The paper introduces molecular *graph convolutions*, a machine learning architecture designed to process undirected graphs, specifically small molecules. Unlike traditional fingerprint representations, which emphasize certain aspects of molecular structure while ignoring others, graph convolutions use a simple encoding of the molecular graph—atoms, bonds, distances, etc.—to allow the model to make data-driven decisions. The authors describe the architecture, which includes operations that maintain properties such as order invariance, permutation invariance, and pair order invariance. They demonstrate that graph convolutions can achieve performance comparable to state-of-the-art multitask neural networks trained on traditional molecular fingerprint representations, as well as alternative methods like "neural fingerprints" and influence relevance voter. The flexibility of graph convolutions is highlighted, as they can utilize any available information from the molecular graph for the task at hand. The authors also discuss potential improvements and future directions, emphasizing the need for further optimization and the extension of deep learning methods to three-dimensional biology.The paper introduces molecular *graph convolutions*, a machine learning architecture designed to process undirected graphs, specifically small molecules. Unlike traditional fingerprint representations, which emphasize certain aspects of molecular structure while ignoring others, graph convolutions use a simple encoding of the molecular graph—atoms, bonds, distances, etc.—to allow the model to make data-driven decisions. The authors describe the architecture, which includes operations that maintain properties such as order invariance, permutation invariance, and pair order invariance. They demonstrate that graph convolutions can achieve performance comparable to state-of-the-art multitask neural networks trained on traditional molecular fingerprint representations, as well as alternative methods like "neural fingerprints" and influence relevance voter. The flexibility of graph convolutions is highlighted, as they can utilize any available information from the molecular graph for the task at hand. The authors also discuss potential improvements and future directions, emphasizing the need for further optimization and the extension of deep learning methods to three-dimensional biology.