AI Insight
Researchers developed Creator 1.0, a 2-billion-parameter foundation model for plant genomics trained on 1,000 plant species using a novel phylogeny-guided stagewise pretraining approach. The model achieved top performance among 46 models on plant-related genomic tasks, and the staged training process produced 52 different checkpoints optimized for various data availability scenarios. Notably, different training stages performed best depending on task type and data amount, with earlier checkpoints excelling in low-data situations while later stages were superior with full datasets.
Why it matters
This work provides a powerful computational tool for analyzing plant genomes and demonstrates that matching model checkpoints to specific tasks and data availability can improve performance. The approach could accelerate plant breeding, crop improvement, and understanding of plant gene function, particularly in data-limited scenarios common in agricultural genomics.
Understand the Science
⚠️ Preprint – Noch nicht peer-reviewed
Dieser Artikel wurde noch nicht von unabhängigen Experten begutachtet. Die Ergebnisse sind vorläufig und sollten mit Vorsicht interpretiert werden.
Genomic language models have advanced DNA sequence analysis, but their applicability to plant-specific tasks and representation changes during pretraining remain poorly characterized. Here, we present Creator 1.0, a 2-billion-parameter parameter plant genome foundation model pretrained on 1,000 plant genomes using phylogeny-guided stagewise pretraining (PGSP). Training comprised 83 consecutive stages and produced 52 cumulative checkpoints associated with distinct plant orders. On GENEB, the stage 83 checkpoint ranked first among 46 models across 10 plant-related tasks, with a mean Matthews correlation coefficient (MCC) of 0.565, and fourth across all 100 tasks. Evaluation of 52 checkpoints on six plant lncRNA tasks identified different optimal stages under full-data, 10-shot and 1-shot settings, corresponding to stages 71, 72 and 29, respectively. Checkpoint performance varied across tasks, without consistent advantages for corresponding plant orders. These findings demonstrate the competitive performance of Creator 1.0 and highlight the potential of PGSP to provide task-adaptable checkpoints for plant genomic applications.