Abstract
We present a nonparametric Bayesian method for disease subtype discovery in multi-dimensional cancer data. Our method can simultaneously analyse a wide range of data types, allowing for both agreement and disagreement between their underlying clustering structure. It includes feature selection and infers the most likely number of disease subtypes, given the data. We apply the method to 277 glioblastoma samples from The Cancer Genome Atlas, for which there are gene expression, copy number variation, methylation and microRNA data. We identify 8 distinct consensus subtypes and study their prognostic value for death, new tumour events, progression and recurrence. The consensus subtypes are prognostic of tumour recurrence (log-rank p-value of $3.6 \times 10^{-4}$ after correction for multiple hypothesis tests). This is driven principally by the methylation data (log-rank p-value of $2.0 \times 10^{-3}$) but the effect is strengthened by the other 3 data types, demonstrating the value of integrating multiple data types. Of particular note is a subtype of 47 patients characterised by very low levels of methylation. This subtype has very low rates of tumour recurrence and no new events in 10 years of follow up. We also identify a small gene expression subtype of 6 patients that shows particularly poor survival outcomes. Additionally, we note a consensus subtype that shows a highly distinctive data signature and suggest that it is therefore a biologically distinct subtype of glioblastoma.
Introduction: The Challenge of Multi-Omic Integration
Cancer is a complex disease rarely driven by a single biological mechanism. To gain a holistic understanding, researchers collect "multi-omic" data, including gene expression (transcriptomics), DNA mutations or copy number changes (genomics), and chemical modifications like methylation (epigenomics).
The primary challenge lies in the integration of these heterogeneous datasets. A common "early fusion" approach involves concatenating all data tables into a single matrix. However, this method is fundamentally flawed because different data types possess distinct statistical properties, noise levels, and biological relevance. Furthermore, biological layers do not always align; for instance, a change in DNA copy number may not manifest as a change in gene expression. This research addresses the need for a principled framework to identify consensus across diverse datasets without forcing artificial agreement where biological divergence exists.
Methodology: Multiple Data Integration (MDI)
The authors utilize and extend the Multiple Data Integration (MDI) framework, a nonparametric Bayesian approach designed for simultaneous analysis of multiple data types. Key technical features include:
- Dirichlet-Multinomial Allocation (DMA): Each data type is modeled using its own mixture model, allowing the algorithm to automatically infer the optimal number of clusters (subtypes) from the data.
- Coupled Priors: Rather than forcing a single global partition, MDI employs association parameters ($\phi$) that control the strength of dependency between pairs of data types. This allows the model to learn which omic layers share structural information and which provide independent signals.
- Integrated Feature Selection: The model incorporates binary indicator variables to distinguish "informative" features from noise, enhancing the robustness of the clustering process.
- MCMC Enhancements: A "split-merge" Markov Chain Monte Carlo (MCMC) sampler is implemented to improve the reliability of the results by avoiding local optima.
Data and Implementation
The study analyzed 277 glioblastoma samples from The Cancer Genome Atlas (TCGA). The integrated dataset comprised:
- Gene Expression: 1,011 features.
- Copy Number Variation (CNV): 1,000 significant probes.
- DNA Methylation: 769 binarized features.
- microRNA (miRNA): 104 features.
- Clinical Follow-up: Data regarding mortality, new tumor events, and recurrence.
Results and Discussion
The MDI algorithm identified eight distinct consensus subtypes across the patient cohort. A critical finding was the value of allowing for "disagreement" between data types; the model confirmed that while a consensus structure exists, it is only partial. Forcing a single partition would have obscured the complex interplay between different biological layers.
The consensus subtypes proved highly prognostic for tumor recurrence, achieving a log-rank p-value of $3.6 \times 10^{-4}$ after Bonferroni correction. As shown in the table below, the integrated "all" approach provided a more significant prognostic signal for recurrence than any individual data type, with the exception of methylation.
Key Biological Insights
- Methylation as a Prognostic Driver: Methylation data was identified as the strongest individual predictor of recurrence ($p = 2.0 \times 10^{-3}$).
- Low-Methylation Subtype (Cluster 7): A particularly noteworthy subtype of 47 patients was characterized by exceptionally low methylation levels. This group exhibited a significantly better prognosis, with zero new tumor events recorded over a 10-year follow-up period.
- Rare High-Risk Subtype: The model identified a small cluster of 6 patients with a unique gene expression signature and exceptionally poor survival outcomes, highlighting the sensitivity of the Bayesian approach to small but clinically relevant groups.
- Biological Validation: Gene Ontology (GO) and KEGG pathway analysis linked the identified subtypes to known cancer pathways, such as MAPK signaling and cell cycle regulation, confirming the biological relevance of the statistical clusters.
Conclusion
While the MDI framework provides powerful insights, certain limitations remain. Some identified subtypes are very small, necessitating validation in larger cohorts. Additionally, missing clinical data for some patients limits the power of survival analyses. Future work could expand this framework to include additional data types, such as proteomics, or apply it to other complex diseases requiring multi-omic integration.
The code for this study is available at: https://sites.google.com/site/multipledatafusion/
Summary updated