Comparison of structure- and ligand-based scoring functions for deep generative models: a GPCR case study.

Area of Science:

• Computational Chemistry
• Medicinal Chemistry
• Drug Discovery

Background:

• Deep generative models accelerate the discovery of bioactive compounds.
• Current methods often rely on ligand-based predictors, limiting novelty and applicability to data-rich targets.
• Ligand-based approaches can bias generation towards known chemical space, hindering the identification of novel chemotypes.

Purpose of the Study:

• To evaluate structure-based molecular docking (Glide) as a scoring function for the deep generative model REINVENT.
• To compare the performance of structure-based versus ligand-based scoring functions in guiding molecule generation.
• To assess the novelty and chemical space exploration of generated molecules.

Main Methods:

• Utilized the REINVENT deep generative model guided by Glide docking scores.
• Compared structure-based (Glide) and ligand-based scoring functions.
• Modified the MOSES dataset to remove bias and proposed a new diversity metric.
• Optimized the model against the DRD2 target.

Main Results:

• The structure-based approach improved predicted ligand affinity beyond known active molecules for DRD2.
• Generated molecules explored novel chemical and physicochemical space compared to ligand-based methods and known actives.
• The model learned to generate molecules satisfying crucial protein-ligand interactions.

Conclusions:

• Molecular docking is advantageous for de novo molecule generation, improving predicted affinity and novelty.
• Structure-based guidance enables the generation of molecules with key target interactions, crucial for drug discovery.
• This approach is valuable for hit generation and novelty-focused projects, especially with limited prior ligand data.