Evaluating Self-Supervised Learning for Molecular Graph Embeddings