prediction.py

import streamlit as st
import torch
import torch.nn as nn
import numpy as np
import joblib
from transformers import AutoTokenizer, AutoModel
from rdkit import Chem
from rdkit.Chem import Descriptors
from rdkit.Chem import AllChem
from datetime import datetime
from db import get_database  # This must be available in your repo
import random

# ------------------------ Ensuring Deterministic Behavior ------------------------
random.seed(42)
np.random.seed(42)
torch.manual_seed(42)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
# Check if CUDA is available for GPU acceleration
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# ------------------------ Load ChemBERTa Model + Tokenizer ------------------------
@st.cache_resource
def load_chemberta():
    tokenizer = AutoTokenizer.from_pretrained("seyonec/ChemBERTa-zinc-base-v1")
    model = AutoModel.from_pretrained("seyonec/ChemBERTa-zinc-base-v1")
    model.eval()
    model.to(device)  # Send model to GPU if available
    return tokenizer, model


# ------------------------ Load Scalers ------------------------
scalers = {
    "Tensile Strength": joblib.load("scaler_Tensile_strength_Mpa_.joblib"),
    "Ionization Energy": joblib.load("scaler_Ionization_Energy_eV_.joblib"),
    "Electron Affinity": joblib.load("scaler_Electron_Affinity_eV_.joblib"),
    "logP": joblib.load("scaler_LogP.joblib"),
    "Refractive Index": joblib.load("scaler_Refractive_Index.joblib"),
    "Molecular Weight": joblib.load("scaler_Molecular_Weight_g_mol_.joblib")
}

# ------------------------ Transformer Model ------------------------
class TransformerRegressor(nn.Module):
    def __init__(self, input_dim=2058, embedding_dim=768, ff_dim=1024, num_layers=2, output_dim=6):
        super().__init__()
        self.feat_proj = nn.Linear(input_dim, embedding_dim)
        encoder_layer = nn.TransformerEncoderLayer(
            d_model=embedding_dim,
            nhead=8,
            dim_feedforward=ff_dim,
            dropout=0.0,  # No dropout for consistency
            batch_first=True
        )
        self.transformer_encoder = nn.TransformerEncoder(encoder_layer, num_layers=num_layers)
        self.regression_head = nn.Sequential(
            nn.Linear(embedding_dim, 256),
            nn.ReLU(),
            nn.Linear(256, 128),
            nn.ReLU(),
            nn.Linear(128, output_dim)
        )

    def forward(self, x):
        x = self.feat_proj(x)
        x = self.transformer_encoder(x)
        x = x.mean(dim=1)
        return self.regression_head(x)

@st.cache_resource
def load_model():
    # Initialize the model architecture first
    model = TransformerRegressor()

    # Load the state_dict (weights) from the saved model file
    state_dict = torch.load("transformer_model.bin", map_location=device)  # Ensure loading on the correct device

    # Load the state_dict into the model
    model.load_state_dict(state_dict)

    # Set the model to evaluation mode
    model.eval()
    model.to(device)  # Send model to GPU if available
    return model
# ------------------------ Descriptors ------------------------
def compute_descriptors(smiles: str):
    mol = Chem.MolFromSmiles(smiles)
    if mol is None:
        raise ValueError("Invalid SMILES string.")
    
    descriptors = [
        Descriptors.MolWt(mol),
        Descriptors.MolLogP(mol),
        Descriptors.TPSA(mol),
        Descriptors.NumRotatableBonds(mol),
        Descriptors.NumHDonors(mol),
        Descriptors.NumHAcceptors(mol),
        Descriptors.FractionCSP3(mol),
        Descriptors.HeavyAtomCount(mol),
        Descriptors.RingCount(mol),
        Descriptors.MolMR(mol)
    ]
    return np.array(descriptors, dtype=np.float32)

# ------------------------ Fingerprints ------------------------
def get_morgan_fingerprint(smiles, radius=2, n_bits=1280):
    mol = Chem.MolFromSmiles(smiles)
    if mol is None:
        raise ValueError("Invalid SMILES string.")
    fp = AllChem.GetMorganFingerprintAsBitVect(mol, radius, nBits=n_bits)
    return np.array(fp, dtype=np.float32).reshape(1, -1)

# ------------------------ Embedding ------------------------
def get_chemberta_embedding(smiles: str):
    inputs = tokenizer(smiles, return_tensors="pt")
    with torch.no_grad():
        outputs = chemberta(**inputs)
    return outputs.last_hidden_state.mean(dim=1)  # Use average instead of CLS token

# ------------------------ Save to DB ------------------------
def save_to_db(smiles, predictions):
    predictions_clean = {k: float(v) for k, v in predictions.items()}
    doc = {
        "smiles": smiles,
        "predictions": predictions_clean,
        "timestamp": datetime.now()
    }
    db = get_database()
    db["polymer_predictions"].insert_one(doc)

# ------------------------ Streamlit App ------------------------
def show():
    st.markdown("<h1 style='text-align: center; color: #4CAF50;'>🔬 Polymer Property Prediction</h1>", unsafe_allow_html=True)
    st.markdown("<hr style='border: 1px solid #ccc;'>", unsafe_allow_html=True)

    smiles_input = st.text_input("Enter SMILES Representation of Polymer")

    if st.button("Predict"):
        try:
            mol = Chem.MolFromSmiles(smiles_input)
            if mol is None:
                st.error("Invalid SMILES string.")
                return

            descriptors = compute_descriptors(smiles_input)
            descriptors_tensor = torch.tensor(descriptors, dtype=torch.float32).unsqueeze(0)

            fingerprint = get_morgan_fingerprint(smiles_input)
            fingerprint_tensor = torch.tensor(fingerprint, dtype=torch.float32)

            embedding = get_chemberta_embedding(smiles_input)

            combined_input = torch.cat([embedding, descriptors_tensor, fingerprint_tensor], dim=1)
            combined = combined_input.unsqueeze(1)

            with torch.no_grad():
                preds = model(combined)

            preds_np = preds.numpy()
            keys = list(scalers.keys())

            preds_rescaled = np.concatenate([
                scalers[keys[i]].inverse_transform(preds_np[:, [i]])
                for i in range(6)
            ], axis=1)

            results = {key: round(val, 4) for key, val in zip(keys, preds_rescaled.flatten())}

            st.success("Predicted Properties:")
            for key, val in results.items():
                st.markdown(f"**{key}**: {val}")

            save_to_db(smiles_input, results)

        except Exception as e:
            st.error(f"Prediction failed: {e}")