TACO produces robust multisample transcriptome assemblies from RNA-seq

Abstract

Accurate transcript structure and abundance inference from RNA sequencing (RNA-seq) data is foundational for molecular discovery. Here we present TACO, a computational method to reconstruct a consensus transcriptome from multiple RNA-seq data sets. TACO employs novel change-point detection to demarcate transcript start and end sites, leading to improved reconstruction accuracy compared with other tools in its class. The tool is available at http://tacorna.github.io and can be readily incorporated into RNA-seq analysis workflows.

Figure 1

Schematic detailing the transcriptome meta-assembly workflow for TACO. Reads are initially aligned to the genome. Ab initio assembly is then performed, generating a transcriptome assembly for each input sample. These transcriptome assemblies are then merged into a meta-assembly using the TACO tool, which leverages change point detection and a dynamic programming algorithm to generate robust transcript isoforms from the underlying network of splicing patterns.

Figure 2

Assessment of TACO performance. a. Performance metrics for TACO, Cuffmerge, and Stringtie when merging different numbers of input assemblies. Recall (i.e., sensitivity), precision and the F-measure for all three tools were assessed for splicing patterns, splice junctions, and bases. Points represent the mean statistic across the 20 runs, error bars represent the 95% confidence interval. (Data to make this panel can be found in Supplementary Table 3) b,c. Precision-recall plots (left) and bar plots depicting the average precision (right) depicting performance for the three tools merging 55 CCLE breast cancer cell lines (a) at 50 different isoform fraction cutoffs ranging from 0.001-0.999, and (b) for the highest expressed transcripts in the meta-assemblies. Points represent statistics for the top N transcripts, with N ranging from 500-30,000. (Data to make panels b and c can be found in Supplementary Tables 4 and 6, respectively).

Figure 3

Examples of TACO performance. The 3p21 genomic locus is depicted. Assembly of the genes SLC26A6, CELSR3, and NCKIPSD are shown for 50, 100, and 500 samples merged. The assembly produced by Cuffmerge is shown in blue, Stringtie in orange, and TACO in green. The Refseq reference annotation is shown above in red.