Compositional uncertainty in high-throughput sequencing data analysis should not be ignored.
- Autores: Gregory B. Gloor, Andrew D. Fernandes, Jean M. Macklaim, Michael Vu
- Localización: Proceedings of the 6th International Workshop on Compositional Data Analysis: Girona, 1-7 de juny de 2015 / coord. por Santiago Thió Fernández de Henestrosa, Josep Antoni Martín Fernández, 2015, ISBN 978-84-8458-451-3
- Idioma: inglés
- Enlaces
-
Resumen
High throughput sequencing is used for many different types of experimental analyses including 16S rRNA gene sequencing, transcriptomics and metagenomics. All of these experiments represent random sampling of sequence fragments derived from an underlying unknown distribution of fragments and attempt to determine which, if any, of the sampled fragments are differentially abundant between predefined groups. These data typically are composed of thousands of fragments in only a small number of samples. The number of sequence reads binned to each fragment represents a single estimate of the underlying frequency of the fragment in each sample. The values obtained are compositional because the total number of sequences obtained is trivially determined by the instrument itself and provides no information about the number of random fragments in the input pool. In addition, in many instances, the investigator is interested in the relative, and not the absolute abundance, of the fragments. We show that the variation due to sampling alone (technical variation) in compositional datasets derived from high-throughput sequencing is large and inversely related to the number of reads mapping to a fragment. Ignoring this technical variation can lead to false positive inference regardless if the data is treated as a composition or not. We find that a two-step procedure that incorporates a Baysian estimate of fragment abundance along with the centred log-ratio transformation markedly improves specificity with no loss of sensitivity
