Compositional features of eukaryotic genomes for checking predicted genes

Gene prediction relies on the identification of characteristic features of coding sequences that distinguish them from non-coding DNA. The recent large-scale sequencing of entire genomes from higher eukaryotes, in conjunction with currently used gene prediction algorithms, has provided an abundance of putative genes that can now be analysed for their compositional properties. Strong, systematic differences still exist, in several species, between the compositional properties of sets of ex novo predicted genes and genes that have been experimentally detected and/or verified. This is particularly evident in the estimated gene set (>45,000 genes) of the recently sequenced rice genome, where roughly half the predicted genes are compositionally unusual and have no known orthologues in the dicot Arabidopsis. In a few cases such differences might suggest a bias in experimental gene-finding protocols, but the quasi-random nature of the compositionally aberrant predicted genes is a strong indication that many, if not most, of them are false positives. It therefore appears that some important features of coding regions have not yet been taken into account in existing gene prediction programs. Statistical base compositional properties of curated gene data sets from vertebrates, which we briefly review here, should therefore provide a useful benchmark for fine-tuning probabilistic gene models and model parameters that are currently in use.