A reproducible workflow for exploratory consensus clustering of gene expression data

초록

Introduction: This protocol describes a consensus clustering workflow for grouping samples by gene expression similarity and assessing the stability of those groupings. It was developed for use with small case series where formal subtype validation is not yet feasible. Practice: A processed logCPM expression matrix comprising 11 case samples was used as an example dataset. The 500 most variable genes were selected using median absolute deviation (MAD) and standardized using gene-wise Z-scores. Consensus clustering was performed with ConsensusClusterPlus for K = 2–6 using 1,000 resampling iterations. Candidate solutions were evaluated using consensus matrices, cumulative distribution function curves, delta area, proportion of ambiguous clustering, cluster consensus scores, and cluster sizes. Solutions with K ≥ 3 contained singleton clusters, whereas K = 2 separated the samples into two groups without singleton clusters. K = 2 was therefore retained only as an exploratory and illustrative solution for demonstrating the workflow rather than as evidence of biologically validated molecular subtypes. The stability measures did not consistently support the same K, highlighting the importance of using multiple complementary criteria for cluster selection. Conclusion: No single stability metric pointed to one clear K in this dataset. Reporting several measures side by side, as done here, makes it easier to judge how much confidence a given cluster solution deserves.

키워드