Utilizing Amino Acid Composition and Entropy of Potential Open Reading Frames to Identify Protein-Coding Genes.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Prokaryotic gene finding requires distinguishing protein-coding open reading frames (ORFs) from chance occurrences.
• Current homology-based methods are limited to identifying previously discovered genes.
• Existing gene prediction programs often rely on protein-coding training models, with some allowing *ab initio* model creation.

Purpose of the Study:

• To introduce GOODORFS, a new method for identifying protein-coding genes from all possible ORFs.
• To enhance the accuracy of prokaryotic gene prediction by improving training model creation.
• To evaluate GOODORFS against established gene prediction tools.

Main Methods:

• GOODORFS utilizes amino acid frequencies to calculate an entropy density profile (EDP) for each ORF.
• KMeans clustering is applied to EDPs to group potential coding regions.
• The cluster with the lowest variation is selected as the set of coding ORFs.

Main Results:

• GOODORFS achieved the highest accuracy (0.94) and F1-score (0.85) across 14,179 annotated phage genomes.
• Compared to Glimmer, MED2, PHANOTATE, and Prodigal, GOODORFS demonstrated superior overall performance.
• Glimmer showed the highest precision (0.92), while PHANOTATE had the highest recall (0.96).

Conclusions:

• GOODORFS provides a more accurate and effective approach for identifying protein-coding genes in prokaryotes.
• The entropy density profile and KMeans clustering method offers an improved strategy for training model creation in gene prediction.
• This method addresses limitations of homology-based approaches by enabling *de novo* gene discovery.