This paper introduces Message Passing Neural Networks (MPNNs) as a unified framework for supervised learning on molecular graphs, aiming to address the challenges of predicting chemical properties. MPNNs are designed to be invariant to graph isomorphism, making them suitable for molecules with similar structures. The authors reformulate several existing neural network models into this framework and explore novel variations. Using the QM9 dataset, they demonstrate state-of-the-art results in predicting quantum mechanical properties of small organic molecules, achieving chemical accuracy on 11 out of 13 targets. They also develop variants that can predict DFT calculations within chemical accuracy on 5 out of 13 targets without spatial information. Additionally, they propose a method to train MPNNs with larger node representations without increasing computational costs. The paper highlights the importance of allowing long-range interactions between nodes and suggests future directions for improving generalization to larger graphs.This paper introduces Message Passing Neural Networks (MPNNs) as a unified framework for supervised learning on molecular graphs, aiming to address the challenges of predicting chemical properties. MPNNs are designed to be invariant to graph isomorphism, making them suitable for molecules with similar structures. The authors reformulate several existing neural network models into this framework and explore novel variations. Using the QM9 dataset, they demonstrate state-of-the-art results in predicting quantum mechanical properties of small organic molecules, achieving chemical accuracy on 11 out of 13 targets. They also develop variants that can predict DFT calculations within chemical accuracy on 5 out of 13 targets without spatial information. Additionally, they propose a method to train MPNNs with larger node representations without increasing computational costs. The paper highlights the importance of allowing long-range interactions between nodes and suggests future directions for improving generalization to larger graphs.