Exploring Dimensionality Reduction Techniques for Deep Learning Driven QSAR Models of Mutagenicity.

Area of Science:

• Computational chemistry
• Toxicology
• Machine learning

Background:

• Deep learning quantitative structure-activity relationship (QSAR) models require dimensionality reduction for high-dimensional toxicological data.
• The selection of dimensionality reduction techniques in QSAR is often arbitrary and lacks thorough investigation.

Purpose of the Study:

• To compare the effectiveness of six dimensionality reduction techniques (linear and non-linear) in enhancing deep learning QSAR models for mutagenicity prediction.
• To assess the impact of these techniques on model performance and chemical space navigation.

Main Methods:

• Applied six dimensionality reduction techniques (e.g., PCA, kernel PCA, autoencoders) to a high-dimensional mutagenicity dataset.
• Trained a deep learning QSAR model using each technique, optimizing hyperparameters via grid search.
• Analyzed chemical space using XLogP and molecular weight to define the applicability domain.

Main Results:

• Principal Component Analysis (PCA), a linear technique, achieved optimal QSAR model performance, suggesting approximate linear separability of the data.
• Non-linear methods like kernel PCA and autoencoders performed comparably, offering broader applicability to potentially non-linear datasets.
• Analysis revealed most data within the applicability domain, with specific regions negatively impacting performance.

Conclusions:

• Linear dimensionality reduction techniques are sufficient for QSAR models when data is linearly separable.
• Non-linear techniques provide valuable alternatives for complex datasets and can facilitate unique navigations of chemical space.
• Understanding data distribution within the applicability domain is crucial for robust QSAR model development.