prediction.py

import torch
import torch.nn as nn
import joblib
from rdkit import Chem
from rdkit.Chem import Descriptors
from transformers import AutoTokenizer, AutoModel
import numpy as np

# Load tokenizer and model for embeddings
tokenizer = AutoTokenizer.from_pretrained("seyonec/ChemBERTa-zinc-base-v1")
embedding_model = AutoModel.from_pretrained("seyonec/ChemBERTa-zinc-base-v1")

# Load saved scalers (for inverse_transform)
scaler_tensile_strength = joblib.load("scaler_Tensile_strength_Mpa_.joblib")
scaler_ionization_energy = joblib.load("scaler_Ionization_Energy_eV_.joblib")
scaler_electron_affinity = joblib.load("scaler_Electron_Affinity_eV_.joblib")
scaler_logp = joblib.load("scaler_LogP.joblib")
scaler_refractive_index = joblib.load("scaler_Refractive_Index.joblib")
scaler_molecular_weight = joblib.load("scaler_Molecular_Weight_g_mol_.joblib")

# Descriptor function with exact order from training
def compute_descriptors(smiles):
    mol = Chem.MolFromSmiles(smiles)
    if mol is None:
        raise ValueError("Invalid SMILES")

    return np.array([
        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)
    ], dtype=np.float32)

# Define your model class exactly like in training
class TransformerRegressor(nn.Module):
    def __init__(self, input_dim, hidden_dim, num_layers, output_dim):
        super().__init__()
        self.feat_proj = nn.Linear(input_dim, hidden_dim)
        encoder_layer = nn.TransformerEncoderLayer(d_model=hidden_dim, nhead=8)
        self.transformer_encoder = nn.TransformerEncoder(encoder_layer, num_layers=num_layers)
        self.regression_head = nn.Sequential(
            nn.Linear(hidden_dim, 128),
            nn.ReLU(),
            nn.Linear(128, 64),
            nn.ReLU(),
            nn.Linear(64, 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)

# Set model hyperparameters (must match training config)
input_dim = 768  # ChemBERTa embedding size
hidden_dim = 256
num_layers = 2
output_dim = 6   # Number of properties predicted

# Load model
model = TransformerRegressor(input_dim, hidden_dim, num_layers, output_dim)
model.load_state_dict(torch.load("transformer_model.pt", map_location=torch.device("cpu")))
model.eval()

# Main prediction function
def predict_properties(smiles):
    try:
        descriptors = compute_descriptors(smiles)
        descriptors_tensor = torch.tensor(descriptors).unsqueeze(0)

        # Get embedding
        inputs = tokenizer(smiles, return_tensors="pt")
        with torch.no_grad():
            outputs = embedding_model(**inputs)
        emb = outputs.last_hidden_state[:, 0, :]  # CLS token output (1, 768)

        # Forward pass
        with torch.no_grad():
            preds = model(emb)

        preds_np = preds.numpy()
        preds_rescaled = np.concatenate([
            scaler_tensile_strength.inverse_transform(preds_np[:, [0]]),
            scaler_ionization_energy.inverse_transform(preds_np[:, [1]]),
            scaler_electron_affinity.inverse_transform(preds_np[:, [2]]),
            scaler_logp.inverse_transform(preds_np[:, [3]]),
            scaler_refractive_index.inverse_transform(preds_np[:, [4]]),
            scaler_molecular_weight.inverse_transform(preds_np[:, [5]])
        ], axis=1)

        keys = ["Tensile Strength", "Ionization Energy", "Electron Affinity", "logP", "Refractive Index", "Molecular Weight"]
        results = dict(zip(keys, preds_rescaled.flatten().round(4)))

        return results

    except Exception as e:
        return {"error": str(e)}