A comparative Bayesian and classical framework for time-series clustering of high-dimensional transcriptomic data
Abstract
High-dimensional time-series clustering is still an open problem since it involves temporal dependence, noise and heterogeneous cluster structures. Here we propose a comparison computational strategy that combines classical fixed-partition clustering and Bayesian nonparametric modeling to disentangle temporal signatures in transcriptomic time-series. The framework includes variance-based feature selection and multi-resolution K-means clustering (k = 4,6) with a Dirichlet Process Gaussian Mixture Model to estimate data-adaptive temporal modules without requiring a priori knowledge of the number of clusters. As a case study, we started from an osteoblast differentiation transcriptome in Mus musculus (at days 3, 4, 5, 6 and 8) for which the whole data structure was first characterized by principal component and hierarchical clustering, retaining accordingly a set of highly dynamical genes (621 with standard deviation>0.9) for further analysis. Quantitative assessment by complementary clustering indexes demonstrated 0.350, 0.358, and 0.357 Silhouette, ARI, and NMI scores for the k-means approach with k = 4; 0.289, 0.238 and 0.309 for k = 6; and retained solution applied also to DPGMM (12 detailed temporal modules). application level Functional coherence calculated at the application level showed better capture of the fine-grained, more interpretable temporal patterns enriched for biologically relevant signaling processes as measured by Gene Ontology, while KEGG-based comparisons were mostly lower across all clustering strategies. In summary, these results indicate that Bayesian nonparametric clustering does indeed extend classical approaches by presenting more refined views on complicated longitudinal transcriptomic data.