Tiberius: end-to-end deep learning with an HMM for gene prediction.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Eukaryotic gene prediction has historically relied on Hidden Markov Models (HMMs) for over 25 years.
• Recent advancements show deep learning combined with HMMs can enhance ab initio gene prediction accuracy.

Purpose of the Study:

• To introduce Tiberius, a novel deep learning-based ab initio gene predictor.
• To improve the accuracy and efficiency of eukaryotic gene prediction.

Main Methods:

• Tiberius integrates convolutional and long short-term memory layers with a differentiable HMM layer.
• A custom gene prediction loss function was utilized.
• The model was trained on mammalian genomes and evaluated on human and other genomes.

Main Results:

• Tiberius significantly outperforms existing ab initio methods, achieving a 62% F1 score on the human genome.
• It accurately predicts the exon-intron structure of human genes in de novo mode.
• Tiberius's accuracy rivals methods using external data, and it is the fastest state-of-the-art gene prediction tool.

Conclusions:

• Tiberius represents a significant advancement in ab initio eukaryotic gene prediction.
• Its deep learning architecture and integrated HMM offer superior accuracy and speed.
• This tool has the potential to accelerate genomic research by improving gene structure identification.