Fail to reproduce Fig.5 results for human data

[jonathan-f]

Hello, First thanks for your impressive work!

I am facing issues in reproducing the results of Fig.5 of the paper for the human datasets using the STRING ground truth net, while I manage to reproduce the results for the mouse datasets.

1) I used the script generateExpInputs.py, getting the following statistics:

 hHep STRING:  #TFs: 409, #Genes: 656, #Edges: 15046, Density: 0.056
 hESC STRING:  #TFs: 343, #Genes: 517, #Edges: 8514, Density: 0.048

These densities are different from fig. 5, why?

2) I ran PIDC several times using BLRunner.py and then I used BLEvaluator.py to compute the early precision (EP), I always get:

 hHep STRING: EP=0.08, while it should be 0,105 according to Fig. 5 (multiplying the EPR by the net density from fig. 5=0.03). 
 hESC STRING: EP=0.074, which seems correct with the density indicated in Fig.5, but not with the one reported above.

Also for the other algorithms I get an EP smaller than that shown in Fig.5. Do you have an idea of the reason? Could you provide your rankedEdges.csv file for PIDC for the hHep dataset for comparison please?

Thank you Best wishes Jonathan

[adyprat]

Hi @jonathan-f,

Thank you for bringing this to our notice. For some strange reason, all the lines in STRING-network.csv are duplicated and the #Edges is 2x / the density is 2x (i.e., EPR is 0.5x) the one you see in Fig 5. While we make changes to generateExpInputs.py to compute the correct number of of edges; please use len(netDF.drop_duplicates()) instead of net.shape[0] to get the correct "#Edges".

Best, Aditya

[tmmurali]

@adyprat thanks for looking into this issue. Can we upload the correct version of the STRING network to prevent this issue for recurring again? I presume this file is on Zenodo.

[jonathan-f]

Dear @adyprat , thanks for the swift reply and for finding the error.

Best Jonathan

[JaneJiayiDong]

Hello, sorry for bothering. I am also facing some issues in reproducing the results of Fig.5 of the paper. I downloaded the data (BEELINE-data and Networks) from Zenodo and used the the generateExpInputs.py.

1. I used the expression data (mESC) and the network(Non-Specific-ChIP-seq-network.csv), and set other parameter as default. The mistake is as follows:

   Traceback (most recent call last):
   File "generateExpInputs_raw.py", line 171, in <module>
   print("\n#TFs: %d, #Genes: %d, #Edges: %d, Density: %.3f" % (nTFs,nGenes,netDF.shape[0],netDF.shape[0]/((nTFs*nGenes)-nTFs)))
   ZeroDivisionError: division by zero

   I found that the Gene names in Non-Specific-ChIP-seq-network.csv are uppercase, which is different from ExpressionData.csv, so I add expr_df.index = expr_df.index.to_series().apply(lambda x:x.upper()) before expr_df.to_csv(opts.outPrefix+'-ExpressionData.csv') The result is:

   TFs: 27, #Genes: 144, #Edges: 264, Density: 0.068

2. After looking this page, I try to reproduce the results for the hESC datasets using the STRING ground truth net, and the result is:

   TFs: 28, #Genes: 82, #Edges: 112, Density: 0.049

I need some help for these problems. Maybe there are some steps for data preprocessing and I ignore them, please give me some advice.

Thank you Best wishes Jiayi Dong