Abstract
Molecules are frequently represented as graphs, but the underlying 3D molecular geometry (the locations of the atoms) ultimately determines most molecular properties. However, most molecules are not static and at room temperature adopt a wide variety of geometries or conformations. The resulting distribution on geometries $p(x)$ is known as the Boltzmann distribution, and many molecular properties are expectations computed under this distribution. Generating accurate samples from the Boltzmann distribution is therefore essential for computing these expectations accurately. Traditional sampling-based methods are computationally expensive, and most recent machine learning-based methods have focused on identifying modes in this distribution rather than generating true samples. Generating such samples requires capturing conformational variability, and it has been widely recognized that the majority of conformational variability in molecules arises from rotatable bonds. In this work, we present VonMisesNet, a new graph neural network that captures conformational variability via a variational approximation of rotatable bond torsion angles as a mixture of von Mises distributions. We demonstrate that VonMisesNet can generate conformations for arbitrary molecules in a way that is both physically accurate with respect to the Boltzmann distribution and orders of magnitude faster than existing sampling methods.
Introduction
In computational chemistry, a molecule is far more than a static 2D graph; it is a dynamic 3D entity. At room temperature, molecules undergo constant thermal fluctuations, shifting between various shapes known as conformations. To accurately predict physical properties, researchers must sample from the Boltzmann distribution, which describes the probability density of these shapes based on their potential energy.
Existing approaches face a significant trade-off. Traditional physics-based sampling, such as Molecular Dynamics or Markov Chain Monte Carlo, is computationally prohibitive, often requiring minutes or hours for a single molecule. Conversely, modern machine learning models often prioritize finding the single most stable, lowest-energy state (the mode). However, calculating many ensemble properties requires a diverse set of samples that represent the entire distribution. VonMisesNet addresses this by focusing on the primary source of conformational variability: rotations around specific chemical bonds.
Methodology: VonMisesNet
VonMisesNet is a Graph Neural Network (GNN) designed to predict the parameters of circular probability distributions for every rotatable bond in a molecule. Instead of predicting Cartesian (XYZ) coordinates directly—which is challenging due to rotational and translational symmetries—the model utilizes internal coordinates: bond lengths, bond angles, and torsion angles.
Von Mises Mixtures
Because torsion angles are periodic (ranging from 0 to 360 degrees), the authors employ the von Mises distribution, a circular analogue to the Gaussian distribution. By modeling each bond as a mixture of four von Mises distributions, the network can capture complex, multi-modal probability densities, such as the staggered and eclipsed positions common in organic chemistry.
The Independence Assumption
To achieve high inference speeds, the model assumes that the rotation of one bond is approximately independent of others. The researchers validated this "rank-1 analysis" by showing that for many lead-like molecules, this simplification holds sufficiently well to produce accurate Boltzmann samples without the need for iterative refinement.
Key Innovations: Chirality Inversion
A standout feature of VonMisesNet is its explicit modeling of chirality inversion. Certain atoms, such as nitrogen, can "flip" their geometry at room temperature, effectively changing their chirality (e.g., from an S to an R configuration).
Previous machine learning models typically ignored this phenomenon, leading to physically inaccurate geometries for up to 60% of molecules in common datasets. VonMisesNet is the first ML approach to explicitly predict the probability of these flips ($p(R)$) and condition the predicted bond angles on the resulting chirality, ensuring physical realism.
Results and Performance
VonMisesNet was evaluated against major chemical databases, including NMRShiftDB and GDB-17, using ground-truth samples generated via Parallel Tempering Hamiltonian Monte Carlo (PT-HMC).
- Efficiency: VonMisesNet generates a single conformation in approximately 3 milliseconds on a GPU. This represents a 47x speedup over Torsional Diffusion on a GPU and over a 200x speedup on a CPU.
- Accuracy: Using KL Divergence to measure distributional similarity, VonMisesNet consistently outperformed baselines like ETKDG, GeoMol, and Torsional Diffusion. It successfully identifies multiple likely "peaks" in torsion angle distributions where other models often collapse to a single average value.
Conclusion
VonMisesNet demonstrates that modeling molecular conformations through a variational approach with circular distributions offers a powerful path toward fast, physically accurate Boltzmann sampling. While the current model assumes bond independence—a limitation for very large, constrained molecules—it sets a new benchmark for speed and introduces critical handling of chirality inversion. Future work may involve incorporating Markov Random Fields to model joint distributions between bonds while maintaining the model's characteristic efficiency.
Summary updated