Selectivity-aware dual-target cholinesterase optimization using a novel model SPaR-GNN

We developed SPaR-GNN (Selectivity Pair-aware Ranking Graph Neural Network), a multitask molecular learning framework for dual-target cholinesterase optimization. SPaR-GNN jointly predicts acetylcholinesterase (AChE) activity, butyrylcholinesterase (BuChE) activity, and molecular selectivity while incorporating explicit pair-aware supervision from structurally similar analogs and selectivity-cliff pairs. In addition, we integrated SPaR-GNN into a goal-driven molecular optimization workflow for candidate generation and prioritization. Experimental results showed that SPaR-GNN performed satisfactorily in selectivity-oriented prediction, and particularly in local analog ranking and direction-aware pair-sensitive benchmarks. The optimization framework further demonstrated that template-based candidate generation combined with SPaR-GNN scoring can be used to prioritize candidate molecules according to predefined dual-target optimization objectives. Together, these results demonstrate that SPaR-GNN provides an effective and traceable computational framework for dual-target cholinesterase inhibitor discovery and optimization in Alzheimer's disease.