A great example of why bioinformaticians need to understand the biological context behind their data, rather than just follow a standard pipeline.

In practice, we use a combination of a dataset-dependent threshold (e.g., the 98th percentile) and a fixed cutoff (e.g., 10% ), taking the more stringent of the two for each sample.

We implemented this approach in our LinQ-View paper. BTW, The 98th percentile and 10% cutoff are just our defaults for PBMC data, and both are adjustable.

Interestingly, this also gives us a useful warning signal: if a sample ends up hitting that 10% ceiling, I take it as a reminder to double-check what’s going on.

And here’s one simple trick I always recommend: plot violin plots of nCount, nFeature, and percent.mt before and after QC.

It only takes a few seconds to glance at them, but it can save you from accidentally throwing away biologically meaningful cells.

Good QC isn’t just about removing low-quality cells. It’s also about knowing what you’re keeping — and what you might be losing.


Reposted source: Alessandro De Santis on mitochondrial QC thresholds.