Hierarchical Mechanistic Modeling of Complex Toxicity Endpoints from Public Concentration–Response Data

TL;DR

This paper introduces a new framework that organizes and interprets high-throughput screening data to predict toxicity outcomes and understand biological pathways.

Contribution

A hierarchical mechanistic modeling framework that integrates HTS data with biological pathways to predict toxicity and enhance interpretability.

Findings

Integrated 455 PubChem assays with 216 protein targets and 103 biological pathways.

Generated pathway-level toxicity scores for five in vivo toxicity endpoints.

Linked chemical bioactivity to adverse outcomes for compound ranking and hazard prediction.

Abstract

High-throughput screening (HTS) programs have generated
abundant
data on numerous chemicals, supporting the discovery of toxicity mechanisms
and advancing understanding of adverse outcome pathways (AOPs) in
chemical toxicity. However, organizing and interpreting these data
for predictive modeling remain challenging due to inconsistent repository
formats, varied program objectives, heterogeneous assay targets, and
differences in experimental protocols, including concentration ranges.
To address these limitations, we developed a hierarchical mechanistic
modeling framework that systematically structures and interprets concentration
response HTS data. The model integrated curated data sets by mapping
metadata from 455 PubChem assays to 216 protein targets and 103 biological
pathways in WikiPathways. Assay-level concentration–response
data were organized within a biologically layered hierarchy to construct
AOP-based models. The resulting models generated pathway-level toxicity
scores that quantified compound potency by integrating inferred protein
activity and downstream pathway perturbations. In total, 103 pathways
were statistically associated with five in vivo toxicity
endpoints: acute systemic, maternal, developmental, human hepatotoxicity,
and preclinical hepatotoxicity. This hierarchical framework leverages
HTS metadata to predict diverse in vivo toxicity
outcomes and enhance mechanistic interpretability of pathway-level
effects. It links chemical bioactivity to adverse outcomes, providing
a quantitative basis for compound ranking, potency assessment, and
hazard prediction. Overall, this framework offers a structured, scalable
method for integrating large bioactivity data sets into computational
toxicology, supporting chemical risk assessment and early-stage drug
discovery.

Figures