README.md

---
license: cc-by-nc-sa-4.0
tags:
- biology
- protein
- protein language model
- protein embedding
datasets:
- agemagician/uniref50
---

# ANKH2-Large model

Pretrained model on protein sequences using a masked language modeling (MLM) objective. It was introduced in
[this paper](https://arxiv.org/abs/2301.06568) and first released in
[this repository](https://github.com/agemagician/Ankh). This model is trained on uppercase amino acids: it only works with capital letter amino acids.


## Model description

ANKH2-Large is based on the `ANKH-Large` model and was pretrained on a large corpus of protein sequences in a self-supervised fashion.
This means it was pretrained on the raw protein sequences only, with no humans labelling them in any way (which is why it can use lots of
publicly available data) with an automatic process to generate inputs and labels from those protein sequences.

One important difference between this ANKH2-Large model and the original ANKH-Large version is that the model was trained with more number of epochs.

It has been shown that the features extracted from this self-supervised model (LM-embeddings) captured important biophysical properties governing protein shape.
shape.
This implied learning some of the grammar of the language of life realized in protein sequences.

## Intended uses & limitations

The model could be used for protein feature extraction or to be fine-tuned on downstream tasks.
We have noticed in some tasks you can gain more accuracy by fine-tuning the model using lora method rather than using it as a feature extractor.
We have also noticed that for feature extraction, its better to use the feature extracted from the encoder rather than from the decoder.

### How to use

Here is how to use this model to extract the features of a given protein sequence in PyTorch:

```python
sequence_examples = ["PRTEINO", "SEQWENCE"]
# tokenize sequences and pad up to the longest sequence in the batch
ids = tokenizer.batch_encode_plus(sequence_examples, add_special_tokens=True, padding="longest")
input_ids = torch.tensor(ids['input_ids']).to(device)
attention_mask = torch.tensor(ids['attention_mask']).to(device)
# generate embeddings
with torch.no_grad():
    embedding_repr = model(input_ids=input_ids,attention_mask=attention_mask)
# extract embeddings for the first ([0,:]) sequence in the batch while removing padded & special tokens ([0,:7]) 
emb_0 = embedding_repr.last_hidden_state[0,:7] # shape (7 x 1536)
print(f"Shape of per-residue embedding of first sequences: {emb_0.shape}")
# do the same for the second ([1,:]) sequence in the batch while taking into account different sequence lengths ([1,:8])
emb_1 = embedding_repr.last_hidden_state[1,:8] # shape (8 x 1536)
# if you want to derive a single representation (per-protein embedding) for the whole protein
emb_0_per_protein = emb_0.mean(dim=0) # shape (1536)
print(f"Shape of per-protein embedding of first sequences: {emb_0_per_protein.shape}")
```

## Training data

The ANKH2-Large model was pretrained on [UniRef50](https://www.uniprot.org/help/uniref), a dataset consisting of 60 million protein sequences.

## Training procedure

### Preprocessing

The protein sequences are uppercased and tokenized using a single space and a vocabulary size of 25.
The inputs of the model are then of the form:

```
Protein Sequence </s>
```

The preprocessing step was performed on the fly, by cutting and padding the protein sequences up to 512 tokens.

The details of the masking procedure for each sequence are as follows:
- 20% of the amino acids are masked.
- In 100% of the cases, the masked amino acids are replaced by `<extra_id_num>` token, where "num" is a number in range 0 and 115.

### Pretraining

The model was trained on a single TPU Pod V4-256 for 45 epochs in total, using sequence length 512 (batch size 1k).
It was trained using ANKH-Large model as an initial checkpoint, rather than training from scratch.  
It has a total of approximately 2B parameters and was trained using the encoder-decoder architecture.
The optimizer used is Adafactor with linear warmup with linear decay learning rate schedule for pre-training.


## Evaluation results

When the model is used for feature extraction "FE" and parameter efficient fine-tuning "Lora", this model achieves the following results:

Test results :

| Task/Dataset | Method | secondary structure (3-states) | secondary structure (8-states)  |  Localization | Membrane  | Solubility | Fluorescence |
|:-----:|:-----:|:-----:|:-----:|:-----:|:-----:|:-----:|:-----:|
|   CASP12  | FE | comming soon | comming soon |    |    |    |    |
|   CASP12  | Lora | comming soon | comming soon |    |    |    |    |
|   TS115   | FE | comming soon | comming soon |    |    |     |    |
|   TS115   | Lora | comming soon | comming soon |    |    |     |    |
|   CB513   | FE | comming soon | comming soon |    |    |    |    |
|   CB513   | Lora | comming soon | comming soon |    |    |    |    |
|  DeepLoc  | FE |    |    | comming soon | comming soon |    |
|  DeepLoc  | Lora |   |    | comming soon | comming soon |    |    |
|  Solubility  | FE |   |    |   |   |  comming soon |    |
|  Solubility  | Lora |   |    |   |   |  74%  |    |
|  Fluorescence  | FE |   |    |   |   |   |  Comming Soon  |
|  Fluorescence  | Lora |   |    |   |   |    |  68%  |

### BibTeX entry and citation info

```bibtex
@article{elnaggar2023ankh,
  title={Ankh☥: Optimized protein language model unlocks general-purpose modelling},
  author={Elnaggar, Ahmed and Essam, Hazem and Salah-Eldin, Wafaa and Moustafa, Walid and Elkerdawy, Mohamed and Rochereau, Charlotte and Rost, Burkhard},
  journal={bioRxiv},
  pages={2023--01},
  year={2023},
  publisher={Cold Spring Harbor Laboratory}
}
```

> Created by [Ahmed Elnaggar/@Elnaggar_AI](https://twitter.com/Elnaggar_AI) | [LinkedIn](https://www.linkedin.com/in/prof-ahmed-elnaggar/)