Abstract
Gene expression analysis often struggles with high dimensionality, noise, and the overlapping nature of biological processes. Traditional clustering methods like k-means are often unstable and produce mutually exclusive groups. This paper presents a robust biclustering framework that combines Nonnegative Matrix Factorization (NMF) with Positive Tensor Factorization (PTF). By treating multiple NMF runs as a tensor and applying PTF as a meta-clustering step, the authors extract stable, overlapping biological signals. Evaluation on synthetic and lung cancer datasets demonstrates superior recovery of ground truth and high histological accuracy, reducing algorithmic variability from 31% to just 3%.
Introduction
The analysis of microarray gene expression data is a cornerstone of modern genomics, yet it faces the "curse of dimensionality"—thousands of genes measured across relatively few samples. Biological reality is complex; genes often participate in multiple overlapping pathways, a feature that standard "hard" clustering algorithms like k-means fail to capture. Furthermore, these algorithms are notoriously unstable, yielding different results based on random initialization. To address these challenges, this research introduces a two-stage approach using Nonnegative Matrix Factorization (NMF) and Positive Tensor Factorization (PTF) to achieve stable "biclustering"—the simultaneous grouping of genes and samples.
Methodology
The proposed framework operates in two primary stages:
1. Object-Level Clustering with NMF
The researchers first apply NMF to decompose the gene expression matrix into two lower-rank matrices: one representing gene membership in clusters and the other representing the influence of these clusters across samples. NMF is particularly well-suited for biological data because it enforces non-negativity and allows for additive overlaps, reflecting how a gene's total expression can be the sum of multiple active biological processes.
2. Two-Way Meta-Clustering with PTF
To overcome the instability of individual NMF runs, the authors perform multiple independent runs with different initializations. These results are stacked into a 3D tensor (Samples × Genes × Runs). PTF is then applied to this tensor to identify "bicluster prototypes" that are consistent across all runs. This meta-clustering step acts as a filter, distilling persistent biological signals from the noise and local minima of individual algorithmic attempts.
Experimental Results
Synthetic Data Performance
The framework was first tested on synthetic data designed to mimic the characteristics of real microarrays, including log-normal noise and overlapping clusters.
As shown in Figure 1, the PTF-based meta-clustering consistently outperformed NMF, fuzzy k-means, and standard k-means in recovering the underlying ground truth, particularly as the complexity (number of biclusters) increased.
Lung Cancer Classification
The algorithm was applied to a large-scale lung cancer dataset containing 12,600 genes across 203 samples. The framework successfully recovered known histological classifications, including adenocarcinomas, squamous cell carcinomas, and small cell lung cancers.
The heatmap in Figure 4 illustrates the clear separation of these groups. Notably, the model identified "adeno-squamous" samples as mixtures of subtypes, demonstrating its sensitivity to complex biological phenotypes. Quantitatively, while individual NMF runs showed a 31% variability in results, the PTF meta-clustering reduced this to just 3%, significantly enhancing reliability for clinical applications.
Discussion and Conclusion
The integration of NMF and PTF provides a powerful tool for extracting stable, biologically relevant patterns from noisy genomic data. By leveraging consensus across multiple runs, the framework mitigates the instability inherent in traditional clustering. However, challenges remain. The selection of the optimal number of clusters ($n_c$) is still heuristic-based, and the computational intensity of tensor factorization may pose scaling issues for massive single-cell datasets. Future work will focus on refining parameter selection and optimizing the preprocessing pipeline to further improve the robustness of genomic biclustering.
Summary updated