Skip to content

Homework 2

Homework 2: Parameter Parcheesi

Deadline: Friday, Feb 13th 2026 11:59pm

Background

This homework will ask you to to explore the impact of varying steps in the scRNA-seq analysis pipeline (as applied in HW1) on your estimate of cell type composition.

Data

Some data is the same as in HW1 but there are additional samples, too.

Data

Sample 1

This is the same Sample 1 as we used in Lecture 4. Here are some more details:

  • 10k PBMCs from a Healthy Donor (v3 chemistry) Single Cell Gene Expression Dataset by Cell Ranger 3.0.0
  • Peripheral blood mononuclear cells (PBMCs) from a healthy donor (the same cells were used to generate pbmc_1k_v2, pbmc_10k_v3). PBMCs are primary cells with relatively small amounts of RNA (~1pg RNA/cell).
  • 11,769 cells detected by CellRanger
  • Sequenced on Illumina NovaSeq with approximately 54,000 reads per cell
  • 10X Genomics web page for this data
  • pbmc_10k_v3_filtered_feature_bc_matrix.h5

Sample 2

  • 20k Human PBMCs, 3’ HT v3.1, Chromium X
  • Sourced from a healthy female donor
  • 23,837 cells
  • 35,000 reads per cell
  • link to downlad

Sample 3

  • Single Cell 3' v2
  • 2,909 cells
  • 115,061 reads per cell
  • healthy donor
  • download

Your mission

Analyze the provided data to address the following questions:

  1. You are curious about some of the steps in the processing pipeline. While some seem reasonable, others seem highly arbitrary. How does altering the pipeline impact cell composition estimates? If so, in what ways? Only use Sample 1 data here.
    • Specifically, determine how each of the following changes (independently) impacts the estimated cell composition:
      • Don't filter cells based on MT content
      • Don't filter cells based on total counts
      • Don't normalize expression (so skip both sc.pp.normalize_total() and sc.pp.log1p()
      • Don't use PCs for kNN graph
      • Test extreme values of n_pcs (e.g. n_pcs = 3, n_pcs = 50) when computing kNN graph
      • Test extreme values of n_neighbors when computing kNN graph
      • In what ways do these parameter adjustments alter the cell type gene signatures that you defined?
    • Deliverables:
      • Efficient visualisation of the cell type proportions across pipeline variant. Ideally the students will use their barplot function and pass it the adatas that result from annotation each of the outcomes of each variation to the pipeline.
      • Written summary detailing effect of each change to the pipeline and an explanation (or speculation) for this result.
  2. To what extent do cell type compositions as estimated by scRNA-seq vary between samples? From your work on qustion 1 above, you have arrived at an opinion as to the best pipeline. (by 'pipeline' here we mean the set of processing steps and parameter choices that lead to final clustering results). Apply that to pipeline the Samples 2 and 3. Compare all three samples to the cell type proportion estimates from the literature (see HW1). Do you think that the variation that you observe in the scRNAseq-based estimates primarily reflect real biological variation, or do they reflect technical variation due to the vicissitudes of scRNAseq data generation and analysis? Justify your answer.
    • Deliverables:
      • Stacked barplots showing cell type compositions across samples and literature
      • Written answer to questions about source of variation plus justification.
  3. Adult stem cells are typically not as numerous as their downstream progeny. This makes you wonder about the factors that impact scRNA-seq's sensitivty for detecting rare cell populations. Some factors include read depth (the number of UMIs per cell), size of the 'rare' population, the degree of distinctness of its transcriptome, and the analysis pipeline. Use existing Scanpy functions to explore the impact of read depth, population size, and analysis pipeline steps/parameters on the power to detect rare populations.
    • Deliverables:
      • Written report including the strategy (including what cell type(s) you treated as 'rare' population'), rationale, and general lessons derived from the analyses with relevant figures. What is the most important factor? What is the minimum population size (expressed as fraction of total number of cells in sample) that yields a detectable distinct cluster? How does this vary with read depth? Which, if any, of the pipeline parameters substantially impact power to detect a rare population?

ALL SUBMISSIONS MUST...

... include all code necessary to replicate the results, and must adhere to posted guidelines for code documentation, cogent written reports, and meaningful figures.