Synthesis-native language modelling unifies target-aware molecular generation and optimization

Designing molecules for disease-relevant protein targets is a central goal in drug discovery, yet existing target-aware generative models often rely on high-quality three-dimensional protein structures and overlook synthesizability. We present OmniSyn, a synthesis-native molecular language model that unifies de novo design, synthesizability projection, and hit-to-lead (H2L) optimization while jointly generating molecules and their synthesis routes. OmniSyn employs a two-stage training strategy that combines route self-distillation with task-specific reinforcement learning. On 35 unseen targets from the MolGenBench de novo ligand-design benchmark, OmniSyn achieves state-of-the-art performance across key metrics, attaining the highest scaffold- and SMILES-level hit fractions, SMILES-level target hit rate, and Target-Aware (TA) score among all evaluated methods. Independent retrosynthetic evaluation further shows that OmniSyn outperforms the strongest de novo and H2L baselines by 61.3% and 184.2% in AiZynthFinder success rate, respectively. OmniSyn also projects molecules generated by baseline ligand-design models into close, synthesizable analogs while preserving molecular similarity and docking compatibility. These results establish OmniSyn as a unified, synthesis-native framework for target-aware molecular design and synthesis planning.