You can deploy this model and start earning money today!

magicslabnu/
DNABERT-2

Name: DNABERT-2
Availability: InStock

Views

No views yet

This is the official pre-trained baseline model introduced in Fast and Low-Cost Genomic Foundation Models via Outlier Removal .

We sincerely appreciate the MosaicML team for the MosaicBERT implementation, which serves as the base of DNABERT-2 development.

DNABERT-2 is a transformer-based genome foundation model trained on multi-species genome.

To load the model from huggingface:

import torch
from transformers import AutoTokenizer, AutoModel

tokenizer = AutoTokenizer.from_pretrained("zhihan1996/DNABERT-2-117M", trust_remote_code=True)
model = AutoModel.from_pretrained("zhihan1996/DNABERT-2-117M", trust_remote_code=True)

To calculate the embedding of a dna sequence

dna = "ACGTAGCATCGGATCTATCTATCGACACTTGGTTATCGATCTACGAGCATCTCGTTAGC"
inputs = tokenizer(dna, return_tensors = 'pt')["input_ids"]
hidden_states = model(inputs)[0] # [1, sequence_length, 768]

# embedding with mean pooling
embedding_mean = torch.mean(hidden_states[0], dim=0)
print(embedding_mean.shape) # expect to be 768

# embedding with max pooling
embedding_max = torch.max(hidden_states[0], dim=0)[0]
print(embedding_max.shape) # expect to be 768

magicslabnu/DNABERT-2

magicslabnu/
DNABERT-2