Guided optimization of ToxPi model weights using a Semi-Automated approach

Area of Science:

• Toxicology and computational biology
• Data science and statistical modeling

Background:

• The Toxicological Prioritization Index (ToxPi) is a tool for visualizing and analyzing complex, high-dimensional toxicological data.
• ToxPi synthesizes data into weighted profiles for assessing chemical toxicity and other sample entities, but relies on user-defined feature weights.
• Current methods for predicting feature weights are often statistically flawed and lack global distributional expectations.

Purpose of the Study:

• To develop and evaluate a novel semi-supervised method for predicting feature weights in ToxPi models.
• To address the challenge of user-defined feature weights by leveraging inherent knowledge of reference samples.
• To improve the accuracy of ToxPi-based sample ranking and chemical prioritization.

Main Methods:

• A semi-supervised approach combining ordinal regression with a customized genetic algorithm to predict feature weights.
• Ordinal regression is used to predict weights based on known response levels of reference samples.
• A genetic algorithm fine-tunes these weights for optimal model performance.

Main Results:

• Simulation studies demonstrate that the proposed method can improve upon standard ordinal regression for feature weight prediction.
• The method allows for accurate sample ranking even with limited reference data.
• Testing on published data validates the approach against expert-assigned weights.

Conclusions:

• The developed semi-supervised method offers a robust approach to determining feature weights for ToxPi.
• This technique enhances the reliability and accuracy of ToxPi models for chemical assessment and other profiling applications.
• The method provides a data-driven solution for weight assignment, reducing reliance on subjective expert input.