Hello,
I’m an undergraduate student working on identifying and analyzing differentially expressed proteins (DEPs) from a DIA proteomics dataset (urine, Schistosoma haematobium infected vs. control, n=28 infected / 22 control, raw peak intensity data) using MetaboAnalyst. I’d appreciate feedback on whether my pipeline makes sense.
What I did:
Data check: 4,756 missing values detected (12.2% of the dataset) on upload.
Filtering:
-
Missing value exclusion: features with >50% missing values removed
-
Low variance filter: IQR-based, set to remove bottom 10%
Missing value imputation: Left-censored (LOD) estimation, using 1/5 of the minimum positive value per feature
Normalization:
-
Sample normalization: Median
-
Data transformation: Log2
-
Data scaling: Autoscaling
Statistical analysis:
-
T-test: Student’s t-test, raw (unadjusted) p-value
-
Also ran Limma (moderated t-test) in parallel for comparison
-
Fold-change threshold: 1.5
-
Volcano plot: combined FC (1.5) + raw p-value (0.05), run separately with Classical and Limma methods to compare
Questions:
-
Given that I confirmed a strong relationship between missingness and low abundance in this dataset (proteins with more missing values had dramatically lower average intensity among detected values), does LOD-based imputation seem like a reasonable, defensible choice here, versus simple mean imputation?
-
Is median normalization + autoscaling a sensible combination for peak-intensity proteomics data, or would you recommend skipping the autoscaling step?
-
Is it reasonable/common practice to run both Classical and Limma fold-change methods and compare, or should I commit to just one for a cleaner final analysis?
-
Any other red flags in this pipeline I should reconsider?
Thanks so much for any input!