gLM2_650M
gLM2 is a mixed-modality genomic language model, trained on the OMG Dataset.
The model encodes a genomic scaffold with both both amino-acid and DNA tokens.
gLM2 is trained at two scales: 150M (available at tattabio/gLM2_150M) and 650M parameters.
See https://github.com/TattaBio/gLM2 for inference scripts.
Model Description
gLM2 is a transformer encoder trained with the masked language modeling objective.
It encodes a genomic contig as a sequence of protein coding sequences (CDS) and DNA inter-genic sequences (IGS).
CDS elements are tokenized using per-amino acid tokens, and IGS elements are tokenized using per-nucleotide tokens.
- To encode the genomic strand, we prepended each genomic element with a special token, either
<+>or<->to indicate the positive and negative strands. - To avoid collision between amino acid and nucleotide tokens, the tokenizer expects all amino acids to be uppercase, and all nucleotides to be lowercase.
UPDATE(09/2024): We updated the model with longer context length (4096 tokens vs. 2048 tokens) and per-nucleotide IGS tokenization instead of BPE.
Getting Started
import torch
from transformers import AutoModel, AutoTokenizer
model = AutoModel.from_pretrained('tattabio/gLM2_650M', torch_dtype=torch.bfloat16, trust_remote_code=True).cuda()
tokenizer = AutoTokenizer.from_pretrained('tattabio/gLM2_650M', trust_remote_code=True)
# A contig with two proteins and an inter-genic sequence.
# NOTE: Nucleotides should always be lowercase, and prepended with `<+>`.
sequence = "<+>MALTKVEKRNRIKRRVRGKISGTQASPRLSVYKSNK<+>aatttaaggaa<->MLGIDNIERVKPGGLELVDRLVAVNRVTKVTKGGRAFGFSAIVVVGNED"
# Tokenize the sequence.
encodings = tokenizer([sequence], return_tensors='pt')
# Extract embeddings.
with torch.no_grad():
embeddings = model(encodings.input_ids.cuda(), output_hidden_states=True).last_hidden_state
Training Data
gLM2 is trained on the OMG dataset.
To improve the dataset balance and remove near-duplicate examples, the data is tokenized and pruned by applying Semantic Deduplication SemDedup.
We use an embedding distance threshold of 2e-3, resulting in 49% of the dataset being pruned.
Training Details
- Pretraining tokens: 315B
- Context length: 4096
- Masking rate: 30%
- Learning rate: 1e-3
- Optimizer: AdamW (betas = (0.9, 0.95))
- Mixed precision training: bfloat16
- Weight decay: 0.1
Citation
BioRxiv: https://www.biorxiv.org/content/10.1101/2024.08.14.607850
BibTeX:
author = {Cornman, Andre and West-Roberts, Jacob and Camargo, Antonio Pedro and Roux, Simon and Beracochea, Martin and Mirdita, Milot and Ovchinnikov, Sergey and Hwang, Yunha},
title = {The OMG dataset: An Open MetaGenomic corpus for mixed-modality genomic language modeling},
elocation-id = {2024.08.14.607850},
year = {2024},
doi = {10.1101/2024.08.14.607850},
publisher = {Cold Spring Harbor Laboratory},
URL = {https://www.biorxiv.org/content/early/2024/08/17/2024.08.14.607850},
eprint = {https://www.biorxiv.org/content/early/2024/08/17/2024.08.14.607850.full.pdf},
journal = {bioRxiv}
}