ReadSeeker: A DNABERT based de-novo read-level gene predictor

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Next-generation sequencing (NGS) generates vast amounts of short reads.
• Differentiating protein-coding (CDS) from non-protein-coding (non-CDS) regions is crucial for genomic annotation.
• Current methods often rely on reference sequences, limiting application in novel or unannotated genomes.

Purpose of the Study:

• To develop and evaluate ReadSeeker, a DNABERT-based model for classifying NGS short reads.
• To enable accurate CDS/non-CDS differentiation without reliance on reference genomes.
• To assess the model's performance across diverse biological datasets.

Main Methods:

• Fine-tuning a DNABERT model named ReadSeeker.
• Training on approximately 3 million synthetic reads from annotated viral, bacterial, and mammalian genomic elements.
• Evaluating performance on real-world human, viral, and bacterial sequencing data.

Main Results:

• ReadSeeker achieved high accuracy exceeding 94% in differentiating NGS short reads.
• Receiver Operating Characteristic Area Under the Curve (ROC-AUC) scores surpassed 98% in most evaluations.
• The model demonstrated robust performance across diverse sample types (human, viral, bacterial).

Conclusions:

• ReadSeeker provides a robust, reference-free method for classifying DNA sequencing reads.
• The model significantly advances genomic annotation capabilities, particularly for uncharacterized or novel sequences.
• High accuracy and AUC scores indicate ReadSeeker's potential for broad application in genomic research.