Annotating the genome at single-nucleotide resolution with DNA foundation models.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Genome annotation is crucial for biological research, but current tools are limited.
• Existing methods often focus on specific elements and require training from scratch on small datasets.

Purpose of the Study:

• To develop a versatile genome annotation model using deep learning.
• To segment multiple genic and regulatory elements at single-nucleotide resolution.
• To leverage pretrained foundation models for improved annotation accuracy and generalization.

Main Methods:

• Framed genome annotation as multilabel semantic segmentation.
• Fine-tuned pretrained DNA foundation models, including Nucleotide Transformer, Enformer, and Borzoi.
• Developed SegmentNT, a general segmentation model processing up to 50-kb DNA sequences.

Main Results:

• SegmentNT achieved state-of-the-art performance in gene annotation, splice site detection, and regulatory element identification.
• Integrating Enformer and Borzoi extended sequence context to 500 kb, enhancing regulatory element detection.
• A SegmentNT model trained on human data generalized to other species, and a multispecies model showed strong cross-species generalization.

Conclusions:

• The proposed methodology offers a powerful and extensible approach for genome annotation.
• Fine-tuning foundation models significantly improves the accuracy and scope of genomic element segmentation.
• The developed model demonstrates robust generalization capabilities across different species, advancing comparative genomics.