Generative molecular design inverts the screening logic. Instead of ranking a fixed catalog, a model learns the statistical patterns of valid, drug-like molecules and then proposes new structures that fit those patterns while satisfying a desired property — for example, complementing a particular binding pocket. The output is a molecule that has never been synthesized and may not appear in any catalog.
The process is constructive rather than selective. The model assembles a structure piece by piece, adding fragments or atoms in a sequence, and at each step it is guided toward the target property. Because the space of possible drug-like molecules is vastly larger than any library, generation can in principle reach regions that screening cannot. But that reach comes with a cost: nothing guarantees the proposed molecule can actually be made, and nothing guarantees it is stable or safe. Generation expands the candidate space; it does not filter it for practicality.