Development and application of consensus in silico models for advancing high-throughput toxicological predictions

Area of Science:

• Computational toxicology
• In silico modeling
• Structure-activity relationships

Background:

• Numerous in silico (quantitative) structure-activity relationship ([Q]SAR) models exist for predicting toxicological endpoints.
• Discrepancies among models arise from variations in training sets, algorithms, and methodologies, leading to data conflicts in high-throughput screening.
• A need exists for strategies to reconcile diverse predictions and expand chemical space coverage.

Purpose of the Study:

• To develop a consensus modeling strategy for combining predictions from multiple in silico (Q)SAR models.
• To create consensus models for nine toxicological endpoints: estrogen receptor (ER) and androgen receptor (AR) interactions (binding, agonism, antagonism), and genotoxicity (bacterial mutation, in vitro chromosomal aberration, in vivo micronucleus).
• To identify optimal consensus models using Pareto fronts for multi-objective decision-making.

Main Methods:

• Developed consensus models by integrating predictions from various existing (Q)SAR models.
• Employed diverse weighting schemes to combine model predictions.
• Utilized Pareto front analysis to determine optimal consensus models that simultaneously optimize multiple criteria for each endpoint.

Main Results:

• Consensus models demonstrated improved predictive power and expanded chemical space coverage compared to individual models.
• Pareto fronts identified sets of optimal consensus models for each of the nine toxicological endpoints.
• Analysis revealed trends between the performance of consensus models and their constituent component models.

Conclusions:

• The developed consensus modeling approach is flexible and adaptable for various toxicological endpoints.
• These consensus models enhance chemical prioritization and support the transition to non-animal testing methods in risk assessment.
• The strategy effectively addresses data conflicts and expands applicability in computational toxicology frameworks.