This paper introduces Message Passing Neural Networks (MPNNs) as a general framework for supervised learning on graphs, which abstracts commonalities among several promising neural models for graph-structured data. The authors demonstrate that MPNNs can achieve state-of-the-art results on a molecular property prediction benchmark, QM9, which contains 130,000 molecules with 13 properties each. These results are strong enough to suggest that future work should focus on larger molecules or more accurate ground truth labels. MPNNs operate on graph-structured data and are invariant to graph isomorphism, making them suitable for molecules. The paper explores various MPNN variants, including those that predict DFT to within chemical accuracy on 11 out of 13 targets, and those that operate on the topology of the molecule alone without spatial information. The authors also develop a general method to train MPNNs with larger node representations without a corresponding increase in computation time or memory. The paper discusses several existing models that can be described within the MPNN framework, including Convolutional Networks for Learning Molecular Fingerprints, Gated Graph Neural Networks, Interaction Networks, Molecular Graph Convolutions, and Deep Tensor Neural Networks. The authors also explore different message functions, readout functions, and input representations for MPNNs. The QM9 dataset is used to evaluate the performance of MPNNs. The dataset contains 130,000 molecules with 13 properties each, and the results show that MPNNs can achieve chemical accuracy on 11 out of 13 targets. The authors also explore the use of multiple towers to improve scalability and generalization performance. The paper concludes that MPNNs with appropriate message, update, and output functions have a useful inductive bias for predicting molecular properties, outperforming several strong baselines and eliminating the need for complicated feature engineering. The authors suggest that future work should focus on designing MPNNs that can generalize effectively to larger graphs and work with benchmarks designed to expose issues with generalization across graph sizes.This paper introduces Message Passing Neural Networks (MPNNs) as a general framework for supervised learning on graphs, which abstracts commonalities among several promising neural models for graph-structured data. The authors demonstrate that MPNNs can achieve state-of-the-art results on a molecular property prediction benchmark, QM9, which contains 130,000 molecules with 13 properties each. These results are strong enough to suggest that future work should focus on larger molecules or more accurate ground truth labels. MPNNs operate on graph-structured data and are invariant to graph isomorphism, making them suitable for molecules. The paper explores various MPNN variants, including those that predict DFT to within chemical accuracy on 11 out of 13 targets, and those that operate on the topology of the molecule alone without spatial information. The authors also develop a general method to train MPNNs with larger node representations without a corresponding increase in computation time or memory. The paper discusses several existing models that can be described within the MPNN framework, including Convolutional Networks for Learning Molecular Fingerprints, Gated Graph Neural Networks, Interaction Networks, Molecular Graph Convolutions, and Deep Tensor Neural Networks. The authors also explore different message functions, readout functions, and input representations for MPNNs. The QM9 dataset is used to evaluate the performance of MPNNs. The dataset contains 130,000 molecules with 13 properties each, and the results show that MPNNs can achieve chemical accuracy on 11 out of 13 targets. The authors also explore the use of multiple towers to improve scalability and generalization performance. The paper concludes that MPNNs with appropriate message, update, and output functions have a useful inductive bias for predicting molecular properties, outperforming several strong baselines and eliminating the need for complicated feature engineering. The authors suggest that future work should focus on designing MPNNs that can generalize effectively to larger graphs and work with benchmarks designed to expose issues with generalization across graph sizes.