MSGAT: Multi-Scale Graph Attention Network

A lightweight graph neural network (242K parameters) for predicting 10 quantum-chemical properties from molecular SMILES strings.

Model Details

• Architecture: Edge-aware multi-head attention + MPNN + cross-scale fusion
• Parameters: 242,407
• Hidden dim: 64, 4 attention heads, 3 layers
• Node features: 58-dim (atomic number, degree, charge, Hs, hybridization, aromaticity)
• Bond features: 12-dim (bond type, conjugation, ring, stereo)
• Training data: QuantumChem/QuantumChem_200k

Properties Predicted

Property	Unit	MAE	R²
Sigma at 780 nm	GM	10.10	0.953
Max sigma	GM	10.19	0.957
ISC energy	eV	0.0055	0.933
Toxicity score	—	0.017	0.924
SA score	—	0.0069	0.967
Boiling point	°C	5.77	0.984
logP	—	0.057	0.993
Aromaticity	—	0.0077	0.998
Solubility	ug/ml	71,953	0.053
Molecular weight	g/mol	1.61	0.806

Mean R² across 9/10 properties (excl. solubility): 0.946

Usage

import torch
import torch.nn as nn

# Load the model
model_state = torch.load("msgat_model.pt", weights_index=None)
norm_stats = torch.load("norm_stats.pt")

# Reconstruct model architecture (see model.py in the repo)
from model import create_model  # clone https://github.com/devansh0703/MSGAT
model = create_model()
model.load_state_dict(model_state)
model.eval()

# mean/std for inverse transform (shape: [10])
mean = norm_stats["mean"]
std = norm_stats["std"]

Inverse normalization

The model outputs z-score normalized predictions. To get raw values:

# solubility uses log1p before normalization — must invert with expm1
raw = preds * std + mean
sol_idx = 8  # solubility index in active props
raw[:, sol_idx] = torch.expm1(raw[:, sol_idx])

Files

File	Description
msgat_model.pt	Trained model state dict (242K params)
norm_stats.pt	Z-score normalization stats (mean, std) for the 10 active properties

Training

git clone https://github.com/devansh0703/MSGAT
cd MSGAT
pip install torch rdkit-pypi datasets pandas tqdm matplotlib
python train.py  # train/val split
python train_final.py  # full data retrain

Citation

@article{raulo2025msgat,
  title={MSGAT: Multi-Scale Graph Attention Network for Efficient Molecular Property Prediction},
  author={Raulo, Devansh},
  year={2025}
}

Acknowledgements

Dataset: QuantumChem/QuantumChem_200k by Zeng et al.