AMODO-EO: A Practical Framework for Adaptive Objective Discovery in Multi-Objective Drug Optimization

Multi-objective optimization (MOO) in drug discovery typically relies on a fixed set of predefined objectives, such as potency, drug-likeness, and synthetic accessibility. However, this assumption may overlook emergent molecular trade-offs that arise during the optimization process. This study presents AMODO-EO (Adaptive Multi-Objective Drug Optimization with Emergent Objectives), a computational framework designed to discover and integrate new chemically meaningful objectives during active optimization. AMODO-EO builds on established principles from symbolic regression and automated feature engineering. It operates by generating candidate objective functions from molecular descriptors using mathematical transformations such as ratios, products, and differences. These candidates are evaluated for statistical independence from existing objectives, sufficient variance across the population, and chemical interpretability. Validated objectives are incorporated into the optimization process using adaptive weighting and conflict resolution mechanisms. The framework was tested on benchmark datasets and two ChEMBL subsets targeting the dopamine D2 receptor. Initial objectives included binding affinity (pIC₅₀), a drug-likeness composite score, and synthetic accessibility. AMODO-EO consistently identified emergent objectives such as the hydrogen bond acceptor to rotatable bond ratio (HBA/RTB), MW/TPSA ratio, and LogP × aromatic ring count. These objectives passed statistical filters and were chemically interpretable, suggesting they capture relevant trade-offs not encoded in the original objective set. Optimization performance was evaluated using hypervolume indicators and compared against baseline NSGA-II runs. AMODO-EO maintained competitive performance on original objectives, with a modest reduction in hypervolume (1.5–2.9%), attributed to the computational cost of discovery. Ablation studies confirmed that statistical filtering and interpretability checks are essential for maintaining optimization quality and avoiding spurious objectives. The inclusion of emergent objectives expanded the Pareto front into higher-dimensional spaces, revealing new solution clusters with distinct chemical profiles. For example, the HBA/RTB ratio highlighted a polarity-flexibility trade-off relevant to ligand design. This objective was consistently discovered across all runs and parameter settings, indicating robustness of the discovery mechanism. Other emergent objectives, such as the Selectivity Score and Synthetic Complexity metric, provided additional design axes that could inform compound prioritization. Sensitivity analysis showed that AMODO-EO is stable across a range of parameter values, including correlation thresholds, learning rates, and weight-mixing parameters. The framework’s reliance on interpretable descriptor relationships ensures that discovered objectives are scientifically grounded and potentially actionable in medicinal chemistry workflows. While AMODO-EO currently uses predefined functional forms and semi-automated interpretability checks, future work may incorporate grammar-based symbolic regression and ontology-driven validation to enhance discovery scope and automation. The framework is applicable to various chemical optimization contexts and can be integrated with generative models or molecular simulation pipelines. In summary, AMODO-EO provides a structured approach to discovering emergent objectives in drug optimization. It extends existing MOO techniques by enabling dynamic expansion of the objective space, offering a more comprehensive view of molecular trade-offs. The framework demonstrates that adaptive objective discovery can be achieved without compromising optimization performance, and that chemically interpretable objectives can be systematically identified from descriptor data.