fukui-net 0.1.0

A neural network for predicting Fukui indices using Kernel-based Attention Networks (KAN) with Chebyshev graph convolutions

---

license: mit language:

• en tags:
• chemistry
• biology
• DFT
• molecular-property-prediction
• graph-neural-networks
• fukui-indices
• reactivity
• pytorch library_name: transformers

---

Fukui_Net

Neural network for predicting Fukui indices using Kernel-based Attention Networks (KAN) with Chebyshev graph convolutions.

Installation

# Clone and install
git clone https://huggingface.co/Nikolenko-Sergei/FukuiNet
cd FukuiNet
uv sync

Usage

CLI Interface

The CLI provides a simple interface for molecular analysis:

# Check available devices and model info
uv run fukui_net info

# Predict single molecule
uv run fukui_net predict "CCO" --device cuda:1

# Batch prediction from CSV file
uv run fukui_net predict --csv molecules.csv --output predictions.csv --device cuda:1

CLI Options:

• --device: Specify device (cpu, cuda:0, cuda:1, etc.)
• --csv: Input CSV file with SMILES column
• --output: Output CSV file for batch predictions
• --column: Name of SMILES column in CSV (default: "smiles")

Input CSV format:

smiles,name
CCO,Ethanol
c1ccccc1,Benzene

Output CSV format:

smiles,fukui_indices
CCO,"[-0.322, -0.122, -0.935, ...]"
c1ccccc1,"[-0.280, -0.280, ...]"

Python API

from transformers import AutoModel

# Load model from Hugging Face Hub
model = AutoModel.from_pretrained(
    "Nikolenko-Sergei/FukuiNet",
    trust_remote_code=True
)

# Predict Fukui indices
fukui_indices = model.predict("CCO")
print(f"Fukui indices: {fukui_indices}")

# Batch prediction
results = model.predict_batch(["CCO", "c1ccccc1"])

Direct Usage

from fukui_net.predictor import FukuiNetPredictor

# Load predictor
predictor = FukuiNetPredictor("models/final_model.ckpt", device="cuda:1")

# Predict
fukui_indices = predictor.predict_smiles("CCO")

Model Architecture

• Graph Neural Network: Molecular structure as graphs
• Kernel-based Attention Networks (KAN): Advanced attention mechanisms
• Chebyshev Convolutions: Efficient graph operations
• RDKit Integration: Molecular featurization

Input/Output Format

Input: SMILES strings (e.g., "CCO", "c1ccccc1")

Output: List of Fukui indices for each atom

• Positive values: Electrophilic sites
• Negative values: Nucleophilic sites
• Magnitude: Reactivity strength

License

MIT License