Toward efficient generation, correction, and properties control of unique drug-like structures.

Area of Science:

• Computational chemistry and cheminformatics
• Artificial intelligence in drug discovery

Background:

• Designing novel molecules with specific properties is a significant challenge in drug and material development.
• Current methods often lack efficiency in exploring vast chemical spaces and ensuring desired characteristics.

Purpose of the Study:

• To develop and validate a multi-stage deep neural network pipeline for generating novel molecular structures with controlled properties.
• To enhance the efficiency and accuracy of molecule design for applications in drug discovery and material science.

Main Methods:

• A pipeline combining deep neural network models to map discrete molecular representations into a continuous vector space.
• Utilizing an Attention-based Sequence-to-Sequence model for correcting generated molecular structures.
• Employing oversampling in the continuous space to ensure desired property distributions, even with limited data.
• Validating generated molecular properties through computer simulations and numerical experiments.

Main Results:

• Successfully generated novel molecular structures with controllable properties, including Synthetic Accessibility Score and drug-likeness metrics.
• Demonstrated the pipeline's effectiveness in generating candidate structures with desired distributions for properties and molecular descriptors.
• Validated the predicted properties of generated molecules using computer simulations.

Conclusions:

• The proposed deep learning pipeline offers an efficient and effective approach for the design and screening of novel molecules.
• This method shows significant promise for accelerating drug discovery and material design by enabling the generation of molecules with tailored characteristics.