From Target to Drug: Generative Modeling for the Multimodal Structure-Based Ligand Design

Area of Science:

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

Background:

• The vastness of chemical space hinders traditional virtual screening for identifying bioactive molecules.
• Structure-based drug design requires efficient methods to explore potential drug candidates.

Purpose of the Study:

• To propose a novel generative adversarial network (GAN) for de novo drug design.
• To generate diverse three-dimensional (3D) ligand shapes complementary to protein binding pockets.
• To enable structure-based de novo drug design by decoding generated shapes into SMILES sequences.

Main Methods:

• Utilized a generative adversarial network (GAN) to generate 3D ligand shapes.
• Employed a shape-captioning network to translate generated shapes into SMILES strings.
• Evaluated generated molecules using structure-based (docking) and ligand-based (quantitative structure-activity relationship - QSAR) virtual screening.

Main Results:

• The GAN successfully generated diverse 3D ligand shapes.
• The shape-captioning network effectively decoded shapes into SMILES, facilitating de novo design.
• Enrichment was observed in both docking and QSAR virtual screening compared to random sampling.

Conclusions:

• The proposed GAN-based approach offers an effective strategy for structure-based de novo drug design.
• This method expands the possibilities for exploring chemical space and discovering novel drug candidates.
• The integration of generative models and shape decoding represents a significant advancement in computational drug discovery.