
+            nconc += np.sum(np.sign(diff_r) == -1, axis=0)
+            mask_neq = diff_r != 0
+            ndisc += np.sum((np.sign(diff_r) != -1) & mask_neq, axis=0)
+    return nconc, ndisc
+
+
+def cell_to_gamma(
+    covariable: np.ndarray,
+    groups: np.ndarray,
+    label_order: Sequence[str],
+    indexes: Sequence[str],
+    cell_crowd: Dict[str, int],
+    rank_index: np.ndarray,
+    rng: np.random.Generator,
+) -> Tuple[np.ndarray, np.ndarray]:
+    """Random null distribution: mix all covariables, then subsample per-group.
+
+    Returns (n_concordant, n_discordant) per covariable column (length n_cov).
+    """
+    pieces_cov, pieces_grp = [], []
+    for label in label_order:
+        n = int(cell_crowd[label])
+        sample = rng.choice(covariable, size=n, replace=True)
+        pieces_cov.append(sample)
+        pieces_grp.append(np.repeat(label, n))
+    cov = np.concatenate(pieces_cov)
+    grp = np.concatenate(pieces_grp)
+    tab = _table(grp, cov)
+    ranks = _ranked_proportions(tab, label_order, indexes)
+    return _concordant_discordant(ranks, rank_index, original=False)
+
+
+def cell_to_gamma_original(
+    covariable: np.ndarray,
+    groups: np.ndarray,
+    label_order: Sequence[str],
+    indexes: Sequence[str],
+    cell_crowd: Dict[str, int],
+    rank_index: np.ndarray,
+    rng: np.random.Generator,
+) -> Tuple[np.ndarray, np.ndarray]:
+    """Original-data subsampling: subsample within each group preserving labels."""
+    pieces_cov, pieces_grp = [], []
+    for label in label_order:
+        n = int(cell_crowd[label])
+        pool = covariable[groups == label]
+        if len(pool) == 0:
+            continue
+        replace = n > len(pool)
+        sample = rng.choice(pool, size=n, replace=replace)
+        pieces_cov.append(sample)
+        pieces_grp.append(np.repeat(label, n))
+    cov = np.concatenate(pieces_cov)
+    grp = np.concatenate(pieces_grp)
+    tab = _table(grp, cov)
+    ranks = _ranked_proportions(tab, label_order, indexes)
+    return _concordant_discordant(ranks, rank_index, original=True)
+
+
+# --- Monte Carlo for 2-condition fold-change -----------------------------------------
+
+def _proportions_from_table(
+    tab: pd.DataFrame,
+    label_order: Sequence[str],
+    indexes: Sequence[str],
+    pseudo: bool = True,
+) -> np.ndarray:
+    """`pseudo_count(tab)` then column-wise proportions, indexed by label_order/indexes."""
+    tab = _ensure_rows(tab, label_order)
+    tab = _ensure_cols(tab, indexes)
+    vals = tab.values.astype(float)
+    if pseudo:
+        vals = pseudo_count(vals)
+    col_sums = vals.sum(axis=0) + 1.0
+    return vals / col_sums[np.newaxis, :]  # (n_labels, n_cov)
+
+
+def cell_to_montecarlo(
+    covariable: np.ndarray,
+    groups: np.ndarray,
+    label_order: Sequence[str],
+    indexes: Sequence[str],
+    cell_crowd: Union[Dict[str, int], Dict[str, List[int]]],
+    rng: np.random.Generator,
+) -> Tuple[np.ndarray, np.ndarray]:
+    """Return (mixed-pool fold change, original-resampled fold change).
+
+    Both are arrays of length len(indexes), holding
+    log2( asrt(p1) / asrt(p2) ).
+    """
+    def _build_mixed(crowd_for_label):
+        if isinstance(crowd_for_label, (list, np.ndarray)):
+            sizes = np.asarray(crowd_for_label, dtype=int)
+            return np.concatenate(
+                [rng.choice(covariable, size=int(s), replace=True) for s in sizes]
+            )
+        return rng.choice(covariable, size=int(crowd_for_label), replace=True)
+
+    def _build_orig(crowd_for_label, label):
+        pool = covariable[groups == label]
+        if len(pool) == 0:
+            return np.array([], dtype=covariable.dtype)
+        if isinstance(crowd_for_label, (list, np.ndarray)):
+            sizes = np.asarray(crowd_for_label, dtype=int)
+            return np.concatenate(
+                [rng.choice(pool, size=int(s), replace=True) for s in sizes]
+            )
+        n = int(crowd_for_label)
+        return rng.choice(pool, size=n, replace=True)
+
+    mixed_cov, mixed_grp, orig_cov, orig_grp = [], [], [], []
+    for label in label_order:
+        cm = _build_mixed(cell_crowd[label])
+        co = _build_orig(cell_crowd[label], label)
+        mixed_cov.append(cm)
+        mixed_grp.append(np.repeat(label, len(cm)))
+        orig_cov.append(co)
+        orig_grp.append(np.repeat(label, len(co)))
+
+    mixed_tab = _table(np.concatenate(mixed_grp), np.concatenate(mixed_cov))
+    orig_tab = _table(np.concatenate(orig_grp), np.concatenate(orig_cov))
+
+    p_mixed = _proportions_from_table(mixed_tab, label_order, indexes, pseudo=True)
+    p_orig = _proportions_from_table(orig_tab, label_order, indexes, pseudo=True)
+
+    fc_mixed = np.log2(asrt(p_mixed[0]) / asrt(p_mixed[1]))
+    fc_orig = np.log2(asrt(p_orig[0]) / asrt(p_orig[1]))
+    return fc_mixed, fc_orig

lotsofcells/_utils.py

@@ -0,0 +1,213 @@
+"""Internal helpers: metadata extraction and color palette."""
+from __future__ import annotations
+
+from typing import Optional, Sequence, Union
+
+import numpy as np
+import pandas as pd
+
+# Default categorical palette — scanpy `default_20` (Vega-20 with
+# yellow/green hues swapped for higher contrast). The de facto standard in
+# the single-cell community, so colours feel native to scanpy figures.
+_DEFAULT_PALETTE = [
+    "#1F77B4", "#FF7F0E", "#279E68", "#D62728", "#AA40FC",
+    "#8C564B", "#E377C2", "#B5BD61", "#17BECF", "#AEC7E8",
+    "#FFBB78", "#98DF8A", "#FF9896", "#C5B0D5", "#C49C94",
+    "#F7B6D2", "#DBDB8D", "#9EDAE5", "#AD494A", "#8C6D31",
+]
+
+
+def _is_anndata(obj) -> bool:
+    """Return True if obj quacks like an AnnData (has .obs)."""
+    try:
+        import anndata  # noqa: F401
+    except Exception:
+        anndata = None  # type: ignore
+    if anndata is not None and isinstance(obj, anndata.AnnData):
+        return True
+    return hasattr(obj, "obs") and isinstance(getattr(obj, "obs"), pd.DataFrame)
+
+
+def _is_spatialdata(obj) -> bool:
+    try:
+        import spatialdata  # type: ignore
+        return isinstance(obj, spatialdata.SpatialData)
+    except Exception:
+        return False
+
+
+def _is_mudata(obj) -> bool:
+    try:
+        import mudata  # type: ignore
+        return isinstance(obj, mudata.MuData)
+    except Exception:
+        return False
+
+
+def get_metadata(sc_object, table: Optional[str] = None) -> pd.DataFrame:
+    """Return a metadata DataFrame from a scanpy/spatial/dataframe object.
+
+    Parameters
+    ----------
+    sc_object
+        One of: ``pandas.DataFrame``, ``anndata.AnnData``, ``mudata.MuData``,
+        or ``spatialdata.SpatialData``. AnnData/Mu/Spatial objects expose their
+        cell-level metadata via ``.obs``; this is the analogue of
+        ``Seurat[[]]`` / ``SingleCellExperiment::colData``.
+    table
+        Only used when ``sc_object`` is a ``SpatialData`` (the name of the
+        table whose ``.obs`` should be returned) or ``MuData`` (the modality
+        name). If ``None`` and the object has a single table/modality, that
+        one is used.
+    """
+    if sc_object is None:
+        raise ValueError("At least an AnnData/SpatialData/DataFrame is required.")
+
+    if isinstance(sc_object, pd.DataFrame):
+        return sc_object.copy()
+
+    if _is_spatialdata(sc_object):
+        tables = dict(sc_object.tables)
+        if not tables:
+            raise ValueError("SpatialData object has no tables.")
+        if table is None:
+            if len(tables) > 1:
+                raise ValueError(
+                    f"SpatialData has multiple tables {list(tables)}; "
+                    "specify `table=...`."
+                )
+            table = next(iter(tables))
+        return tables[table].obs.copy()
+
+    if _is_mudata(sc_object):
+        if table is None:
+            return sc_object.obs.copy()
+        return sc_object[table].obs.copy()
+
+    if _is_anndata(sc_object):
+        return sc_object.obs.copy()
+
+    raise TypeError(
+        "Unsupported object type for metadata extraction. "
+        "Pass a pandas.DataFrame or AnnData/MuData/SpatialData."
+    )
+
+
+def get_numerical_variable(
+    sc_object, numerical_variable: str, metadata: pd.DataFrame
+) -> np.ndarray:
+    """Resolve a numerical variable from .obs OR feature counts (gene name).
+
+    Mirrors the R behaviour of `density_chart`: if the column is in
+    metadata, return it; otherwise look for a feature in the AnnData and
+    return its expression vector aligned to ``metadata.index``.
+    """
+    if numerical_variable in metadata.columns:
+        return metadata[numerical_variable].to_numpy()
+
+    if _is_anndata(sc_object):
+        adata = sc_object
+        if numerical_variable in adata.var_names:
+            idx = adata.var_names.get_loc(numerical_variable)
+            X = adata.X
+            col = X[:, idx]
+            if hasattr(col, "toarray"):
+                col = col.toarray().ravel()
+            else:
+                col = np.asarray(col).ravel()
+            # Align to metadata row order
+            obs_idx = metadata.index
+            full = pd.Series(col, index=adata.obs_names)
+            return full.loc[obs_idx].to_numpy()
+
+    raise ValueError(
+        f"Variable '{numerical_variable}' not found in metadata columns "
+        "or feature names."
+    )
+
+
+def get_palette(
+    use_palette: Optional[Sequence[str]] = None, n_colors: int = 20
+) -> list:
+    """Return a list of `n_colors` colors.
+
+    If `use_palette` is None, the default lotsOfCells palette is used.
+    If more colors than provided are requested, a linear interpolation in RGB
+    space (analogue of `colorRampPalette`) is performed.
+    """
+    base = list(use_palette) if use_palette is not None else list(_DEFAULT_PALETTE)
+    if n_colors <= len(base):
+        return base[:n_colors]
+    return _ramp_palette(base, n_colors)
+
+
+def _hex_to_rgb(h: str) -> np.ndarray:
+    h = h.lstrip("#")
+    return np.array([int(h[i : i + 2], 16) for i in (0, 2, 4)], dtype=float) / 255.0
+
+
+def _rgb_to_hex(rgb: Union[np.ndarray, Sequence[float]]) -> str:
+    rgb = np.clip(np.asarray(rgb), 0, 1)
+    return "#{:02X}{:02X}{:02X}".format(*(int(round(c * 255)) for c in rgb))
+
+
+def _ramp_palette(colors: Sequence[str], n: int) -> list:
+    """Equivalent of grDevices::colorRampPalette in linear RGB."""
+    rgbs = np.stack([_hex_to_rgb(c) for c in colors])  # (k, 3)
+    if n == 1:
+        return [_rgb_to_hex(rgbs[0])]
+    src = np.linspace(0, 1, len(colors))
+    tgt = np.linspace(0, 1, n)
+    interp = np.stack(
+        [np.interp(tgt, src, rgbs[:, c]) for c in range(3)], axis=1
+    )
+    return [_rgb_to_hex(rgb) for rgb in interp]
+
+
+def lighten(color: str, amount: float = 0.2) -> str:
+    """Lighten an HSV-based color by `amount` (0..1). Analogue of colorspace::lighten."""
+    import colorsys
+
+    r, g, b = _hex_to_rgb(color)
+    h, l, s = colorsys.rgb_to_hls(r, g, b)
+    l = l + amount * (1 - l)
+    r, g, b = colorsys.hls_to_rgb(h, l, s)
+    return _rgb_to_hex((r, g, b))
+
+
+def darken(color: str, amount: float = 0.2) -> str:
+    """Darken color by `amount` (0..1). Analogue of colorspace::darken."""
+    import colorsys
+
+    r, g, b = _hex_to_rgb(color)
+    h, l, s = colorsys.rgb_to_hls(r, g, b)
+    l = l * (1 - amount)
+    r, g, b = colorsys.hls_to_rgb(h, l, s)
+    return _rgb_to_hex((r, g, b))
+
+
+def desaturate(color: str, amount: float = 0.16) -> str:
+    """Reduce saturation. Analogue of colorspace::desaturate."""
+    import colorsys
+
+    r, g, b = _hex_to_rgb(color)
+    h, l, s = colorsys.rgb_to_hls(r, g, b)
+    s = max(0.0, s * (1 - amount))
+    r, g, b = colorsys.hls_to_rgb(h, l, s)
+    return _rgb_to_hex((r, g, b))
+
+
+def save_to_pdf(fig, pdf_file: Optional[str]) -> None:
+    """Save a matplotlib Figure to PDF if `pdf_file` is provided.
+
+    Used by every plotting function when the user passes ``pdf_file=...``.
+    Uses ``bbox_inches="tight"`` so that legends placed outside the axes
+    are included and not clipped.
+    """
+    if pdf_file is None:
+        return
+    if fig is None:
+        import matplotlib.pyplot as plt
+
+        fig = plt.gcf()
+    fig.savefig(pdf_file, format="pdf", bbox_inches="tight")