Enhancing Generalizability in Molecular Conformation Generation with METRIZATION-Informed Geometric Diffusion Pretraining

Xiaozhuang Song, Yuzhao Tu, Hangting Ye, Wei Fan, Qingquan Zhang, Xiaoxue Wang, Tianshu Yu

Introduction & Problem Statement

Predicting the 3D molecular conformation from a 2D graph is a foundational task in drug discovery and material science. Currently, the field faces a significant "generalization gap" between traditional computational methods and modern deep learning models. While traditional distance geometry methods generalize well across molecular sizes, they are computationally slow. Conversely, deep learning diffusion models are fast but often fail to generate physically plausible structures for molecules outside their training distribution. For instance, a model trained on small molecules (QM9) typically produces "meaningless" structures when applied to larger, drug-like molecules (GEOM-Drugs).

The generalization gap in molecular conformation generation. Models trained on small molecules (QM9) often fail to produce physically valid structures for larger, drug-like molecules (GEOM-Drugs).

Methodology: MiGDiff

To address this lack of generalizability, we propose MiGDiff (Metrization-Informed Geometric Diffusion), a hybrid framework that injects "geometric common sense" into diffusion models through a two-phase training paradigm.

  1. Metrization-Driven Pretraining: Instead of training from scratch on limited real-world data, the model is first trained on a massive synthetic dataset generated via the classical METRIZATION algorithm. This algorithm uses universal geometric rules—such as bond lengths, angles, and triangle inequalities—to sample valid 3D coordinates. This teaches the network the fundamental physics of 3D space.
  2. Data-Driven Finetuning: The pretrained model is then refined on real-world datasets to learn the specific energy distributions (Boltzmann distributions) of actual molecules.

The underlying architecture utilizes a roto-translational equivariant network that combines Message Passing Neural Networks (MPNN) for bonded atoms and SchNet for non-bonded spatial interactions.

Experimental Results

The effectiveness of MiGDiff was evaluated in a "zero-shot" generalization scenario, where models were trained on small molecules and tested on the GEOM-DRUGS dataset without prior exposure to large molecules.

Performance comparison on the GEOM-DRUGS dataset. MiGDiff-P, pretrained on synthetic geometric data, significantly outperforms the baseline GeoDiff in zero-shot generalization.

As shown in the results, the vanilla GeoDiff model achieves a Coverage Recall (COV-R) of only 18.93% on out-of-distribution molecules. In contrast, the model pretrained solely on synthetic geometric data (MiGDiff-P) achieves a COV-R of 73.51%. When fully finetuned and combined with force-field (FF) optimization, the framework reaches a state-of-the-art COV-R of 89.85%.

Furthermore, conformations generated by MiGDiff result in more accurate predictions of molecular properties, such as energy and the HOMO-LUMO gap, compared to other generative methods. This indicates that the model captures higher physical fidelity, not just geometric validity.

Conclusion

MiGDiff demonstrates that traditional rule-based algorithms can serve as ideal "teachers" for deep generative models. By leveraging geometry as a universal language through pretraining, the model gains a level of generalizability that purely data-driven approaches lack. This paradigm of "rule-informed pretraining" holds promise for other domains with well-defined physical constraints, such as protein folding or crystal structure prediction.

Summary updated

BibTeX
@inproceedings{song2025enhancing,
  author = {Xiaozhuang Song and Yuzhao Tu and Hangting Ye and Wei Fan and Qingquan Zhang and Xiaoxue Wang and Tianshu Yu},
  title = {{Enhancing Generalizability in Molecular Conformation Generation with METRIZATION-Informed Geometric Diffusion Pretraining}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i1.32058},
}
Figure