Using machine learning methods to predict experimental high-throughput screening data.

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning in drug discovery

Background:

• High-throughput screening (HTS) is essential for drug discovery but remains expensive.
• Technological advancements in biotechnology have not fully mitigated the cost of HTS.
• Predictive modeling offers a cost-effective alternative to experimental HTS.

Purpose of the Study:

• To evaluate machine learning methods for predicting experimental HTS measurements.
• To compare the predictive performance of different machine learning algorithms.
• To identify optimal molecular and atomic descriptors for virtual HTS.

Main Methods:

• Analysis of the Test assay dataset from the McMaster University Data Mining and Docking Competition.
• Application of six machine learning algorithms: binary decision trees, neural networks, support vector machines (SVM), linear discriminant analysis, k-nearest neighbors, and partial least squares.
• Evaluation of molecular and atomic descriptors, method sensitivity, false positive/negative rates, and enrichment factor.

Main Results:

• Comparison of six machine learning methods based on predictive performance metrics.
• Identification of the most informative molecular and atomic descriptors for HTS prediction.
• Implementation and application of a variable selection procedure to enhance prediction sensitivity, particularly with polynomial SVM.

Conclusions:

• Machine learning methods can effectively predict HTS assay results, enabling virtual HTS.
• The choice of descriptors and machine learning algorithm significantly impacts prediction accuracy.
• Variable selection can further improve the sensitivity of predictive models for drug discovery applications.