PlantCaduceus (PlantCad)

Learns from All Plants

Improve Yield

Genome Modeling

Discover Genes

PlantCaduceus, abbreviated as PlantCAD, is a DNA language model trained to capture 160 million years of plant genome evolution. Built on the state-of-the-art Mamba architecture, it models diverse plant genomes at single-nucleotide resolution using a masked language modeling objective. By learning patterns of evolutionary conservation and divergence, PlantCAD enables cross-species genome annotation and deleterious mutation prediction. Its versatility makes it a powerful AI tool for both basic plant genomics and applied crop improvement.

Deleterious mutations are genetic changes that reduce organismal fitness, including traits like yield. PlantCAD is pre-trained to predict masked nucleotides from sequence context, where high probability indicates evolutionary conservation. A zero-shot score—the log-ratio of ALT to REF probabilities—quantifies disruption caused by a mutation (left). As shown (right), variants with known functional impacts—such as start_lost, stop_gained, and splice_site mutations—tend to have strongly negative scores. This enables PlantCAD to prioritize deleterious mutations directly from sequence.

Transcription and translation are core processes of the central dogma—and accurate prediction of their junction sites is essential for genome annotation. When fine-tuned on limited data from Arabidopsis, PlantCAD demonstrates exceptional cross-species performance, maintaining high accuracy even in maize, which diverged over 160 million years ago. It outperforms other DNA language models (gray) by up to 7-fold and surpasses supervised deep learning baselines (light gray).

Back to top

For more information on how to get started, please visit our GitHub repository.