Randomized SMILES strings improve the quality of molecular generative models.

Area of Science:

• * Computational chemistry and cheminformatics.
• * Machine learning applications in drug discovery.

Background:

• * Recurrent Neural Networks (RNNs) can generate novel molecular structures using SMILES strings.
• * Canonical SMILES representations may limit the diversity and accuracy of generated chemical spaces.

Purpose of the Study:

• * To benchmark RNN performance using different SMILES variants (canonical, randomized, DeepSMILES) and cell types (LSTM, GRU).
• * To develop new metrics for evaluating the generalization capabilities of chemical space generation models.
• * To assess model performance on different dataset sizes (GDB-13 subsets) and chemical databases (ChEMBL).

Main Methods:

• * Trained RNN models (LSTM, GRU) on varying sizes of GDB-13 and ChEMBL datasets.
• * Utilized canonical, randomized, and DeepSMILES string representations for molecular input.
• * Developed and applied metrics for chemical space uniformity, closedness, and completeness.

Main Results:

• * Models trained with randomized SMILES and LSTM cells demonstrated superior generalization and accuracy in representing chemical spaces.
• * A model trained with randomized SMILES on GDB-13 generated molecules with quasi-uniform probability, covering a significant portion of the dataset.
• * Randomized SMILES training on ChEMBL yielded models that generated double the unique drug-like molecules compared to canonical SMILES, with similar property distributions.

Conclusions:

• * Randomized SMILES representations enhance the ability of RNNs to generate diverse and accurate chemical spaces.
• * LSTM cells combined with randomized SMILES offer a powerful approach for exploring and representing chemical diversity.
• * This method shows significant promise for improving drug discovery by enabling better exploration of the chemical space.