Home

<--

Answers to Polls

14_FirstSteps

• How many genes and cells does this dataset have?

• dim(pbmc.data)

• 32738 genes, 2700 cells

• How many genes are not expressed in any cell?

• rowSums(pbmc.data) == 0 %>% sum()

• 16104

• Which are the top 3 genes with the highest total count?

• rowSums(pbmc.data) %>% sort(decreasing = TRUE) %>% head(3)

• MALAT1, TMSB4X, B2M

• In cell "AAATTCGATTCTCA-1", how many reads map to gene "ACTB"?

• pbmc.data["ACTB","AAATTCGATTCTCA-1"]

• 10

• How many cells have less than 2000 counts?

• (colSums(pbmc.data) <= 2000) %>% sum()

• 1025

• What's the current number of cells after this step?

• dim(pbmc)

• 2638

• Which are the 3 most highly variable genes?

• VariableFeatures(pbmc) %>% head(3)

• LYZ, S100A9, PPBP

• What's the variance of the gene PYCARD?

• HVFInfo(pbmc)["PYCARD",]

• 5.05

• How many components should we choose to include?

• 7

• Which is the default value of the parameter K?

• k.param = 20

• How many clusters did we find? pbmc$seurat_clusters %>% unique() %>% length() 9

16_UMAP

• Is t-SNE affected by the seed?

• Yes. And there's always one seed set because the algorithm is stochastic (the S in SNE!).

• (perplexity constraint) Is nrow(X) the number of genes or cells?

• nrow(X) is the number of cells ("observations" in the data matrix are in rows.) Meanwhile, each column of the data matrix is a variable (gene).

• Is UMAP affected more than t-SNE by the seed?

• t-SNE is more sensitive. UMAP emphasis on global structure helps (a lot.)

• Would decreasing the number of PCs fed to the UMAP algorithm change our visualization? Would you say the results are 'better'?

• Yes, and it’s better with few PCs. That's why we look for a "sweet" spot with Elbow or JackStraw. The variability in each PC is a mixed proportion (not always 1:1) of relevant, biological variability or technical variability.

• Could you have the UMAP projection onto 3 axes instead of 2?

• All answers are correct, except for "no".

21_DE

• What would happen if we ran FindVariableFeatures after ScaleData?

• Nothing, the resulting list of genes would be the same. Data scaling is important for ML (e.g. dim reduc); the method(s) to finding variable features are more akin to 'classic' statistical modelling techniques (e.g. regression).

• What would happen if we skip ScaleData? Would PCA be affected?

• The resulting PCA has biases introduced by the absolute values of our count matrix. Genes with higher counts would rule the components, even if their variability is not big.

• Would it be any difference if we ran FindNeighbors and FindClusters but only after UMAP? Depends! If we wanted to FindNeighbors using reduction='umap'. The first option is incorrect in the context of single cell data and seurat processing pipeline.

• Given that genes used for clustering are the same genes tested for differential expression, Would you interpret the (adjusted) p-values without concerns? Yeah. There's a bit of chicken-and-egg problem... but that's just how it is.

• Which marker gene is the most expressed in cluster 1 when comparing it to cluster 2?

• What would happen if we used ident.1 = 2, and ident.2 = 1 instead?

• How would you extract the 'gene signature' of a given cluster? By DE testing with… Any test, against the remaining clusters

Day 3

• paste question here

Day 4

• paste question here

-->