Creator 1.0: Phylogeny-Guided Stagewise Pretraining for Plant Genome Foundation Models across 1,000 Species

Creator 1.0, a 2-billion-parameter plant genome foundation model pretrained on 1,000 plant genomes with PGSP, achieves a stage 83 best rank on GENEB with mean MCC 0.565.

• Why it matters: Plant genomic language models’ effectiveness for plant-specific tasks and how representations shift during pretraining are not well characterized, limiting task-adaptable model selection.
• What they did: The model was trained in 83 consecutive PGSP stages, yielding 52 cumulative checkpoints tied to distinct plant orders, then evaluated across 10 plant-related tasks and 100 total tasks.
• The result: On GENEB, stage 83 ranked first among 46 models (mean MCC 0.565) and fourth across all 100 tasks; for six plant lncRNA tasks, optimal stages differ by data regime (stages 71, 72, 29).