TL;DR
This paper addresses the challenge of analyzing DNA methylation data, which is inherently non-Gaussian and bounded. The authors propose and compare several non-Gaussian statistical models for dimension reduction and unsupervised clustering, demonstrating their superiority over traditional Gaussian-based methods. Notably, models that explicitly account for the bounded nature of DNA methylation data yield the best clustering performance.
What Problem Does This Paper Address?
DNA methylation is a crucial epigenetic mechanism that regulates gene expression and is frequently altered in diseases like cancer. This data is typically quantified as "beta-values," representing the fraction of methylated DNA, which are naturally bounded between 0 and 1. This bounded, non-Gaussian nature makes standard statistical methods, which often assume Gaussian distributions and unbounded data, ill-suited for accurate analysis. Specifically, existing dimension reduction techniques like PCA and ICA, and clustering methods based on Gaussian assumptions, can lead to suboptimal results when applied to DNA methylation data. The paper aims to bridge this gap by introducing and evaluating non-Gaussian statistical models that can better capture the unique properties of this data.
Methods
The study focuses on two main stages: dimension reduction and unsupervised clustering.
Dimension Reduction: * PCA (Principal Component Analysis): A standard Gaussian-based method used as a baseline. * BG-NMF (Beta-Gamma Nonnegative Matrix Factorization): A novel method proposed by the authors that explicitly models the bounded nature of DNA methylation data by assuming beta-distributed elements. This method decomposes the data into components that retain the bounded support property. * SC (Spectral Clustering): Another dimension reduction technique that maps data into a lower-dimensional space, resulting in features with unit length.
Unsupervised Clustering: * VBGMM (Variational Bayesian Gaussian Mixture Model): A Gaussian-based clustering method used in conjunction with PCA. * RPBMM (Recursive Partitioning Beta Mixture Model): A non-Gaussian clustering method that uses beta distributions and a recursive partitioning strategy to determine the optimal number of clusters. This is used as a benchmark for non-Gaussian methods. * VBBMM (Variational Bayesian Beta Mixture Model): A variation of the beta mixture model that uses variational inference to automatically determine model complexity and prune redundant components, offering computational efficiency. * VBvMM (Variational Inference Framework-based Bayesian Analysis of the vMF Mixture Model): A non-Gaussian clustering method designed for directional data (unit vectors), suitable for features derived from SC. * VBWMM (Variational Bayesian Estimation of WMM): A non-Gaussian clustering method for axially symmetric data, also used with SC-derived features.
The core innovation lies in applying non-Gaussian models, particularly those that respect the bounded nature of the data (like BG-NMF and beta mixture models), to both dimension reduction and clustering.
Datasets
The study utilizes a publicly available DNA methylation dataset from the Gene Expression Omnibus (GEO) website, specifically dataset GSE32393. This dataset contains DNA methylation profiles for approximately 27,578 CpGs from 113 breast cancer samples and 23 non-neoplastic breast tissue samples. For their analysis, the authors selected 5000 CpGs with the highest variance, resulting in a data matrix of size 5000 x 136.
Key Insights and Ideas
- Boundedness Matters: The most significant insight is that the bounded nature of DNA methylation data is not just a characteristic but a crucial property that must be preserved during analysis. Models that explicitly account for this (like BG-NMF) outperform those that ignore it.
- Non-Gaussian is Superior: Conventional Gaussian-based methods (like PCA + VBGMM) are demonstrably less effective than non-Gaussian approaches for DNA methylation data. This highlights the importance of choosing statistical models that align with the data's underlying distribution.
- Dimension Reduction Aids Clustering: The study confirms that effective dimension reduction can significantly improve the performance of subsequent unsupervised clustering tasks, making it easier to discern meaningful patterns and subgroups.
- Computational Efficiency Gains: Variational Bayesian methods (like VBBMM) offer a more computationally efficient approach to model selection and clustering compared to methods requiring cross-validation or recursive partitioning (like RPBMM).
Key Results
- Clustering Performance: Non-Gaussian methods consistently outperformed the Gaussian-based PCA + VBGMM approach.
- BG-NMF + VBBMM Best Performer: Among the non-Gaussian methods evaluated, BG-NMF combined with VBBMM demonstrated the best clustering performance.
- Misclassification Rates:
- PCA + VBGMM misclustered 9 samples (all cancer samples classified as normal).
- BG-NMF + RPBMM and BG-NMF + VBBMM misclustered 5 samples (4 cancer as normal, 1 normal as cancer).
- SC + VBvMM misclustered 7 samples (7 cancer as normal).
- SC + VBWMM misclustered 6 samples (6 cancer as normal).
- Computational Speed: BG-NMF + VBBMM was faster than BG-NMF + RPBMM, with a computational time of 124 seconds compared to 139 seconds for the latter, and significantly faster than applying RPBMM directly on the original data (1275 seconds).
- Feature Properties: BG-NMF yields features that retain the bounded property, while SC produces unit-length features, which are less directly reflective of the original data's bounded support.
Limitations and Future Work
- SC Method Limitations: While SC methods did not misclassify normal samples as cancer, their overall accuracy was lower. The authors speculate that this might be due to SC embedding data in a "tighter" manner, creating a clearer boundary but potentially losing some nuance.
- Improvement Needed in Clustering: The authors acknowledge that the consequent unsupervised clustering methods applied to the reduced features (from both BG-NMF and SC) could be further improved to achieve even better clustering accuracy.
- Extension to Other Data: The methodology presented is suggested to be easily extendable to analyze other DNA methylation datasets, and other non-Gaussian statistical models could also be explored for similar purposes.
- Optimal σ in SC: The choice of the parameter σ in spectral clustering is noted as an open topic for future research.
Summary updated