Exploring Molecular Heteroencoders with Latent Space Arithmetic: Atomic Descriptors and Molecular Operators

Area of Science:

• Computational chemistry
• Machine learning in chemistry
• Drug discovery

Background:

• Developing novel atomic descriptors is crucial for advancing chemical property prediction.
• Recurrent neural networks (RNNs) and variational autoencoders (VAEs) offer powerful tools for learning molecular representations.
• SMILES notation provides a concise way to represent molecular structures for computational analysis.

Purpose of the Study:

• To develop and validate a new set of atomic descriptors, delta latent space vectors (DLSVs), using a variational heteroencoder.
• To assess the utility of DLSVs for predicting 19F NMR chemical shifts.
• To explore the application of DLSVs as molecular operators for de novo molecule generation.

Main Methods:

• A variational heteroencoder based on RNNs was trained using SMILES notations.
• Delta latent space vectors (DLSVs) were generated by comparing latent space representations of molecules with and without specific atomic modifications.
• Machine learning models, including random forests and gradient-boosting regressors, were trained using DLSVs to predict 19F NMR chemical shifts.
• DLSVs were employed as latent space operators to perform virtual chemical reactions, such as halogenation.

Main Results:

• Unsupervised mapping of DLSVs revealed clustering based on atomic properties like element, hybridization, and aromaticity.
• Machine learning models trained on DLSVs achieved high accuracy in predicting 19F NMR chemical shifts (R² up to 0.89, MAE up to 5.5 ppm).
• Using DLSVs as molecular operators successfully generated novel fluorinated molecules with high validity (99%) and a significant proportion incorporating fluorine (75%).

Conclusions:

• DLSVs are effective atomic descriptors that capture fundamental chemical information.
• DLSVs enable accurate prediction of 19F NMR chemical shifts using machine learning.
• DLSVs represent a promising approach for generative chemistry, facilitating the design of novel molecules with desired properties.