Abstract
Enzyme commission (EC) numbers play a vital role in classifying enzymes and understanding their functions in enzyme-related research. Although accurate and informative encoding of EC numbers is essential for enhancing the effectiveness of machine learning applications, simple EC encoding approaches suffer from limitations such as false numerical order and high sparsity. To address these issues, we developed EC2Vec, a multimodal autoencoder that preserves the categorical nature of EC numbers and leverages their hierarchical relationships, resulting in more meaningful and informative representations. EC2Vec encodes each digit of the EC number as a categorical token and then processes these embeddings through a 1D convolutional layer to capture their relationships. Comprehensive benchmarking against a large collection of EC numbers indicates that EC2Vec outperforms simple encoding methods. The t-SNE visualization of EC2Vec embeddings revealed distinct clusters corresponding to different enzyme classes, demonstrating that the hierarchical structure of the EC numbers is effectively captured. In downstream machine learning applications, EC2Vec embeddings outperformed other EC encoding methods in the reaction-EC pair classification task, underscoring its robustness and utility for enzyme-related research and bioinformatics applications.