Sc2Mol: a scaffold-based two-step molecule generator with variational autoencoder and transformer

Area of Science:

• Computational chemistry
• Medicinal chemistry
• Artificial intelligence in drug discovery

Background:

• Drug discovery requires identifying molecules with specific pharmaceutical properties.
• Generative models offer an efficient approach to molecule generation by learning data distributions.
• Existing models often struggle with scaffold inefficiency or bias due to predefined patterns.

Purpose of the Study:

• To develop a generative model for molecule generation that does not require prior scaffold patterns.
• To improve the efficiency and reduce bias in de novo molecule design.
• To enable the generation of sophisticated drug candidates from basic scaffolds.

Main Methods:

• Proposed Sc2Mol, a generative model-based molecule generator.
• Utilized SMILES strings for molecular representation.
• Employed a two-step process: scaffold generation (variational autoencoder) and scaffold decoration (transformer).

Main Results:

• Sc2Mol successfully learned molecular distributions and optimized molecules on drug-like datasets.
• The model demonstrated the ability to generate molecules without prior scaffold constraints.
• Empirical evaluations confirmed the model's effectiveness in random molecule generation and optimization.

Conclusions:

• Sc2Mol provides a novel scaffold-free approach to generative molecule design.
• The model automatically learns rules for transforming scaffolds into drug candidates, aligning with lead optimization strategies.
• This method offers a promising direction for efficient and unbiased drug discovery.