During the past decade, with the significant progress of computational power as well as ever-rising data availability, deep learning techniques became increasingly popular due to their excellent performance on computer vision problems. The size of the Protein Data Bank (PDB) has increased more than 15-fold since 1999, which enabled the expansion of models that aim at predicting enzymatic function via their amino acid composition. Amino acid sequence, however, is less conserved in nature than protein structure and therefore considered a less reliable predictor of protein function. This paper presents EnzyNet, a novel 3D convolutional neural networks classifier that predicts the Enzyme Commission number of enzymes based only on their voxel-based spatial structure. The spatial distribution of biochemical properties was also examined as complementary information. The two-layer architecture was investigated on a large dataset of 63,558 enzymes from the PDB and achieved an accuracy of 78.4% by exploiting only the binary representation of the protein shape. Code and datasets are available at .

The number of protein structures in the PDB database has been increasing more than 15-fold since 1999. The creation of computational models predicting enzymatic function is of major importance since such models provide the means to better understand the behavior of newly discovered enzymes when catalyzing chemical reactions. Until now, single-label classification has been widely performed for predicting enzymatic function limiting the application to enzymes performing unique reactions and introducing errors when multi-functional enzymes are examined. Indeed, some enzymes may be performing different reactions and can hence be directly associated with multiple enzymatic functions. In the present work, we propose a multi-label enzymatic function classification scheme that combines structural and amino acid sequence information. We investigate two fusion approaches (in the feature level and decision level) and assess the methodology for general enzymatic function prediction indicated by the first digit of the enzyme commission (EC) code (six main classes) on 40,034 enzymes from the PDB database. The proposed single-label and multi-label models predict correctly the actual functional activities in 97.8% and 95.5% (based on Hamming-loss) of the cases, respectively. Also the multi-label model predicts all possible enzymatic reactions in 85.4% of the multi-labeled enzymes when the number of reactions is unknown. Code and datasets are available at .

International audienceThe massive expansion of the worldwide Protein Data Bank (PDB) provides new opportunities for computational approaches which can learn from available data and extrapolate the knowledge into new coming instances. The aim of this work is to apply machine learning in order to train prediction models using data acquired by costly experimental procedures and perform enzyme functional classification. Enzymes constitute key pharmacological targets and the knowledge on the chemical reactions they catalyze is very important for the development of potent molecular agents that will either suppress or enhance the function of the given enzyme, thus modulating a pathogenicity, an illness or even the phenotype. Classification is performed on two levels: (i) using structural information into a Support Vector Machines (SVM) classifier and (ii) based on amino acid sequence alignment and Nearest Neighbor (NN) classification. The classification accuracy is increased by fusing the two classifiers and reaches 93.4% on a large dataset of 39,251 proteins from the PDB database. The method is very competitive with respect to accuracy of classification into the 6 enzymatic classes, while at the same time its computational cost during prediction is very small