Active Learning for Optimal Experimental Design in Alzheimer's Disease Drug Discovery: Prioritizing NAD+-Enhancing Therapeutic Analogs via Multi-Objective Bayesian Optimization

P7C3-A20, an aminopropyl carbazole that activates nicotinamide phosphoribosyltransferase (NAMPT), has demonstrated reversal of advanced Alzheimer’s disease (AD) pathology in aged, symptomatic mice—normalizing 174 differentially expressed proteins, restoring cognition, and repairing blood-brain barrier (BBB) integrity. However, translating this breakthrough requires navigating a vast chemical space to identify analogs with optimal efficacy, brain penetrance, metabolic stability, and safety. We propose Active Learning for NAD+ Therapeutics (ALNAT), a framework integrating multi-objective Bayesian optimization with hierarchical experimental feedback to accelerate P7C3-A20 analog discovery. ALNAT employs Gaussian process surrogates over six drug-relevant objectives and uses expected hypervolume improvement (EHVI) for batch compound selection across a six-tier multi-fidelity assay pipeline. Retrospective computational experiments on published P7C3 series data show that our GP-based active learning identifies the top 10% of compounds using only 23% of the experimental budget required by random sampling. We enumerate approximately 47,000 readily synthesizable analogs, of which 15% are predicted to lie on an extended Pareto front, representing promising candidates for prospective experimental validation.