Modelling the Tox21 10 K chemical profiles for in vivo toxicity prediction and mechanism characterization

Abstract

Target-specific, mechanism-oriented in vitro assays post a promising alternative to traditional animal toxicology studies. Here we report the first comprehensive analysis of the Tox21 effort, a large-scale in vitro toxicity screening of chemicals. We test ∼ 10,000 chemicals in triplicates at 15 concentrations against a panel of nuclear receptor and stress response pathway assays, producing more than 50 million data points. Compound clustering by structure similarity and activity profile similarity across the assays reveals structure-activity relationships that are useful for the generation of mechanistic hypotheses. We apply structural information and activity data to build predictive models for 72 in vivo toxicity end points using a cluster-based approach. Models based on in vitro assay data perform better in predicting human toxicity end points than animal toxicity, while a combination of structural and activity data results in better models than using structure or activity data alone. Our results suggest that in vitro activity profiles can be applied as signatures of compound mechanism of toxicity and used in prioritization for more in-depth toxicological testing.

Figure 1. Concentration response data of the positive control compounds for the 30 Tox21 phase II assays.

(a) Agonist mode assays; (b) antagonist mode assays. The positive control compound is plated as 16-pt. titrations in duplicate in the control columns of every assay plate. In the figure, each concentration response curve is from one plate with a total of 408 plates per assay. The consistency of the control response curves is an indicator of good assay performance.

Figure 2. Activity distribution of the Tox21 10 K library screened against the 30 assays.

(a) Activity outcome distribution; (b) potency distribution.

Figure 3. Clustered activity profiles of the Tox21 10 K library.

In the heat map, each row is a compound and each column is an assay readout. The heat map is coloured by the compound activity outcome, such that darker red or blue colours indicate more confident activators (red) or inhibitors (blue). Compounds are grouped into clusters of similar activity profiles. Each cluster of compounds is labelled by the most significantly enriched MeSH PA term in that cluster measured by a Fisher's exact test. (a) All 10 K compounds; (b) example clusters of oestrogenic compounds.

Figure 4. AUC-ROC values of predictive models built for 72 in vivo toxicity end points using the Tox21 10 K in vitro assay activity data, compound structure data and a combination of the assay and structure data.

Columns of AUC-ROC values are shown as mean±s.d. (n=100).

Figure 5. Correlation between reproducibility of in vivo data and model performance.

A positive correlation (Pearson correlation: r=0.61, P=1.51 × 10⁻⁵) is found between the reproducibility of the compounds tested in replicates and the model performance (AUC-ROC). Models built for more reproducible data showed better predictive power.