Towards Causal Analysis of Genetic Factors for Colorectal Cancer

David C. Miller, Hana Chockler, Andrew D. Lawrence, Daniel McNamee, Nicholas ChiaResearch 2021 · DOI

Abstract We present preliminary results and outline future work for the causal analysis of genetic factors influencing the occurrence and severity of colorectal cancer. The findings are based on publicly available datasets. We show that using relatively simple methods we are able to detect meaningful dependencies, reflecting the current biomedical knowledge about the causes of cancer and related conditions (such as infection and sepsis). We also argue that a deeper analysis, taking into account gene regulation, would require a significantly larger dataset.

Introduction

Colorectal cancer (CRC) remains a leading cause of global mortality. While screening colonoscopies are effective, they are invasive and increasingly strained by the rise of early-onset cases and "interval cancers"—malignancies that emerge between scheduled screenings. Consequently, there is an urgent clinical need for non-invasive early detection methods, such as blood or stool-based biomarkers.

Identifying these biomarkers presents a significant data science challenge known as the "curse of dimensionality." Researchers must navigate over 20,000 RNA expressions and protein levels across relatively small patient cohorts. Standard statistical methods often yield correlations that lack causal relevance. This study, a collaboration between causaLens and the Mayo Clinic, explores causal discovery techniques to move beyond simple correlation and identify the true drivers of CRC.

Methodological Approach

To avoid observer bias, the authors adopted a strictly data-driven pipeline utilizing several key techniques:

  • Bivariate Filtering: Spearman’s rank correlation was employed to identify significant relationships between RNA and protein expression. This method provides greater robustness to noise and outliers than standard Pearson correlation.
  • Agglomerative Hierarchical Clustering: To manage the high-dimensional feature space, the Ward method was used to group RNA molecules with similar expression patterns, allowing for the analysis of gene clusters rather than individual variables.
  • Bimodal Activation: The researchers tested binarization—categorizing gene expression as "active" or "inactive" using Mixed Gaussian Models—to simplify the data.
  • Causal Counterfactual Reasoning: The long-term objective is to develop models capable of predicting the effects of interventions (e.g., drug treatments) by identifying specific candidate causes within the genomic data.

Data-Driven Insights and Redundancy

The study utilized the COADREAD cohort from The Cancer Genome Atlas (TCGA), focusing on individuals with overlapping RNAseq and RPPA protein data. Despite the initial size of the TCGA project, only 283 individuals met these criteria, highlighting the limitations of current public datasets for complex causal modeling.

A key finding was the high degree of redundancy among RNA molecules. In intercorrelation heatmaps, these molecules often appear in distinct "blocks."

Heatmap of RNA intercorrelation p-values related to the ATM protein, illustrating high redundancy and block structures.

While redundancy is often viewed as a hurdle, the authors leveraged it to demonstrate that these genes likely share common functional pathways or hidden confounders. This redundancy justifies the use of hierarchical clustering to reduce 20,500 RNA features to approximately 4,100 clusters—a 20% reduction that still retains over 50% of the cumulative variance.

Case Study: The ATM Protein

To validate the approach, the researchers analyzed the ATM protein, which is critical for DNA damage response. Hierarchical clustering successfully grouped genes such as CCL3 and CCL4, which are biologically known to function together. This alignment with established biomedical knowledge suggests that data-driven clustering can accurately capture functional genetic relationships.

Hierarchical clustering dendrogram for RNA molecules related to the ATM protein.

However, the study also revealed the dangers of over-simplification. While binarizing gene expression is common in bioinformatics, it proved too aggressive for this dataset, discarding roughly 15% of the data as noise and rendering most significant correlations uncalculable.

Limitations and Future Directions

The primary constraint facing this research is the lack of healthy control data within the TCGA cohort, which exclusively contains cancer patients. This imbalance can lead to "spurious correlations" where variables appear related only because of the underlying disease state.

Future work will focus on aggregating multiple datasets to increase sample size and implementing "coarse-to-fine" causal models. These models will first identify relationships between broad gene clusters before refining the analysis to pinpoint specific genetic drivers of colorectal cancer.

Summary updated

Abstract

Abstract We present preliminary results and outline future work for the causal analysis of genetic factors influencing the occurrence and severity of colorectal cancer. The findings are based on publicly available datasets. We show that using relatively simple methods we are able to detect meaningful dependencies, reflecting the current biomedical knowledge about the causes of cancer and related conditions (such as infection and sepsis). We also argue that a deeper analysis, taking into account gene regulation, would require a significantly larger dataset.
BibTeX
@inproceedings{miller2021towards,
  author = {David C. Miller and Hana Chockler and Andrew D. Lawrence and Daniel McNamee and Nicholas Chia},
  title = {{Towards Causal Analysis of Genetic Factors for Colorectal Cancer}},
  booktitle = {Research},
  year = {2021},
  doi = {10.21203/rs.3.rs-967255/v1},
}
Figure