README.md

---
license: cc-by-nc-sa-4.0
widget:
- text: AAAAGCGACATGACCAAACTGCCCCTCACCCGCCGCACTGATGACCGA
inference: false
tags:
- DNA
- biology
- genomics
datasets:
- zhangtaolab/plant_reference_genomes
---
# 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](https://github.com/zhangtaolab/plant_DNA_LLMs)
- **Manuscript:** [PDLLMs: A group of tailored DNA large language models for analyzing plant genomes](https://doi.org/10.1016/j.molp.2024.12.006)

### Architecture

The model is trained based on the State-Space Mamba-130m model with modified tokenizer specific for DNA sequence.

### How to use

Install the runtime library first:
```bash
pip install transformers
```

Here is a simple code for inference (Note that Mamba model requires NVIDIA GPU for inference):
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_name = 'plant-dnamamba-4mer'
# load model and tokenizer
model = AutoModelForCausalLM.from_pretrained(f'zhangtaolab/{model_name}', trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained(f'zhangtaolab/{model_name}', trust_remote_code=True)

# example sequence and tokenization
sequences = ['ATATACGGCCGNC','GGGTATCGCTTCCGAC']
tokens = tokenizer(sequences,padding="longest")['input_ids']
print(f"Tokenzied sequence: {tokenizer.batch_decode(tokens)}")

# inference
device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
model.to(device)
inputs = tokenizer(sequences, truncation=True, padding='max_length', max_length=512, 
                   return_tensors="pt")
inputs = {k: v.to(device) for k, v in inputs.items()}
outs = model(
    **inputs,
    output_hidden_states=True
)

# get the final layer embeddings and prediction logits
embeddings = outs['hidden_states'][-1].detach().numpy()
logits = outs['logits'].detach().numpy()
```


### Training data
We use CausalLM method to pre-train the model, the tokenized sequence have a maximum length of 512.  
Detailed training procedure can be found in our manuscript.


#### Hardware
Model was pre-trained on a NVIDIA RTX4090 GPU (24 GB).