Locally weighted learning methods for predicting dose-dependent toxicity with application to the human maximum...

Area of Science:

• Computational toxicology
• Medicinal chemistry
• Pharmacology

Background:

• Animal toxicological experiments are standard for predicting human effects but lack cross-species correlation.
• Existing quantitative structure-activity relationship (QSAR) methods often fail with structurally diverse compounds due to varied molecular mechanisms.
• Predicting human toxicity directly from human data of related compounds is more advantageous.

Purpose of the Study:

• To evaluate locally weighted learning methods for predicting human toxicological effects.
• To address limitations of the k-nearest neighbors (kNN) method in handling diverse compounds.
• To introduce and validate a variable number nearest neighbor (VNN) method for improved QSAR predictions.

Main Methods:

• Application of locally weighted learning, specifically k-nearest neighbors (kNN), to human maximum recommended daily dose data.
• Development and implementation of a variable number nearest neighbor (VNN) method to overcome kNN's constant neighbor limitation.
• Comparison of VNN with other QSAR methods for predicting toxicological effects.

Main Results:

• Locally weighted learning methods, like kNN, are suitable for predicting toxicological effects of structurally diverse compounds.
• The standard kNN method's flaw is using a fixed number of neighbors, regardless of structural similarity and shared mechanisms.
• The proposed VNN method allows for tighter molecular distance thresholds and automatically identifies when predictions are unreliable due to out-of-domain compounds.

Conclusions:

• The variable number nearest neighbor (VNN) method offers more reliable toxicological predictions for diverse chemical compounds.
• VNN enhances QSAR by ensuring that predictions are based on structurally similar compounds with likely shared mechanisms of action.
• VNN provides a mechanism to detect and flag unreliable predictions, improving the applicability domain assessment in toxicological modeling.