ElnaggarLab / ankh3-xl

Model Details:

Ankh3 is a protein language model that is jointly optimized on two objectives:

• Masked language modeling with multiple masking probabilities
• Protein sequence completion.

1. Masked Language Modeling:

• The idea of this task is to intentionally 'corrupt' an input protein sequence by masking a certain percentage (X%) of its individual tokens (amino acids), and then train the model to reconstruct the original sequence.

• Example on a protein sequence before and after corruption:

  Original protein sequence: MKAYVLINSRGP

  This sequence will be masked/corrupted using sentinel tokens as shown below: Sequence after corruption: M A Y L I S R G

  The decoder learns to correspond each sentinel token to the actual amino acid that was masked. In this example: K means that corresponds to the "K" amino acid and so on.

  Decoder output: K V N P

1. Protein Sequence Completion:

• The idea of this task is to cut the input sequence into two segments, where the first segment is fed to the encoder and the decoder is tasked to auto-regressively generate the second segment conditioned on the first segment representation outputted from the encoder.

• Example on protein sequence completion:

  Original sequence: MKAYVLINSRGP

  We will pass "MKAYVL" of it to the encoder, and the decoder is trained that given the representation of the first part provided by the encoder, it should output the second part which is: "INSRGP"

How to use:

For Embedding Extraction:

from transformers import T5ForConditionalGeneration, T5Tokenizer, T5EncoderModel
import torch

# Random sequence from uniprot, most likely Ankh3 saw it during pre-training.
sequence = "MDTAYPREDTRAPTPSKAGAHTALTLGAPHPPPRDHLIWSVFSTLYLNLCCLGFLALAYSIKARDQKVVGDLEAARRFGSKAKCYNILAAMWTLVPPLLLLGLVVTGALHLARLAKDSAAFFSTKFDDADYD"

ckpt = "ElnaggarLab/ankh3-xl"

# Make sure that you must use `T5Tokenizer` not `AutoTokenizer`.
tokenizer = T5Tokenizer.from_pretrained(ckpt)

# To use the encoder representation using the NLU prefix:
encoder_model = T5EncoderModel.from_pretrained(ckpt).eval()


# For extracting embeddings, consider trying the '[S2S]' prefix.
# Since this prefix was specifically used to denote sequence completion
# during the model's pre-training, its use can sometimes
# lead to improved embedding quality.

nlu_sequence = "[NLU]" + sequence
encoded_nlu_sequence = tokenizer(nlu_sequence, add_special_tokens=True, return_tensors="pt", is_split_into_words=False)

with torch.no_grad():
  embedding = encoder_model(**encoded_nlu_sequence)

For Sequence Completion:

from transformers import T5ForConditionalGeneration, T5Tokenizer
from transformers.generation import GenerationConfig
import torch

sequence = "MDTAYPREDTRAPTPSKAGAHTALTLGAPHPPPRDHLIWSVFSTLYLNLCCLGFLALAYSIKARDQKVVGDLEAARRFGSKAKCYNILAAMWTLVPPLLLLGLVVTGALHLARLAKDSAAFFSTKFDDADYD"

ckpt = "ElnaggarLab/ankh3-xl"
tokenizer = T5Tokenizer.from_pretrained(ckpt)
# To use the sequence to sequence task using the S2S prefix:
model = T5ForConditionalGeneration.from_pretrained(ckpt).eval()


half_length = int(len(sequence) * 0.5)
s2s_sequence = "[S2S]" + sequence[:half_length]
encoded_s2s_sequence = tokenizer(s2s_sequence, add_special_tokens=True, return_tensors="pt", is_split_into_words=False)
# + 1 to account for the start of sequence token.
gen_config = GenerationConfig(min_length=half_length + 1, max_length=half_length + 1, do_sample=False, num_beams=1)
generated_sequence = model.generate(encoded_s2s_sequence["input_ids"], gen_config, )
predicted_sequence = sequence[:half_length] + tokenizer.batch_decode(generated_sequence)[0]