Question 1

How should cell filtering thresholds be chosen?

Accepted Answer

Look at the joint distribution of counts, detected features, and mitochondrial fraction before setting any cutoff, since a threshold sensible for one tissue removes real cells in another. Apply thresholds per sample rather than globally, report the number of cells removed at each step, and show the diagnostic plots the cutoffs came from.

Question 2

Which clustering resolution is correct?

Accepted Answer

None is. Resolution sets how finely the graph is partitioned, so run several values and examine how populations split as it rises. A cluster that survives across settings and carries distinct markers is more defensible than one that appears at a single value. State the chosen value and show the sensitivity.

Question 3

When does batch correction do more harm than good?

Accepted Answer

When the batches differ biologically as well as technically. Aggressive integration can merge cell states that genuinely differ between conditions, which destroys the comparison the study was designed to make. Check that well characterised populations remain distinct after correction, and compare results with and without integration before deciding.

Question 4

Can trajectory inference be trusted?

Accepted Answer

Only with corroboration. These methods impose a continuous ordering whether or not the underlying process is continuous, so a smooth path is not evidence of one. Support the ordering with independent markers, RNA velocity, or a time course sampled at known stages, and state which cells were used as the root.

Question 5

How many cells and how many samples are needed?

Accepted Answer

Biological replication matters more than cell number. Many cells from one animal give precision about that animal and nothing about the population, so plan several independent samples per condition even if each yields fewer cells. Where a rare population is the target, estimate its expected frequency and size the run accordingly.

Question 6

How should differential expression between conditions be tested?

Accepted Answer

Aggregate counts to pseudobulk per sample and test at the sample level. Treating individual cells as replicates inflates significance badly, because cells from one sample are not independent. Cell level tests remain useful for marker identification within a dataset, but comparisons across conditions belong at the level of the biological unit.

Bioinformatics

Bioinformatics

Research Frontiers in Single Cell RNA Sequencing Analysis