TL;DR
Existing diffusion models for molecular conformation generation either operate on 3D coordinates or assume that inter-atomic distances follow a simple Gaussian distribution under noise. SDDiff (Shifting Distance Diffusion) recognizes that as perturbation forces increase, inter-atomic distance distributions naturally transition from Gaussian (Brownian motion) to Maxwell-Boltzmann (freely moving particles). By deriving a closed-form "shifting score" kernel to bridge these physical regimes, SDDiff generates higher-fidelity 3D molecular conformations while strictly preserving $\mathrm{SE}(3)$-equivariance.
What Problem Does This Paper Address?
Predicting a molecule's stable 3D shape (conformation) from its 2D chemical graph is essential for drug discovery and material design. Generative diffusion models have shown great promise here, but they face a core geometric challenge: $\mathrm{SE}(3)$-equivariance (invariance to 3D rotation and translation).
To preserve symmetry, recent models perform diffusion on rotation-invariant inter-atomic distances rather than raw coordinates. However, existing methods (such as GeoDiff) rely on heuristic assumptions—specifically, that perturbed distances always follow a standard Gaussian distribution. In physical reality, corrupting Cartesian coordinates with Gaussian noise does not induce Gaussian shifts on pairwise distances across all noise scales, leading to mismatched score estimation and unrealistic geometries.
Methods
The authors introduce SDDiff, modeling molecular disintegration through a physically grounded thermodynamic lens:
- Shifting Distribution Mechanics: Under small perturbation forces ($\sigma \to 0$), chemical bonds restrict atom movement into localized Brownian motion, making the distance perturbation Gaussian: $\mathcal{N}(d, 2\sigma^2)$. Under large perturbation forces ($\sigma \to \infty$), bonds are disrupted, and atoms move freely as an ideal gas where relative atomic speeds follow a Maxwell-Boltzmann (MB) distribution: $\text{MB}(\sqrt{2}\sigma)$.
- Closed-Form Shifting Score Kernel: To smoothly bridge these two regimes across intermediate diffusion timesteps without expensive numerical solvers, the authors propose a closed-form conditional score function: $$\nabla_{\tilde{d}} \log q_\sigma(\tilde{d} \mid d) = \left(1 - e^{-\sigma/d}\right) \frac{2}{\tilde{d}} - \frac{\tilde{d} - d}{2\sigma^2}$$
- Preserving 3D Feasibility (Isomorphism): Not all pairwise distance matrices can be embedded into valid 3D Euclidean coordinates. To ensure the distance space is isomorphic to $\mathbb{R}^{n \times 3} / \mathrm{SE}(3)$, each atom requires at least four reference distances. SDDiff restricts distance modeling to 3-hop graph neighborhoods, providing sufficient geometric constraints without over-constraining the model into infeasible distance spaces.
- Architecture: The scoring model $s_\theta$ utilizes a Message Passing Neural Network (combining SchNet for global physical representations and GIN for local graph topology) and updates conformations via annealed Langevin dynamics.
Datasets
The authors evaluated SDDiff on standard molecular conformation benchmarks:
- GEOM-QM9: Small molecules with an average of 11 atoms (40k training, 5k validation, 200 test molecules).
- GEOM-Drugs: Larger, drug-like molecules with an average of 44 atoms (40k training, 5k validation, 200 test molecules).
- Note: No new benchmark datasets were introduced.
Key Insights and Ideas
- The Gaussian-to-Maxwell-Boltzmann Shift: The authors prove mathematically and verify empirically that isotropic Gaussian coordinate noise translates to a Maxwell-Boltzmann distribution for inter-atomic distances at large $\sigma$, not a Gaussian.
- Failure of Entropic Optimal Transport: While Schrödinger bridges / regularized Wasserstein barycenters (via the Sinkhorn algorithm) could theoretically interpolate intermediate distributions, they numerically collapse (yielding
NaNs) at small regularizations or severely blur the distribution at larger ones. The closed-form analytical score avoids this numerical instability entirely. - Long-Tailed Score Targets: As $\sigma$ increases, the score function target transforms into a long-tailed Gaussian distribution, which standard neural network backbones can still effectively learn.
Key Results
Evaluated on conformation coverage (COV, diversity) and matching error (MAT, accuracy in Ångströms via RMSD alignment), SDDiff outperforms prior distance- and coordinate-based diffusion baselines (using an architecture directly comparable to GeoDiff):
- GEOM-QM9:
- COV (Mean / Median): 91.07% / 94.69% (vs. GeoDiff's 90.54% / 94.61% and ConfGF's 88.49% / 94.31%).
- MAT (Mean / Median): 0.2048 Å / 0.1941 Å (vs. GeoDiff's 0.2090 Å / 0.1988 Å).
- GEOM-Drugs:
- COV (Mean / Median): 90.68% / 98.48% (vs. GeoDiff's 89.13% / 97.88%).
- MAT (Mean / Median): 0.8564 Å / 0.8503 Å (vs. GeoDiff's 0.8629 Å / 0.8529 Å).
- Adding a standard Force Field (FF) post-processing relaxation further boosts SDDiff to 93.07% COV and 0.7465 Å MAT.
Limitations and Future Work
- Planar Molecular Motifs: The conversion from distance scores to 3D coordinate scores assumes non-coplanar configurations. Strictly planar local sub-structures (such as aromatic benzene rings) can violate this assumption, occasionally distorting local ring geometry and requiring rule-based post-processing (force-field optimization).
- High-Dimensional Manifold: Unlike torsional diffusion models that restrict diffusion solely to a handful of rotatable torsion angles using pre-built RDKit templates, SDDiff predicts within an ~130-dimensional distance manifold without built-in local chemical priors. Investigating better ways to inject chemical domain priors into purely data-driven distance diffusion remains an open direction.
Summary updated