Molecular hypergraph neural networks.

Area of Science:

• Computational chemistry
• Machine learning for materials science

Background:

• Graph neural networks (GNNs) excel in chemistry but struggle with higher-order molecular connections like multi-center bonds.
• Existing GNNs fail to fully represent complex structural features crucial for predicting material properties.

Purpose of the Study:

• Introduce molecular hypergraphs to represent complex chemical structures.
• Develop Molecular Hypergraph Neural Networks (MHNNs) for predicting optoelectronic properties of organic semiconductors.
• Design a general algorithm for irregular high-order connections in molecular hypergraphs.

Main Methods:

• Representing molecules as hypergraphs where hyperedges capture conjugated structures.
• Developing and applying MHNNs for property prediction tasks.
• Utilizing a general algorithm for efficient operation on molecular hypergraphs with varying hyperedge orders.

Main Results:

• MHNN outperforms baseline models on organic photovoltaic, OCELOT chromophore v1, and PCQM4Mv2 datasets.
• MHNN achieves superior performance without 3D geometric information, surpassing models that use atom positions.
• MHNN demonstrates excellent data efficiency, outperforming pretrained GNNs with limited training data.

Conclusions:

• Molecular hypergraphs offer a more general and effective molecular representation.
• MHNNs provide a powerful new strategy for predicting properties of organic semiconductors and other materials with high-order connections.
• This approach enhances molecular property prediction accuracy and data efficiency in machine learning for chemistry.