slug = "human-colon"
release = "2026-10-28"
use_gpu = True
out_dir = "bundles"
smoke = 0
primers_ref = "feat-primers"
cohorts_path = None
catalog_path = None
representatives = 1
dev_note = ""

# Parameters
slug = "human-pancreatic-cancer"
release = "2026-11-01"
out_dir = "bundles"
primers_ref = "4d7af340849c00077f9d21c5b54af9e689a4a070"
dev_note = ""

pip install "sddb-primers @ git+https://github.com/theislab/spatialdata-db-client@4d7af340849c00077f9d21c5b54af9e689a4a070#subdirectory=primers"

import inspect
import json
import platform
import time

import anndata as ad
import numpy as np
import pandas as pd
import scanpy as sc
import spatialdata_plot  # noqa: F401  (registers the .pl accessor)
import zarr
from primers.cohort import hd_table, load_cohort, smoke_spec
from primers.embed import CAVEAT, MAX_FLOOR_LOSS, MIN_SHARED_GENES, N_HVG, N_NEIGHBORS, N_PCS, SEED, embed_views
from primers.export import export_bundle, mixing_by_view, with_qc_status
from primers.markers import load_markers, panel_id
from primers.qc import MIN_GENES_FOR_QC, apply_qc, load_thresholds, qc_floors
from spatialdata.transformations import get_transformation

import sddb
from sddb.remote import storage_options

start = time.time()
out_dir = Path(out_dir)
if use_gpu:
    import cupy as cp
    import rapids_singlecell as rsc
    import rmm
    from rmm.allocators.cupy import rmm_cupy_allocator

    rmm.reinitialize(managed_memory=True)
    cp.cuda.set_allocator(rmm_cupy_allocator)


def show_call(func, what):
    """Show what a primers function does and link its source at ``primers_ref``."""
    _, line = inspect.getsourcelines(func)
    module = func.__module__.split(".")[-1]
    url = f"{REPO}/blob/{primers_ref}/primers/primers/{module}.py#L{line}"
    display(Markdown(f"`{func.__module__}.{func.__name__}`: {what} [Source at `{primers_ref}`]({url})."))


def show_table(df, digits=3):
    """Display ``df`` rounded, with missing values written as n/a, and long text columns in full."""
    shown = df.round(digits).astype(object)
    with pd.option_context("display.max_colwidth", None):
        display(shown.where(df.notna(), "n/a"))


catalog = sddb.Catalog.from_file(catalog_path) if catalog_path else sddb.Catalog(version=release)

if smoke:
    show_call(
        smoke_spec,
        "Keeps only the smallest kept member of each of a few papers, taken in turn from each family, for a "
        "quick end-to-end check.",
    )
    n_spec_excluded = len(spec.get("exclude_uids") or [])
    spec = smoke_spec(spec, catalog, smoke)
    print(
        f"Smoke run on {smoke} papers: {len(spec['exclude_uids']) - n_spec_excluded} more uids are excluded "
        f"({len(spec['exclude_uids'])} in all)."
    )
for key in ("slug", "query", "include_uids", "exclude_uids", "families", "marker_set", "rationale"):
    value = spec.get(key)
    if smoke and key == "exclude_uids":
        value = f"{len(value)} uids"
    print(f"{key}: {value}")

slug: human-pancreatic-cancer
query: {'organism': 'human', 'tissue': 'pancreas', 'disease': ['pancreatic ductal adenocarcinoma', 'pancreatic adenocarcinoma', 'exocrine pancreatic carcinoma', 'malignant pancreatic neoplasm', 'pancreatic cancer'], 'n_obs__gt': 0}
include_uids: None
exclude_uids: None
families: ['spot', 'imaging']
marker_set: pancreas
rationale: Pancreatic ductal adenocarcinoma on Visium (one large 107-section study plus smaller ones) and Xenium; one disease context.

show_call(
    load_cohort,
    "Runs the query on the catalog, adds the included uids, removes the excluded ones, and loads each kept "
    "member's table keyed by gene symbol, joined within its technology family.",
)
adatas, cohort = load_cohort(spec, catalog=catalog)

columns = ["uid", "title", "technology", "family", "paper", "paper_source", "license", "n_obs", "status"]
show_table(cohort[columns].rename(columns={"license": "licence"}))
show_table(cohort["status"].value_counts().rename_axis("status").reset_index(name="members"))
print(f"{int((cohort['status'] == 'kept').sum())} kept, {int((cohort['status'] != 'kept').sum())} dropped.")

178 kept, 1 dropped.

show_call(load_thresholds, "Reads the QC threshold groups from thresholds.yaml.")
thresholds = load_thresholds()
show_table(
    pd.DataFrame(
        [
            {
                "group": name,
                "technologies": ", ".join(g["technologies"]),
                "min_counts": g["min_counts"],
                "min_genes": g["min_genes"],
                "max_pct_mt": g.get("max_pct_mt"),
                "flag: min median counts": g["flag"]["min_median_counts"],
                "flag: min share kept": g["flag"]["min_share_kept"],
                "reason": g["reason"],
            }
            for name, g in thresholds.items()
        ]
    ).set_index("group")
)
print(f"MIN_GENES_FOR_QC = {MIN_GENES_FOR_QC}: QC, and every view, needs at least this many genes.")

MIN_GENES_FOR_QC = 50: QC, and every view, needs at least this many genes.

show_call(
    apply_qc,
    "Leaves out control probes, computes counts, genes and the mitochondrial share per observation, keeps "
    "the observations that pass their technology's filters, and describes each dataset.",
)
qcs = []
for fam in list(adatas):
    adatas[fam], q = apply_qc(adatas[fam], thresholds)
    qcs.append(q.assign(family=fam))
qc = pd.concat(qcs, ignore_index=True)
show_table(qc)

MAX_VIOLIN_DATASETS = 40
for fam, adata in adatas.items():
    obs = adata.obs[["dataset_uid", "total_counts", "n_genes"]].copy()
    sizes = obs["dataset_uid"].value_counts()
    if len(sizes) > MAX_VIOLIN_DATASETS:
        print(f"{fam}: {len(sizes)} datasets, so the {MAX_VIOLIN_DATASETS} largest are shown.")
        obs = obs[obs["dataset_uid"].isin(sizes.index[:MAX_VIOLIN_DATASETS])]
    obs["dataset_uid"] = obs["dataset_uid"].astype(str).astype("category")
    display(Markdown(f"Counts and genes per observation after QC, family `{fam}`."))
    sc.pl.violin(ad.AnnData(obs=obs), ["total_counts", "n_genes"], groupby="dataset_uid", rotation=90, stripplot=False)

spot: 169 datasets, so the 40 largest are shown.

show_call(load_markers, "Reads the marker sets from markers.yaml, each with its citation and fallbacks.")
marker_set = load_markers()[spec["marker_set"]]
print(f"Marker set {spec['marker_set']!r}, cited from {', '.join(marker_set['citation'])}.")
show_table(
    pd.DataFrame({"markers": {c: ", ".join(g) for c, g in marker_set["compartments"].items()}}).rename_axis(
        "compartment"
    )
)
kept_after_qc = set(qc.loc[qc["status"] == "kept", "dataset_uid"].astype(str))
technology_of = cohort.set_index("uid")["technology"].astype(str)
panel_of = {
    uid: panel_id(technology_of[uid], genes)
    for adata in adatas.values()
    for uid, genes in adata.uns.get("member_genes", {}).items()
    if uid in kept_after_qc
}
used = [
    {
        "panel": pid,
        "compartment": comp,
        "genes": "not measured" if fb.get("not_measured") else ", ".join(fb.get("genes") or []),
        "note": fb.get("note", ""),
    }
    for pid, entries in (marker_set.get("fallback") or {}).items()
    if pid in set(panel_of.values())
    for comp, fb in entries.items()
]
if used:
    show_table(pd.DataFrame(used))
else:
    print("No fallback of this marker set applies to the panels of this cohort.")

Marker set 'pancreas', cited from 10.1038/s41422-019-0195-y.

show_call(with_qc_status, "Marks the members that have no cells left after QC as dropped.")
kept_members = with_qc_status(cohort, qc).query("status == 'kept'")
ENSEMBL = r"ENS[A-Z]*G\d{6}"
SYMBOL_COLUMNS = ("symbol", "gene_symbols", "gene_symbol", "gene_name", "feature_name")


def colour_gene(table, panel):
    """(var name, symbol) of the first gene of the first compartment the table measures.

    A compartment's genes are its markers, then the fallback genes for this section's own panel.
    """
    names = pd.Index(table.var_names.astype(str))
    symbols = names
    if names.str.match(ENSEMBL).mean() >= 0.5:
        column = next((c for c in SYMBOL_COLUMNS if c in table.var), None)
        if column is None:
            return None, None
        symbols = pd.Index(table.var[column].astype(str))
    by_symbol = {s.upper(): (n, s) for n, s in zip(names, symbols, strict=True)}
    fallback = (marker_set.get("fallback") or {}).get(panel) or {}
    for comp, genes in marker_set["compartments"].items():
        for gene in [*genes, *((fallback.get(comp) or {}).get("genes") or [])]:
            if gene.upper() in by_symbol:
                return by_symbol[gene.upper()]
    return None, None


def table_region(zarr_url, table_name):
    """The shapes element a table annotates, read from the table's zarr attributes without loading it."""
    group = zarr.open_group(zarr_url, mode="r", storage_options=storage_options(zarr_url) or None)
    attrs = dict(group[f"tables/{table_name}"].attrs)
    if "region" not in attrs and f"tables/{table_name}/uns/spatialdata_attrs" in group:
        attrs = ad.io.read_elem(group[f"tables/{table_name}/uns/spatialdata_attrs"])
    region = attrs.get("region")
    return region if isinstance(region, str) or region is None else (list(region) or [None])[0]


def representative_uids(members):
    """The ``representatives`` kept sections closest to the median number of observations."""
    n_obs = members.set_index("uid")["n_obs"].astype(float)
    return list((n_obs - n_obs.median()).abs().sort_values(kind="stable").index[:representatives])

for tech, members in kept_members.groupby("technology"):
    for uid in representative_uids(members):
        ds = catalog.dataset(uid)
        stored = ds.elements()
        table_name = hd_table(stored) if tech == "VisiumHD" else (getattr(ds, "default_table", None) or "table")
        region = table_region(ds.zarr_url, table_name)
        if region is None or f"shapes/{region}" not in stored:
            display(Markdown(f"{tech}, section `{uid}`: the table annotates no shapes element, so no plot."))
            continue
        # a small tissue image when there is one: lowres, else hires (never the full-resolution image)
        images = [k.split("/", 1)[1] for k in stored if k.startswith("images/")]
        lowres = next((k for k in images if "lowres" in k), None) or next((k for k in images if "hires" in k), None)
        wanted = [f"tables/{table_name}", f"shapes/{region}"] + ([f"images/{lowres}"] if lowres else [])
        sdata = ds.to_sdata(elements=wanted)
        gene, symbol = colour_gene(sdata[table_name], panel_of.get(uid))
        section = members.set_index("uid").loc[uid, "title"]
        title = f"{tech}, {section}: {symbol}" if gene else f"{tech}, {section}"
        display(Markdown(f"{tech}, section `{uid}` ({int(members.set_index('uid').loc[uid, 'n_obs'])} observations)."))
        systems = list(get_transformation(sdata.shapes[region], get_all=True))
        system = "global" if "global" in systems else systems[0]
        # spatialdata-plot 0.4.2 rasterises shapes with datashader in "global" only; elsewhere use matplotlib
        method = None if system == "global" else "matplotlib"
        # fmt: off
        if lowres:
            (
                sdata
                .pl.render_images(lowres)
                .pl.render_shapes(region, color=gene, table_name=table_name, method=method, colorbar_params={"label": symbol})
                .pl.show(
                    coordinate_systems=system,
                    title=title,
                )
            )
        else:
            (
                sdata
                .pl.render_shapes(region, color=gene, table_name=table_name, method=method, colorbar_params={"label": symbol})
                .pl.show(
                    coordinate_systems=system,
                    title=title,
                )
            )
        # fmt: on

2026-11-04 04:15:46 | [INFO] root_attr: version

2026-11-04 04:15:46 | [INFO] root_attr: multiscales

2026-11-04 04:15:46 | [INFO] root_attr: omero

2026-11-04 04:15:46 | [INFO] datasets [{'path': 's0', 'coordinateTransformations': [{'type': 'scale', 'scale': [1.0, 1.0, 1.0]}, {'type': 'translation', 'translation': [0.0, 0.0, 0.0]}]}]

2026-11-04 04:15:46 | [INFO] resolution: s0

2026-11-04 04:15:46 | [INFO]  - shape ('c', 'y', 'x') = (3, 1806, 2000)

2026-11-04 04:15:46 | [INFO]  - chunks =  ['3', '1024 (+ 782)', '1024 (+ 976)']

2026-11-04 04:15:46 | [INFO]  - dtype = uint8

2026-11-04 04:15:46 | [INFO] root_attr: version

2026-11-04 04:15:46 | [INFO] root_attr: multiscales

show_call(
    qc_floors, "Gives each technology's count floor, the min_counts of its QC group, for gate 3 and the joint view."
)
floors = qc_floors(thresholds)
show_table(
    pd.DataFrame(
        {
            "value": [str(x) for x in (MIN_SHARED_GENES, MAX_FLOOR_LOSS, N_HVG, N_PCS, N_NEIGHBORS, SEED)],
            "meaning": [
                "genes a view needs before a family is split per technology, and the joint view's gate 1",
                "largest share of a technology's observations that may fall below its floor on the shared genes (gate 3)",
                "highly variable genes kept when a view has more genes",
                "principal components",
                "neighbours per observation",
                "random seed",
            ],
        },
        index=["MIN_SHARED_GENES", "MAX_FLOOR_LOSS", "N_HVG", "N_PCS", "N_NEIGHBORS", "SEED"],
    )
)
show_table(pd.Series(floors, name="count floor").rename_axis("technology").to_frame())

show_call(
    embed_views,
    "Builds one view per family (or per technology when a family shares too few genes), plus the joint view "
    "when the gates pass, and embeds each with PCA, UMAP and Harmony on the section.",
)
views = embed_views(adatas, marker_set, floor=floors, use_gpu=use_gpu)
for name, v in views.items():
    p = v.uns["primer"]
    print(f"{name}: {p['n_obs']} observations, {p['n_sections']} sections; {p['genes']['rule']}. {p['note']}".strip())
first = next(iter(views.values())).uns["primer"]
for uid, reason in first["members_dropped_panel"].items():
    print(f"member {uid} {reason}")
for c in {tuple(c["uids"]): c for v in views.values() for c in v.uns["primer"]["panel_costs"]}.values():
    print(f"{c['technology']} panel of {', '.join(c['uids'])}: {c['reason']} shared genes")
for name, reason in first["views_skipped"].items():
    print(f"no view {name}: {reason}")
gates = next((v.uns["primer"]["gates"] for v in views.values() if "gates" in v.uns["primer"]), None)
if gates is not None:
    print(f"Joint view: {'yes' if gates['joint'] else 'no'}. {gates['reason']}".strip())
    for drop in gates["drops"]:
        print(f"  left out {drop['technology']} ({drop['step']}): {drop['reason']}")

spot: 524995 observations, 169 sections; 2000 highly variable genes of the 17242 measured, expressed genes (seurat_v3 on raw counts, without batch, because the per-section fit failed); its 5 gene panels have 17,943 to 36,601 genes each and share 17,242.
imaging_cell: 1220850 observations, 9 sections; all 239 measured, expressed genes of the view; its 3 gene panels have 377 to 5,101 genes each and share 239.
Joint view: no. fewer than two families keep a technology
  left out Visium (gate3): loses 92% of its observations below 500 counts on the 229 shared genes (limit 50%)

for name, v in views.items():
    colours = [c for c in ("paper", "technology", "tissue") if c in v.obs and v.obs[c].nunique() > 1]
    display(Markdown(f"View `{name}`, before Harmony."))
    sc.pl.umap(v, color=colours or None, ncols=1)
    if "X_umap_harmony" in v.obsm:
        display(Markdown(f"View `{name}`, after Harmony on the section."))
        sc.pl.embedding(v, basis="X_umap_harmony", color=colours or None, ncols=1)
    else:
        display(Markdown(f"View `{name}` has one section, so it has no Harmony embedding."))

show_call(
    mixing_by_view,
    "Measures how evenly the batches mix among each observation's nearest neighbours, per view, before and "
    "after Harmony; the bundle's manifest records the same numbers.",
)
mixing = mixing_by_view(views)
rows, definitions = [], {}
for name, by_batch in mixing.items():
    for label, by_basis in by_batch.items():
        for basis, m in by_basis.items():
            for stat in ("entropy", "other_share"):
                definitions.setdefault(stat, m[stat]["definition"].replace(m["batch_key"], "batch"))
                rows.append(
                    {
                        "view": name,
                        "batch": label,
                        "embedding": "after Harmony" if basis == "X_pca_harmony" else "before Harmony",
                        "statistic": stat,
                        "mean": m[stat]["mean"],
                        "random reference": m[stat]["random"],
                        "limit from proportions": m[stat]["random_limit"],
                        "note": m["note"],
                    }
                )
show_table(pd.DataFrame(rows))
for stat, text in definitions.items():
    print(f"{stat}: {text}")
print(CAVEAT)

entropy: How evenly the batch labels are spread among each observation's 30 nearest neighbours, divided by log(min(30, number of labels)): 0 when all neighbours share one label, 1 when the labels are as evenly spread as the neighbours allow.
other_share: The share of each observation's 30 nearest neighbours that carry another batch.
Well mixed is not the same as biologically right: a correction can also mix cells that differ for a real reason, such as tissue or disease.

fallback_genes = {}
for entries in (marker_set.get("fallback") or {}).values():
    for comp, fb in (entries or {}).items():
        fallback_genes.setdefault(comp, set()).update(str(g).upper() for g in (fb or {}).get("genes") or [])
compartment_genes = {
    comp: {str(g).upper() for g in genes} | fallback_genes.get(comp, set())
    for comp, genes in marker_set["compartments"].items()
}
readings = {}
for name, v in views.items():
    key = "nn_harmony" if "nn_harmony_distances" in v.obsp else "nn_pca"
    leiden = {"resolution": 0.5, "random_state": SEED, "neighbors_key": key, "key_added": "leiden"}
    if use_gpu:  # rapids-singlecell's n_iterations counts aggregation levels, so its default is kept
        rsc.tl.leiden(v, **leiden)
    else:
        sc.tl.leiden(v, flavor="igraph", n_iterations=2, **leiden)
    expr = v.obsm["marker_expr"]
    m = ad.AnnData(
        X=expr.to_numpy(dtype=np.float32),
        obs=v.obs[["leiden"]].copy(),
        var=pd.DataFrame(index=expr.columns.astype(str)),
    )
    display(Markdown(f"View `{name}`: {m.obs['leiden'].nunique()} clusters on `{key}`, {m.n_vars} marker genes."))
    if m.n_vars == 0 or m.obs["leiden"].nunique() < 2:
        display(Markdown("This view has fewer than two clusters or no marker gene, so there is no dotplot."))
        continue
    by_compartment = {
        comp: [g for g in expr.columns if g.upper() in genes] for comp, genes in compartment_genes.items()
    }
    sc.pl.dotplot(m, {k: g for k, g in by_compartment.items() if g}, groupby="leiden", standard_scale="var")
    scores = (
        pd.DataFrame(
            {
                comp: expr[[g for g in expr.columns if g.upper() in genes]].mean(axis=1)
                for comp, genes in compartment_genes.items()
                if any(g.upper() in genes for g in expr.columns)
            }
        )
        .groupby(m.obs["leiden"].to_numpy())
        .mean()
    )
    ranked = scores.apply(lambda r: r.sort_values(ascending=False), axis=1)
    table = pd.DataFrame(
        {
            "observations": m.obs["leiden"].value_counts().reindex(scores.index),
            "top compartment": scores.idxmax(axis=1),
            "top mean": scores.max(axis=1),
            "runner-up": [r.dropna().index[1] if r.notna().sum() > 1 else "" for _, r in ranked.iterrows()],
            "runner-up mean": [r.dropna().iloc[1] if r.notna().sum() > 1 else np.nan for _, r in ranked.iterrows()],
        }
    ).rename_axis("cluster")
    readings[name] = table
    show_table(table)

for name, table in readings.items():
    lines = [f"View `{name}`:"]
    for cluster, r in table.iterrows():
        text = (
            f"- Cluster {cluster} ({int(r['observations'])} observations) has its highest mean marker expression "
            f"for {r['top compartment']} ({r['top mean']:.2f})"
        )
        if r["runner-up"]:
            text += f", followed by {r['runner-up']} ({r['runner-up mean']:.2f})"
        lines.append(text + ".")
    display(Markdown("\n".join(lines)))

show_call(
    export_bundle,
    "Writes the manifest, umap.parquet, qc.parquet, cohort.parquet and embedding.h5ad for this cohort. A cohort whose embedding is over 1 GiB gets parts of whole datasets instead (embedding-1.h5ad, embedding-2.h5ad, ...), and anndata.concat over the parts gives the whole cohort.",
)
bundle = export_bundle(
    views,
    cohort,
    qc,
    spec,
    out_dir,
    release,
    runtime_s=time.time() - start,
    mixing=mixing,
    primers_ref=primers_ref,
)
for f in sorted(bundle.iterdir()):
    print(f"{f.relative_to(out_dir)}  {f.stat().st_size} bytes")
manifest = json.loads((bundle / "manifest.json").read_text())
for key in (
    "slug",
    "release",
    "primers_ref",
    "n_datasets",
    "n_papers",
    "technologies",
    "n_cells_before",
    "n_cells_after",
    "n_flagged",
    "path",
    "hardware",
    "runtime_s",
):
    print(f"{key}: {manifest[key]}")

ImplicitModificationWarning: Trying to modify attribute `.obs` of view, initializing view as actual.

ImplicitModificationWarning: Trying to modify attribute `.obs` of view, initializing view as actual.

ImplicitModificationWarning: Trying to modify attribute `.obs` of view, initializing view as actual.

ImplicitModificationWarning: Trying to modify attribute `.obs` of view, initializing view as actual.

ImplicitModificationWarning: Trying to modify attribute `.obs` of view, initializing view as actual.

ImplicitModificationWarning: Trying to modify attribute `.obs` of view, initializing view as actual.

ImplicitModificationWarning: Trying to modify attribute `.obs` of view, initializing view as actual.

ImplicitModificationWarning: Trying to modify attribute `.obs` of view, initializing view as actual.

ImplicitModificationWarning: Trying to modify attribute `.obs` of view, initializing view as actual.

ImplicitModificationWarning: Trying to modify attribute `.obs` of view, initializing view as actual.

command = (
    f"papermill primers/notebooks/primer.ipynb primer.ipynb -p slug {slug} -p release {release} "
    f"-p primers_ref {primers_ref} -p use_gpu {use_gpu}" + (f" -p smoke {smoke}" if smoke else "")
)
versions = "\n".join(f"  - {k} {v}" for k, v in manifest["versions"].items())
display(
    Markdown(
        f"- Catalog release: `{release}`\n- primers reference: `{primers_ref}`\n- Versions:\n{versions}\n"
        f"- Hardware: {manifest['hardware']} ({platform.system()})\n- Path: {manifest['path']}\n\n"
        "To run this primer again:"
    )
)
show_command(command)

papermill primers/notebooks/primer.ipynb primer.ipynb -p slug human-pancreatic-cancer -p release 2026-11-01 -p primers_ref 4d7af340849c00077f9d21c5b54af9e689a4a070 -p use_gpu True

	uid	title	technology	family	paper	paper_source	licence	n_obs	status
0	hODmVoyNFahbWd1X0000	GSM8641107	Visium	spot	10.1016/j.celrep.2025.116827	doi	ncbi-geo-policy	3571	kept
1	hD3XfYwTQkHOJJ3D0000	GSM8641108	Visium	spot	10.1016/j.celrep.2025.116827	doi	ncbi-geo-policy	3996	kept
2	L2p8xQ6uDeIrtam50000	GSM8641074	Visium	spot	10.1016/j.celrep.2025.116827	doi	ncbi-geo-policy	2691	kept
3	lnP5VqQCyDjNfHHl0000	human-pancreas-preview-data-xenium-human-multi-tissue-and-cancer-panel-1-standard	Xenium	imaging_cell	10x Genomics demo datasets	10x_demo	CC-BY-4.0	103901	kept
4	GdkN2quGouCTy2AT0000	GSM9051173	Visium	spot	10.1016/j.canlet.2026.218348	doi	ncbi-geo-policy	4611	kept
...	...	...	...	...	...	...	...	...	...
174	7VBtB072n0akCvCa0000	GSM8641056	Visium	spot	10.1016/j.celrep.2025.116827	doi	ncbi-geo-policy	2247	kept
175	uX6VZ6cou3rvFSnM0000	GSM8641039	Visium	spot	10.1016/j.celrep.2025.116827	doi	ncbi-geo-policy	3207	kept
176	m5ngpxS7i9eHWZdX0000	GSM8443452	Visium	spot	geo-gse274103	geo_series	ncbi-geo-policy	4983	kept
177	1SOfUUKmmUMNFwj90000	GSM8443451	Visium	spot	geo-gse274103	geo_series	ncbi-geo-policy	4134	kept
178	kcsfSI6N1ESWcR8P0000	GSM8452875	Visium	spot	geo-gse274557	geo_series	ncbi-geo-policy	3177	kept

	markers
compartment
ductal	KRT19, KRT7, SOX9, MUC1
acinar	PRSS1, CPA1, CELA3A
endocrine	INS, GCG, CHGA
fibroblast	COL1A1, LUM, DCN
stellate_pericyte	RGS5, PDGFRB
endothelial	PECAM1, VWF, CDH5
t_cell	CD3E, CD3D, CD2
b_plasma	MS4A1, CD79A, MZB1
myeloid	CD68, CD14, LYZ

	panel	compartment	genes	note
0	Xenium-377-9193d582	acinar	AQP8	the only acinar gene on this panel
1	Xenium-377-9193d582	fibroblast	FBLN1, SFRP2	fibroblast genes on this panel
2	Xenium-5101-5714379c	acinar	CTRL, SERPINI2, CUZD1	acinar genes on this panel
3	Xenium-5101-5714379c	fibroblast	POSTN, FAP, CXCL12	fibroblast and CAF genes on this panel

	count floor
technology
Visium	500.0
SpatialTranscriptomics	500.0
VisiumHD	200.0
Xenium	10.0
Merscope	10.0
MERFISH	10.0

	observations	top compartment	top mean	runner-up	runner-up mean
cluster
0	104429	fibroblast	1.812	b_plasma	0.071
1	40508	fibroblast	1.772	b_plasma	0.143
10	30947	acinar	3.498	b_plasma	0.032
11	396	fibroblast	1.772	b_plasma	0.027
12	1553	stellate_pericyte	0.54	b_plasma	0.001
13	646	fibroblast	1.049	b_plasma	0.027
2	119502	ductal	1.697	b_plasma	0.044
3	94649	fibroblast	1.775	b_plasma	0.213
4	49650	acinar	1.754	b_plasma	0.114
5	35581	fibroblast	1.666	b_plasma	0.592
6	25931	endocrine	3.198	b_plasma	0.186
7	2787	myeloid	1.023	b_plasma	0.27
8	16666	fibroblast	2.42	b_plasma	0.055
9	1750	myeloid	1.236	b_plasma	0.003

Cohort primer¶

Human pancreatic cancer¶

Setup¶

Query and cohort¶

Quality control¶

Marker set¶

Spatial overview¶

Embedding¶

Biology¶

Draft reading, to be reviewed by the Lead¶

Export and reproduce¶

	status	members
0	kept	178
1	dropped: no paper (no study_id, series or collection)	1

	technologies	min_counts	min_genes	max_pct_mt	flag: min median counts	flag: min share kept	reason
group
spot	Visium, SpatialTranscriptomics	500	250	30.0	1000	0.5	A Visium or ST spot holds about 1 to 10 cells, so a spot with under 500 UMIs or 250 genes is mostly background or tissue edge; solid tissues often reach 20 to 30% mitochondrial reads, so the cap is 30%.
binned_spot	VisiumHD	200	50	30.0	200	0.5	On the 5 Visium HD sections of oliveira-2024-colorectal, 16 um bins at 200 counts kept 74% of bins with the same marker signal as 8 um bins at Oliveira et al.'s 100 (Nat Genet 57:1512-1523), which kept 60%, and found T cells in all 5 sections against 3 (web tasks/2026-11-01-t094-hd-bin-size.md).
imaging_cell	Xenium, Merscope, MERFISH	10	3	n/a	30	0.5	Targeted panels give tens to hundreds of decoded transcripts per segmented cell, and a cell with under 10 transcripts or 3 genes cannot be given a type; the panels have no mitochondrial genes.

	dataset_uid	technology	n_before	n_after	median_counts	median_genes	pct_mt_median	flagged	flag_reason	status	family
0	0877Bxvk1AFUpoD90000	Visium	2261	2153	2411.0	1336.0	0.0	False		kept	spot
1	0OZKzzcWGDkeLfdg0000	Visium	3249	3238	6461.0	2789.0	0.0	False		kept	spot
2	0fNGoWEF8xCnfaN30000	Visium	3091	3091	13164.0	4966.0	0.0	False		kept	spot
3	0fpFB3kKbfLq5fNx0000	Visium	2141	2104	10415.0	4319.0	0.0	False		kept	spot
4	0gBwEBel4oHwTfbm0000	Visium	4261	4261	13239.0	4889.0	0.0	False		kept	spot
...	...	...	...	...	...	...	...	...	...	...	...
173	TCU9OYSnDceV2OiJ0000	Xenium	169765	162147	128.0	105.0	n/a	False		kept	imaging_cell
174	UFIoalxHktLSI5fo0000	Xenium	238408	224908	144.0	119.0	n/a	False		kept	imaging_cell
175	lnP5VqQCyDjNfHHl0000	Xenium	103901	103862	109.0	39.0	n/a	False		kept	imaging_cell
176	mzxQ72GZIk9tyOpx0000	Xenium	95675	64548	58.0	49.0	n/a	False		kept	imaging_cell
177	n9ZxbmGR8CCbUhgN0000	Xenium	133272	132229	215.0	167.0	n/a	False		kept	imaging_cell

	value	meaning
MIN_SHARED_GENES	200	genes a view needs before a family is split per technology, and the joint view's gate 1
MAX_FLOOR_LOSS	0.5	largest share of a technology's observations that may fall below its floor on the shared genes (gate 3)
N_HVG	2000	highly variable genes kept when a view has more genes
N_PCS	30	principal components
N_NEIGHBORS	30	neighbours per observation
SEED	0	random seed

	view	batch	embedding	statistic	mean	random reference	limit from proportions
0	spot	section	before Harmony	entropy	0.175	0.964	1.493
1	spot	section	before Harmony	other_share	0.265	0.96	0.994
2	spot	section	after Harmony	entropy	0.639	0.963	1.493
3	spot	section	after Harmony	other_share	0.768	0.96	0.994
4	spot	paper	before Harmony	entropy	0.036	0.474	0.538
5	spot	paper	before Harmony	other_share	0.042	0.52	0.537
6	spot	paper	after Harmony	entropy	0.263	0.473	0.538
7	spot	paper	after Harmony	other_share	0.336	0.519	0.537
8	imaging_cell	section	before Harmony	entropy	0.478	0.905	0.971
9	imaging_cell	section	before Harmony	other_share	0.561	0.848	0.875
10	imaging_cell	section	after Harmony	entropy	0.547	0.905	0.971
11	imaging_cell	section	after Harmony	other_share	0.589	0.848	0.875
12	imaging_cell	paper	before Harmony	entropy	0.117	0.608	0.639
13	imaging_cell	paper	before Harmony	other_share	0.078	0.385	0.397
14	imaging_cell	paper	after Harmony	entropy	0.169	0.606	0.639
15	imaging_cell	paper	after Harmony	other_share	0.136	0.385	0.397

	observations	top compartment	top mean	runner-up	runner-up mean
cluster
0	378689	stellate_pericyte	0.274	b_plasma	0.002
1	153040	t_cell	0.276	b_plasma	0.004
10	15930	stellate_pericyte	0.078	b_plasma	0.003
100	212	ductal	0.0	b_plasma	0.0
101	113	ductal	0.0	b_plasma	0.0
...	...	...	...	...	...
95	167	endothelial	0.002	b_plasma	0.0
96	80	endocrine	0.003	b_plasma	0.0
97	656	ductal	0.0	b_plasma	0.0
98	218	endothelial	2.624	b_plasma	0.0
99	298	endocrine	0.001	b_plasma	0.0