app.py

import gradio as gr
import joblib
import torch
import numpy as np
from transformers import AutoTokenizer, AutoModel, EsmTokenizer, EsmModel
from rdkit import Chem

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# Load the pretrained models
esm = "facebook/esm2_t12_35M_UR50D" # generate protein embeddings
esm_tokenizer = EsmTokenizer.from_pretrained(esm)
esm_model = EsmModel.from_pretrained(esm).to(device)
chemberta = "DeepChem/ChemBERTa-10M-MTR" # generate ligand embeddings
chemberta_tokenizer = AutoTokenizer.from_pretrained(chemberta)
chemberta_model = AutoModel.from_pretrained(chemberta).to(device)
scaler = joblib.load("scaler.pkl")
pca = joblib.load("pca.pkl")
svr = joblib.load("svr_model.pkl")

def generate_protein_embedding(protein_input, input_type):
    # Generate FASTA string from file
    if input_type == "File":
        mol = Chem.MolFromPDBFile(protein_input)
        if not mol:
            return None
        fasta = Chem.MolToFASTA(mol).splitlines()[1]
    # The input was FASTA
    else:
        fasta = protein_input.strip()
    # Generate protein embedding
    esm_model.eval()
    with torch.no_grad():
        inputs = esm_tokenizer(fasta, return_tensors="pt", padding=True, truncation=True, max_length=1024).to(device)
        outputs = esm_model(**inputs)
        embedding = outputs.last_hidden_state.mean(dim=1).cpu().numpy()
    return embedding

def generate_ligand_embedding(ligand_input, input_type):
    if input_type == "File":
        mol = Chem.MolFromMol2File(ligand_input)
        if not mol:
            return None
        smiles = Chem.MolToSmiles(mol)
    else:
        smiles = ligand_input.strip()
    # Generate compounds embeddings from SMILES
    chemberta_model.eval()
    with torch.no_grad():
        inputs = chemberta_tokenizer(smiles, return_tensors="pt", padding=True, truncation=True).to(device)
        outputs = chemberta_model(**inputs)
        embedding = outputs.last_hidden_state.mean(dim=1).cpu().numpy()
    return embedding

# Convert -logKd predicted value to Kd
def value_conversion(logKa):
    logKd = logKa * -1
    Kd = 10 ** (logKd)
    if Kd >= 1e-3:  
        return f"{Kd * 1e3:.4f} mM"  # Millimolar
    elif Kd >= 1e-6:  
        return f"{Kd * 1e6:.4f} µM"  # Micromolar
    elif Kd >= 1e-9:  
        return f"{Kd * 1e9:.4f} nM"  # Nanomolar
    else:  
        return f"{Kd * 1e12:.4f} pM"  # Picomolar

def predict_affinity(protein_file, protein_fasta, protein_type, ligand_file, ligand_smiles, ligand_type):
    # Determine protein input
    if protein_file is not None:
        protein_input = protein_file
        protein_type = "File"
    elif protein_fasta is not None:
        protein_input = protein_fasta.strip()
        protein_type = "FASTA"
    else:
        return "Error: No valid protein input provided."

    # Determine ligand input
    if ligand_file is not None:
        ligand_input = ligand_file
        ligand_type = "File"
    elif ligand_smiles is not None:
        ligand_input = ligand_smiles.strip()
        ligand_type = "SMILES"
    else:
        return "Error: No valid ligand input provided."

    # Get embeddings
    protein = generate_protein_embedding(protein_input, protein_type)
    ligand = generate_ligand_embedding(ligand_input, ligand_type)

    if protein is None:
        return "Unable to parse protein .pdb file"
    if ligand is None:
        return "Unable to parse ligand .pdb file"
        
    embedding = np.concatenate((protein, ligand), axis=1)
    # Apply scaling and PCA
    svr_input = scaler.transform(embedding)
    svr_input = pca.transform(svr_input)
    # Predict the log binding affinity
    log_prediction = svr.predict(svr_input)[0]
    affinity_value = value_conversion(log_prediction)
    return f"Predicted Binding Affinity:\nlogKa = {log_prediction:.4f}\nKd = {affinity_value}"

def update_inputs(protein_type, ligand_type):
    # Updates visibility and interactivity dynamically
    return (
        gr.update(visible=(protein_type == "File"), interactive=(protein_type == "File")),  
        gr.update(visible=(protein_type == "FASTA"), interactive=(protein_type == "FASTA")),  
        gr.update(visible=(ligand_type == "File"), interactive=(ligand_type == "File")),  
        gr.update(visible=(ligand_type == "SMILES"), interactive=(ligand_type == "SMILES"))  
    )

with gr.Blocks() as iface:
    gr.Markdown("# Predict Protein-Ligand Binding Affinity")
    gr.Markdown("Upload protein and compound files or enter FASTA/SMILES strings to predict binding affinity.")

    with gr.Row():
        protein_type = gr.Radio(["File", "FASTA"], label="Protein Input Type", value="File")
        ligand_type = gr.Radio(["File", "SMILES"], label="Ligand Input Type", value="File")

    protein_file = gr.File(label="Protein .pdb file", visible=True, interactive=True)
    protein_fasta = gr.Textbox(label="Protein FASTA sequence", visible=False, interactive=True)  

    ligand_file = gr.File(label="Ligand .mol2 file", visible=True, interactive=True)
    ligand_smiles = gr.Textbox(label="Ligand SMILES string", visible=False, interactive=True)  

    output = gr.Textbox(label="Prediction Result", lines=3)

    submit_btn = gr.Button("Predict")
    submit_btn.click(
        predict_affinity,
        inputs=[protein_file, protein_fasta, protein_type, ligand_file, ligand_smiles, ligand_type],
        outputs=output
    )

    protein_type.change(update_inputs, inputs=[protein_type, ligand_type], outputs=[protein_file, protein_fasta, ligand_file, ligand_smiles])
    ligand_type.change(update_inputs, inputs=[protein_type, ligand_type], outputs=[protein_file, protein_fasta, ligand_file, ligand_smiles])

iface.launch()