Alzheimer’s disease (AD) remains a major therapeutic challenge, characterized by high clinical trial failure rates and limited efficacy of current treatments. Drug repurposing offers a faster, lower-risk route to new therapies; however, existing computational approaches often prioritize predictive accuracy over mechanistic novelty and interpretability, both of which are critical for clinical translation. We introduce a quality-diversity Automated Machine Learning (AutoML) framework that integrates biologically informed graph neural network (GNN) embeddings with a MAP-Elites-guided search to discover predictive yet mechanistically distinct therapeutic hypotheses. Drugs and genes are embedded using GraphSAGE and variational graph autoencoders trained on the Alzheimer’s Knowledge Base (AlzKB), with a clustering loss used to anchor known AD entities and define dimensions of biological novelty. In an AD case study using matched ADSP GWAS-derived features, our framework successfully recovered known drug–gene relationships and identified robust consensus candidates across independent validation runs. Most notably, the search consistently prioritized Triclosan—a recently identified environmental risk factor for AD neuroinflammation—and the Ketamine/Quazepam pair, suggesting a model-driven preference for restoring synaptic E/I balance. Furthermore, exploratory leads such as Exemestane and Felodipine were identified in underrepresented biological niches, supported by enrichment in oxidative stress and autophagy pathways. The framework demonstrated high stability across multiple random seeds and a 48% reduction in computational cost compared to standard multi-objective evolutionary baselines. Beyond AD, this framework offers a generalizable strategy for integrating biomedical knowledge graphs with diversity-enhancing AutoML to accelerate the discovery of mechanistically novel drug candidates across complex polygenic diseases.
Shao et al. (Mon,) studied this question.