Abstract
Background: The gene shaving algorithm and many other clustering algorithms identify gene clusters showing high variation across samples. However, gene expression in many signaling pathways show only modest and concordant changes that fail to be identified by these methods. The increasingly available signaling pathway prior knowledge provide new opportunity to solve this problem. Results: We propose an innovative semi-supervised gene clustering algorithm, where the original gene shaving algorithm was extended and generalized so that prior knowledge of signaling pathways can be incorporated. Different from other methods, our method identifies gene clusters showing concerted and modest expression variation as well as strong expression correlation. Using available pathway gene sets as prior knowledge, whether complete or incomplete, our algorithm is capable of forming tightly regulated gene clusters showing modest variation across samples. We demonstrate the advantages of our algorithm over the original gene shaving algorithm using two microarray data sets. The stability of the gene clusters was accessed using a jackknife approach. Conclusion: Our algorithm represents one of the first clustering algorithms that is particularly designed to identify signaling pathways of low and concordant gene expression variation. The discriminating power is achieved by manufacturing a principal component enriched by signaling pathways.
Introduction: The Challenge of "Quiet" Signals
In transcriptomics, researchers typically seek clusters of genes that behave similarly across conditions. However, traditional unsupervised methods like K-means or hierarchical clustering suffer from a "loudest signal" bias. They prioritize genes with high variance, often overlooking critical signaling pathways that involve subtle, tightly regulated, and highly coordinated (concordant) changes.
Furthermore, standard algorithms often assume mutual exclusivity, whereas biological reality involves overlapping pathways where a single gene may participate in multiple processes. The original "gene shaving" algorithm addressed overlapping clusters using Principal Component Analysis (PCA), but still focused on the leading principal component with the most variance.
Methods: Semi-Supervised Gene Shaving
The proposed semi-supervised gene shaving algorithm addresses these gaps by incorporating prior biological knowledge to guide the discovery process. The method uses Singular Value Decomposition (SVD) to deconvolve data into principal components (PCs). Instead of defaulting to the first PC, the algorithm calculates an "enrichment score" for every PC using a small set of known genes from a pathway of interest.
A key innovation is the separation of positive and negative loadings for each PC. This allows the algorithm to distinguish between up-regulated and down-regulated genes within the same biological signal, significantly boosting the signal-to-noise ratio. The algorithm starts with a large gene set and iteratively "shaves off" the 10% of genes least correlated with the most enriched PC until a core, tightly regulated cluster is identified.
Results and Discussion
The method was validated using yeast cell cycle data and mouse somitogenesis microarray data. In the yeast benchmark, the semi-supervised method successfully recovered low-variation gene sets that the original gene shaving algorithm missed.
In the mouse embryo study, the most biologically relevant signals for the WNT and NOTCH pathways were found in the second principal component, not the first. Standard unsupervised methods would have ignored these signals entirely. The algorithm identified 90 genes in the predicted NOTCH cluster and 45 in the WNT cluster. Many of these candidates were not in the initial "prior knowledge" set but were later validated through literature searches and Gene Ontology (GO) analysis as being related to embryonic development. Robustness was assessed using a "jackknife" approach, demonstrating that the algorithm's performance is stable even when the initial knowledge base is incomplete.
Conclusion
This semi-supervised approach represents a significant advancement in identifying signaling pathways characterized by low and concordant gene expression variation. By using prior knowledge as a compass, the algorithm can uncover subtle biological signals that are otherwise lost in the noise of high-variance data, providing a more nuanced view of the transcriptome.
Summary updated