A complete R analysis pipeline for MD Anderson RPPA Core Excel output
Author: Amar Kumar
Version: 0.1.0
License: MIT
Note: The CI badge may show a build failure as the GitHub Actions environment does not have all RPPA dependencies pre-installed. The package installs and runs correctly locally following the instructions below.
RPPAnalyzeR is an R package that takes the Excel workbook produced by the MD Anderson Functional Proteomics RPPA Core and turns it into a complete, reproducible analysis — with one function call or step by step.
It was built and fully tested on a MCF7 human breast cancer cell serum-starvation time-course experiment (15 samples × 497 antibodies, Set208) but works with any MD Anderson RPPA Core output file.
Reverse Phase Protein Array (RPPA) is a high-throughput antibody-based proteomics platform. Lysates from your samples are printed onto nitrocellulose slides and probed with hundreds of validated antibodies simultaneously.
The MD Anderson RPPA Core returns results as a multi-sheet Excel workbook containing raw signal, normalised values at multiple levels, quality control metrics, and pairwise comparison sheets.
| Parameter | Value |
|---|---|
| Cell line | MCF7 human breast cancer cells |
| Experiment type | Serum starvation time-course |
| Condition | Cells switched from serum-containing → serum-free media |
| Time points | 0h (baseline), 2h, 4h, 8h, 24h |
| Replicates | 3 per time point |
| Total samples | 15 |
| Antibodies | 497 |
| RPPA Set | Set208 |
| Analyst | Amar Kumar |
Why serum starvation? Removing serum (growth factors) from cell culture media activates major signalling pathways as cells respond to nutrient deprivation. This is a classic model for studying PI3K/Akt/mTOR, MAPK/ERK, and cell cycle regulation dynamics in real time.
The MD Anderson RPPA Core delivers a single .xlsx file containing 12 sheets.
Here is what each sheet contains and whether you need it:
| Sheet | Contents | Use |
|---|---|---|
| L4 (log_2) | Fully normalised log2 values — loading + batch corrected | ✅ PRIMARY — use for all analysis |
| L4 (CHM) | L4 + antibody median centering | ✅ Use for heatmaps |
| L4 (linear) | L4 converted to linear scale | ✅ Use for bar graphs only |
| L4 CHM 0h vs 2h/4h/8h/24h | Pairwise comparisons with Log2FC and pre-computed P values | ✅ Use for quick DE results |
| Antibody QC Scores | QC score per antibody (0–1 scale, min = 0.8) | ✅ Filter low-quality antibodies |
| Sample QC metrics | Total protein content per sample | ✅ Flag failed samples |
| L3 (log_2) | Loading-normalised only (intermediate step) | |
| L2 (log_2) | Raw RPPASPACE output, no normalisation | |
| Report Software Versions | Pipeline version info | ℹ️ Reference only |
The MD Anderson Excel file has a non-standard structure that confuses standard Excel readers:
Row 1 : Completely empty
Rows 2-9 : Antibody metadata — OFFSET to columns 9-506 only
(columns 1-8 are empty in these rows)
Col 9 = metadata label (e.g. "Antibody Name")
Cols 10-506 = values for each of the 497 antibodies
Row 10 : TRUE column headers (Order, Sample Name, ..., 497 protein names)
Rows 11-25 : 15 sample data rows (your actual expression values)
Row 26+ : May contain stray values (e.g. "RI" annotation in L4 linear)
— these must be ignored
RPPAnalyzeR handles this automatically. It uses skip and n_max
arguments in readxl to read exactly the rows it needs and ignores everything
else.
The RPPA Core uses THREE different naming systems across sheets:
| Sheet | Name format | Example |
|---|---|---|
| Antibody QC Scores | Human-readable | Cyclin B1 |
| L4 expression sheets | Heatmap label with suffix | Cyclin-B1-R-V |
| Metadata rows | Short antibody name | Cyclin-B1 |
These never match directly — "Cyclin B1" != "Cyclin-B1-R-V".
The only shared key across all sheets is the Antigen ID
(e.g. AGID00024), which is consistent everywhere.
RPPAnalyzeR joins all sheets via Antigen ID to ensure correct QC filtering.
Protein column names in L4 sheets follow the pattern:
AntibodyName-Species-ValidationStatus
| Code | Meaning |
|---|---|
| R | Rabbit antibody |
| M | Mouse antibody |
| G | Goat antibody |
| V | Validated — performs well in all RPPA assays |
| C | Use with Caution — mostly reliable, some edge cases |
| Q | Tissue-reactive — detects non-specific components |
Example: Akt_pS473-R-V = phospho-Akt Ser473, rabbit antibody, validated.
L2 (log2) Raw RPPASPACE curve-fitting output. No normalisation.
|
v
L3 (log2) + Loading normalisation
(bidirectional median centering: by antibody, then by sample)
|
v
L4 (log2) + Set-to-Set batch correction <-- USE THIS for analysis
|
v
L4 (CHM) + Antibody median centering <-- USE THIS for heatmaps
Always use L4 for analysis. L2 and L3 are included for transparency only.
- R version ≥ 4.0.0
- RStudio (recommended)
Find your R version: In RStudio Console type R.version$version.string
install.packages(c(
"readxl", "dplyr", "tidyr", "ggplot2", "pheatmap",
"stringr", "tibble", "purrr", "openxlsx",
"ggrepel", "RColorBrewer", "scales", "writexl",
"remotes", "rmarkdown"
))remotes::install_github("akumar901/RPPAnalyzeR")install.packages(
"/path/to/RPPAnalyzeR",
repos = NULL,
type = "source"
)library(RPPAnalyzeR)
results <- run_pipeline(
xlsx_path = "your_rppa_file.xlsx",
output_dir = "rppa_output"
)library(RPPAnalyzeR)
# 1. Import all 12 sheets
rppa <- import_rppa("your_rppa_file.xlsx")
# 2. Quality control
rppa <- run_qc(rppa, min_qc_score = 0.8)
print_qc_summary(rppa)
# 3. Differential expression — all time points vs 0h baseline
de_all <- diff_expression_all(rppa)
# Top 10 hits at 24h
top_proteins(de_all, n = 10, timepoint = "24h")
# 4. Plots
plot_sample_qc(rppa) # sample total protein bar chart
plot_antibody_qc(rppa) # QC score histogram
plot_volcano(de_all, timepoint = "24h") # volcano plot
plot_heatmap(rppa) # clustered heatmap top 50 proteins
plot_timecourse(rppa, proteins = c( # time-course line plot
"Akt_pS473-R-V",
"p70-S6K_pT389-R-V",
"MAPK_pT202_Y204-R-C"
))
# 5. Export
export_rppa_excel(rppa, de_results = de_all,
output_path = "results.xlsx")
export_de_csv(de_all, significant_only = TRUE)Open RPPAnalyzeR_Analysis.Rmd, update the two path lines at the top,
and click Knit. This generates a single self-contained HTML report
with all plots, tables, and results embedded.
These are real issues encountered during development — documented here so other users don't have to debug them.
Cause: readxl with col_names = FALSE assigns internal names
(...1, ...2) to columns. Renaming by column name string then fails
because the original names no longer exist.
Fix: Assign column names by position using a hardcoded vector of 9 fixed metadata names, not by matching strings.
Cause: A stray value "RI" (Reference Interval annotation) sits
in row 30, column 316 of the L4 (linear) sheet — 5 rows below the
actual data. When all rows were read, this triggered column count
mismatches leading to recycled (duplicate) column names.
Fix: Use n_max = 15 when reading expression sheets to read
exactly the 15 sample rows and nothing beyond.
Cause: The metadata rows (rows 2–9) have columns 1–8 completely
empty. readxl detects the used range as starting at column 9,
returning a tibble where the label column becomes column 1 instead
of column 9. Transposing this matrix then produced empty column names.
Fix: Read each of the 8 metadata rows individually using
skip = i, n_max = 1 and build the data frame directly with
hardcoded clean column names — no transpose needed.
Cause: The Antibody QC sheet uses human-readable names
("Cyclin B1") while the expression sheets use heatmap-label names
("Cyclin-B1-R-V"). These share zero exact matches, so name-based
joining drops every protein.
Fix: Join QC scores to expression sheet proteins via Antigen ID
(e.g. AGID00024) — the only field that is identical across all sheets.
All 497 Antigen IDs match perfectly.
Cause: plot_timecourse() was building a lookup map using
suffix-stripped names as keys (e.g. "Akt_pS473") but then looking
up full names (e.g. "Akt_pS473-R-V") — a key mismatch.
Fix: plot_timecourse() now tries exact match first, then
falls back to suffix-stripped match, accepting protein names in
either format.
Cause: R keeps loaded packages in memory for the session even after reinstalling. Knitting an RMD uses the in-memory version.
Fix: After reinstalling, always go to Session → Restart R in RStudio before knitting.
Cause: The MD Anderson CHM pairwise comparison sheets contain
pre-computed Log2FC and P values for all 497 antibodies, including
any that failed QC filtering (e.g. B7-H4, QC score < 0.8). The
diff_expression() function reads these pre-computed values directly
from the CHM sheets, meaning a QC-failed antibody could theoretically
appear as a significant hit in a volcano plot even though it was removed
from the expression matrix by run_qc().
Fix: diff_expression() now cross-references the list of
QC-passing proteins from rppa$l4_log2 and filters DE results to
exclude any antibody that did not pass QC. This ensures complete
consistency between the QC-filtered expression data and all downstream
DE analysis and visualisations. Download the fixed 03_differential.R
from the repository and reinstall the package.
Note: In practice this only affects antibodies with a QC score below 0.8. In the tested dataset (Set208), only B7-H4 was removed, and its low QC score means it is very unlikely to produce a strong DE signal. However the fix ensures rigorous consistency for all future datasets.
Cause: The protein name matching regex used in diff_expression()
to cross-reference QC-passing proteins was -[RMG]-[VCE]$, which
handles Validated (V) and Caution (C) antibody suffixes but missed
the Q (tissue-reactive) suffix. This caused antibodies like
Snail-M-Q to be incorrectly excluded from DE results even though
they passed the QC score threshold, because "Snail" (CHM sheet name)
did not match "Snail-M-Q" (expression sheet name).
Impact: In the tested dataset (Set208), Snail was incorrectly
absent from the 2h and 8h DE results when the Issue 7 fix was first
applied.
Fix: The regex was updated to -[RMG]-[VCEQq]$ to correctly
handle all three validation codes:
- V — Validated
- C — Use with Caution
- Q — Tissue-reactive
Important note for users: If you modify the protein name matching regex, ensure it covers all three validation codes (V, C, Q). Missing any one code will silently exclude the affected antibodies without any error message.
This regex appears in more than one file. It is used in
03_differential.R (DE filtering) and twice in 06_pipeline.R (top-50
selection for the heatmap, and automatic key-protein selection for the
time-course). Fixing only one location leaves the others broken —
run_pipeline() carried the old -[RMG]-[VCE]$ pattern for one release
after diff_expression() was corrected. Search the whole R/ directory
when changing it.
Cause: plot_heatmap() wraps pheatmap, which does not return a
ggplot object and therefore cannot be saved with save_plot(). It takes
its own filename argument instead. The heatmap chunk in
RPPAnalyzeR_Analysis.Rmd called plot_heatmap() without that argument,
so the figure was drawn into the HTML report but never written to
plots/. run_pipeline() was unaffected — it always passed filename.
Symptom: every documented plot appears in plots/ except
04_heatmap_top50.png, with no error or warning.
Fix: the template now calls plot_heatmap() twice — once without
filename to draw the inline figure, once with it to write the PNG.
Two calls are required because pheatmap writes to file instead of
drawing when filename is supplied; a single call with filename set
would save the PNG but remove the heatmap from the HTML report.
Related: the same omission affected the two QC figures in the
template — plot_sample_qc() and plot_antibody_qc() were drawn inline
but had no save_plot() call, so the RMarkdown path wrote 8 of its 10
figures to disk. Both now save as sample_QC.png and antibody_QC.png.
If you add a plot to the template, remember that drawing it and saving it
are separate steps.
Cause: get_expression_matrix() separates protein columns from sample
metadata by matching column names against a fixed list. That list contained
"Sample Source" with a space, but import_rppa() renames the column to
"Sample_Source" with an underscore. The names never matched, so the
metadata column was treated as a protein.
Symptoms: a NAs introduced by coercion warning on every call, because
the column holds text; one entirely NA column in the returned matrix; and
ncol() reporting 497 after QC had retained 496 — the 496 real proteins plus
the stray metadata column.
Impact: results were not affected in practice. sort() drops NA when
ranking proteins by variance, and plot_heatmap() receives its protein list
explicitly. However, any user-side rowMeans(), cor(), or apply() over
the full matrix would silently return NA, and the reported dimensions were
misleading.
Fix: corrected to "Sample_Source". The same list is duplicated in
03_differential.R and 04_visualize.R; both were fixed. If you add a
metadata column to the importer, add it to both lists — a name that does not
match is silently treated as protein data rather than raising an error.
Find your Mac username:
Sys.getenv("USER")Confirm the fix is in your installed version:
grep("stripped_map", deparse(body(plot_timecourse)))
# Should return line numbers — if character(0), old version is still loadedFind where R installed the package:
find.package("RPPAnalyzeR")Check what proteins are available after QC:
expr <- get_expression_matrix(rppa, "l4_log2")
head(colnames(expr$matrix), 20)Check protein names before plotting:
# Search for a protein
colnames(expr$matrix)[grep("Cyclin", colnames(expr$matrix))]Running the full pipeline produces:
output_dir/
├── RPPAnalyzeR_results.xlsx All key sheets + DE results (colour-coded)
├── DE_significant_hits.csv Significant proteins across all time points
└── plots/
├── 01_sample_QC.png Bar chart of total protein per sample
├── 02_antibody_QC.png Antibody QC score histogram
├── 03_volcano_0h_vs_2h.png Volcano plot: 0h vs 2h
├── 03_volcano_0h_vs_4h.png Volcano plot: 0h vs 4h
├── 03_volcano_0h_vs_8h.png Volcano plot: 0h vs 8h
├── 03_volcano_0h_vs_24h.png Volcano plot: 0h vs 24h
├── 04_heatmap_top50.png Clustered heatmap top 50 proteins
└── 05_timecourse_key_proteins.png Time-course of the top 6 hits at the
final time point
run_pipeline() selects the time-course proteins automatically (the six most
significant hits at the last time point). To plot specific pathways instead,
pass them explicitly:
run_pipeline(
xlsx_path = "your_rppa_file.xlsx",
output_dir = "rppa_output",
key_proteins = c("Akt_pS473", "p70-S6K_pT389", "MAPK_pT202_Y204")
)The RMarkdown template (vignettes/RPPAnalyzeR_Analysis.Rmd) instead produces
three curated pathway panels — timecourse_mTOR_PI3K.png,
timecourse_MAPK_ERK.png and timecourse_CellCycle.png — because the protein
sets are hard-coded in that file. Edit those chunks to change the panels.
MIT © Amar Kumar — see LICENSE for details.
You are free to use, modify, share, and build on this package. If you use it in a publication, a citation is appreciated.
Kumar A (2025). RPPAnalyzeR: A complete R analysis pipeline for
MD Anderson RPPA Core output. R package version 0.1.0.
https://github.com/akumar901/RPPAnalyzeR
Please also acknowledge the MD Anderson RPPA Core:
The Functional Proteomics RPPA Core is supported by MD Anderson Cancer Center Support Grant # 5 P30 CA016672-40.
Amar Kumar
Email: amarcompbio@gmail.com
Issues and pull requests welcome:
https://github.com/akumar901/RPPAnalyzeR/issues