Bayesian Biclustering for Patient Stratification

Sahand Khakabimamaghani, Martin Ester

Modern oncology is shifting from empirical, "one-size-fits-all" treatments toward personalized medicine. A critical intermediate step is Stratified Medicine, where patients are grouped into subtypes based on the underlying molecular mechanisms of their disease. However, current stratification methods face significant hurdles, including the difficulty of integrating diverse data types and the rigidity of deterministic algorithms that struggle with biological noise.

The B2PS Framework

To address these challenges, researchers developed B2PS (Bayesian Biclustering for Patient Stratification), a novel probabilistic graphical model (PGM). Unlike standard deterministic tools like Non-negative Matrix Factorization (NMF) or Singular Value Decomposition (SVD), B2PS uses Bayesian inference to simultaneously cluster patients and genetic features across multiple data types.

Comparison of B2PS with existing patient stratification methods.

As illustrated in the comparison above, B2PS is unique in its ability to perform biclustering within a probabilistic framework while handling multiple data types.

Model Architecture and Inference

B2PS is designed to be integrative and flexible. It can ingest somatic point mutations, copy number variations (CNV), and gene expression data. By employing "priors," the model incorporates existing biological knowledge to effectively filter out noise.

The B2PS Probabilistic Graphical Model.

The model utilizes Gibbs sampling for inference. A key advantage of B2PS is its ability to decouple the number of patient clusters from the number of gene clusters—a flexibility many standard models lack. Furthermore, the algorithm naturally discovers the "natural" number of patient subtypes by emptying unnecessary clusters during execution, reducing the need for manual parameter tuning.

Experimental Validation

The study utilized data from The Cancer Genome Atlas (TCGA), focusing on Glioblastoma Multiform (GBM) and Breast Invasive Carcinoma (BRCA). For each patient, three data types were integrated: somatic point mutations, discrete copy number variations, and discretized gene expression levels.

Clinical Relevance and Biological Coherence

B2PS significantly outperformed NMF in identifying clinically meaningful patient subgroups. In GBM survival analysis, B2PS achieved a log-rank p-value of 0.004, indicating distinct survival outcomes between clusters. In contrast, NMF failed to find statistically significant differences (p-value 0.458).

Beyond clinical outcomes, B2PS also demonstrated superior biological coherence. Using the Gene Ontology Term Overlap (GOTO) score, the model identified gene clusters that were more biologically related than those found by NMF (e.g., a GOTO score of 3.408 for B2PS vs. 2.535 for NMF in GBM).

Insights into Data Integration

A key finding was that gene expression data is the most informative for stratification, likely because it is closer to the final disease phenotype. Interestingly, the researchers noted that simply adding more data types isn't always beneficial; if data types are inconsistent, they can introduce more noise than signal, potentially lowering the quality of the resulting clusters.

Conclusion

B2PS represents a powerful new approach to patient stratification, offering a stable and biologically grounded framework for identifying disease subtypes. Future work may involve moving to continuous distributions for gene expression data to provide higher resolution and integrating additional "omics" layers, such as DNA methylation and miRNA expression, to further refine personalized treatment strategies.

Summary updated

Abstract

The move from Empirical Medicine towards Personalized Medicine has attracted attention to Stratified Medicine (SM). Some methods are provided in the literature for patient stratification, which is the central task of SM, however, there are still significant open issues. First, it is still unclear if integrating different datatypes will help in detecting disease subtypes more accurately, and, if not, which datatype(s) are most useful for this task. Second, it is not clear how we can compare different methods of patient stratification. Third, as most of the proposed stratification methods are deterministic, there is a need for investigating the potential benefits of applying probabilistic methods. To address these issues, we introduce a novel integrative Bayesian biclustering method, called B2PS, for patient stratification and propose methods for evaluating the results. Our experimental results demonstrate the superiority of B2PS over a popular state-of-the-art method and the benefits of Bayesian approaches. Our results agree with the intuition that transcriptomic data forms a better basis for patient stratification than genomic data.
BibTeX
@inproceedings{khakabimamaghani2016bayesian,
  author = {Sahand Khakabimamaghani and Martin Ester},
  title = {{Bayesian Biclustering for Patient Stratification}},
  booktitle = {PSB},
  year = {2016},
  url = {https://openreview.net/forum?id=KQKNhC463d},
}
Figure