Pre-training a lettuce foundation model
Year 5 Research Project

Pre-training a lettuce foundation model

Description

Lettuce is worth more than $6 billion per year, and California is its major producer. For many people in the U.S., it may be the only green vegetable they eat. PlantCAD2 is a plant-specific DNA language model with long-range context that supports functional annotation of genomes and the study of regulatory structures across species (Zhai et al. 2025). Its training set included only one lettuce genotype. The Michelmore lab has near telomere-to-telomere genome assemblies for more than 100 diverse cultivated and wild lettuce accessions, plus extensive short-read and long-read transcriptomic data. This project will continue pre-training PlantCAD2 on these genomes, fine-tune it with additional -omics data for high-quality annotation of genes and regulatory elements, and integrate the model with existing phenotypic data for downstream breeding tasks. We will evaluate the lettuce foundation model against baseline PlantCAD2 models on lettuce-specific gene annotation, regulatory annotation and variant-effect prediction. The work extends AIFS contributions in molecular breeding from AI-enabled phenotyping to genomics, and gives the field a much-needed benchmark for plant genomic foundation models.

Team

Portrait of Christine Diepenbrock

Christine Diepenbrock

Principal Investigator

Portrait of Xin Liu

Xin Liu

Co Principal Investigator

Richard Michelmore

Co Principal Investigator