Systematic artifacts in support vector regression-based compound potency prediction revealed by statistical and...

Area of Science:

• Computational Chemistry
• Chemoinformatics
• Machine Learning in Drug Discovery

Background:

• Support vector machines (SVMs) are widely used for classification in biology and chemistry.
• Support vector regression (SVR) is a popular machine learning technique for numerical property predictions.
• SVR is extensively applied in chemoinformatics and pharmaceutical research for modeling non-linear structure-activity relationships (SARs) and predicting compound potency.

Purpose of the Study:

• To systematically generate and analyze SVR prediction models across diverse compound datasets with varying SAR characteristics.
• To evaluate the performance of SVR models, particularly their accuracy in predicting highly potent compounds and capturing local SAR discontinuities.

Main Methods:

• Generation of multiple SVR prediction models for various compound datasets.
• Systematic analysis of model performance using global prediction statistics.
• Assessment of model behavior in regions of local SAR discontinuity and comparison of predicted vs. observed activity landscapes.

Main Results:

• SVR models demonstrated global accuracy and were not prone to overfitting.
• Consistent misprediction of highly potent compounds was observed.
• SVR models exhibited limitations in regions of local SAR discontinuity, leading to flattened activity landscapes and loss of activity cliff information.

Conclusions:

• While SVR models are generally accurate, they possess inherent limitations in predicting highly potent compounds.
• The flattening of activity landscapes and loss of activity cliff information can impact drug discovery efforts.
• Prospective SVR-based potency predictions require caution, as artificially low predictions are likely for the most potent candidate compounds.