Shallow Representation Learning via Kernel PCA Improves QSAR Modelability.

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning in drug discovery

Background:

• Linear models for quantitative structure-activity relationships (QSARs) are efficient but often outperformed by nonlinear methods.
• Support vector machines (SVMs) and neural networks excel by learning data representations, improving model accuracy.
• Existing QSAR methods face challenges in balancing performance, computational efficiency, and flexibility.

Purpose of the Study:

• To improve the performance of L1 regularized logistic regression (LASSO) using shallow representation learning.
• To achieve performance comparable to nonlinear methods like Tanimoto SVM for QSAR prediction.
• To evaluate the computational efficiency of enhanced linear models compared to nonlinear alternatives.

Main Methods:

• Embedding chemical fingerprints into Euclidean space via Tanimoto similarity kernel principal component analysis (KPCA).
• Applying LASSO and SVM classifiers to predict binding activities of chemical compounds against 102 virtual screening targets.
• Comparing model performance, training times, and cross-validation efficiency between LASSO and SVM.

Main Results:

• LASSO, enhanced with KPCA, demonstrated performance comparable to Tanimoto SVM.
• Similar performance improvements were observed for both LASSO and SVM when using KPCA.
• KPCA combined with LASSO classification showed significantly faster computation than linear SVM across various training set sizes.

Conclusions:

• Powerful linear QSAR methods can achieve performance levels rivaling nonlinear methods through representation learning.
• A modular approach to nonlinear classification enhances QSAR model prototyping, flexibility, and transferability.
• This study highlights a computationally efficient strategy for improving QSAR model accuracy and applicability.