Deep Learning to Generate in Silico Chemical Property Libraries and Candidate Molecules for Small Molecule...

Area of Science:

• Computational Chemistry
• Metabolomics
• Drug Discovery

Background:

• Small molecule identification is crucial for understanding biological systems and drug discovery.
• Current methods are limited by incomplete chemical reference libraries and lack of direct feature-to-candidate matching.
• Existing technologies can measure molecular properties but struggle with comprehensive identification.

Purpose of the Study:

• To develop a novel computational method for expanding small molecule identification capabilities.
• To overcome limitations of existing chemical reference libraries and identification processes.
• To enable the generation of candidate molecular structures with desired chemical properties.

Main Methods:

• Developed a variational autoencoder (VAE) to learn a latent representation of molecular structure.
• Extended the VAE with a chemical property decoder trained via multitask learning.
• Employed a cascade of transfer learning iterations using structural, in silico, and experimental data.

Main Results:

• The VAE, named DarkChem, can predict chemical properties (m/z, CCS) directly from molecular structures.
• DarkChem generates candidate structures with desired chemical properties, expanding identification possibilities.
• The method is significantly faster than first-principles simulations for property prediction.

Conclusions:

• DarkChem offers a powerful approach to accelerate small molecule identification and expand chemical reference libraries.
• The ability to generate novel molecules with specific properties has broad applications in metabolomics, drug discovery, and forensics.
• This computational framework enhances the discovery pipeline for new therapeutics and chemical probes.