Endocrine disruption: the noise in available data adversely impacts the models' performance

Area of Science:

• Computational toxicology
• Molecular modeling
• Endocrine disruption research

Background:

• Endocrine-disrupting chemicals (EDCs) pose risks by interfering with hormone systems.
• Oestrogen (ER) and androgen (AR) receptors are key targets for EDC activity.
• Reliable predictive models are crucial for EDC screening.

Purpose of the Study:

• To analyze experimental data for ER and AR receptor binding affinity.
• To develop and evaluate predictive classification and regression models for EDC activity.
• To identify challenges in existing experimental data for model development.

Main Methods:

• Data compilation from multiple sources (CERAPP, CoMPARA, Tox21, ChEMBL, PubChem).
• Generative topographic mapping for data analysis.
• Development of classification and regression models for ER and AR binding.
• External validation of model performance.

Main Results:

• Significant low agreement observed between experimental data from different sources.
• Classification models showed poor sensitivity (ER: 0.34, AR: 0.49).
• Regression models exhibited limited predictive power (RBA/IC50 R-squared: 0.44-0.76).

Conclusions:

• Poor model performance is attributed to misinterpreted or erroneous experimental data.
• Data quality issues confirmed findings from CERAPP and CoMPARA studies.
• Freely available datasets (6215 ER, 3789 AR compounds) are provided for future research.