ESC

Models

We develop DNA language models for plant genomes — from large-scale foundation models to tailored model suites. This page collects the lab’s model zoo. All model weights are openly available on Hugging Face and ModelScope.

Featured
PlantHelixSeek paper Figure 1: pretraining corpus, model architecture and benchmarks

PlantHelixSeek

Foundation Model · MoE
GitHub Hugging Face ModelScope

A plant DNA foundation model pretrained with bidirectional masked language modeling at single-nucleotide resolution, combining interleaved attention, Delta linear-attention, and MLA layers with a sparse Mixture-of-Experts feed-forward network.

params: 470M total / ~240M active

context: 8,192 bp

tokenizer: 1 nucleotide = 1 token

zoo: 1 base + 11 fine-tuned + 2 validation models

Tasks: promoter & terminator strength, lncRNA, open chromatin, conservation, H3K27ac/H3K27me3/H3K4me3, translation on/off, gene structure (BILOU), cis-regulatory elements

Manuscript in preparation.

PDLLMs paper Figure 1: model overview

PDLLMs — Plant DNA Large Language Models

Model Suite
GitHub Docs Demo Paper

A group of tailored DNA large language models for analyzing plant genomes (Molecular Plant 2025, 18(2):175–178) — a suite of pretrained and fine-tuned checkpoints for regulatory element prediction and downstream tasks.

More models: Plant DNAGemma · Plant DNAGPT · Plant DNAMamba · CrossDNA — see the DNALLM documentation for details.

dnallm Toolkit

All models are accessible through the unified dnallm toolkit — install with pip install dnallm — see DNALLM on GitHub and the documentation.