Plant foundation DNA large language models

The plant DNA large language models (LLMs) contain a series of foundation models based on different model architectures, which are pre-trained on various plant reference genomes.
All the models have a comparable model size between 90 MB and 150 MB, BPE tokenizer is used for tokenization and 8000 tokens are included in the vocabulary.

Developed by: zhangtaolab

Model Sources

Repository: Plant DNA LLMs
Manuscript: Versatile applications of foundation DNA large language models in plant genomes

Architecture

The model is trained based on the InstaDeepAI/nucleotide-transformer-v2-100m-multi-species model with modified tokenizer that replaces k-mer to BPE.

This model is fine-tuned for predicting active core promoters.

How to use

Install the runtime library first:

pip install transformers

Here is a simple code for inference:

from transformers import AutoModelForSequenceClassification, AutoTokenizer, pipeline

model_name = 'plant-nucleotide-transformer-singlebase-promoter'
# load model and tokenizer
model = AutoModelForSequenceClassification.from_pretrained(f'zhangtaolab/{model_name}', trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained(f'zhangtaolab/{model_name}', trust_remote_code=True)

# inference
sequences = ['TTACTAAATTTATAACGATTTTTTATCTAACTTTAGCTCATCAATCTTTACCGTGTCAAAATTTAGTGCCAAGAAGCAGACATGGCCCGATGATCTTTTACCCTGTTTTCATAGCTCGCGAGCCGCGACCTGTGTCCAACCTCAACGGTCACTGCAGTCCCAGCACCTCAGCAGCCTGCGCCTGCCATACCCCCTCCCCCACCCACCCACACACACCATCCGGGCCCACGGTGGGACCCAGATGTCATGCGCTGTACGGGCGAGCAACTAGCCCCCACCTCTTCCCAAGAGGCAAAACCT',
             'GACCTAATGATTAACCAAGGAAAAATGCAAGGATTTGACAAAAATATAGAAGCCAATGCTAGGCGCCTAAGTGAATGGATATGAAACAAAAAGCGAGCAGGCTGTCTATATATGGACAATTAGTTGCATTAATATAGTAGTTTATAATTGCAAGCATGGCACTACATCACAACACCTAAAAGACATGCCGTGATGCTAGAACAGCCATTGAATAAATTAGAAAGAAAGGTTGTGGTTAATTAGTTAACGACCAATCGAGCCTACTAGTATAAATTGTACCTCGTTGTTATGAAGTAATTC']
pipe = pipeline('text-classification', model=model, tokenizer=tokenizer,
                trust_remote_code=True, top_k=None)
results = pipe(sequences)
print(results)

Training data

We use EsmForSequenceClassification to fine-tune the model.
Detailed training procedure can be found in our manuscript.

Hardware

Model was trained on a NVIDIA GTX1080Ti GPU (11 GB).