WGCNA error: "Error in cutreeDynamic"

Question

[giuseppe0525]

Hi,

An error occurred to me when I used the function recutBlockwiseTrees of WGCNA package. The error message is:

> blockwiseModules: cutreeDynamic returned the following error:
>         Error in cutreeDynamic(dendro = dendrograms[[blockNo]], deepSplit = deepSplit,  :    Argument dendro must have class hclust.

However, the class of dendro is indeed correct, as is shown:

> class(net$dendrograms[[1]])   
> [1] "hclust"

I used the following codes:

> net <- recutBlockwiseTrees(   datExprPFC,  
> good_Samplegenes$goodSamples, good_Samplegenes$goodGenes,   rep(1,
> times=ncol(datExprPFC)),   "total_blockwiseTOM-block.1.RData",  
> **net$dendrograms[[1]]**,   corType = "bicor",   networkType = "signed",  
> deepSplit = 4,   detectCutHeight = 0.99, minModuleSize = min(20,
> ncol(datExprPFC)/2 ),   maxCoreScatter = NULL, minGap = NULL,  
> maxAbsCoreScatter = NULL, minAbsGap = NULL,   minSplitHeight = NULL,
> minAbsSplitHeight = NULL,   useBranchEigennodeDissim = FALSE,  
> minBranchEigennodeDissim = mergeCutHeight,   pamStage = TRUE,
> pamRespectsDendro = TRUE,   minCoreKME = 0.5, minCoreKMESize =
> minModuleSize/3,   minKMEtoStay = 0.3,   reassignThreshold = 1e-6,  
> mergeCutHeight = 0.15, impute = TRUE,   trapErrors = FALSE,
> numericLabels = FALSE,   verbose = 0, indent = 0)

Previously, I ran these codes to create the associated objects, and there's nothing wrong.

    > net = blockwiseModules(
   >   datExprPFC, 
   >   loadTOM = FALSE,
   >   corType="bicor", 
   >   networkType="signed", power=12,
   >   maxBlockSize=50000, 
   >   minModuleSize=30, pamRespectsDendro=FALSE,
   >   mergeCutHeight=mergingThresh,
   >   numericLabels=TRUE,
   >   TOMType = "signed", saveTOMs=TRUE, 
    >   saveTOMFileBase="total_blockwiseTOM",
    >   nThreads = 8)

The R environment is 3.5.0 on x86_64-redhat-linux-gnu (64-bit).

Do you know what might cause the error? Thank you!

[Kevin Blighe]

How does your network actually appear prior to recutBlockwiseTrees?, i.e., plot the dendrogram

[giuseppe0525]

Thanks for your reply. That's a weird plot where the majority of the genes don't fall into any module, and that's why I want to re-cut the trees. The plot was build on a single-cell RNA-seq dataset containing around 2,000 cells. The input genes were transformed to TPM and log2(x+1) transformed.! The plot is linked here. enter image description here

[Kevin Blighe]

That looks very 'atypical', which may indirectly be the cause of the error...

You may consider trying to do some pre-processing of the data before WGCNA, such as removing samples with a high number of NA values and / or those with many genes of low expression.

[giuseppe0525]

Unfortunately, it didn't work. I have filtered many samples with many low expressed genes and the dendrogram looks similar. Have you done WGCNA on single-cell RNA-seq data? Is there any difference that I should pay attention to? Maybe I missed something.

[Kevin Blighe]

Okay, can you explain the processing that you performed prior to WGCNA? - normalisation,. transformation(s), et cetera? I have worked with scRNA-seq data, but never with WGCNA.

Generally, your data-points look 'too similar', which renders your clustering 'flat' and not amenable to a network analysis.

[giuseppe0525]

Sure. The data has been transformed to TPM. What I performed was: 1. filter samples with many genes with low expression; 2. log2(x+1) transformation; 3. quantile normalisation.

The codes I used were shown below:

> ## filter sample with too many low expressed genes 
> sample_filter <- function(ExprData=ExprData,
>                           filter.gene=3,
>                           filter.gene.level=3){   del_col=c()   for ( i in 1:ncol(ExprData) ){
>     if ( sum(ExprData[ ,i] >= filter.gene.level) < filter.gene)
>       del_col = c(i, del_col)   }   new_dat=ExprData[, -del_col]   return(new_dat) }
> 
> DatExp_filter2 <- sample_filter(ExprData = DatExp_filter,
>                                 filter.gene = nrow(DatPFC_filter)/8,
>                                 filter.gene.level = 1)
> ## log2 transform 
> DatExp_filter_lg<- log2(DatExp_filter2  + 1)
> 
> ## quantile normalisation 
> library(preprocessCore) 
> DatExp_filter_lg_qn <- DatExp_filter_lg 
> DatExp_filter_lg_qn  <- as.matrix(DatExp_filter_lg_qn ) 
> DatExp_filter_lg_qn <- normalize.quantiles(DatExp_filter_lg_qn )

Additionally, I tried removing the 'bad samples' while not doing log2 and quantile transformation, and the dendrogram shows that slightly more genes were clustered. However, the majority were labeled grey.

[Kevin Blighe]

Hi, but it does not really make sense to quantile normalise after you have log-transformed the data. From where does ExprData derive?

[giuseppe0525]

I got ExprData here. I want to test whether WGCNA can be performed on sc-RNA-seq data.

[Kevin Blighe]

Hi! Well, there lies the issue. That file already contains TPM, which is already normalised. For WGCNA, just use the logged TPM data. You do not have to perform quantile normalisation.

[giuseppe0525]

Thanks for suggestion. But unluckily, It didn't work. I skipped the quantile normalisation step and the cluster dendrogram looks similar. It seems it is the data issue.

[Kevin Blighe]

You can try without logging, too. WGCNA is based on correlation, so, the data does not strictly have to be logged. Even the authors state this.

[giuseppe0525]

Yes, I did try and the plot was similar. Maybe I can try the variance Stabilizing Transformation included in DESeq2, which takes a lot of time.

[Kevin Blighe]

Variance stabilisation should, actually, be quick. It is rlog that can take a long time.

As a final suggestion, perhaps you could convert the TPM data to Z-scores, and then try WGCNA. Not ideal at all, but TPM data is neither ideal starting off.

[giuseppe0525]

Thanks for your suggestion.

[rodd]

What happened to your analysis, giuseppe0525?

[Kevin Blighe]

Having some issues, rodd?

[Dio]

Hi, have you fixed it? I met the same problem.