Overview test dataset.

ReadSeeker, a newly fine-tuned, DNABERT-based model, differentiates NGS short reads into protein-coding (CDS) and non-protein-coding (non-CDS) categories without requiring known reference sequences. For model training, extensive datasets encompassing viral, bacterial, and mammalian sequences where used. Training involved generating approximately 3 million synthetic reads from annotated genomic elements.
Performance evaluation on real-world datasets, including human, viral, and bacterial samples, revealed ReadSeeker’s high accuracy, exceeding 94%, with ROC-AUC scores above 98% in most cases.