2016/01/19 by Saverio Ranciati, Cinzia Viroli, Ranciati, Saverio +4
Biochemistry, Genetics and Molecular Biology · Computer Science · Mathematics · #Applications (stat.AP) #Bayesian Methods and Mixture Models #Bioinformatics and Genomic Networks #FOS: Computer and information sciences #Gene expression and cancer classification #stat.AP
paper · pdf · doi:10.48550/arxiv.1601.04879
25 pages; 3 tables, 6 figures
openalex publication_date 2016/01/19 · arxiv created 2016/05/12 · arxiv updated 2016/05/13 · openalex created_date 2022/10/01 · openalex updated_date 2026/07/28
Model-based clustering is a technique widely used to group a collection of units into mutually exclusive groups. There are, however, situations in which an observation could in principle belong to more than one cluster. In the context of Next-Generation Sequencing (NGS) experiments, for example, the signal observed in the data might be produced by two (or more) different biological processes operating together and a gene could participate in both (or all) of them. We propose a novel approach to cluster NGS discrete data, coming from a ChIP-Seq experiment, with a mixture model, allowing each unit to belong potentially to more than one group: these multiple allocation clusters can be flexibly defined via a function combining the features of the original groups without introducing new parameters. The formulation naturally gives rise to a `zero-inflation group' in which values close to zero can be allocated, acting as a correction for the abundance of zeros that manifest in this type of data. We take into account the spatial dependency between observations, which is described through a latent Conditional Auto-Regressive process that can reflect different dependency patterns. We assess the performance of our model within a simulation environment and then we apply it to ChIP-seq real data.