!!! clock "time"
* Teaching: 20 minutes
!!! circle-info "Objectives and Key points"
#### Objectives
* Understand the limitations of protein coding predictions when working with annotations
#### Keypoints
* *De novo* protein sequence prediction is a good starting point, but proper annotation will require significant manual curation
* Not all tools are able to predict gene boundaries over splice junctions - be careful when interpreting predictions
* Protein sequence prediction does not predict all informative genetic elements - additional tools are required for features such as rRNA, tRNA, or ncRNA elements
Prediction of genes in a genome assembly is a complicated process - there are many tools which can perform good initial predictions from assembled contigs, but there are often many biological features which confound the prediction process and make it more complicated than simply finding start and stop codons within a sequence.
At the most basic level, searching for proteins is simply looking for open reading frames (ORF) within a contig, but in practice there are many factors which confound the process. At the biological level a number of features complicate the process of predicting protein coding regions:
- Alternate coding schemes, including the amber, umber, and ochre stop codons
- Stop codon read-through
- Splicing of intronic regions
Simply translating the nucleotide sequence between a start/stop pairing is not sufficient to correctly identify the complete protein complement of the genome.
Furthermore, if our genome assembly is not complete we run the risk of encountering partial coding sequeences in which either the 5' or 3' region of the sequence were not assembled. In these cases, a simple search for ORFs will fail to detect the partial sequence. The prediction of protein coding sequences must be achieved using more complicated techniques than a simple grep search for the start and stop codon(s).
Similar to assembly we can perform gene prediction in either a reference-guided manner or through the use of ab initio prediction tools. We will not be covering the reference-guided approach, as it is quite simple to perform in Geneious, but it is not to be underestimated as a technique - particularly when working with viruses or other organisms with complex read-through or splicing properties.
Ab initio prediction is akin to de novo assembly - the tool is created with some internal models for what coding regions look like, which are then applied to query sequences to find putative coding regions.
Depending on the intended use of the tool, each prediction tool may be better tuned for partiular assumptions of the data. We are going to use two different tools today, one designed for prediction of prokaryotic coding sequences (which generally lack introns) and one designed primarily for eukaryotic sequences, where splicing is common.
As you will see from both examples above, protein coding prediction is at best a good starting point for identifying genes. Careful validation of each sequence needs to be performed if you are trying to produce a comprehensive annotation.
In addition, these tools are only for prediction of protein coding sequences so if you're trying to recover a particular element from your data make sure that the tool you are using is suitable for the job.
There are many other genomic features you may need to look for - some additional tools to examine if you are looking for a complete annotation:
- Metaxa2 (Bengtsson-Palme et al, 2015) - Prediction of small and large subunit ribosomal RNA sequences
- Barrnap - Prediction of small and large subunit ribosomal RNA sequences
- ARAGORN (Laslett & Canback, 2004) - Prediction of tRNA and tm RNA sequences
- Infernal (Nawrocki & Eddy et al, 2013) - Prediction of non-coding RNA sequences, requires rfam database