Processing plasma proteomics for SPRINT

This project contains pre-processing and processing for plasma proteomics for SPRINT. It involved performing a number of pre-processing pipelines followed by DE and GSEA for plasma data generated for Sputum and plasma proteomes in response to CFTR therapy (SPRINT).

Pre-processing:

Running analysis

• main.R performs all analysis, initially:
• All proteins > 20% missing are filtered out.
• Abundance values are log2 transformed.
• Analysis pipelines are run on this data, these steps include: performing batch correction using Combat or "housekeeping" proteins (the most abundant proteins with the lowest coefficient of variation within plates), performing normalisation of data - either within batches (prior to batch correction) or across all batches (post-batch correction), and performing these steps with and without removing depleted proteins.
• model_metadata.R performs variance partition analysis and saves figures.

Results

• Mean IQR of all proteins in QC samples was calculated following running each pipeline, the results of this are in sorted_IQR.csv.
• Each folder in figures contains the plots and .csv's for an analysis pipeline.
• Each folder name indicates the pipeline steps and order performed. e.g.
  • folder depl_withinvsn_combat contains data and plots following the steps of: depleted proteins being removed, then VSN being performed within batches, before combat is run across all batches.
  • folder hk_vsn contains data and plots following the steps of: depleted proteins are NOT removed, "housekeeping" proteins are identified across plates and used to batch correct, before VSN is performed on all data to normalise.
• log_filter contains data/ figures from filtered and log2 transformed data only, prior to any other processing steps.
• *_df.csv is the processed data, including missing values at a stage.
• *_mf_df.csv is the processed data, imputed using MissForest.
• Within each folder the prefix to each .csv or figure refers to the data or figure produced following that specific step during the pipeline. e.g. in the folder depl_withinvsn_combat, depl_df.csv is the processed data following the step of removing depleted proteins, however combat_df.csv is the processed data following the final step in the pipeline of performing combat.
• For pipelines where depleted proteins were not removed during pre-processing, "*_rm_depl_df.csv" is the same data as *_df.csv processed data, without depleted proteins. "*_rm_depl_mf_df.csv" is the equivalent with the data imputed using MissForest.

Figures

• *_boxplot_cohort.png / *_boxplot_plate.png are boxplots of all proteins in each sample coloured by cohort/ plate respectively.
• *_heatmap_cohort.png / *_heatmap_plate.png are heatmaps of all proteins in each sample coloured by cohort/ plate respectively. Missing values are imputed using MissForest.
• *_boxplot_selected*.png are boxplots of selected proteins defined in line 52 onwards of main.R.
• NB. in cases where the proteins shown on *_boxplot_selected_least.png are the same as *_boxplot_selected_most.png. Then the proteins in *_boxplot_selected_least.png have been filtered from the dataset, due to missingness, and R is saving the most recent figure.
• *_deplete.png are boxplots of depleted proteins in each sample (if they have not been removed).
• *_pca_cohort.png / *_pca_plate.png are PCA plots of each sample coloured by cohort/ plate respectively. Missing values are imputed using MissForest.
• *_pca_var_imp.png are barplots of top contributing proteins to the PCA, coloured by whether these proteins were depleted or not.
• p(1-3)_low_cv.png are plot if "housekeeping" proteins were used to batch correct. These plots are the log2 filtered values of the proteins used for this purpose within plate 1, 2 and 3 respectively.
• A linear mixed model was fit for each protein to estimate the contribution of Cohort and Plate. In *_var_part.png, A) shows the variance explained by Cohort and Plate of the top 30 proteins ordered by Cohort. B) shows violin plots representing the distribution of variance explained of all proteins.

QC-RSC for Signal Drift and Batch Correction

• check_signal_drift.R runs testing of using the QC-RSC algorithm to correct batch effect and signal drift.
• Folder QCRSC contains the results.
• The algorithm involves fitting a curve to QC samples to check for, and then attempt to remove, signal drift and batch effect at the same time.
• There were not enough QC samples in Plate 3 (4 are required) to run the algorithm as is intended. Instead, fitting the curve to healthy samples only and then to all samples was tested.
• Using healthy samples generally over-corrected for signal drift, worsening data quality.
• Using all samples, often did not improve signal drift and potentially removed biological variation.

Recommendations

• Attempting to correct for signal drift using QC-RSC seemed to result in poorer data quality.
• Combat appears to have corrected batch sufficiently, using "housekeeping" proteins does not.
• Removing depleted proteins (deplr), or not, does not seem to have substantially impacted pre-processing, though it may be best to include them during pre-processing to reduce the risk of biasing results.
• Performing VSN prior to combat, compared to afterwards, does not impact batch correction, but results in slightly reduced QC protein IQR and therefore may be preferred.
• Therefore, the withinvsn_combat pipeline would be recommended for further processing. Within these folders, the combat_rm_depl_df.csv contains the processed data, including missing values, and combat_rm_depl_mf_df.csv includes the processed data with missing values imputed using MissForest.

Processing:

• differential_expression.R, funcs_DE.R and DE_figs.R contain differential expression analysis and figures for pre vs post ETI and pre ETI vs healthy and post ETI vs healthy.
• fpsea.R runs functional gene set enrichment analysis and produces figures.