Hello all,
I am performing a weighted gene co-expression network analysis (WGCNA) on label-free quantitative proteomics data and would appreciate some advice regarding data preprocessing and network construction.
Data information
- Number of samples: 25 samples
- Number of proteins after filtering: 410 proteins
Data processing workflow
1- Raw LC–MS/MS Orbitrap data were processed using MaxQuant with Label-Free Quantification (LFQ) enabled.
2- The resulting proteinGroups.txt file was imported into Perseus.
3- In Perseus, I performed:
- Removal of potential contaminants, reverse hits, and proteins identified only by site.
- Log2 transformation of LFQ intensities.
- Filtering to retain proteins with at least 70% valid values.
- Missing value imputation using Perseus' normal distribution-based imputation method.
4- The processed protein expression matrix was then imported into R for WGCNA (unsigned network).
5- Prior to WGCNA, I applied removeBatchEffect() (limma) to adjust for batch effects and additional covariates.
After running WGCNA, the scale-free topology fit plot appears unusual (shown below).
Questions
1- Does the preprocessing workflow described above appear appropriate for WGCNA of LFQ proteomics data?
2- Could the combination of missing value imputation and batch correction be affecting the correlation structure required for WGCNA?
3- Are there alternative preprocessing strategies for proteomics data that may improve network construction?
4- Are there any known issues with applying WGCNA to datasets with relatively high sparsity or extensive imputation?
•
written
by
Vinesh •
0 answers
No answers yet.
Log in to answer this question.
-
Tool: Pipette.bio - a conversational AI agent that runs real bioinformatics analyses
written by chirag •
-
Analysis of a batch-normalized expression dataset
written by dave •
-
Statistical analysis for proteomic data with more than one effect in the model
written by TCC •
-
Pre-processing of label-free MS proteomics data
written by JessieLenwell •
To answer your questions, the core issue here likely isn't your preprocessing pipeline.
Without seeing how your samples behave (via a PCA of the normalized data), my main guess is that very few of your 410 proteins show strong correlations.
WGCNA penalizes weak connections by raising the correlation matrix to a power. Because the bulk of your proteins lack strong co-expression, raising the power effectively wipes out the vast majority of the network (most correlations reach 0). The connectivity values that do survive are maintained by a very small, isolated subset of strongly correlated proteins. This is exactly what your plots show: at a power of 6, the mean connectivity crashes almost immediately with the "unusual" drop in the scale-free topology fit.
Thank you for your response.
The PCA of my data after batch correction and covariate adjustment is shown below.
Your explanation regarding the lack of strong correlations among proteins is very helpful. I had one additional question regarding normalization. You mentioned examining the behavior of the normalized data. In my workflow, I did not perform any additional normalization after obtaining the LFQ intensity matrix from MaxQuant. My understanding is that MaxQuant's LFQ algorithm already incorporates normalization across samples during LFQ intensity calculation. Therefore, the data imported into Perseus were the LFQ intensities generated by MaxQuant, which were subsequently log2-transformed, filtered, imputed, batch-corrected, and used for WGCNA.
Given this workflow, would you recommend any additional normalization steps before network construction, or are the LFQ intensities from MaxQuant generally considered normalized for WGCNA?
Hello,
If the majority of my 410 proteins do not exhibit strong co-expression patterns, would that suggest that WGCNA may not be the most suitable approach for identifying meaningful modules in this dataset? In that case, would it be more appropriate to focus on direct protein–trait correlations rather than attempting network-based clustering?
Also, would there be any alternative network or clustering approaches that are less dependent on strong co-expression structure and might perform better in a dataset like this?
Thank you for your insights.
Hi Vinesh,
If your traits are discrete variables, I would run a differential abundance analysis. It will allow you to directly identify proteins that are significantly up- or down-regulated between your groups
I am not aware of any method