Identification of a prognostic signature in colorectal cancer using combinatorial algorithm-driven analysis

Abdo Alnabulsi, Teihui Wang, Wei Pang, Marius Ionescu, Stephanie G. Craig, Matthew P. Humphries, Kris McCombe, Manuel Salto Tellez, Ayham Alnabulsi, Graeme MurrayJournal of Pathology: Clinical Research 2022 · OpenReview

TL;DR

This paper introduces a novel computational approach to identify robust prognostic biomarker signatures in colorectal cancer by intelligently exploring complex combinatorial interactions. They discovered a six-biomarker signature that accurately predicts patient survival, independent of existing clinical factors, offering a promising tool for personalized treatment and early diagnosis.

What Problem Does This Paper Address?

Colorectal cancer (CRC) remains a significant cause of cancer-related deaths globally. While current clinicopathological parameters offer some prognostic insight, they often fall short due to the inherent molecular heterogeneity of CRC. This heterogeneity leads to variations in tumor progression, prognosis, and treatment response, making it challenging to accurately stratify patients. The ideal solution would involve identifying molecular signatures that can provide more precise prognostic information. However, discovering such signatures is a formidable task due to the sheer number of potential biomarker combinations (the "combinatorial explosion") and the statistical challenges of multiple hypothesis testing. Existing methods struggle to efficiently navigate this vast search space to find truly optimal and robust combinations.

Methods

The authors developed an innovative, multi-stage methodology to overcome the combinatorial explosion problem.

  1. Screening: Initially, 66 biomarkers were screened for their individual prognostic power. Biomarkers lacking significant association with survival were discarded, reducing the search space.
  2. Grouping: The remaining biomarkers were then grouped into subsets based on various interrelationships: biological (e.g., protein families, pathways), clinical (e.g., association with prognosis), and statistical (e.g., correlation). This step aimed to provide a biological and clinical framework for analysis and further reduce the number of combinations to evaluate.
  3. Iterative Combinatorial Algorithm: A custom iterative algorithm was employed to evaluate all possible combinations of biomarkers within each subset. This algorithm uses a "fitness function" based on multivariate Cox regression models (incorporating established prognostic parameters) to assess the prognostic performance of each combination. Key metrics like hazard ratio, coefficient, and model score were used. The algorithm iteratively refines combinations, aiming to find the optimal set with the highest prognostic power.
  4. Validation: The identified optimal signature was then validated on an independent external cohort to confirm its prognostic utility.

This approach is novel in its systematic screening, biologically-informed grouping, and iterative algorithmic exploration of biomarker combinations, designed to be more efficient and robust than brute-force methods.

Datasets

  • Discovery Patient Cohort: A retrospective cohort of primary colorectal cancers (n=650) from the Grampian Biorepository (NHS Grampian, UK), collected between 1994 and 2009. Patients with stage I-III disease and no evidence of distant metastasis or prior neoadjuvant/radiotherapy were included.
  • Validation Patient Cohort: A publicly available dataset (accession number GSE39582) from the NCBI Gene Expression Omnibus. After exclusions (stage IV and missing survival data), 502 patients were included.

Key Insights and Ideas

  • The "Combinatorial Explosion" is a Major Hurdle: The paper highlights that the sheer number of potential biomarker combinations makes finding truly predictive signatures incredibly difficult with traditional methods.
  • Intelligent Subsetting is Key: Grouping biomarkers based on biological and clinical relationships is crucial for making the combinatorial search manageable and biologically meaningful.
  • A Six-Biomarker Signature Emerges: The study successfully identified a specific combination of six biomarkers (STAT1, UCP1, p-cofilin, LIMK2, FOXP3, and ICOS) that demonstrates significant prognostic power.
  • Prognostic Independence is Crucial: The identified signature's ability to predict survival independently of established clinical factors (age, tumor stage, lymph node status, etc.) is a major strength, suggesting it offers novel information.
  • Potential for Early Detection: The signature's strong performance even at the biopsy stage is particularly exciting, as it could inform treatment decisions earlier in the disease course.

Key Results

  • Discovery Cohort:
    • A signature of six biomarkers (STAT1, UCP1, p-cofilin, LIMK2, FOXP3, and ICOS) was identified as the optimal prognostic combination.
  • This signature was significantly associated with overall survival (χ² = 53.183, p < 0.001 as a linear variable; χ² = 67.625, p < 0.001 as a cluster variable).
  • The signature was independently prognostic of age, extramural vascular invasion, tumor stage, and lymph node metastasis (Wald = 32.898, p < 0.001).
  • Hierarchical clustering identified five prognostic groups with significant survival differences.
  • Validation Cohort:
    • The signature remained significantly associated with prognosis (χ² = 14.217, p = 0.007).
  • It was prognostically independent in multivariate models including age, tumor stage, and lymph node stage.
  • Biomarker Function: The identified biomarkers are associated with immune response, cytoskeletal organization, and metabolic pathways, suggesting complex interactions within the tumor microenvironment.

Limitations and Future Work

  • Scoring Method Variability: The authors acknowledge that using different scoring methods (quantitative vs. semi-quantitative) for biomarker assessment could introduce variability and affect the accuracy of dichotomized scores.
  • Algorithmic Refinement: While the developed algorithm is effective, further customization, optimization, and automation are suggested to improve efficiency and applicability. This includes exploring alternative methods for composite variable calculation (e.g., factor analysis, structural equation modeling) and optimizing group identification and cut-off points.
  • External Validation Scope: While an external cohort was used, the authors note the need for further validation in larger, diverse patient populations.
  • Therapeutic Potential: The identified signature offers potential therapeutic targets, but further research is needed to explore how these biomarkers can be manipulated for treatment interventions.
  • MMR Status: The signature was found to be prognostically significant in MMR-proficient tumors, which represent the majority of CRC. This opens avenues for developing targeted therapies for this specific patient subgroup, as many current immunotherapies focus on MMR-deficient tumors.
  • Code Availability: The authors have made their algorithm and code publicly available on GitHub, encouraging further development and application by the research community.

Summary updated

Abstract

Colorectal carcinoma is one of the most common types of malignancy and a leading cause of cancer-related death. Although clinicopathological parameters provide invaluable prognostic information, the accuracy of prognosis can be improved by using molecular biomarker signatures. Using a large dataset of immunohistochemistry-based biomarkers (n = 66), this study has developed an effective methodology for identifying optimal biomarker combinations as a prognostic tool. Biomarkers were screened and assigned to related subsets before being analysed using an iterative algorithm customised for evaluating combinatorial interactions between biomarkers based on their combined statistical power. A signature consisting of six biomarkers was identified as the best combination in terms of prognostic power. The combination of biomarkers (STAT1, UCP1, p-cofilin, LIMK2, FOXP3, and ICOS) was significantly associated with overall survival when computed as a linear variable (χ2 = 53.183, p < 0.001) and as a cluster variable (χ2 = 67.625, p < 0.001). This signature was also significantly independent of age, extramural vascular invasion, tumour stage, and lymph node metastasis (Wald = 32.898, p < 0.001). Assessment of the results in an external cohort showed that the signature was significantly associated with prognosis (χ2 = 14.217, p = 0.007). This study developed and optimised an innovative discovery approach which could be adapted for the discovery of biomarkers and molecular interactions in a range of biological and clinical studies. Furthermore, this study identified a protein signature that can be utilised as an independent prognostic method and for potential therapeutic interventions.
BibTeX
@article{alnabulsi2022identification,
  author = {Abdo Alnabulsi and Teihui Wang and Wei Pang and Marius Ionescu and Stephanie G. Craig and Matthew P. Humphries and Kris McCombe and Manuel Salto Tellez and Ayham Alnabulsi and Graeme Murray},
  title = {{Identification of a prognostic signature in colorectal cancer using combinatorial algorithm-driven analysis}},
  journal = {Journal of Pathology: Clinical Research},
  year = {2022},
  url = {https://openreview.net/forum?id=u89W85mSct},
}
Figure