State of the art in eukaryotic gene prediction

PhyloCSF: a comparative genomics method to distinguish protein coding and non-coding regions

“…Accurate computational methods are needed to classify these transcripts and the corresponding genomic exons as protein coding or non-coding, even if the transcript models are incomplete or if they only reveal novel exons of already-known genes. In addition to classifying novel transcript models, such methods also have applications in evaluating and revising existing gene annotations ( Butler et al , 2009 ; Clamp et al , 2007 ; Kellis et al , 2003 ; Lin et al , 2007 ; Pruitt et al , 2009 ), and as input features for de novo gene structure predictors ( Alioto and Guigó, 2009 ; Brent, 2008 ). We have previously ( Lin et al , 2008 ) compared numerous methods for determining whether an exon-length nucleotide sequence is likely to be protein coding or non-coding, including single-sequence metrics that analyze the genome of interest only and comparative genomics metrics that use alignments of orthologous regions in the genomes of related species.…”

Mining for Small Translated ORFs

“…34 Finally, codon and amino acid preference metrics, codon pair preferences, and hidden Markov models could be used (Figure 1B). 34,36 Comparative genomics approaches might be combined with ab initio coding region identification using nucleotide periodicities, codon, and codon pair frequencies to increase the precision of prediction. 37 Experimental methods are needed to support translated sORF predictions made purely by computational sequence analysis.…”

“…Another approach to predict the functionality of a putative sORF is to analyze its nucleotide and codon composition (Figure B). , Because a fragment of meaningful written text differs from a random set of letters, a similar difference in compositional statistics applies to functional protein coding sequences. Parameters as simple as nucleotide frequencies are different for coding and noncoding DNA; for example, human coding sequences are generally more GC-rich than noncoding .…”

“…Parameters as simple as nucleotide frequencies are different for coding and noncoding DNA; for example, human coding sequences are generally more GC-rich than noncoding . The likelihood of a sequence to be coding may be assessed more precisely using periodicity in nucleotide frequencies because the first, the second, and the third nucleotides of codons have different average composition . Finally, codon and amino acid preference metrics, codon pair preferences, and hidden Markov models could be used (Figure B). , Comparative genomics approaches might be combined with ab initio coding region identification using nucleotide periodicities, codon, and codon pair frequencies to increase the precision of prediction …”