GeneCAD: Plant Genome Annotation with a DNA Foundation Model.

Area of Science:

• Genomics
• Bioinformatics

Background:

• Accurate plant genome annotation is challenging due to polyploidy and repetitive sequences.
• Existing homology- and RNA-based annotation methods struggle with these complexities.

Purpose of the Study:

• To develop a sequence-only method for accurate plant gene model prediction directly from DNA.
• To overcome limitations of current annotation pipelines in plants.

Main Methods:

• GeneCAD utilizes a plant DNA foundation model (PlantCAD2) and a ModernBERT encoder.
• A chromosome-wide conditional random field enforces biological constraints.
• A protein language model filters repeat-driven open reading frames.
• Public annotations are refined using a masked-motif score and fine-tuned on high-quality references.

Main Results:

• GeneCAD improves transcript-level F1 scores by 8-10% compared to Helixer and BRAKER3 across five angiosperms.
• The method increases exact match transcripts and refines boundaries at start/stop codons and splice junctions.
• Demonstrated accuracy on the allotetraploid *Nicotiana tabacum*.

Conclusions:

• GeneCAD offers an accurate and scalable approach for plant gene model prediction using DNA sequence alone.
• It eliminates the need for species-matched RNA-seq or proteomics data.
• Provides biologically coherent gene models essential for plant genomics research.