A foundational large language model for edible plant genomes.

Area of Science:

• Plant genomics
• Bioinformatics
• Computational biology

Background:

• High-throughput sequencing enables genome-wide molecular phenotype characterization in plants.
• Understanding genetic mechanisms of plant traits is crucial for crop improvement.
• Predictive modeling is key to leveraging genomic data for agricultural applications.

Purpose of the Study:

• Introduce AgroNT, a foundational large language model for plant genomics.
• Evaluate AgroNT's predictive capabilities on regulatory annotations, gene expression, and variant prioritization.
• Establish the Plants Genomic Benchmark (PGB) for deep learning in plant genomics.

Main Methods:

• Training a large language model (AgroNT) on genomes from 48 plant species, focusing on crops.
• Conducting in silico saturation mutagenesis analysis on cassava to assess regulatory mutation impacts.
• Compiling diverse plant genomic datasets to create the Plants Genomic Benchmark (PGB).

Main Results:

• AgroNT achieves state-of-the-art predictions for regulatory annotations, promoter/terminator strength, and tissue-specific gene expression.
• AgroNT effectively prioritizes functional variants.
• A comprehensive resource of predicted regulatory effects for over 10 million cassava mutations is generated.
• The Plants Genomic Benchmark (PGB) is established for evaluating deep learning models.

Conclusions:

• AgroNT demonstrates significant potential for advancing crop genomic improvement through accurate predictive modeling.
• The AgroNT model and PGB dataset provide valuable resources for the plant genomics research community.
• Future research can leverage AgroNT and PGB for enhanced understanding and manipulation of plant genomes.