CLASS2 – accurate and efficient splice variant annotation from RNA-Seq reads

Next generation sequencing of cellular RNA is making it possible to characterize genes and alternative splicing in unprecedented detail. However, designing bioinformatics tools to accurately capture splicing variation has proven difficult. Current programs can find major isoforms of a gene but miss lower abundance variants, or are sensitive but imprecise.

CLASS2 is a novel open source tool for accurate genome-guided transcriptome assembly from RNA-seq reads based on the model of splice graph. An extension of our program CLASS, CLASS2 jointly optimizes read patterns and the number of supporting reads to score and prioritize transcripts, implemented in a novel, scalable and efficient dynamic programming algorithm. When compared against reference programs, CLASS2 had the best overall accuracy and could detect up to twice as many splicing events with precision similar to the best reference program. Notably, it was the only tool to produce consistently reliable transcript models for a wide range of applications and sequencing strategies, including ribosomal RNA-depleted samples. Lightweight and multi-threaded, CLASS2 requires <3GB RAM and can analyze a 350 million read set within hours, and can be widely applied to transcriptomics studies ranging from clinical RNA sequencing, to alternative splicing analyses, and to the annotation of new genomes.

The CLASS2 transcript assembly algorithm

Step 1 (A) Exon and introns. Infer exons from the read coverage levels, using linear programming, and introns from spliced alignments. Step 2 (B) Splice graph. Build a splice graph to represent the gene, connecting exons by introns. Shown is a section from a splice graph, with a skipped exon event and a 2-intron retention event, encoding two possible paths (transcripts). Step 3 (C) Constraints. Cluster reads into classes (constraints) by their splicing and interval patterns. Step 4 (D) Transcript selection. Build and solve the bipartite constraint graph and associated transcript selection problem, shown here for four read pairs c1, c2, c3 and c4, and three transcripts t1, t2 and t3.

1. CLASS2 offers the best tradeoff between sensitivity and precision in reconstructing full transcripts. In its default setting, CLASS2 detects 10–70% more transcripts than Cufflinks, which is the most popular and most precise of these programs, with higher or comparable precision. In its sensitive settings, CLASS2 detects up to twice as many transcripts as Cufflinks for a relatively small drop in precision.

2. It is the best suited to capture local alternative splicing variation. In particular, it can detect up to twice as many alternative splicing events as Cufflinks, with high precision. CLASS2 finds slightly fewer events than Scripture, which is the most sensitive of the programs, but its precision is considerably (70–80%) higher.

3. It employs a combined gene-level and genome-level model of intronic ‘noise’ that allows more accurate detection of intron retention events.

4. The amount of novel alternative splicing variation detected by CLASS2 increases with increasingly large data sets.

5. CLASS2 is multi-threaded and scales well with the amount of data, requiring <3GB RAM for all of our tests, and can complete most regular tasks in a few hours.

6. Lastly, since CLASS2 can produce annotations from RNA-seq data alone, without requiring an existing set of gene annotations, it is very well suited for the annotation of newly sequenced organisms.

Availability – The software is available free of charge for all and under a GNU GPL license from http://sourceforge.net/projects/Splicebox.