Variational autoencoder-based chemical latent space for large molecular structures with 3D complexity.

Area of Science:

• Computational Chemistry
• Drug Discovery
• Machine Learning

Background:

• Chemical latent space represents structural diversity in compound libraries for drug candidate generation.
• Existing methods struggle with large molecular structures and complex datasets, including natural compounds with chirality.

Purpose of the Study:

• To develop a deep learning method, NP-VAE (Natural Product-oriented Variational Autoencoder), for managing and analyzing complex chemical datasets.
• To construct and explore a chemical latent space for large-sized compounds, including natural products with chirality.

Main Methods:

• Developed NP-VAE, a deep learning model based on variational autoencoder.
• Applied NP-VAE to analyze datasets from DrugBank and large molecular structures, including natural compounds with chirality.

Main Results:

• NP-VAE successfully constructed chemical latent spaces from large compounds previously unmanageable by existing methods.
• Achieved higher reconstruction accuracy and demonstrated stable performance as a generative model.
• Enabled comprehensive analysis of compound libraries and generation of novel structures with optimized functions.

Conclusions:

• NP-VAE offers a robust solution for analyzing complex chemical structures and exploring chemical diversity.
• The method facilitates the generation of novel drug candidates with optimized functions through latent space exploration.