Generating valid Euclidean distance matrices

Noe Frank, Moritz Hoffmann, Frank NoéarXiv 2019 · arXiv

Introduction

The generation of 3D molecular structures is a central challenge in computational chemistry and drug discovery. Traditional deep learning approaches often struggle with the fundamental symmetries of 3D space: translation, rotation, and permutation. Autoregressive models, which place atoms sequentially, frequently fail to capture global constraints like ring closures. Conversely, direct coordinate prediction forces models to learn arbitrary orientations, wasting representational capacity. This paper introduces a "one-shot" method that generates entire 3D geometries simultaneously by predicting Euclidean Distance Matrices (EDMs), ensuring inherent invariance to rotation and translation.

Methodology

The core of the proposed architecture is the generation of EDMs—matrices where each entry $D_{ij}$ represents the squared distance between atoms $i$ and $j$.

Gram Matrix Reparameterization

Valid EDMs must satisfy specific geometric properties: they must be symmetric, non-negative, have a zero diagonal, and be embeddable in 3D space. To navigate the complex "cone" of valid EDMs, the authors employ a Gram matrix reparameterization. The network predicts an arbitrary symmetric matrix, which is then transformed into a Positive Semi-Definite (PSD) Gram matrix $L$ using eigenvalue decomposition and a softplus activation. By explicitly constraining the matrix to rank 3, the model guarantees that the resulting distances correspond to a structure in three-dimensional space.

Permutation Invariance

To achieve permutation invariance, the model is embedded within a Wasserstein GAN (WGAN-GP) framework. The "critic" (discriminator) utilizes SchNet, a message-passing neural network that operates solely on relative distances and pooling. This design ensures the model is indifferent to the ordering of atoms in the input data.

Results and Discussion

The model was evaluated using a subset of the QM9 database, specifically focusing on 6,095 isomers of $C_7O_2H_{10}$. This controlled dataset allows for testing the model's ability to explore diverse arrangements of a fixed set of atoms.

Comparison of pairwise distance distributions between the training data and generated samples.

As shown in the distance distribution plots, the model accurately captures the characteristic inter-atomic distances present in the training data. This statistical alignment suggests that the rank-constrained EDM approach effectively learns the underlying geometric priors of organic molecules.

Validity and Diversity

While the chemical validity rate (correct valency and connectivity) was approximately 7.5%, the model demonstrated significant creative capacity. It successfully generated 31 entirely new molecular topologies (unique SMILES strings) not present in the training or test sets. Furthermore, the model discovered new spatial conformations for known structures, indicating it had learned a continuous representation of the "geometry space."

Visual comparison of generated molecular structures (top row) and their corresponding energy-relaxed conformations (bottom row).

Visual inspection of the generated samples confirms that the model produces physically plausible 3D structures. Energy relaxation using MOPAC software showed that the generated geometries reside in energy ranges comparable to real molecules in the QM9 database, further validating their physical realism.

Conclusion

This work demonstrates that generating Euclidean Distance Matrices with explicit rank constraints is a viable path toward one-shot 3D molecular generation. By baking physical symmetries directly into the architecture, the model avoids the pitfalls of coordinate-based systems. Future research will focus on improving the chemical validity rate through additional constraints and extending the framework to handle molecules of varying sizes and compositions.

Summary updated

BibTeX
@misc{frank2019generating,
  author = {Noe Frank and Moritz Hoffmann and Frank Noé},
  title = {{Generating valid Euclidean distance matrices}},
  year = {2019},
  eprint = {1910.03131},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1910.03131},
}
Figure