Unsupervised Learning in Drug Design from Self-Organization to Deep Chemistry.

Area of Science:

• Computational Chemistry
• Cheminformatics
• Artificial Intelligence in Drug Design

Background:

• Computers have revolutionized drug design, with early applications of neural networks (NN) in cheminformatics.
• Interest in NN for drug design waned for nearly two decades before a resurgence driven by deep learning (DL).
• Deep learning enables direct data analysis, offering potential for advanced applications in chemistry.

Purpose of the Study:

• To review the evolution and application of self-organizing maps (SOM) in mapping molecular representations from the 1990s to current deep chemistry.
• To explore the efficiency of unsupervised learning, specifically SOM, within the context of deep learning.
• To assess the current state and future prospects of deep learning in chemistry, particularly in drug design.

Main Methods:

• Review of literature on deep learning projects, focusing on unsupervised architectures.
• Analysis of self-organizing maps (SOM) for mapping molecular representations.
• Comparison of SOM capabilities with human chemist's expectations and current deep learning trends.

Main Results:

• Self-organizing maps (SOM) demonstrate significant efficiency in identifying both expected and non-trivial molecular features.
• Unsupervised learning methods like SOM show promise for deep chemistry applications.
• Deep learning excels in pattern recognition tasks (e.g., Deep Face, Deep Blue).

Conclusions:

• Deep learning holds considerable promise for advancing deep chemistry and drug design.
• The practical application of deep learning in chemistry is currently limited by the scarcity of measured property data.
• Self-organizing maps (SOM) represent a valuable tool for molecular representation mapping in the era of deep chemistry.