Metadata annotation issues

[WTX-to-be]

Hi, I have come across this wonderful integrated dataset and done several analyses such as DEG and survival. You really made great contributions to PCa dataset collection and integration. However, I have also found some mistakes, maybe I assume, in the integrated dataset. For example, the PSA variable in dataset tcga is annotated as at the time point of follow-up according to the cBioPortal raw metadata. However, in curatedPCaData the psa in the metadata is annotated as psa at diagnosis according to the user guide at https://bioconductor.org/packages/devel/data/experiment/vignettes/curatedPCaData/inst/doc/overview.html. You may check the PSA values and find most of the PSA values in tcga dataset are normally below 4. Really the metadata in different datasets are hard to integrate considering the sharply heterogeneously organized metadata information in different datasets. So I suggest re-check the organized data in the curatedPCaData, particularly metadata and their annotations. But in summary curatedPCaData is a wonderful integrated dataset for PCa bioinformatic analysis and thanks to your great efforts.

[Syksy]

Hello @WTX-to-be
thanks for continuing to check out the package,

I've been trying to replicate this issue, but have not been able to track down a mismatch - partly because our old parsing of metadata code is no longer supported by the data repositories:

> mycgds <- cgdsr::CGDS("http://www.cbioportal.org/")
> uncurated <- cgdsr::getClinicalData(mycgds, caseList = "prad_tcga_sequenced")
Error in processURL.CGDS(x, url) : Not Found (HTTP 404).

Looks like they're not offering this metadata any more via cBioPortal.
So it's a bit difficulty for me, i.e. trying to follow our previous steps that we did >3 years ago in lines 1187 onward in:
https://github.com/Syksy/curatedPCaData/blob/master/data-raw/download-clinical.R

It's true that the metadata from studies are very heterogeneous and not entirely comparable. We always used our best judgment and tried to read a consensus among the authors on how to best consolidate the differences between the cohorts, to make the metadata tables as comparable as possible straight off the bar. This may very well be case also for PSA measurements, as it's not always clearly reported which PSA was closest to the actual diagnosis - we did our best to try always use as close definition as possible

TCGA in particular is difficult in its metadata, because we've retrieved data from two sources; cBioPortal (above code), and then supplemented it through XenaBrowser's TCGA metadata fields available at https://tcga-xena-hub.s3.us-east-1.amazonaws.com/download/TCGA.PRAD.sampleMap%2FPRAD_clinicalMatrix
... the latter's code is on lines 1343 onwards in our above download-clinical.R file.

PSA by definition is tricky because it's a longitudinal marker. Conventionally, PCa is "confirmed" by a prostate biopsy (Gleason / Grade Group), and PSA usually precedes or follows it a bit. So "PSA at diagnosis" is a bit loose definition.

Could you provide me a few examples of where you have the "correct" PSA values and give me the code for fetching it, and show where the mismatches may be in our currently offered metadata?
Feel free to also point out if the interpretation is wrong in some other fields, but it'd be easier to track it down with concrete examples (correct answer and where it's from vs. what we currently offer in the package's metadata fields).

Thanks,
Daniel

[WTX-to-be]

@Syksy Hi, Daniel, thanks for your reply and attention.

Here is how I found metadata mis-match or mis-annotation in curatedPCaData. I initially wanted to explore the clinical and expression differences in high-risk PCa, defined by gleason score>=8, T stage>=3 or PSA at diagnosis>20mg/ml according to NCCN guidelines. When I grouped patients in tcga dataset in curatedPCaData, I found most of the PSA values are normally below 4ng/ml.

library(curatedPCaData)

maes=base::readRDS(file = "~/curatedPCaData.rds")

tcga=maes[["tcga"]]

metadata=as.data.frame(tcga@colData)

library(ggplot2)

ggplot(data = metadata)+
  geom_violin(mapping = aes(x = "PSA",
                            y = psa))+
  thisplot::theme_this(aspect.ratio = 1)+
  geom_jitter(mapping = aes(x = "PSA",
                            y = psa))+
  geom_hline(yintercept = 4,
             linetype = "dashed",
             color = "grey",
             size = 1)

You can see the PSA values in curatedPCaData are mostly below 4ng/ml and assumed to be post-radical therapy follow-up values. But in curatedPCaData user guideline at "https://bioconductor.org/packages/devel/data/experiment/vignettes/curatedPCaData/inst/doc/overview.html", psa is annonated as "Prostate specific antigen level (numeric) at diagnosis".

Regarding the TCGA raw metadata, you can access it in cBioPortal website online at "https://www.cbioportal.org/study/summary?id=prad_tcga" and click the download button as follow, or directly via link "https://datahub.assets.cbioportal.org/prad_tcga.tar.gz",

After you download it, unzip it and open "data_clinical_patient.txt" metadata file, you can find the "PSA_MOST_RECENT_RESULTS" and it is annotated as "Psa most recent results", or PSA value at follow-up after radical therapy I assume. Unfortunately, this tcga dataset, even in the raw metadata supplied by cBioPortal, does not seem to supply PSA information at diagnosis. And I guess you fetch this metadata version or something alike and wrongly annotated it as "Prostate specific antigen level (numeric) at diagnosis."

Yes I agree PSA is something tricky and needs longitudinal inspection, and I may not focus on this parameter in future analysis. And I found this issue by accident and reported this to you curatedPCaData developers. About other parameters or columns in your metadata, I have not noticed other issues so apparent as PSA values yet. I think more detailed re-check for more secret errors, if there is any, needs manual mapping between curatedPCaData data and raw data, including their values, NA, annotations, etc.

Still thanks to your contributions to PCa dataset integration!