Incongruent results when performing DME Analysis and Module Trait Correlation

[sandcell10]

Hello @smorabit,

Thank you for creating and maintaining such an excellent package!

I have a categorical variable called "Invasive" describing the invasiveness of a tumor (invasive vs non-invasive). When running DME analysis in a cell cluster of interest, cluster 8, I see that the tan and yellow modules have a high average log2FC in invasive samples in this cluster (see lollipop plot).

However, when I conduct module trait correlation analysis, while the yellow module is positively correlated with Invasion status, the tan module is negatively correlated (see heatmaps). Given the DME results showing high differential ME activity in invasive cells in this cluster, I would expect both modules to be positively correlated with invasion status. I was wondering how this could be and if I might be misinterpreting these results. I have included the code used to create the plots below. Any information you could provide would be greatly appreciated and thank you for your consideration!

DME Analysis

group1 <- obj@meta.data %>% subset(test_clusters_2 == '8' & Invasive == 'Invasive') %>% rownames
group2 <- obj@meta.data %>% subset(test_clusters_2 == '8' & Invasive == 'Non-Invasive') %>% rownames

stem_inv_DMEs <- FindDMEs(
  obj,
  barcodes1 = group1,
  barcodes2 = group2,
  test.use='wilcox',
  wgcna_name='hdWGCNA'
)

head(stem_inv_DMEs)


#Visualize
pdf('Clin_Graphs/Clust8_Invasive_DME_lolli.pdf')
PlotDMEsLollipop(
  obj, 
  stem_inv_DMEs, 
  wgcna_name='hdWGCNA', 
  pvalue = "p_val_adj")+
  ggtitle('Invasive Cluster 8 Differential Expression')+
  theme(
    plot.title = element_text(size = 18, face = "bold"),  # title size
  )
dev.off()

Module Trait Correlation

#Convert to Numeric or Factor 
obj$Max_dimension_cm <- as.numeric(obj$Max_dimension_cm)
obj$Ki67 <- as.numeric(obj$Ki67)
obj$Invasive <- as.factor(obj$Invasive)


# list of traits to correlate
cur_traits <- c('Max_dimension_cm','Ki67', 'Invasive')

obj <- ModuleTraitCorrelation(
  obj,
  traits = cur_traits,
  group.by='test_clusters_2'
)


# get the mt-correlation results
mt_cor <- GetModuleTraitCorrelation(obj)

#Plot
pdf('Clin_Graphs/Module_Trait_hmap2.pdf', height = 14, width = 12)
PlotModuleTraitCorrelation(
  obj,
  label = 'fdr',
  label_symbol = 'stars',
  text_size = 2,
  text_digits = 2,
  text_color = 'black',
  high_color = 'red',
  mid_color = 'white',
  low_color = 'blue',
  plot_max = 0.2,
  combine=TRUE)
dev.off()

[smorabit]

Hi,

It is possible to get different results since these are two different analyses. It could be more useful to understand why there is a discrepancy if you look at the DME results table instead of the plot.

[thgirb-matt]

Hi both,

I have found this incongruence too and I am also confused. I have a module that DME reported as significantly upregulated in my diseased cells compared to control cells, but correlation analysis would suggest that the module is negatively correlated with disease status.

Looking at the DME table does reveal that the module shows an avg_log2FC of 7.92 despite pct.1 being 0.289 and pct.2 being 0.444. I assume that this is likely to explain the discrepancy? The much lower percentage of cells 'expressing' the module in disease, even if when they express the module they do so at much higher levels could make the correlation appear negative? In this case which analysis is more valid?

Even more confusingly, I have some modules which show a significant positive correlation with disease status in every cell type but in the all_cells group the module is negatively correlated, how can this be?