Exercises and Answers

Exercises

1. Experimental units

Four mice per condition each contribute 3,000 cells. What is the sample size for a condition comparison? What does the large cell count improve?

2. Matrix orientation

A 10x Matrix Market header reports 20,000 rows and 5,000 columns. What do those usually represent? What orientation does sc.load() expose?

3. QC sensitivity

Run the teaching workflow with maximum mitochondrial percentages of 3.0, 5.0, and 10.0. Record retained cells, clusters, and ARI.

4. Resolution sensitivity

Use sc.sweep() across [0.2, 0.5, 0.8, 1.2]. Which result has the most clusters? Does more detail automatically mean more truth?

5. Annotation

For one real cluster, prepare a table of positive markers, negative markers, state programs, reference evidence, and uncertainty. Which label is the narrowest one the evidence supports?

6. Validation

Why can BioLang cluster 0 correspond to Scanpy cluster 2 while ARI remains 1.0?

7. Differential expression

Write a pseudobulk table design for six patients measured before and after treatment. What are the rows, columns, pairing variable, and primary contrast?

8. Clinical claim

Rewrite “Cluster 4 proves this patient will respond to therapy” as an accurate exploratory statement.

Answers

1

The independent sample size is four mice per condition. More cells improve the description of within-mouse heterogeneity and rare populations, not the number of independently treated mice.

2

Rows usually represent genes/features and columns represent barcodes/cells. sc.load() exposes a cells-by-genes matrix.

3

The exact numbers are empirical. The important answer is the recorded change and whether removed cells concentrate in a population. A threshold is not justified because it gives the preferred ARI.

4

Higher resolution generally creates more clusters. More clusters can split states or technical variation and does not automatically improve biological truth.

5

The answer must use multiple genes and conflicting evidence. unknown or a broad lineage is acceptable when evidence is insufficient.

6

Cluster IDs are arbitrary labels. ARI compares co-membership: which cells are grouped together.

7

Rows are genes and columns are patient-by-timepoint pseudobulk profiles within a specified cell type. Metadata contains patient and timepoint; the model uses patient pairing and tests the post-vs-pre contrast.

8

“In this exploratory dataset, a cell population labeled cluster 4 is associated with response and requires validation in independent patients before predictive or clinical interpretation.”