Gene prediction in novel fungal genomes using an ab initio algorithm with unsupervised training

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Accurate identification of protein-coding genes is crucial for understanding genome function.
• Existing gene prediction algorithms often rely on supervised training, requiring extensive curated datasets.
• Fungal genomes present unique challenges due to variations in gene organization and splicing mechanisms.

Purpose of the Study:

• To introduce GeneMark-ES version 2, a novel ab initio algorithm for protein-coding gene identification in fungal genomes.
• To develop an unsupervised approach that eliminates the need for predetermined training sets.
• To enhance fungal gene annotation efficiency and accuracy.

Main Methods:

• Developed an ab initio algorithm, GeneMark-ES version 2, utilizing iterative unsupervised training on genomic sequences.
• Enhanced the hidden Markov model (HMM) intron submodel to accommodate variations in fungal splicing, including branch point sites.
• Applied the algorithm to fungal genomes, comparing its performance against supervised training methods.

Main Results:

• GeneMark-ES version 2 achieves high accuracy in identifying protein-coding genes at both exon and whole-gene levels.
• The algorithm demonstrates favorable accuracy compared to existing gene finders that use supervised training.
• Application to known fungal genomes reveals substantial improvements in annotation quality over current methods.

Conclusions:

• GeneMark-ES version 2 offers a robust and accurate solution for fungal gene prediction without requiring training data.
• The unsupervised, self-training nature of the algorithm streamlines and accelerates the annotation process for fungal genome sequencing projects.
• This method is applicable across diverse fungal phyla, accommodating variations in splicing mechanisms.