Using GNN property predictors as molecule generators.

Area of Science:

• Computational chemistry and materials science
• Machine learning applications in drug discovery and materials design

Background:

• Graph neural networks (GNNs) are increasingly vital for predicting material and molecular properties.
• Automated discovery pipelines require efficient methods for generating novel molecular structures.

Purpose of the Study:

• To leverage the invertible nature of GNNs for direct generation of molecular structures with targeted electronic properties.
• To optimize molecular graphs towards desired properties using gradient ascent on fixed GNN weights.

Main Methods:

• Utilized gradient ascent on molecular graph inputs of pre-trained GNNs to optimize for target properties.
• Ensured strict adherence to valence rules through careful graph construction.
• No additional training on molecular structures was required, relying solely on the property predictor.

Main Results:

• Successfully generated molecules with specific energy gaps (verified by DFT) and octanol-water partition coefficients (logP).
• Achieved target property prediction rates comparable to or better than state-of-the-art generative models.
• Generated a dataset of 1617 new molecules and their DFT-calculated properties for out-of-distribution testing.

Conclusions:

• Invertible GNNs offer a powerful, training-free approach for de novo molecular generation with desired properties.
• The method demonstrates high efficiency, accuracy, and superior molecular diversity compared to existing models.
• The generated dataset provides a valuable resource for benchmarking and validating QM9-trained models.