Do we need different machine learning algorithms for QSAR modeling? A comprehensive assessment of 16 machine learning...

Area of Science:

• Computational Chemistry
• Machine Learning in Drug Discovery

Background:

• Quantitative Structure-Activity Relationship (QSAR) modeling is crucial for predicting biological activity and properties of chemical compounds.
• A diverse range of machine learning (ML) algorithms are applied to QSAR, but no single algorithm is universally optimal.
• Understanding the performance characteristics of different ML algorithms in QSAR is essential for effective model development.

Purpose of the Study:

• To comprehensively evaluate and compare the performance of various popular ML algorithms for regression-based QSAR modeling.
• To identify the most effective ML algorithms for QSAR tasks across different data set sizes and endpoint types.
• To investigate the potential of ensemble methods in enhancing QSAR prediction accuracy.

Main Methods:

• Employed 16 ML algorithms across linear, analogizer, symbolist, and connectionist categories for QSAR regression.
• Trained and tested models on 14 public datasets covering nine physicochemical properties and five toxicity endpoints.
• Assessed algorithm performance based on prediction accuracy and computational efficiency.

Main Results:

• Radial basis function support vector machine (rbf-SVM), radial basis function Gaussian process regression (rbf-GPR), extreme gradient boosting (XGBoost), and deep neural network (DNN) generally demonstrated superior performance.
• A performance ranking of algorithms was established, with rbf-SVM and XGBoost identified as top performers.
• Ensemble models integrating predictions from multiple ML algorithms showed improved predictive accuracy over individual best algorithms.

Conclusions:

• rbf-SVM and XGBoost are recommended for regression learning in QSAR, particularly for small datasets, while XGBoost is excellent for large datasets.
• Ensemble methods offer a viable strategy to boost the predictive power of QSAR models.
• The study provides valuable insights into algorithm selection for QSAR modeling, aiding researchers in choosing appropriate tools for their specific needs.