Undersampling Techniques for Nonlinear Chemical Space Visualization

Area of Science:

• Computational chemistry
• Cheminformatics
• Data visualization

Background:

• Visualizing high-dimensional chemical space is crucial for understanding molecular diversity and structure-property relationships.
• Non-linear dimensionality reduction (DR) techniques (e.g., t-SNE, UMAP, GTM) are powerful but sensitive to hyperparameters and computationally expensive for large datasets.

Purpose of the Study:

• To investigate the impact of undersampling methods on hyperparameter selection for non-linear DR techniques.
• To assess if undersampling can mitigate computational costs and improve the performance of DR methods.

Main Methods:

• Applied undersampling techniques to create representative subsets of large chemical datasets.
• Trained and evaluated non-linear DR methods (t-SNE, UMAP, GTM) using both full and undersampled datasets.
• Analyzed the effect of undersampling on hyperparameter optimization and the quality of resulting chemical space projections.

Main Results:

• Undersampling significantly reduces the computational cost of hyperparameter training for non-linear DR methods.
• Selecting small, representative subsets enables effective hyperparameter tuning.
• Projections generated from undersampled data demonstrate improved preservation of local chemical structure.

Conclusions:

• Undersampling is an effective strategy for optimizing hyperparameter selection in non-linear DR methods.
• This approach offers a computationally efficient and innovative way to visualize complex chemical spaces.
• Undersampling enhances the utility of DR techniques for guiding compound selection and understanding molecular diversity.