VNFlow: integration of variational autoencoders and normalizing flows for novel molecular design

Area of Science:

• Computational Chemistry
• Drug Discovery
• Artificial Intelligence

Background:

• Generative Artificial Intelligence (AI) is revolutionizing molecular discovery by enabling exploration of vast chemical spaces.
• Existing methods like normalizing flows face challenges in balancing complex objective optimization with sampling speed, especially for intricate molecular scaffolds.
• Generating specific compound classes and complex structures, such as aromatic rings, remains a hurdle in current generative models.

Purpose of the Study:

• To develop a generative model capable of efficiently sampling novel molecules.
• To optimize key molecular properties including drug-likeness, synthetic accessibility, and chemical reactivity.
• To address limitations in inverse molecular design, particularly for scenarios with limited training data.

Main Methods:

• Employed normalizing flows integrated with variational autoencoders for molecular generation.
• Evaluated generated molecules using Quantitative Estimate of Drug-likeness (QED) and Synthetic Accessibility (SA) scores.
• Calculated electronic density on phosphorus atoms in organofluorine-phosphates using Hirschfeld charges derived from density functional theory (DFT).
• Utilized SELFIES (Simplified Molecular Input Line Entry System) and group-SELFIES for direct integration with normalizing flows.

Main Results:

• The framework efficiently generated a diverse set of novel organofluorine-phosphates.
• Demonstrated successful optimization of drug-likeness, synthetic accessibility, and electronic properties.
• Showcased the effectiveness of combining normalizing flows with SELFIES/group-SELFIES for inverse molecular design.
• Overcame limitations of variational autoencoders when training data is scarce.

Conclusions:

• The developed generative model efficiently samples novel molecules while optimizing complex objectives.
• Combining normalizing flows with SELFIES/group-SELFIES addresses key limitations in inverse molecular design, especially with limited data.
• This approach enables holistic capture of chemical structures, paving the way for targeted therapies with optimized molecular objectives.