Generative Model for Proposing Drug Candidates Satisfying Anticancer Properties Using a Conditional Variational...

Area of Science:

• Computational chemistry
• Drug discovery
• Artificial intelligence in medicine

Background:

• Deep learning models excel at identifying drug candidates but often generate invalid molecular structures.
• Syntactically incorrect molecules limit the practical application of these models in drug discovery pipelines.

Purpose of the Study:

• To develop a generative model capable of proposing drug candidates with specific desired properties, even those outside the initial training data range.
• To address the challenge of generating syntactically valid and property-specific molecules for drug discovery.

Main Methods:

• A conditional variational autoencoder (CVAE) was trained using molecular fingerprints and GI50 (inhibition of growth by 50%) values for breast cancer cell lines.
• The model was trained on property-specific data rather than a broad range of physical properties for each molecule.

Main Results:

• The CVAE successfully generated molecular fingerprints representing desired properties not present in the training dataset.
• Generated fingerprints were confirmed to accurately reflect the desired biological activity.
• The method demonstrated utility as a query expansion technique for database searching.

Conclusions:

• The proposed CVAE model effectively generates novel molecular fingerprints with desired properties, enhancing drug candidate identification.
• This approach overcomes the syntactic invalidity issue in deep learning-generated molecules and offers a new strategy for database querying in drug discovery.