Molecular optimization by capturing chemist's intuition using deep neural networks.

Area of Science:

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

Background:

• * Drug discovery faces challenges in balancing multiple molecular properties.
• * Molecular optimization aims to improve a starting molecule's characteristics.
• * Traditional methods rely on chemists' intuition and matched molecular pairs.

Purpose of the Study:

• * To frame molecular optimization as a machine translation task using SMILES representations.
• * To capture chemists' intuition from matched molecular pairs using AI.
• * To develop models that generate molecules with user-specified desirable properties.

Main Methods:

• * Employed sequence-to-sequence models with attention and Transformer models.
• * Utilized SMILES (Simplified Molecular Input Line Entry System) representation for molecules.
• * Incorporated user-specified property changes as conditional inputs.
• * Compared performance against a graph-to-graph translation model (HierG2G).

Main Results:

• * The Transformer model generated molecules with improved ADMET properties (logD, solubility, clearance).
• * Transformer-based optimization involved modifications intuitive to chemists.
• * The Transformer model showed competitive performance against HierG2G.
• * Ensemble modeling enhanced the diversity of generated molecules.

Conclusions:

• * Machine translation, particularly the Transformer model, is effective for molecular optimization.
• * User-guided property optimization is feasible and beneficial.
• * AI approaches can augment chemists' intuition in drug discovery.
• * Ensemble methods can increase the diversity of optimized molecules.