Abstract:Supervised learning on molecules has incredible potential to be useful in chemistry, drug discovery, and materials science. Luckily, several promising and closely related neural network models invariant to molecular symmetries have already been described in the literature. These models learn a message passing algorithm and aggregation procedure to compute a function of their entire input graph. At this point, the next step is to find a particularly effective variant of this general approach and apply it to chemical prediction benchmarks until we either solve them or reach the limits of the approach. In this paper, we reformulate existing models into a single common framework we call Message Passing Neural Networks (MPNNs) and explore additional novel variations within this framework. Using MPNNs we demonstrate state of the art results on an important molecular property prediction benchmark; these results are strong enough that we believe future work should focus on datasets with larger molecules or more accurate ground truth labels.

What problem does this paper attempt to address?

The main goal of this paper is to develop a machine learning model that can effectively predict molecular properties, particularly for chemical prediction problems. Such models can learn features directly from molecular graphs and are invariant to the graph's isomorphism. The authors construct a general framework—Message Passing Neural Networks (MPNNs)—to abstract and unify several promising neural network models previously proposed, in order to better understand and propose new variants. Specifically, the paper addresses the following key issues: 1. **Proposing the MPNNs framework**: To achieve the aforementioned goal, the paper proposes a general framework called "Message Passing Neural Networks," which integrates various neural network models from previous literature and allows for comparison and improvement among these models. 2. **Addressing chemical prediction problems**: The paper focuses on solving chemical prediction problems, especially predicting the quantum mechanical properties of small organic molecules using machine learning methods. This includes predicting various properties of molecules, such as atomization energy, vibrational frequencies, etc. 3. **Achieving chemical accuracy**: One of the authors' goals is to achieve so-called "chemical accuracy" on 13 different molecular property prediction tasks, meaning that the prediction error is close to or lower than the estimated error of quantum chemical calculations (such as Density Functional Theory, DFT). 4. **Experimental results**: By exploring different variants of the MPNN family, the authors not only achieved state-of-the-art performance on the QM9 dataset but also reached chemical accuracy on all but two of the other property prediction tasks. 5. **Summary of contributions**: - Developed an MPNN that achieved state-of-the-art levels on all 13 prediction tasks and reached chemical accuracy on 11 tasks. - Developed several MPNN variants that can achieve chemical accuracy based solely on molecular topology (without using spatial information). - Proposed a training method that allows for the use of larger node representations without significantly increasing computation time and memory consumption. In summary, this paper aims to advance research in the field of chemical prediction by proposing the MPNNs framework and its implementation details, particularly in the prediction of quantum properties of small molecules.

Neural Message Passing for Quantum Chemistry

Predictive Modelling of Quantum Process with Neural Networks

Flexible dual-branched message passing neural network for quantum mechanical property prediction with molecular conformation

Efficient Integration of Molecular Representation and Message-Passing Neural Networks for Predicting Small Molecule Drug-like Properties

Simplicial Message Passing for Chemical Property Prediction

Dual Message Passing Neural Network for Molecular Property Prediction

Message Passing Networks for Molecules with Tetrahedral Chirality

Heterogeneous Relational Message Passing Networks for Molecular Dynamics Simulations

Automatic Identification of Chemical Moieties

Neural Message Passing for NMR Chemical Shift Prediction

Could Chemical LLMs benefit from Message Passing

When Do Quantum Mechanical Descriptors Help Graph Neural Networks Predict Chemical Properties?

Predicting Infrared Spectra with Message Passing Neural Networks

Neural Message Passing on High Order Paths

Ewald-based Long-Range Message Passing for Molecular Graphs

Quantum-chemical insights from deep tensor neural networks

MGNN: Moment Graph Neural Network for Universal Molecular Potentials

Ab initio quantum chemistry with neural-network wavefunctions

Unifying machine learning and quantum chemistry -- a deep neural network for molecular wavefunctions

Unifying machine learning and quantum chemistry with a deep neural network for molecular wavefunctions

Graph Convolutional Neural Networks for (QM)ML/MM Molecular Dynamics Simulations