Multi-stage variational autoencoders for hierarchical molecular generation and activity optimization

Area of Science:

• Computational Chemistry
• Artificial Intelligence in Drug Discovery
• Machine Learning for Molecular Design

Background:

• Traditional single-stage variational autoencoders (VAEs) struggle with molecular representation, lacking validity, uniqueness, and biologically meaningful distribution.
• Representing complex global molecular architecture and properties in a single latent space is challenging for VAEs.

Purpose of the Study:

• To develop a multi-stage VAE system for enhanced molecular generation with improved validity, uniqueness, and biological relevance.
• To address the limitations of single-stage VAEs in capturing intricate molecular structures and properties.
• To optimize generative accuracy through adaptive fine-tuning strategies for inner and outer layers.

Main Methods:

• A multi-stage VAE system was designed to sequentially encode and decode molecular representations, improving latent space properties.
• Methodology validated using ChEMBL and polymer datasets, assessing validity, originality, novelty, Fréchet ChemNet Distance (FCD), and KL divergence.
• Adaptive fine-tuning strategies for inner-layer (IL) and outer-layer (OL) were implemented and evaluated.

Main Results:

• The multi-stage VAE demonstrated improved latent space representation, retaining structural integrity while enhancing innovation and distinction.
• Quantitative evaluations showed consistent gains in validity, novelty, and biological activity compared to baseline methods like MoLeR and RationaleRL.
• The bioefficacy of EGFR inhibitors was assessed using computational Chemprop-based quantitative structure-activity relationship (QSAR) models, confirming the model's utility.

Conclusions:

• The multi-stage VAE system offers a robust solution for generative drug discovery, overcoming limitations of traditional VAEs.
• Hierarchical latent models with multi-stage VAEs are recommended for generative drug discovery due to enhanced accuracy and performance.
• The hierarchical training method proved stable for molecular tasks, suggesting potential for cross-domain applicability.