Next generation genome annotation with mGene.ngs

An increasingly large number of novel genomes is being sequenced and the task of automatic genome annotation has never been more important. The current revolution in sequencing technologies also allows us to obtain a detailed picture of the whole complement of expressed RNA transcripts. We have developed a novel de novo gene finding system mGene.ngs that combines the benefits of accurate ab initio gene finding with the rich information obtained in RNA sequencing (RNA-seq) experiments.

The system is based on the recently developed accurate gene finding system mGene [1], which employs state-of-the-art prediction techniques and which has been shown to perform very well compared to established gene finding systems [2]. In contrast to many HMM-based gene finders, mGene has the conceptual advantage of being very flexible in terms of incorporating heterogeneous input data. The employed inference techniques can exploit the transcriptome information already at the learning stage to appropriately adapt to the relevance of the different evidences. We show that these advantages can be translated into more accurate gene predictions. Moreover, we developed extensions of mGene.ngs to predict and quantify alternative RNA transcripts.

To provide de novo genome annotations based on RNA-seq experiments, we first construct a preliminary, highly specific gene set for genes that are well-covered with RNA-seq reads. In a second step, we train predictors for genomic signals on the preliminary gene set. In the third step we train mGene.ngs, using the preliminary gene models while taking advantage of the RNA-seq read coverage and genomic signal predictions.

We illustrate the power of our approach for the C. elegans genome and 50M paired-end RNA-seq reads (Illumina; 76nt). Figure 1 shows transcript level evaluation results for all annotated genes (WS200) as a function of the expression level. The ab initio mGene-based system (blue) trained on the annotation achieves an average transcript-level F-score of 49.9%. We achieve a slightly better performance (51.8%) for the de novo annotation system (green) using RNA-seq reads, but without considering the existing genome annotation. If we use the RNA-seq reads and train on the existing annotation (red), we achieve 57.6%, and can therefore take advantage of the previous annotation. We find it remarkable that for medium to high expressed genes the de novo gene predictions are as similar to the genome annotation as the predictions of the system, that has seen parts of the annotation in training. Comparing these results to predictions from the recently published method cufflinks [3] (black) reveals that cufflinks seems not to be able to appropriately adapt to the RNA-seq data at hand.

Figure 1

F-score on coding transcript level for different gene finding approaches as a function of the expression level. Coding transcript level evaluation counts a transcript as correct if all coding exons match exactly with the all coding exons of a annotated transcript. The F-score combines sensitivity and specificity penalizing large differences in these values.

Investigating the contribution of individual features we found that spliced read alignments suggesting introns help most to increase the gene prediction performance; 91.6% of the achieved total improvement is due to spliced read alignments. The read coverage alone is much less informative and only leads to improvements similar to the ones achieved with transcriptome tiling arrays. We employed the developed annotation strategy for the re-annotation of the C. briggsae genome, for which only few transcriptome sequences are available yet. We can show that the new annotation is considerably more accurate than previous ones and additionally includes alternative RNA isoforms.

mGene.ngs will be released as open source software on http://mgene.org and is already available as

Galaxy-based web-service at http://galaxy.fml.mpg.de.