Stargate GTM: Bridging Descriptor and Activity Spaces.

Area of Science:

• Chemoinformatics
• Machine Learning
• Computational Chemistry

Background:

• Predicting molecular activity and discovering active compounds are key goals in chemoinformatics.
• Bridging molecular descriptor and activity spaces is crucial for these tasks.
• Existing methods like Generative Topographic Mapping (GTM) lack supervised learning capabilities for activity prediction.

Purpose of the Study:

• Introduce the Stargate Generative Topographic Mapping (S-GTM) approach.
• Link structural descriptor space and activity space via a common 2D latent space.
• Develop a supervised method for predicting activity profiles and identifying relevant chemical structures.

Main Methods:

• The S-GTM algorithm trains manifolds simultaneously in descriptor and activity spaces.
• It uses a weighted geometric mean of probability distributions in the latent space.
• The approach is supervised, incorporating activity data during training.

Main Results:

• S-GTM can predict a molecule's entire activity profile from its descriptors.
• It can identify potential chemical structures when given a desired activity profile.
• Performance was assessed using GPCR ligands and their pKi values for eight targets.

Conclusions:

• S-GTM offers a supervised approach to chemoinformatics tasks, unlike conventional GTM.
• The method effectively predicts activity profiles and aids in structure discovery.
• S-GTM shows competitive performance compared to other machine learning algorithms like Lasso and Random Forest.