PlantCAD2 and GeneCAD

Foundation models for plant genomics from Cornell (first author Jingjing Zhai, with Volodymyr Kuleshov, Matt Pennell and Edward Buckler of Cornell and USDA-ARS) and Open Athena, whose engineers Eric Czech and Betsy Cannon are co-authors. PlantCAD2 is a DNA language model at single-nucleotide resolution with an 8,192-bp context, pretrained on 65 angiosperm genomes and released with a set of public benchmarks; it succeeds the 2024 PlantCaduceus models (512-bp context, up to 225M parameters). In zero-shot tests the 676M-parameter model captures evolutionary conservation well enough to beat the 7B-parameter Evo 2 on 10 of 12 tasks, and with parameter-efficient fine-tuning it outperforms the 1B-parameter AgroNT on seven cross-species tasks covering accessible chromatin, gene expression and protein translation. The long context matters for distal regulation: accessible chromatin prediction in maize rises from 0.587 to 0.711 AUPRC. Three sizes are on HuggingFace under Apache 2.0, with LoRA fine-tuned variants, and the paper was published in Cell Genomics in 2026 (online August 7).

GeneCAD (preprint November 2025, revised through May 2026) builds on PlantCAD2 for gene annotation. It feeds PlantCAD2's DNA representations to a transformer encoder and a chromosome-scale conditional random field that enforces splice phase and feature order, predicting complete gene models from sequence alone, without species-matched RNA or protein evidence. Across diverse flowering plants, including a complex allotetraploid, it improves transcript F1 by about 9% over Helixer and BRAKER3 and recovers 86% of classical coding sequences. Swapping in an animal DNA model gives an experimental vertebrate version that finds exons well but struggles to rebuild full transcripts across long introns. Its code is Apache 2.0.