Automatic Chemical Design Using a Data-Driven Continuous Representation of Molecules.

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry

Background:

• Traditional molecular representations limit exploration of chemical space.
• Generating novel molecules with desired properties is challenging.

Purpose of the Study:

• To develop a method for converting discrete molecular representations to continuous ones.
• To enable efficient generation and optimization of novel chemical structures.

Main Methods:

• Utilized a deep neural network with an encoder, decoder, and predictor.
• Trained on hundreds of thousands of existing chemical structures.
• Developed functions to convert between discrete and continuous molecular representations.

Main Results:

• Successfully created a continuous molecular representation.
• Demonstrated generation of novel molecules via latent space operations (decoding, perturbation, interpolation).
• Enabled gradient-based optimization for property-driven molecule design.

Conclusions:

• The continuous representation facilitates open-ended exploration of chemical compound spaces.
• This approach is effective for drug-like molecules and smaller molecular sets.