BiooBang-1.0

BiooBang is an advanced biological language model designed to integrate protein amino acid sequences and mRNA coding sequences (CDS) within a unified framework. Built on a Transformer-based prefix decoder architecture, BiooBang leverages the principles of natural language processing to treat protein and CDS sequences as biological “languages”, enabling self-supervised learning for comprehensive training.

Use Case

The source code corresponding to this model is publicly available on GitHub at: https://github.com/lonelycrab888/BiooBang

After installing BiooBang, you can use:

import torch

# ========== Set device
device = "cuda:0"

# ========== Prepare Data
data = [
    ("Protein", "MASSDKQTSPKPPPSPSPLRNSKFCQSNMRILIS"),
    ("RNA", "ATGGCGTCTAGTGATAAACAAACAAGCCCAAAGCCTCCTCCTTCACCGTCTCCTCTCCGTAATT")
]

# ========== BiooBang Model
from model.modeling_UniBioseq import UniBioseqForEmbedding
from model.tokenization_UniBioseq import UBSLMTokenizer
model = UniBioseqForEmbedding.from_pretrained("lonelycrab88/BiooBang-1.0")
tokenizer = UBSLMTokenizer.from_pretrained("lonelycrab88/BiooBang-1.0")
model.eval()
model.to(device)

# ========== get Embeddings
embeddings = {}
hidden_states = {}
for name,input_seq in data:
    input_ids = tokenizer(input_seq, return_tensors="pt")['input_ids'].to(device)
    with torch.no_grad():
        # get sequence embedding 
        embeddings[name] = model(input_ids).logits
        # get last hidden states (token embeddings)
        hidden_states[name] = model(input_ids).hidden_states[:,1:-1,:]

# ========== generate CDS
from transformers.generation.logits_process import LogitsProcessorList
from model.UBL_utils import CodonLogitsProcessor
from model.modeling_UniBioseq import UniBioseqForCausalLM
tokenizer = UBSLMTokenizer.from_pretrained("lonelycrab88/BiooBang-1.0")
model = UniBioseqForCausalLM.from_pretrained("lonelycrab88/BiooBang-1.0", device_map='auto')

protein_prompt = "MASSDKQTSPKPPPSPSPLRNSKFCQSNMRILIS"
input_ids = torch.tensor([tokenizer.encode(input_protein)+[36]]).to(model.device)
max_length = 4*len(input_protein)+6

logits_processor = LogitsProcessorList()
logits_processor.append(CodonLogitsProcessor(input_protein, tokenizer, len(input_protein)))
result = model.generate(input_ids, max_length = max_length, num_beams = 10, logits_processor=logits_processor, low_memory=True, num_return_sequences=1)
result_CDS_tok = tokenizer.decode(result[0][len(input_protein)+3:].tolist()).replace(" ","").upper()

Citing this Work

Please cite our paper:

@article {Zhao2024.10.24.620004,
	author = {Zhao, Heng-Rui and Cheng, Meng-Ting and Zhu, Jinhua and Wang, Hao and Yang, Xiang-Rui and Wang, Bo and Sun, Yuan-Xin and Fang, Ming-Hao and Chen, Enhong and Li, Houqiang and Han, Shu-Jing and Chen, Yuxing and Zhou, Cong-Zhao},
	title = {Integration of protein and coding sequences enables mutual augmentation of the language model},
	elocation-id = {2024.10.24.620004},
	year = {2024},
	doi = {10.1101/2024.10.24.620004},
	publisher = {Cold Spring Harbor Laboratory},
	URL = {https://www.biorxiv.org/content/early/2024/10/29/2024.10.24.620004},
	eprint = {https://www.biorxiv.org/content/early/2024/10/29/2024.10.24.620004.full.pdf},
	journal = {bioRxiv}
}

Contacts

If you’re interested in other cell lines and open to collaboration, please don’t hesitate to contact us!