Chemical space exploration guided by deep neural networks.

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning

Background:

• Chemical space visualization is crucial for drug discovery and understanding structure-activity relationships.
• Traditional dimensionality reduction techniques like PCA and MDS have limitations in preserving complex chemical data structures.

Purpose of the Study:

• To introduce a parametric t-SNE approach leveraging deep feed-forward neural networks for improved chemical space visualization.
• To evaluate the method's performance against established techniques like PCA and MDS.
• To demonstrate the utility of this approach in chemical space navigation tasks.

Main Methods:

• Implementation of a parametric t-SNE algorithm utilizing deep feed-forward neural networks.
• Comparative analysis of the proposed method with Principal Component Analysis (PCA) and Multidimensional Scaling (MDS).
• Application of the visualization technique to identify activity cliffs and activity landscapes.

Main Results:

• The parametric t-SNE method demonstrated superior information retention compared to PCA and MDS.
• The approach effectively facilitated the identification of key features within the chemical space.
• Successful application in identifying activity cliffs and landscapes, crucial for drug design.

Conclusions:

• Parametric t-SNE offers a powerful and informative method for chemical space visualization.
• This technique enhances the ability to navigate and interpret complex chemical data.
• The developed tool provides a practical demonstration of the method's capabilities.