PlantGFM

Model Card for Model ID

PlantGFM is a genetic foundation model pre-trained on the complete genome sequences of 12 model plants, encompassing 108 billion nucleotides. Using the Hyena framework with 220 million parameters and a context length of 64K bp, PlantGFM models sequences at single-nucleotide resolution. The model employed a length warm-up strategy, starting with 1K bp fragments and gradually increasing to 64K bp, enhancing training stability and accelerating convergence.

Model Sources

• Repository: PlantGFM
• Manuscript: A Genetic Foundation Model for Discovery and Creation of Plant Genes

Developed by: hu-lab

How to use the model

Install the runtime library first:

pip install transformers

To calculate the embedding of a dna sequence:

import torch
from transformers import PreTrainedTokenizerFast
from plantgfm.modeling_plantgfm import PlantGFMForCausalLM
from plantgfm.configuration_plantgfm import PlantGFMConfig

config = PlantGFMConfig.from_pretrained("hu-lab/PlantGFM")
tokenizer = PreTrainedTokenizerFast.from_pretrained("hu-lab/PlantGFM")
model = PlantGFMForCausalLM.from_pretrained("hu-lab/PlantGFM", config=config)


sequences = ["CCCTAAACCCTAAACCCTAAA", "ATGGCGTGGCTG"]

# get single-nucleotide sequences with space between each base
single_nucleotide_sequences = list(map(lambda seq: " ".join(list(seq)), sequences))


tokenized_sequences = tokenizer(single_nucleotide_sequences, padding="longest")["input_ids"]
input_ids = torch.LongTensor(tokenized_sequences)

embd = model(input_ids=input_ids, output_hidden_states=True)["hidden_states"][0]
print(embd)

Hardware

Model was trained for 468 hours on 8 Nvidia A800-80G GPUs.