Bayesian matrix completion for hypothesis testing.

Area of Science:

• Toxicology and Cheminformatics
• Computational Biology and Bioinformatics
• Statistical Modeling in Life Sciences

Background:

• Toxicology data is often sparse, hindering accurate chemical bioactivity inference.
• Existing methods struggle with predicting activity for unassayed chemicals and quantifying prediction uncertainty.

Purpose of the Study:

• To develop a Bayesian hierarchical framework for inferring chemical bioactivity across assay endpoints.
• To address data sparsity and enable out-of-sample predictions for novel chemicals.
• To introduce a novel approach modeling heteroscedastic errors and nonparametric mean functions for a broader definition of activity.

Main Methods:

• A Bayesian hierarchical model was employed to borrow information across chemicals and assay endpoints.
• The framework facilitates prediction of activity for chemicals not yet assayed.
• Simultaneous modeling of heteroscedastic errors and a nonparametric mean function was implemented.

Main Results:

• The proposed framework effectively infers bioactivity despite sparse toxicology data.
• It enables accurate out-of-sample predictions and quantifies the uncertainty associated with these predictions.
• Real-world application identified specific chemicals with high likelihood of activity related to neurodevelopmental disorders and obesity.

Conclusions:

• The Bayesian hierarchical framework offers a robust solution for inferring chemical bioactivity and addressing data sparsity in toxicology.
• This novel approach enhances predictive capabilities and provides a more comprehensive definition of chemical activity.
• The findings aid in identifying potential toxicological risks and prioritizing chemicals for further investigation, particularly for neurodevelopmental and obesity-related endpoints.