De Novo Molecule Design by Translating from Reduced Graphs to SMILES.

Area of Science:

• Computational Chemistry
• Cheminformatics
• Artificial Intelligence

Background:

• Automated molecular design is crucial for drug discovery, with deep learning offering alternatives to traditional methods.
• Deep learning models typically use molecular graphs (SMILES) but controlling generated chemical space remains a challenge.
• Higher-level cheminformatics representations could potentially constrain molecule generation algorithms.

Purpose of the Study:

• To investigate the utility of the Reduced Graph representation for constraining deep learning-based molecule generation.
• To develop and evaluate a novel sequence-to-sequence (seq-to-seq) model for generating molecules from Reduced Graphs.
• To demonstrate the model's ability to generate valid molecules and extrapolate to novel chemical spaces.

Main Methods:

• Explored the Reduced Graph representation, where functional groups are replaced by superatoms, mappable to SMILES strings.
• Developed a novel seq-to-seq deep learning approach to learn the one-to-many mapping from Reduced Graphs to SMILES.
• Trained the model on a large dataset from ChEMBL, enabling single-time training for subsequent generation.

Main Results:

• The seq-to-seq model successfully generated valid molecules based on specified Reduced Graphs.
• The approach demonstrated the ability to extrapolate to Reduced Graphs not present in the training data.
• Generated compounds exhibited the same Reduced Graph as the input molecule, validating the constraint.

Conclusions:

• The Reduced Graph representation can effectively define target chemical space for deep learning molecule generators.
• The novel seq-to-seq method provides an alternative deep learning approach for de novo molecule design, avoiding transfer learning or adversarial networks.
• This method is applicable to scaffold hopping and other cheminformatics tasks in drug discovery.