Prediction of Compound Profiling Matrices Using Machine Learning.

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning in drug discovery

Background:

• Compound profiling matrices from screening data present challenges for computational prediction due to diverse structures and imbalanced active/inactive compound ratios.
• Accurate prediction of active compounds is crucial for efficient drug discovery pipelines.

Purpose of the Study:

• To model large compound profiling matrices using machine learning techniques.
• To compare the performance of different machine learning methods, including deep learning, for predicting compound activity.
• To explore various prediction strategies for screening data.

Main Methods:

• Extracted and modeled large compound profiling matrices from publicly available screening data.
• Compared various machine learning methods, including deep learning, random forests, and support vector machines.
• Explored different prediction strategies and evaluated prediction accuracy across assays with varying numbers of active compounds.

Main Results:

• Prediction accuracy was dependent on the number of active compounds in an assay.
• Standard machine learning methods, such as random forests, often yielded comparable results to more complex approaches.
• Deep learning did not significantly improve prediction accuracy over established methods like random forests or support vector machines.
• Target-based random forest models demonstrated successful prediction of active compounds for numerous assays.

Conclusions:

• Machine learning, particularly random forests, can effectively model compound profiling matrices for activity prediction.
• The choice of machine learning method may not be as critical as other factors influencing prediction accuracy in this context.
• Target-based random forest models show promise for identifying active compounds in large screening datasets.