Compound Structure-Independent Activity Prediction in High-Dimensional Target Space.

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning in drug discovery

Background:

• Compound library profiling against multiple targets is crucial in pharmaceutical research.
• Predicting multi-target compound activities using machine learning offers significant potential for drug discovery.

Purpose of the Study:

• To explore and compare different machine learning models for predicting compound activities across a high-dimensional target space.
• To evaluate the effectiveness of activity profile data versus compound structure information for multi-target activity prediction.

Main Methods:

• Development and comparison of Naïve Bayesian (NB) and Support Vector Machine (SVM) models.
• Models were based on compound structure information, activity profiles, or a hybrid approach.
• Investigation of NB models utilizing only activity profiles, even with incomplete training data.

Main Results:

• NB models based on activity profiles demonstrated superior accuracy in predicting multi-target compound activities compared to structure-based NB and SVM models.
• Structure-based SVM models were not applicable due to data limitations and feature independence assumptions.
• Analysis revealed target correlations within activity profiles, explaining prediction accuracy.

Conclusions:

• Activity profile information is highly effective for predicting compound activity against novel targets.
• NB models leveraging activity profiles represent a powerful approach for multi-target activity prediction in drug discovery.
• This study highlights the value of historical activity data over structural data for certain predictive tasks.