Multi-Endpoint Prediction of Bioactivity and ADMET Properties Coupled with Multi-Objective Optimization in Molecular Descriptor Space for Candidate ERα Antagonists

Objectives: Candidate estrogen receptor alpha (ERα) antagonists should combine strong inhibitory activity with favorable absorption, distribution, metabolism, excretion, and toxicity (ADMET)-related properties. This study developed an integrated framework for multi-endpoint prediction, multi-objective optimization in molecular descriptor space, and prioritization of feasible solutions. Methods: A dataset of 1974 compounds with 729 molecular descriptors, pIC50 values, and five binary ADMET-related endpoints was analyzed. Key descriptors were identified through two-stage cross-method and cross-endpoint screening. One regression model for pIC50 and five independent classification models for the ADMET-related endpoints were developed and evaluated separately. The final selected models were incorporated into a two-phase hybrid adaptive multi-objective evolutionary algorithm (2P-HAMOEA). Feasible Pareto solutions were ranked using objective weighting, multi-criteria decision-making, reliability-based fusion, and a performance uncertainty index-based adjustment. Results: The final dataset comprised 1875 compounds and 24 key molecular descriptors. The three-model stacking ensemble for pIC50 achieved an R2 of 0.650 in the internal holdout set, and the AUC values of the five ADMET models ranged from 0.878 to 0.979. Under the prespecified criteria, 2P-HAMOEA generated 155 feasible Pareto solutions. Its generational distance was significantly lower than that of four comparator algorithms, whereas its inverted generational distance and hypervolume were not superior to those of most comparators. pIC50 received the highest fused weight (0.328), and pairwise rank correlations among TOPSIS, VIKOR, and WASPAS exceeded 0.90. The top 20% of RWS-ranked solutions (n = 31) were used for descriptor-interval analysis. Conclusions: The framework offers an early-stage, descriptor-level way to examine endpoint-specific predictions before alternative multi-endpoint profiles are compared. The Pareto solutions are numerical descriptor profiles rather than explicit or experimentally validated ERα antagonist structures; independent validation and molecular structure generation remain necessary.