Finding Genes in Compositionally Heterogeneous Plants

A mainstream approach to ab initio gene finding in compositionally heterogeneous genomes is reduction of the problem to homogeneous case by clustering together long genomic DNA fragments with similar nucleotide composition under assumption that any given gene will belong to a single fragment and a single cluster. 

In some plant genomes, specifically grasses, GC content of the first exon in a gene can differ from the GC content of the last exon by more than by 20%. Therefore, local compositional variability makes a single gene a member of at least two clusters and ruins the segmentation and clustering approach stated above.

We present here an algorithm for gene finding in locally heterogeneous genomes implementing another approach to estimation of model parameters and to assignment of particular parameters to a given genomic locus and even given sequence position.

Parameters of the algorithm can be estimated from training sets or by unsupervised training from anonymous genomic DNA. We demonstrate a high performance of the algorithm on grass genomes: rice and brachypodium.