ESM-2 for General Protein Binding Site Prediction

This model is trained to predict general binding sites of proteins using only the sequence. This is a finetuned version of esm2_t6_8M_UR50D (see here and also here for more info on the base model), trained on this dataset. The data is not filtered by family, and thus the model may be overfit to some degree. In the Hugging Face Inference API widget to the right there are three protein sequence examples. The first is a DNA binding protein truncated to the first 1022 amino acid residues (see UniProt entry here).

The second and third were obtained using EvoProtGrad a Markov Chain Monte Carlo method of (in silico) directed evolution of proteins based on a form of Gibbs sampling. The mutatant-type protein sequences in theory should have similar binding sites to the wild-type protein sequence, but perhaps with higher binding affinity. Testing this out on the model, we see the two proteins indeed have the same binding sites, which validates to some degree that the model has learned to predict binding sites well (and that EvoProtGrad works as intended).

Training

This model was trained on approximately 70,000 proteins with binding site and active site annotations in UniProt. The training split was a random 85/15 split for this version, and does not consider anything in the way of family or sequence similarity. New iterations of the model have been trained on larger datasets (over 200,000 proteins), with the split such that there are no overlapping families, however they seem to overfit much earlier and have significantly worse performance in terms of the training metrics (precision, recall, and F1). To address this we plan to implement LoRA (and hopefully QLoRA).

Training Metrics for the Model in the form of the trainer_state.json can be found here.

epoch 3:
Training Loss Validation Loss Precision	Recall	 F1	      Auc
0.031100	  0.074720	      0.684798	0.966856 0.801743 0.980853

The hyperparameters are:

wandb: 	lr: 0.0004977045729600779
wandb: 	lr_scheduler_type: cosine
wandb: 	max_grad_norm: 0.5
wandb: 	num_train_epochs: 3
wandb: 	per_device_train_batch_size: 8
wandb: 	weight_decay: 0.025

Using the Model

To use the model, try running:

import torch
from transformers import AutoModelForTokenClassification, AutoTokenizer

def predict_binding_sites(model_path, protein_sequences):
    """
    Predict binding sites for a collection of protein sequences.

    Parameters:
    - model_path (str): Path to the saved model.
    - protein_sequences (List[str]): List of protein sequences.

    Returns:
    - List[List[str]]: Predicted labels for each sequence.
    """

    # Load tokenizer and model
    tokenizer = AutoTokenizer.from_pretrained(model_path)
    model = AutoModelForTokenClassification.from_pretrained(model_path)

    # Ensure model is in evaluation mode
    model.eval()

    # Tokenize sequences
    inputs = tokenizer(protein_sequences, return_tensors="pt", padding=True, truncation=True)

    # Move to the same device as model and obtain logits
    with torch.no_grad():
        logits = model(**inputs).logits

    # Obtain predicted labels
    predicted_labels = torch.argmax(logits, dim=-1).cpu().numpy()

    # Convert label IDs to human-readable labels
    id2label = model.config.id2label
    human_readable_labels = [[id2label[label_id] for label_id in sequence] for sequence in predicted_labels]

    return human_readable_labels

# Usage:
model_path = "AmelieSchreiber/esm2_t6_8M_general_binding_sites_v2"  # Replace with your model's path
unseen_proteins = [
    "MKVEEILEKALELVIPDEEEVRKGREAEEELRRRLDELGVEYVFVGSYARNTWLKGSLEIDVFLLFPEEFSKEELRERGLEIGKAVLDSYEIRYAEHPYVHGVVKGVEVDVVPCYKLKEPKNIKSAVDRTPFHHKWLEGRIKGKENEVRLLKGFLKANGIYGAEYKVRGFSGYLCELLIVFYGSFLETVKNARRWTRRTVIDVAKGEVRKGEEFFVVDPVDEKRNVAANLSLDNLARFVHLCREFMEAPSLGFFKPKHPLEIEPERLRKIVEERGTAVFAVKFRKPDIVDDNLYPQLERASRKIFEFLERENFMPLRSAFKASEEFCYLLFECQIKEISRVFRRMGPQFEDERNVKKFLSRNRAFRPFIENGRWWAFEMRKFTTPEEGVRSYASTHWHTLGKNVGESIREYFEIISGEKLFKEPVTAELCEMMGVKD", 
    "MKVEEILEKALELVIPDEEEVRKGREAEEELRRRLDELGVEYVFVGSYARNTWLKGSLEIAVFLLFPEEFSKEELRERGLEIGKAVLDSYEIRYAEHPYVHGVVKGVEVDVVPCYKLKEPKNIKSAVDRTPFHHKWLEGRIKGKENEVRLLKGFLKANGIYGAEYKVRGFSGYLCELLIVFYGSFLETVKNARRWTRRTVIDVAKGEVRKGEEFFVVDPVDEKRNVAANLSLDNLARFVHLCREFMEAPSLGFFKVKHPLEIEPERLRKIVEERGTAVFAVKFRKPDIVDDNLYPQLERASRKIFEFLERENFMPLRSAFKASEEFCYLLFECQIKEISRVFRRMGPQFEDERNVKKFLSRNRAFRPFIENGRWWAFEMRKFTTPEEGVRSYASTHWHTLGKNVGESIREYFEIISGEKLFKEPVTAELCEMMGVKD", 
    "MKVEEILEKALELVIPDEEEVRKGREAEEELRRRLDELGVEAVFVGSYARNTWLKGSLEIAVFLLFPEEFSKEELRERGLEIEKAVLDSYEIRYAEHPYVHGVVKGVEVDVVPCYKLKEPKNIKSAVDRTPFHHKELEGRIKGKENEVRLLKGFLKANGIYGAEYAVRGFSGYLCELLIVFYGSFLETVKNARRWTRRTVIDVAKGEVRKGEEFFVVDPVDEKRNVAANLSLDNLARFVHLCREFMEAPSLGFFKVKHPLEIEPERLRKIVEERGTAVFMVKFRKPDIVDDNLYPQLRRASRKIFEFLERNNFMPLRSAFKASEEFCYLLFECQIKEISDVFRRMGPLFEDERNVKKFLSRNRALRPFIENGRWWIFEMRKFTTPEEGVRSYASTHWHTLGKNVGESIREYFEIISGEKLFKEPVTAELCRMMGVKD", 
    "MKVEEILEKALELVIPDEEEVRKGREAEEELRRRLDELGVEAVFVGSYARNTWLKGSLEIAVFLLFPEEFSKEELRERGLEIEKAVLDSYGIRYAEHPYVHGVVKGVELDVVPCYKLKEPKNIKSAVDRTPFHHKELEGRIKGKENEYRSLKGFLKANGIYGAEYAVRGFSGYLCELLIVFYGSFLETVKNARRWTRKTVIDVAKGEVRKGEEFFVVDPVDEKRNVAALLSLDNLARFVHLCREFMEAVSLGFFKVKHPLEIEPERLRKIVEERGTAVFMVKFRKPDIVDDNLYPQLRRASRKIFEFLERNNFMPLRRAFKASEEFCYLLFEQQIKEISDVFRRMGPLFEDERNVKKFLSRNRALRPFIENGRWWIFEMRKFTTPEEGVRSYASTHWHTLGKNVGESIREYFEIIEGEKLFKEPVTAELCRMMGVKD"
]  # Replace with your protein sequences
predictions = predict_binding_sites(model_path, unseen_proteins)
predictions