Molecular generation has evolved from unconditioned sampling toward increasingly constraint-aware forms of molecular design. Among these, spatially aware generation has attracted growing attention, giving rise to pocket-aware and profile-aware systems with impressive expressive power. In practice, however, that expressivity can be a double-edged sword, increasing setup complexity, computational overhead, and resource demands in ways that complicate routine use.​

A complementary approach is to use lightweight, physics-grounded conditioning that preserves useful spatial control while reducing setup complexity, deployment cost, and retraining burden. Such models can be integrated efficiently with established workflows for geometry optimization, docking, and free-energy calculations, allowing generative methods to function as flexible components within broader computational pipelines.​

MLConfGen serves as a case study in this design philosophy, illustrating how compact descriptors and flexible inference can support practical constrained-generation workflows.