Machine Learning and Large Language Models for Modeling Complex Toxicity Pathways and Predicting Steroidogenesis

Area of Science:

• Endocrinology
• Computational Toxicology
• Pharmacology

Background:

• Estrogen and androgen receptor interactions are well-modeled, but steroidogenesis prediction remains limited.
• Steroidogenesis is crucial for hormone regulation and is a target for chemical disruption.
• Existing methods for assessing chemical effects on steroidogenesis are insufficient for large-scale screening.

Purpose of the Study:

• To develop and validate computational models for predicting chemical modulation of steroidogenesis.
• To identify specific molecular targets within the steroidogenesis pathway affected by chemicals.
• To provide a scalable system for chemical risk assessment and regulatory evaluation.

Main Methods:

• Utilized data from ~1,800 chemicals screened in H295R cells to build random forest models.
• Developed classification and regression models using IC50 data for key steroidogenic enzymes from ChEMBL.
• Employed a transformer-based model (MolBART) for simultaneous prediction of multiple endpoints.

Main Results:

• Random forest model achieved 80% accuracy in prospective validation for general steroidogenesis modulation.
• Models were developed for key enzymes including CYP17A1, CYP21A2, and CYP19A1.
• Transformer model demonstrated validated performance for predicting all endpoints concurrently.

Conclusions:

• The developed models provide a rapid and scalable approach to assess chemical impacts on steroidogenesis.
• These tools can support chemical risk assessment, product stewardship, and regulatory decision-making.
• The models enable prediction of both general steroidogenesis inhibition and specific enzyme targets.