Undersampling techniques for non-linear chemical space visualization.

Area of Science:

• Computational chemistry
• Data science
• Cheminformatics

Background:

• High-dimensional chemical space visualization is crucial for understanding molecular diversity and structure-property relationships.
• Non-linear dimensionality reduction (DR) methods (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 techniques on hyperparameter selection for non-linear DR methods.
• To assess if undersampling can mitigate computational costs and improve the quality of chemical space projections.

Main Methods:

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

Main Results:

• Undersampling significantly reduces the computational burden of hyperparameter training for non-linear DR methods.
• The choice of undersampling strategy impacts hyperparameter selection and the final visualization.
• Projections generated from undersampled data effectively preserve the local structure of the original chemical space.

Conclusions:

• Undersampling offers an efficient strategy for optimizing non-linear DR methods in cheminformatics.
• This approach enables more accessible and effective exploration of high-dimensional chemical data.
• Undersampling facilitates better compound selection and understanding of molecular diversity through improved visualization.