Reference: Ryu, S. (2018). “Deeply learning molecular structure-property relationships using attention- and gate-augmented graph convolutional network”. arXiv:1805.10988
- Developed a graph convolutional network (GCN) combining molecular structure and atomic features to predict molecular properties.
- Converted the original TensorFlow model into PyTorch.
-
One-Hot Encoding Issue:
- One-hot encoding makes the input matrix size data-dependent, requiring inefficient padding for varying matrix sizes.
- Solution: Numerical representation and normalization of feature values for efficient mapping.
-
Performance Enhancements:
- Replacing one-hot encoded atomic features with atomic feature values and a combination of atomic & molecular feature values improved performance.
- Results:
- Reduced Mean Absolute Error (MAE).
- Increased R² value, indicating better predictive accuracy.