Powerful molecule generation with simple ConvNet.

Area of Science:

• Computational chemistry and cheminformatics
• Artificial intelligence in drug discovery
• Machine learning for molecular design

Background:

• Automated molecule generation is vital for in-silico drug discovery.
• Existing graph-based methods are complex, difficult to train, and struggle with long molecules.
• A simpler, powerful alternative is needed to enhance the practicality of automated drug discovery.

Purpose of the Study:

• To introduce a novel ConvNet-based sequential graph generation algorithm for automated molecule generation.
• To reformulate molecular graph generation as a series of simple classification tasks.
• To develop models (SEEM and SEED) that are easier to implement and train while maintaining high performance.

Main Methods:

• A convolutional neural network (ConvNet) operates sequentially on sub-graphs.
• The model predicts atom/bond adding actions to populate the molecular graph.
• Pretraining (SEEM) involves sequential reconstruction of existing molecules, followed by reinforcement learning fine-tuning (SEED) for improved properties.

Main Results:

• The proposed SEEM and SEED models demonstrate competitive performance against 16 state-of-the-art baselines.
• The models were evaluated on three benchmark datasets (QM9, ZINC250k, ChEMBL).
• The ConvNet-based approach offers a practical and effective alternative for molecular generation.

Conclusions:

• SEEM and SEED provide a simpler and powerful approach to automated molecule generation.
• The ConvNet-based sequential classification strategy effectively addresses limitations of existing graph-based methods.
• These models advance the field of in-silico drug discovery by improving the practicality and performance of molecular generation.