← All projects

Colorectal Cancer Proteomics

Tumour versus healthy, protein by protein.

An analysis of CPTAC proteomic data comparing colorectal tumours with healthy colon tissue. Starting from 54 samples and more than ten thousand proteins, I cleaned, filtered, imputed and normalised the data, tested every protein for a difference between the two groups, and wrote up each decision in a report.

Volcano plot of log2 fold change against -log10 p-value for 9,140 proteins: significantly down proteins in cyan on the left, up in magenta on the right, the rest grey, with GAP43, L1CAM, NCAM1, OGN, ALDH1A1, DDX21, SERPINH1, S100A11, RRP15 and CKS2 labelled
All 9,140 tested proteins: 83 up in tumour, 506 down.

The data

Fifty-four samples, cleaned first.

The dataset holds 21 healthy and 33 tumour samples from CPTAC, run across six TMT batches. Before anything else I removed 69 contaminant proteins, mapped every sample column to Normal or Tumour, and checked that the two groups were mixed across all six batches, so condition and batch aren't fully confounded.

Two histograms of log10 intensity for all 54 samples, Normal in blue and Tumour in red: offset from each other before median normalisation, overlapping after
Every sample's intensities, before and after median normalisation.

Missing values

Missing, but not at random.

Proteins at low abundance fall below the detection limit, so an empty value usually means a small one. About 20% of values were missing per sample. I dropped proteins missing from more than 75% of samples, leaving 9,140, and filled the remaining gaps with draws from a downshifted distribution, the method Perseus uses.

Histograms of log10 intensity for sample 01_CPTAC_TMT_126 before and after imputation, with the imputed values in orange in the lower tail
One sample before and after imputation. Imputed values in orange.

Testing

589 proteins changed.

Each of the 9,140 proteins was tested with Welch's t-test on log2 intensities, 21 healthy samples against 33 tumour. A protein counted as significant only if it changed at least twofold and had p below 0.05. That gave 83 proteins up in tumour and 506 down, a pattern consistent with tumours losing the identity of normal colon tissue.

Proteins of interest

Ten proteins, up close.

I shortlisted the five most significant proteins in each direction. Those up in tumour fit active growth: DDX21 and RRP15 support ribosome biogenesis, SERPINH1 points to matrix remodelling, and S100A11 and CKS2 both have documented links to colorectal cancer. Those down, GAP43, L1CAM, NCAM1, ALDH1A1 and OGN, fit a loss of normal colon tissue identity.

Box plots of log10 intensity in Normal and Tumour samples for ten proteins: DDX21, SERPINH1, S100A11, RRP15 and CKS2 higher in tumour; GAP43, L1CAM, NCAM1, ALDH1A1 and OGN lower
The shortlist, healthy in blue and tumour in red.

And the details.

How it's built

Python, Pandas, NumPy.

The analysis is written in Python: pandas and NumPy for the tables, SciPy for the statistics, and Seaborn and Matplotlib for every figure. The report walks through each step, the alternatives I considered and what changed along the way.

  1. Downshifted imputation

    For each sample on the log10 scale, missing values are drawn from a normal distribution centred 1.8 SD below the mean with a width of 0.3 SD, then transformed back.

  2. Two thresholds, not one

    Requiring both a twofold change and p < 0.05 avoids calling tiny effects with small p-values, or large changes built on noisy measurements.

  3. Checking the calls

    Benjamini-Hochberg correction ran as a check on the uncorrected calls: 544 of the 589 proteins still passed with the same fold-change threshold.

Colorectal Cancer Proteomics. Tumour versus healthy, protein by protein.

NextPersonal Portfolio WebsitePaper, smoke and stars.