I am working with public cell-free RNA (cfRNA) datasets that include ERCC spike-ins to quantify technical biases, such as GC content and transcript length bias. As a preliminary step, I am attempting to reproduce the Spearman correlation results reported in the original study to validate my pipeline. The paper reports high technical consistency with an ERCC Spearman correlation of ~0.95. However, using the same dataset (SRA accessions), my calculated values consistently fall between 0.40 and 0.60. The hyperlink for paper
I followed these steps:
- Generated a merged FASTA/GTF (GRCh38.fa + ERCC.fa and GENCODEv49 + ERCC.gtf) using the ERCC reference from this github repository
- Used a standard pipeline (fastp -> STAR -> featureCounts).
- I got results and calculated spearman correlation coefficient with the "concentration in Mix 1 (attomoles/ul) " or "concentration in Mix 2 (attomoles/ul)" columns of metadata from repository same as above.
The paper does not specify whether Mix 1 or Mix 2 was used. I tested both, and while one mix usually yields a slightly higher 0.5, neither comes close to 0.95.
I am seeking advice on the following:
- Is there "standard" pipeline to quantify the ERCC spike in public SRA data?
- Given the lack of dilution / Mix metdata, is there any way to evaluate the observed counts are appropriate relative to expected concentrations ?
- Furthermore, How do researchers typically measure GC bias and Length bias using ERCC spike in?
Any insights or suggestions would be greatly appreciated.
Thank you.
•
written
by
MJ Cho •
0 answers
No answers yet.
Log in to answer this question.
What correlation are you computing? The mixes have both fold change differences and large concentration differences. Low-concentration mixes will not appear correctly in the data.
In this paper, the author presents a histogram compiling the correlation values of expected ERCC transcripts versus the observed levels from each sample.
I attempted to reproduce these results, but I am consistently obtaining scores significantly lower than those reported in the paper.
While the study only uses 237 samples for its analysis, the publicly available dataset contains approximately 490 samples when including all replicates. It appears they narrowed the data down to those specific 237 samples through an some kind of QC process and they didn't clarify which samples are used for final analysis.
So....now I am looking for other dataset.
Ultimately, my objective is to determine whether the GC and length biases quantified via ERCC spike-ins in cfRNA correlate with the bias estimates derived from traditional computational methods, such as EDASeq.
Before you compute a correlation, what does a simple "sanity eye check" say?
Pick a row from the mix that has both, large abundance and large fold change. Does the expected fold change match?
A few of these will tell you immediately what to expect.