Can Graph Neural Networks Understand Chemical Elements?

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Graph Neural Networks

Background:

• Classical machine learning (ML) requires extensive feature engineering for chemical applications.
• Graph neural networks (GNNs) learn representations from molecular graphs, reducing feature engineering.
• GNNs may need larger datasets and rely on vertex/edge features like chemical elements and bond orders.

Purpose of the Study:

• To investigate if chemically inspired elemental representations can reduce GNN data requirements or improve model performance and generalizability.
• To evaluate nine different elemental representations across various GNN model and dataset sizes.
• To benchmark performance using molecular pKa prediction.

Main Methods:

• Created nine distinct representations encoding elemental reactivity and behavior.
• Tested these representations with GNNs varying in size, dataset size, and featurization.
• Utilized molecular pKa prediction as the benchmark task.

Main Results:

• GNNs largely classify elements, converting input representations into an internal, orthogonal format.
• Minimal performance and generalizability differences were observed between most elemental representations.
• Slight performance improvements were noted with more orthogonal elemental representations.

Conclusions:

• Chemically inspired elemental representations have limited impact on GNN performance due to internal data transformation.
• Increased orthogonality in representations offers marginal benefits, aligning better with internal model processes.
• Atom-typing and novel similarity typing provided modest performance gains but increased computational costs for data preparation.