Molecular Graph Generation

Molecular graph generation models molecules as nodes for atoms and edges for bonds, aiming to produce valid, diverse candidates for drug discovery, materials design, and structure inference. Common approaches include discrete diffusion and flow matching, autoregressive and fragment-based generation, and latent-space models; some jointly generate 3D geometry or condition on properties, spectra, or molecular fragments. Chemical constraints, permutation symmetry, and reinforcement-learning or representation-based guidance help steer generation, with performance often assessed using validity, novelty, diversity, property control, and benchmarks such as QM9, ZINC, and MOSES. A central tension is improving chemical and geometric fidelity while making sampling faster and more controllable, especially for larger molecules and flexible-size or out-of-distribution designs.