Molecular graph convolutions are a machine learning architecture for learning from undirected graphs, specifically small molecules. Unlike traditional molecular fingerprints, which emphasize particular aspects of molecular structure, graph convolutions use a simple encoding of the molecular graph—atoms, bonds, distances, etc.—to allow the model to take greater advantage of information in the graph structure. Although graph convolutions do not outperform all fingerprint-based methods, they represent a new paradigm in ligand-based virtual screening with exciting opportunities for future improvement. The paper introduces molecular graph convolutions, a deep learning system using a representation of small molecules as undirected graphs of atoms. Graph convolutions extract meaningful features from simple descriptions of the graph structure—atom and bond properties, and graph distances—to form molecule-level representations that can be used in place of fingerprint descriptors in conventional machine learning applications. The paper discusses related work in molecular representation, including 2D molecular descriptors, 3D molecular descriptors, and other graph-based approaches. It also describes the use of deep neural networks in cheminformatics, including the use of convolutional networks on non-Euclidean manifolds. The paper presents the methods used in molecular graph convolutions, including the use of deep neural networks, desired invariants of a model, and invariant-preserving operations. It also describes the construction of molecule-level features and the input featurization process. The paper presents results from experiments on various datasets, including comparisons to other methods such as neural fingerprints and influence relevance voter models. The results show that graph convolution models perform well on classification tasks and are comparable to state-of-the-art multitask neural networks trained on traditional molecular fingerprint representations. The paper discusses the flexibility of graph convolution architecture, which allows the model to use any of the available information for the task at hand. It also discusses the limitations of molecular graph representations, including the fact that much of the information required to represent biological systems and the interactions responsible for small molecule activity is not encapsulated in the molecular graph. The paper concludes that graph convolutions present a new direction in computer-aided drug design and cheminformatics, with potential for further optimization and development. It also highlights the need for further research into extending deep learning methods to three-dimensional biology.Molecular graph convolutions are a machine learning architecture for learning from undirected graphs, specifically small molecules. Unlike traditional molecular fingerprints, which emphasize particular aspects of molecular structure, graph convolutions use a simple encoding of the molecular graph—atoms, bonds, distances, etc.—to allow the model to take greater advantage of information in the graph structure. Although graph convolutions do not outperform all fingerprint-based methods, they represent a new paradigm in ligand-based virtual screening with exciting opportunities for future improvement. The paper introduces molecular graph convolutions, a deep learning system using a representation of small molecules as undirected graphs of atoms. Graph convolutions extract meaningful features from simple descriptions of the graph structure—atom and bond properties, and graph distances—to form molecule-level representations that can be used in place of fingerprint descriptors in conventional machine learning applications. The paper discusses related work in molecular representation, including 2D molecular descriptors, 3D molecular descriptors, and other graph-based approaches. It also describes the use of deep neural networks in cheminformatics, including the use of convolutional networks on non-Euclidean manifolds. The paper presents the methods used in molecular graph convolutions, including the use of deep neural networks, desired invariants of a model, and invariant-preserving operations. It also describes the construction of molecule-level features and the input featurization process. The paper presents results from experiments on various datasets, including comparisons to other methods such as neural fingerprints and influence relevance voter models. The results show that graph convolution models perform well on classification tasks and are comparable to state-of-the-art multitask neural networks trained on traditional molecular fingerprint representations. The paper discusses the flexibility of graph convolution architecture, which allows the model to use any of the available information for the task at hand. It also discusses the limitations of molecular graph representations, including the fact that much of the information required to represent biological systems and the interactions responsible for small molecule activity is not encapsulated in the molecular graph. The paper concludes that graph convolutions present a new direction in computer-aided drug design and cheminformatics, with potential for further optimization and development. It also highlights the need for further research into extending deep learning methods to three-dimensional biology.