Biomarker Clustering of Colorectal Cancer Data to Complement Clinical Classification

Chris M. Roadknight, Uwe Aickelin, Alex Ladas, Daniele Soria, John Scholefield, Lindy Durrant

TL;DR

Researchers applied unsupervised machine learning to a dataset of 84 colorectal cancer biomarkers, discovering that patients naturally group into 3–4 distinct clusters that do not align with traditional TNM staging. These biochemical clusters provide survival prognostic value comparable to clinical staging, suggesting that a patient's internal "immunological status" is just as critical for treatment planning as the physical size of their tumor.

What Problem Does This Paper Address?

In the world of oncology, the "gold standard" for deciding how to treat colorectal cancer is TNM staging: measuring the Tumor size, the involvement of lymph Nodes, and the presence of Metastasis. While effective, TNM is primarily anatomical. It tells us what the tumor looks like and where it has moved, but it says very little about the patient’s underlying biochemistry or immune response.

The authors identify a significant gap: patients within the same TNM stage often have wildly different survival outcomes and responses to treatment. This suggests that the current classification system is "blind" to the immunological and biochemical factors that actually drive the disease. The paper explores whether unsupervised clustering—letting the data speak for itself—can reveal a more accurate way to categorize patients.

Methods

The researchers employed an unsupervised learning approach, specifically K-means clustering, to see how patients naturally group together based on their biological data rather than their clinical labels.

Key technical steps included:

  • Feature Reduction: The initial 84 attributes were narrowed down to ~50 using a mix of correlation coefficients and expert clinical knowledge to remove redundant or "hindsight-dependent" variables (like whether a patient received chemo).
  • Optimal Cluster Determination: To avoid arbitrarily picking a number of groups, the team used six different validation indices (such as the Calinski and Hartigan indices) to find the "elbow" in the data.
  • Consensus Clustering: They tested the stability of these groups by varying the number of patients and attributes included in the analysis.

Datasets

The study utilized a unique dataset from City Hospital, Nottingham, consisting of 462 patients. Each patient profile included 84 attributes categorized into:

  • Physical data: Age, sex.
  • Immunological data: Levels of various T-cell subsets.
  • Biochemical data: Levels of specific proteins (e.g., p53, FLIP).
  • Clinical data: Tumor location and size.
  • Retrospective data: Post-operative survival statistics.

The authors paid particular attention to patients in TNM stages 2 and 3, as these are often the most clinically ambiguous cases for determining post-operative treatment.

Key Insights and Ideas

  • The TNM Mismatch: The most striking finding was that biochemical clusters were almost entirely unrelated to TNM stages. A patient in Stage 2 was just as likely to end up in "Cluster A" as a patient in Stage 3. This implies that clinical staging is missing a massive layer of biological reality.
  • The "Magic Number" is 3 or 4: Across multiple validation tests, the data consistently suggested that patients fall into 3 or 4 biological groups. This contradicts the binary "Stage 2 vs. Stage 3" logic often used in clinical decision-making.
  • FLIP as a Driver: The researchers found that certain markers, like FLIP (a regulator of cell death), were significantly elevated in low-survival clusters (60% of patients) compared to high-survival clusters (21%). This highlights specific proteins that could become targets for future personalized therapies.
  • The "Anti-learning" Warning: The authors highlight the difficulty of supervised modeling in high-dimensional spaces with low sample sizes, arguing that unsupervised methods are safer for discovering genuine biological relationships without forcing the data into pre-existing (and potentially flawed) clinical boxes.

Key Results

  • Survival Prediction: While the biomarker clusters didn't match TNM stages, they did match TNM’s ability to predict survival. In a 4-cluster solution, the lowest-survival group averaged 36.87 months, while the highest-survival group averaged 46.68 months.
  • Cluster Stability: The 3-to-4 cluster structure remained the most "optimal" even when the dataset was pruned to remove patients with missing data or when the number of attributes was reduced to the most significant 25–35 markers.
  • Truth Table Analysis: A "truth table" comparison showed that K-means clustering based on biochemistry split TNM Stage 2 and 3 patients almost perfectly down the middle, proving the two systems are measuring completely different aspects of the disease.

Limitations and Future Work

  • Missing Data: The dataset had a high rate of missingness (11.32% overall, with some attributes missing 40% of values), requiring the use of imputation (averaging), which can introduce bias.
  • Data Granularity: Much of the data was "thresholded" (simplified to 0 or 1) by human observers rather than using raw continuous cell counts. This loss of granularity might hide more subtle sub-clusters.
  • Future Integration: The authors suggest that the next step isn't to replace TNM, but to create a hybrid classification system. By combining physical metrics (TNM) with biochemical clusters, clinicians could potentially create much more accurate, personalized treatment plans.

Summary updated

Abstract

In this paper, we describe a dataset relating to cellular and physical conditions of patients who are operated upon to remove colorectal tumours. This data provides a unique insight into immunological status at the point of tumour removal, tumour classification and post-operative survival. Attempts are made to cluster this dataset and important subsets of it in an effort to characterize the data and validate existing standards for tumour classification. It is apparent from optimal clustering that existing tumour classification is largely unrelated to immunological factors within a patient and that there may be scope for re-evaluating treatment options and survival estimates based on a combination of tumour physiology and patient histochemistry.
BibTeX
@inproceedings{roadknight2012biomarker,
  author = {Chris M. Roadknight and Uwe Aickelin and Alex Ladas and Daniele Soria and John Scholefield and Lindy Durrant},
  title = {{Biomarker Clustering of Colorectal Cancer Data to Complement Clinical Classification}},
  booktitle = {FedCSIS},
  year = {2012},
  url = {https://openreview.net/forum?id=T93DusQaL2},
}

Loading related content...

Figure