SensDiff writes chemical rules into 3D molecule diffusion training in sequence, reaching 96.9% validity on QM9
Synopsis
Diffusion models for 3D molecules learn atomic distributions implicitly from data, so samples can be statistically plausible yet chemically invalid. SensDiff turns four physicochemical rules, steric non-overlap, bond length, bond angle and valence, into differentiable energies that are activated in a coarse-to-fine order with adaptive weighting during training, so the rules enter the generative dynamics rather than only correcting sampling. On the QM9 and GEOM-Drugs benchmarks it reaches 96.9% and 97.2% validity, which the authors report as better than the compared baselines.
Interpretation
3D molecular diffusion models begin to treat physicochemical rules as training signals instead of approximating them implicitly from data. Earlier equivariant architectures, joint discrete-continuous denoising, inference-time guidance and latent-space methods either acquire chemical feasibility only implicitly from data or inject corrective signals at sampling without absorbing them into model parameters. On the QM9 and GEOM-Drugs benchmarks, SensDiff reaches 96.9% and 97.2% validity, and the authors report the best generation quality among the compared baselines, with the largest gains in stability metrics.
The rules act on intermediate denoising states through a differentiable latent-to-soft-molecule interface, so constraint gradients reach the denoising network. Hard decoding fixes atom types as one-hot vectors, which makes the symbolic term constant with zero gradient, so soft decoding that preserves uncertainty in atomic identity is required. Soft decoding yields atom-type probability distributions, a bond posterior head infers a soft bond graph from them, and training-set statistics are interpolated into non-bonded distance lower bounds plus bond-length and bond-angle statistics, from which the four rules become differentiable energies.
The four constraints are activated in stages, steric non-overlap, then bond length, then bond angle, then valence, with each new constraint first ramped linearly and then weighted by how well the rule is satisfied. When all four constraints are enforced simultaneously, the authors report that QM9 molecule stability drops from 88.8% with one constraint to 78.8% with all four, indicating that naive combination introduces competing gradients. Across constraint orders in the ablation, validity ranges from 74.6% to 90.9%, and the authors' chosen order gives the best result.
Perspective
The demonstrated scope is unconditional and conditional generation on the QM9 small-organic-molecule and GEOM-Drugs drug-like benchmarks, plus sampling-time tasks for oral-drug multi-objective constraints and Lipinski Rule-of-Five constraints; the next step the authors propose is learning such rules automatically from data or literature and extending to more diverse molecular distributions and larger chemical spaces.
The authors note that the symbolic constraints are manually specified from established physicochemical knowledge and that the optimal activation order and schedule may vary across domains; how rules and thresholds are set may also bias generation toward over-regularized or undesirable molecules.