AFMOMO: An Agentic Framework for Efficient Multi-Objective Molecular Optimization

Multi-objective molecular optimization requires a single candidate to satisfy several property and structural constraints simultaneously. Existing search methods often rely heavily on random exploration, with limited integration of chemical prior knowledge and limited control over which constraints an edit should address. This can waste oracle evaluations when property assessment is expensive. We present AFMOMO, an agentic framework that couples objective-aware planning, heuristic molecular genetic search, and feedback-driven memory. First, an LLM planner uses objective satisfaction states to select parents, editing operators, and generation-budget allocations. Second, a staged genetic search executes these plans through chemically annotated mutation rules and recombination operators for local refinement, crossover, and broader exploration. Finally, search feedback updates candidate memory and operator weights and enables conditional branch rollback when a search branch regresses and stalls. Across four tasks and two LLM backbones, AFMOMO improves joint success rates over MOMO and MOMO-QMO. With Qwen, it reduces the task-averaged recorded oracle count by 81.13% relative to MOMO on leads successfully optimized by both methods.