ParsEval is a program for comparing two sets of gene annotations for the same sequence. The most common use cases for ParsEval are as follows.
- You are annotating a newly assembled genome. The optimal parameter settings for annotation are not clear initially, so you do some exploratory data analysis and try several different parameter settings. You can use ParsEval to identify the similarities and differences between the different annotations you have produced.
- You are doing a genome-wide analysis of genes in your favorite organism. There is a gene annotation available from the consortium that sequenced and assembled the genome, but there is a different annotation available at NCBI. Again, ParsEval is the best way to compare these two annotations to quickly identify their similarities and differences.
Input for ParsEval is two sets of annotations in GFF3 format. ParsEval uses the GenomeTools GFF3 parser, which strictly enforces syntax rules laid out in the GFF3 specification. ParsEval itself does some additional checks on the data to make sure valid comparisons are possible.
Any features not directly related to protein-coding genes are ignored.
ParsEval will infer feature implicitly encoded in the data. For example, if a gene annotation declares 6 exon features but no intron features, ParsEval will infer the 5 corresponding intron features from the exon boundaries. However, if ParsEval sees any intron features in a gene model it will assume all introns are declared explicitly. Violations of that assumption will likely elicit a program error.
ParsEval is pretty flexible in handling various common conventions for encoding gene structure: exons + start/stop codons, exons + CDS, CDS + UTRs, etc. Any subset of features that completely captures the gene’s exon/intron structure, CDS(s), and UTRs should be handled correctly.
ParsEval requires that gene isoforms be encoded using the feature type mRNA (as opposed to transcript, primary_transcript, or other valid SO terms). For mRNA features lacking an explicitly declared gene parent, ParsEval will create one. Note, however, that ParsEval will treat all such transcripts as belonging to separate distinct genes, which will erroneously inflate summary statistics reported by ParsEval.
ParsEval output includes a variety of similarity statistics that measure the agreement between the two annotations. Our use of agreement here instead of accuracy is intentional: except in a very few rare cases, you will not be comparing a prediction to a true high-quality “gold standard.” It is much more common to compare two annotation sets whose relative quality is unknown. ParsEval uses the terms reference and prediction only to distinguish the two sets: it makes no assumptions as to their relative quality.
Similarity statistics are currently reported at two levels of granularity. First, a report for each individual locus shows the similarity of the annotations at that locus, with the option to also include an embedded graphical representation as well (if using HTML output mode). Second, the similarity statistics are aggregated over the entire data and presented in a single summary report.
For a description of ParsEval’s command-line interface, run the following command (after AEGeAn has been installed).