Aligning Molecules and Fragments in a Shared Embedding Space for Property-Aware Fragment Selection

Abstract Fragment-based molecular design offers a chemically grounded strategy for exploring drug-like chemical space, yet identifying fragments that are both chemically feasible and relevant to a target property remains challenging. Fragment-level attribution is inherently ambiguous because each molecule contains multiple fragments and the same fragment can occur in molecular contexts with different property values. We introduce Molecule–Fragment Representation Alignment for Generation (M-FRAG), a property-aware framework that aligns molecules and fragments in a shared embedding space. M-FRAG organizes fragments according to their local relationships with parent molecules and their global distribution across molecular properties. Candidate fragments are then prioritized by their distance and distributional proximity to property-relevant molecular regions and used to guide reinforcement-learning-based molecular assembly. Across five docking-based multiobjective optimization tasks, M-FRAG improves the novel hit ratio by up to 58.00% over the strongest locally evaluated primary baseline and achieves the best Novel Top-5% docking scores for all five targets while maintaining strong novelty and chemical validity. These results establish embedding-space-guided fragment selection as an effective and traceable strategy for property-optimized molecular generation.