README.md

---
license: cc-by-nc-2.0
library_name: transformers
datasets:
- CCDS
- Ensembl
pipeline_tag: feature-extraction
tags:
- protein language model
- biology
widget:
- text: >-
    ( Z E V L P Y G D E K L S P Y G D G G D V G Q I F s C B L Q D T N N F F G A
    g Q N K % O P K L G Q I G % S K % u u i e d d R i d D V L k n ( T D K @ p p
    ^ v
  example_title: Feature extraction
---

# cdsBERT
<img src="https://cdn-uploads.huggingface.co/production/uploads/62f2bd3bdb7cbd214b658c48/yA-f7tnvNNV52DK2QYNq_.png" width="350">

## Model description

[cdsBERT+](https://doi.org/10.1101/2023.09.15.558027) is a pLM with a codon vocabulary that was seeded with [ProtBERT](https://huggingface.co/Rostlab/prot_bert_bfd) and trained with a novel vocabulary extension pipeline called MELD. cdsBERT+ offers a highly biologically relevant latent space with excellent EC number prediction surpassing ProtBERT.
Specifically, this is the half-precision checkpoint after student-teacher knowledge distillation with Ankh-base.

## How to use

```python
# Imports
import re
import torch
import torch.nn.functional as F
from transformers import BertModel, BertTokenizer

model = BertModel.from_pretrained('lhallee/cdsBERT') # load model
tokenizer = BertTokenizer.from_pretrained('lhallee/cdsBERT') # load tokenizer
device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu') # gather device
model.to(device) # move to device
model.eval() # put in eval mode

sequence = '(ZEVLPYGDEKLSPYGDGGDVGQIFsC#LQDTNNFFGAgQNK%OPKLGQIG%SK%uuieddRidDVLkn(TDK@pp^v]' # CCDS207.1|Hs110|chr1
sequence = ' '.join(list(sequence)) # need spaces in-between codons

example = tokenizer(sequence, return_tensors='pt', padding=False).to(device) # tokenize example
with torch.no_grad():
    matrix_embedding = model(**example).last_hidden_state.cpu()

vector_embedding = matrix_embedding.mean(dim=0)
```

## Intended use and limitations
cdsBERT+ serves as a general-purpose protein language model with a codon vocabulary. Fine-tuning with Huggingface transformers models like BertForSequenceClassification enables downstream classification and regression tasks. Currently, the base capability enables feature extraction. The based checkpoint after MLM, cdsBERT, can conduct mask-filling.

## Our lab
The [Gleghorn lab](https://www.gleghornlab.com/) is an interdisciplinary research group at the University of Delaware that focuses on solving translational problems with our expertise in engineering, biology, and chemistry. We develop inexpensive and reliable tools to study organ development, maternal-fetal health, and drug delivery. Recently we have begun exploration into protein language models and strive to make protein design and annotation accessible.

## Please cite
@article {Hallee_cds_2023,
	author = {Logan Hallee, Nikolaos Rafailidis, and Jason P. Gleghorn},
	title = {cdsBERT - Extending Protein Language Models with Codon Awareness},
	year = {2023},
	doi = {10.1101/2023.09.15.558027},
	publisher = {Cold Spring Harbor Laboratory},
	journal = {bioRxiv}
}