GEN: highly efficient SMILES explorer using autodidactic generative examination networks.

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery

Background:

• Recurrent neural networks (RNNs) are common for de novo molecule generation.
• Existing models often use LSTM/GRU units and canonical SMILES, limiting validity.
• Deep generative models need enhanced architectures for robust molecular design.

Purpose of the Study:

• Introduce Generative Examination Networks (GENs) for improved SMILES generation.
• Enhance the validity and quality of deep generative models for molecular design.
• Develop an early stopping mechanism based on online quality assessment.

Main Methods:

• Utilized concatenated bidirectional RNN units in the GEN architecture.
• Implemented an online statistical quality control (SQC) mechanism for early stopping.
• Employed SMILES augmentation via unrestricted randomization and parallel encoding layers.

Main Results:

• Achieved high validity rates (95-98%) for generated SMILES.
• Demonstrated excellent novelty (85-90%) and property space conservation (95-99%).
• GENs autonomously learned target spaces efficiently within a few epochs.

Conclusions:

• GENs offer a novel and effective approach for de novo molecular generation.
• The integrated examination mechanism ensures high-quality, valid SMILES.
• The GEN framework is adaptable to various architectures and quality metrics.