Source-linked AI summary
Learning Gradient Fields for Molecular Conformation Generation
Chence Shi, Shitong Luo, Minkai Xu, Jian Tang
TL;DR
Molecular conformation generation seeks stable 3D structures from 2D molecular graphs, but distance-based pipelines can introduce reconstruction errors. ConfGF directly learns coordinate log-density gradients, converts distance gradients into equivariant coordinate gradients, and samples conformations with Langevin dynamics. Across multiple tasks, it outperforms previous state-of-the-art baselines by a significant margin.
Problem
Generating valid and stable molecular conformations from molecular graphs remains challenging, while distance-based methods can incur errors during distance-to-coordinate reconstruction.
Method
ConfGF estimates log-density gradients over interatomic distances, converts them into roto-translation-equivariant coordinate gradients, and generates conformations in one Langevin-dynamics stage.
Results
ConfGF outperforms previous state-of-the-art baselines by a significant margin across multiple tasks.
Takeaways & Limitations
Direct coordinate-space sampling with physically inspired gradient fields produces accurate and diverse conformations without the two-stage distance-reconstruction procedure.
Takeaways & Limitations
The current model generates all possible stereoisomers and leaves explicit stereochemistry handling for future work.
Abstract
from arXiv · showhide
We study a fundamental problem in computational chemistry known as molecular conformation generation, trying to predict stable 3D structures from 2D molecular graphs. Existing machine learning approaches usually first predict distances between atoms and then generate a 3D structure satisfying the distances, where noise in predicted distances may induce extra errors during 3D coordinate generation. Inspired by the traditional force field methods for molecular dynamics simulation, in this paper, we propose a novel approach called ConfGF by directly estimating the gradient fields of the log density of atomic coordinates. The estimated gradient fields allow directly generating stable conformations via Langevin dynamics. However, the problem is very challenging as the gradient fields are roto-translation equivariant. We notice that estimating the gradient fields of atomic coordinates can be translated to estimating the gradient fields of interatomic distances, and hence develop a novel algorithm based on recent score-based generative models to effectively estimate these gradients. Experimental results across multiple tasks show that ConfGF outperforms previous state-of-the-art baselines by a significant margin.
1. Introduction
Molecular conformations are intrinsic 3D structures, but generating valid and stable conformations remains challenging. ConfGF directly learns coordinate log-density gradients and uses them to generate conformations with Langevin dynamics while preserving roto-translation equivariance.
- Molecular conformations represent atoms by 3D coordinates and provide a more natural representation than molecular graphs.
- Existing conformation-generation methods use molecular dynamics, with forces derived from expensive DFT calculations or crude hand-designed force fields.
- ConfGF directly learns gradient fields of the log density with respect to atomic coordinates, treating them as pseudo-forces for Langevin dynamics.
- The coordinate gradient fields are roto-translation equivariant: they rotate with the molecular system and remain invariant under translation.
- ConfGF estimates gradients with respect to interatomic distances and recovers coordinate gradients through the chain rule, enabling one-stage coordinate-space sampling.
- ConfGF outperforms previous state-of-the-art baselines by a clear margin across conformation generation, distance modeling, and property-prediction tasks.
2. Related Work
Prior approaches span computationally expensive molecular dynamics, efficient but less accurate empirical rules, and data-driven neural models. ConfGF differs from neural force-field models by generating equilibrium conformations directly from learned pseudo-forces using only equilibrium conformations for training.
- Molecular dynamics sequentially updates atomic coordinates using forces from physical interatomic-potential models but can require expensive quantum-mechanical calculations.
- Empirical methods improve efficiency by fixing molecular distances and angles to idealized values according to predefined rules, at the cost of accuracy.
- CVGAE directly generates 3D coordinates but does not model roto-translation equivariance, producing large differences from ground-truth structures.
- Distance-based neural methods preserve equivariance but require two stages: distance generation followed by distance-geometry reconstruction, which is vulnerable to noisy distances.
- Neural force fields predict molecular energies and forces as alternatives to quantum chemistry, whereas ConfGF generates equilibrium conformations within one stage.
- ConfGF defines gradient fields as non-energy-conserving pseudo-forces that move atoms toward high-density regions and trains only on equilibrium conformations.
3. Preliminaries
The paper formulates conformation generation on extended molecular graphs and introduces score-based sampling alongside roto-translation equivariance. Virtual bonds reduce geometric freedom, while distance-based parameterization provides the basis for equivariant coordinate scores.
- 3.1. Problem Definition: A molecular graph represents atoms as nodes with nuclear charges and coordinates, and bonds as edges carrying bond types and Euclidean distances.
- 3.1. Problem Definition: Virtual bonds connect atoms two or three hops apart to constrain angles and dihedral angles, because bonded edges alone do not characterize conformations.
- 3.1. Problem Definition: The task is to learn a generative model that produces molecular conformations from an extended molecular graph.
- 3.2. Score-Based Generative Modeling: The score function is ∇x log p(x), and score-based models learn it with neural networks before generating samples using Langevin dynamics.
- 3.2. Score-Based Generative Modeling: Noise at multiple intensities enables joint estimation of score functions in regions lacking training data, followed by annealed Langevin dynamics across noise levels.
- 3.3. Equivariance in Molecular Geometry: Equivariance requires transformed inputs to produce correspondingly transformed outputs, a property important for generalization in physical systems.
- 3.3. Equivariance in Molecular Geometry: For conformations, the coordinate score is roto-translation equivariant because the conditional log density is invariant under these transformations.
4. Proposed Method
ConfGF learns distance-based score fields and converts them into roto-translation-equivariant coordinate gradients, enabling direct conformation generation with Langevin dynamics. Its noise-conditional network is trained on perturbed distances, while the resulting pseudo-forces guide iterative sampling.
- 4.1. Overview: ConfGF parameterizes coordinate log density through interatomic distances and a graph neural network over the molecular graph.This distance-based design supports rotation and translation invariance of the energy representation.
- 4.2. Noise Conditional Score Networks for Distances: The distance network jointly estimates scores for Gaussian-perturbed distance distributions at multiple noise levels using denoising score matching.Scores are represented as edge-level scalar predictions from graph-derived edge embeddings and scaled by the noise level.
- 4.1. Overview: The method estimates distance score fields first, then backpropagates them to atomic coordinates through the chain rule without extra parameters.For each atom, neighbor-distance scores are combined with unit vectors between atom pairs.
- 4.1. Overview: The resulting coordinate score network is roto-translation equivariant: vectors rotate with the molecule while remaining unchanged under translation.This follows from distance-based invariance and the coordinate-direction factors in the chain-rule construction.
- 4.3. Conformation Generation via Annealed Langevin Dynamics: ConfGF generates conformations with annealed Langevin dynamics, starting from a random prior and iteratively updating coordinates using the estimated score fields.Noise levels are gradually reduced, and each update uses distances computed from the current conformation before applying the coordinate score.
- 4.4. Discussion: The current treatment generates all possible stereoisomers, while torsional-angle gradients are proposed as a future extension for handling stereochemistry.The stereochemistry study is explicitly left beyond the scope of this work.
5. Experiments
Experiments evaluate ConfGF across conformation generation, distance-distribution modeling, and ensemble property prediction on GEOM and ISO17 benchmarks. ConfGF achieves strong performance across these tasks, while its direct generation strategy avoids post-processing and improves conformation quality over distance-based alternatives.
- Experimental Tasks: The experiments cover conformation generation, distributions over distances, and property prediction, using GEOM-QM9, GEOM-Drugs, and ISO17 data.The property-prediction task aggregates properties across sampled conformations for applications such as drug discovery.
- 5.1. Conformation Generation: ConfGF achieves state-of-the-art performance on all four COV and MAT metrics across GEOM-QM9 and GEOM-Drugs.COV measures diversity, whereas MAT measures accuracy; higher COV and lower MAT are better.
- 5.1. Conformation Generation: ConfGF outperforms RDKit on 7 out of 8 metrics without force-field post-processing.The comparison supports ConfGF’s ability to generate realistic and diverse conformations without the additional post-processing used by several prior methods.
- 5.2. Distributions Over Distances: ConfGF outperforms CGCF on 4 out of 6 distance-distribution metrics and achieves comparable results on the remaining two.Although ConfGF is not designed specifically for molecular distance geometry, it remains competitive on this task.
- 5.3. Property Prediction: ConfGF outperforms machine-learning baselines in ensemble-property prediction and is more accurate than RDKit for Emin and ∆ϵmax.Median absolute errors show the best accuracy for average energy, while outliers negatively affect some prediction results.
- 5.4. Ablation Study: The distance-based ConfGFDist variant performs worse than ConfGF on both datasets and is on par with CGCF.These results indicate that converting generated distances into conformations can harm conformation quality.
6. Conclusion and Future Work
ConfGF generates molecular conformations in one Langevin-dynamics stage using physically inspired gradient fields while preserving roto-translation equivariance. Experiments across multiple tasks show significant improvements over previous state-of-the-art baselines.
- ConfGF generates conformations using Langevin dynamics within one stage, guided by gradient fields of atomic-coordinate log density.
- The method estimates these gradients while preserving roto-translation equivariance.
- ConfGF outperforms previous state-of-the-art baselines by a significant margin across multiple tasks.
- Future work will extend ConfGF to 3D molecular design tasks and many-body particle systems.
A. Proof of Proposition 1
The proposition establishes that the score function is equivariant under translations and rotations. This follows because transformed coordinates preserve interatomic distances, allowing the corresponding score transformation.
- A translation leaves the score output unchanged, while a rotation applies the same rotation to the output gradients.
- Under transformation, each coordinate becomes Dri+t, combining the rotation matrix D with translation t.
- Rotation and translation preserve interatomic distances, so the score network defined in Eq. 3 is roto-translation equivariant.
B. Additional Hyperparameters
The additional hyperparameters specify the noise schedule, sampling procedure, optimization batch size, and training duration of ConfGF.
- The noise schedule includes the biggest noise level σ1, smallest noise level σL, and number of noise levels L.
- Sampling is controlled by the number of steps per noise level T and the smallest step size ϵ.
- Table 6 reports the batch size and training epochs alongside the noise and sampling hyperparameters.
C. Additional Experiments
Additional experiments evaluate ConfGF using COV and MIS across thresholds on GEOM-QM9 and GEOM-Drugs. ConfGF consistently outperforms competing state-of-the-art methods and produces fewer invalid conformations.
- The additional experiments report COV and MIS scores on GEOM-QM9 and GEOM-Drugs at different RMSD thresholds δ.
- ConfGF consistently outperforms previous state-of-the-art baselines, including GraphDG and CGCF, across the additional COV evaluations.
- Lower MIS scores for ConfGF demonstrate fewer invalid conformations than competing existing models.
- MIS measures the fraction of generated conformations unmatched by any reference conformation within threshold δ.
D. More Generated Samples
Figure 4 provides additional visualizations of conformations generated by ConfGF for molecular graphs from the GEOM-QM9 and GEOM-drugs test datasets.
- Figure 4 visualizes conformations generated by ConfGF for models trained on GEOM-QM9 and GEOM-drugs.
- For each molecular graph, the model randomly samples 10 generated conformations.
- The top five rows use molecular graphs from the GEOM-QM9 test dataset, while the bottom five rows use graphs from GEOM-drugs.