Fragment-wise interpretability in graph neural networks via molecule decomposition and contribution analysis