Results of original DROP publication

[AEstebanMar]

The link to plots and results of the test (Geuvadis) dataset provided in the original DROP article (0.1038/s41596-020-00462-5) seems to be broken. Would you be so kind as to provide a working link?

I have contacted Professor Gagneur in regards to this matter. He suggested I post an issue here so that everyone may benefit from the answer.

Thank you in advance.

[vyepez88]

Hi Álvaro, Thanks for noticing and reporting this. At some point the faculty of Informatics changed domain. I added a working link to the Readme (currently in the dev branch only). Also here: The website containing the different reports of the Geuvadis demo dataset described in the paper can be found here. In the different Summary pages (e.g. https://cmm.in.tum.de/public/paper/drop_analysis/webDir/drop_analysis_index.html#Scripts_AberrantExpression_pipeline_OUTRIDER_Datasets.html), you can find the results tables that you can use to compare to your own.

[AEstebanMar]

Hello Vicente,

Thank you for your response, everything seems to be working now! I had a followup question.

Our research group would greatly appreciate some clarification as to what anomalies should be observed within the results (e.g which samples and genes were confirmed as aberrant in your original run, as described for the Kramer dataset in the original DROP article). As we understand it, there might be some slight variation in obtained results (specifically in aberrant expression due to how the OUTRIDER autoencoder works).

To clarify, we would like to know which samples and genes should always be identified as aberrant (be it by expression, splicing or MAE).

[AEstebanMar]

I noticed something rather strange. In the aberrant splicing module, your number of outliers vs sample rank graph looks like this:

Whereas mine looks like this:

I am using the config file linked in the Supplementary Information of the original article (Supplementary Data 2), and our local DROP version is 1.2.2. Is it not the config file I should be using, or is something else going wrong? I am not sure if these differences can be attributed to using a different DROP version, they seem too extreme.

[vyepez88]

Hi, can you share your config file here? If you use the Kremer dataset, the samples that should come up significant are the ones described in the paper, especially 2x TIMMDC1 expression and splicing and ALDH18A1 MAE.

[AEstebanMar]

Sure, here it is: config.txt

We are running DROP on the Geuvadis dataset, not Kremer.

[vyepez88]

I saw the discrepancy in the padj and deltaPsi cut-offs that you're using than the ones that were used for the DROP dataset. We recommend:

    padjCutoff: 0.1 # or 0.05
    ### FRASER1 configuration
    FRASER_version: "FRASER"
    deltaPsiCutoff : 0.3
    quantileForFiltering: 0.95
    ### For FRASER2, use the follwing parameters instead of the 3 lines above:
    # FRASER_version: "FRASER2"
    # deltaPsiCutoff : 0.1
    # quantileForFiltering: 0.75

Running with those cut-offs should give you similar results. Same for expression:

   padjCutoff: 0.05
    zScoreCutoff: 0

[AEstebanMar]

Thank you, I've fixed it in my config and re-run it. However, I have no "quantileForFiltering" field. I am running DROP version 1.2.2, was that field added in a more recent version? If so, would you mind providing the config file used for the DROP run shown in the article (DROP 0.9.2)? Also, could you confirm that the results linked in the original response to my issue correspond to that original DROP run?

[vyepez88]

Yes, the quantileForFiltering parameter was added in DROP 1.3.0, where we introduced FRASER2. I highly recommend you updating to FRASER2 as it provides more specific results than the original FRASER. You can find the config file for the paper here: https://www.nature.com/articles/s41596-020-00462-5#Sec37 (File 2) The results linked in my original response might not correspond to that original DROP run as DROP has been updated since. However, they should not be too different.