Prototype-Based Compound Discovery Using Deep Generative Models.

Area of Science:

• Computational chemistry
• Medicinal chemistry
• Artificial intelligence in drug discovery

Background:

• Drug design is a complex, costly, and time-consuming process.
• The vastness of potential molecular structures necessitates efficient drug discovery strategies.
• Utilizing existing molecules as starting points (prototypes) is a common approach.

Purpose of the Study:

• To develop an unsupervised deep learning model for automatic generation of novel drug molecules.
• To validate the generated molecules for chemical validity and novelty compared to the prototype.
• To assess the potential of the generated molecules by identifying known FDA-approved drugs.

Main Methods:

• Development of a deep learning unsupervised approach for molecule generation.
• Input: a prototype drug molecule.
• Output: a set of novel, chemically valid drug-like molecules.

Main Results:

• The deep learning system generated chemically valid molecules distinct from the prototype.
• Identification of 35 known FDA-approved drugs among the generated compounds.
• Example: the system successfully generated isoniazid, a key tuberculosis drug.

Conclusions:

• The unsupervised deep learning approach offers an efficient method for generating novel drug candidates.
• The model's ability to rediscover known drugs highlights its potential in drug discovery.
• Ranking functions can further refine the selection of promising generated molecules.