MacroSimGNN: Efficient and Accurate Prediction of Macromolecule Pairwise Similarity via A Graph Neural Network

Efficient and accurate prediction of macromolecule pairwise similarity is essential for developing database search engines and is useful for machine learning based predictive tools. Existing methods for calculating macromolecular similarity suffer from significant drawbacks. Graph edit distance is accurate but computationally expensive, and graph kernel methods are computationally efficient but inaccurate. This study introduces a graph neural network model, MacroSimGNN, which significantly improves computational efficiency while maintaining high accuracy on macromolecule pairwise similarity. Furthermore, this approach enables feature embeddings based on macromolecular similarities to a set of landmark molecules, enhancing both unsupervised and supervised learning tasks. This method represents a significant advancement in macromolecular cheminformatics, paving the way for the development of advanced search engines and data-driven design of macromolecules.