cellxgene-census-summary-and-mean

CELLxGENE Census summary and raw-matrix mean

For every task, apply Caveman. Before broad repository navigation, use CocoIndex Code; for web-page research, use Crawl4AI. Read and follow the repository-root docs/agent-tools.md and its linked skills, and retain the required execution evidence.

Pin the Census build and record the chosen organism, measurement, filters and layer. This focused repair covers summary counts and one descriptive matrix mean; it does not validate other workflows in the source skill.

Read a summary count

For an open Census collection census, the summary is a table of label and value rows. Convert the selected count to an integer:

summary = census["census_info"]["summary"].read().concat().to_pandas()
summary_values = summary.set_index("label")["value"]
total_cell_count = int(summary_values.loc["total_cell_count"])

Describe the selected raw matrix

For an open RNA ExperimentAxisQuery named query, this computes one overall mean across all selected cell-by-feature positions, including implicit zero fill values. It is not a per-gene mean. Reject an empty query rather than reporting a misleading zero:

import math
import numpy as np

n_positions = query.n_obs * query.n_vars
if not n_positions:
    raise ValueError("Cannot average an empty cell-by-feature selection")
total = math.fsum(
    float(np.sum(batch["soma_data"].to_numpy(), dtype=np.float64))
    for batch in query.X("raw").tables()
)
mean_expression = total / n_positions

Raw matrices use sparse zero fill: averaging only returned entries computes a nonzero-entry mean instead. This matrix-wide zero-filled mean is a descriptive choice, not automatically a biological expression estimate. Genes absent from source datasets can be represented as zero in the combined matrix. Use the feature dataset presence information and an appropriate measured-cell denominator for a per-gene analysis; do not silently treat unmeasured genes as measured zeros. When grouping per gene, map SOMA join IDs to query positions using the query indexer rather than treating join IDs as array offsets.

See the Census schema and the incremental computation tutorial.