Evigan: a hidden variable model for integrating gene evidence for eukaryotic gene prediction

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Gene annotation quality varies, necessitating robust methods for integrating diverse evidence.
• Automated gene model generation requires a framework to handle inconsistent data.

Purpose of the Study:

• To develop an automated gene annotation program (Evigan) for eukaryotic genomes.
• To integrate multiple, heterogeneous sources of gene evidence using probabilistic inference.

Main Methods:

• Evigan employs a dynamic Bayes network for probabilistic inference.
• Parameters are learned to maximize evidence probability, and gene predictions are derived via maximum likelihood decoding.
• The system accommodates various evidence types, including gene finders, BLAST hits, ESTs, and splice sites.

Main Results:

• Evigan integrates diverse evidence, including gene finders, BLAST, ESTs, and splice sites, to produce accurate gene models.
• The program generates consensus gene predictions as ranked lists of n-best models with associated probabilities.
• Evigan demonstrated superior performance compared to individual evidence sources on human, Plasmodium vivax, and Arabidopsis thaliana genomes.

Conclusions:

• Evigan provides a powerful probabilistic framework for automated gene annotation, especially for novel genomes.
• The ability to generate ranked alternative transcripts aids in identifying alternatively spliced forms.
• Evigan's performance surpasses individual data sources, offering a significant advancement in gene annotation.