DiSCO: Diffusion Schrodinger Bridge for Molecular Conformer Optimization

Danyeong Lee, Dohoon Lee, Dongmin Bang, Sun KimAAAI 2024 · OpenReview

Abstract

The generation of energetically optimal 3D molecular conformers is crucial in cheminformatics and drug discovery. While deep generative models have been utilized for direct generation in Euclidean space, this approach encounters challenges, including the complexity of navigating a vast search space. Recent generative models that implement simplifications to circumvent these challenges have achieved state-of-the-art results, but this simplified approach unavoidably creates a gap between the generated conformers and the ground-truth conformational landscape. To bridge this gap, we introduce DiSCO: Diffusion Schrödinger Bridge for Molecular Conformer Optimization, a novel diffusion framework that enables direct learning of nonlinear diffusion processes in prior-constrained Euclidean space for the optimization of 3D molecular conformers. Through the incorporation of an SE(3)-equivariant Schrödinger bridge, we establish the roto-translational equivariance of the generated conformers. Our framework is model-agnostic and offers an easily implementable solution for the post hoc optimization of conformers produced by any generation method. Through comprehensive evaluations and analyses, we establish the strengths of our framework, substantiating the application of the Schrödinger bridge for molecular conformer optimization. First, our approach consistently outperforms four baseline approaches, producing conformers with higher diversity and improved quality. Then, we show that the intermediate conformers generated during our diffusion process exhibit valid and chemically meaningful characteristics. We also demonstrate the robustness of our method when starting from conformers of diverse quality, including those unseen during training. Lastly, we show that the precise generation of low-energy conformers via our framework helps in enhancing the downstream prediction of molecular properties. The code is available at https://github.com/Danyeong-Lee/DiSCO.

Introduction

Predicting the 3D shape, or conformer, of a molecule is a fundamental task in drug discovery and materials science, as a molecule's biological and chemical function is largely determined by its geometry. While traditional Molecular Dynamics (MD) simulations provide high accuracy, their computational cost makes them impractical for large-scale screening.

Deep learning models have emerged as a faster alternative, yet they often struggle with a "distribution gap." Many current methods simplify the conformer generation task—for instance, by fixing bond lengths or angles—to manage the complexity of the Euclidean search space. Consequently, the resulting conformers often fail to reach the true local energy minima of the molecule. DiSCO addresses this by acting as a refinement framework that "pulls" approximate conformers toward their true, low-energy states, bridging the gap between simplified models and the ground-truth conformational landscape.

Methodology: Diffusion Schrödinger Bridge

The core of DiSCO is the Diffusion Schrödinger Bridge (DSB). Unlike standard diffusion models that typically transform data into Gaussian noise, a Schrödinger bridge identifies the optimal transport path between two arbitrary distributions. In this context, DiSCO connects the distribution of conformers produced by proxy methods (such as RDKit or Torsional Diffusion) to the ground-truth Boltzmann distribution.

Overview of the DiSCO framework: (1) Training the model to predict the ground-truth conformer from intermediate states, and (2) recursively sampling to optimize approximate conformers toward the ground-truth energy landscape.

Key Technical Components

  • Tractable Schrödinger Bridge: The diffusion path is formulated as a convex combination of the starting (approximate) and ending (ground-truth) conformers, making the learning process efficient.
  • SE(3)-Equivariance: To respect physical symmetries, the model incorporates Kabsch alignment and zero-Center-of-Mass (CoM) normalization, ensuring that the generated conformers are invariant to global rotation and translation.
  • Model-Agnostic Optimization: DiSCO is designed as a post-hoc optimizer. It can take the output of any existing conformer generation tool and refine it to improve structural quality and physical validity.

Experimental Results

The framework was evaluated on the GEOM-QM9 and GEOM-Drugs benchmarks, using reference conformers calculated via high-quality semi-empirical MD (CREST).

Performance and Diversity

DiSCO consistently outperformed major baselines, including RDKit ETKDG, Torsional Diffusion, and DMCG. On the GEOM-Drugs dataset, DiSCO improved the median MAT-R (Recall of accuracy) for Torsional Diffusion from 0.70 to 0.65 and achieved a perfect 100% median COV-R (Recall of diversity). By allowing fixed local structures to "relax" during optimization, DiSCO enables conformers to reach lower energy states that constrained models might miss.

Meaningful Intermediate States

Analysis of the diffusion process reveals that intermediate steps are not merely mathematical artifacts. As the process progresses from the input ($t=1$) to the output ($t=0$), the molecule's energy steadily decreases, and structural metrics like Recall and Precision improve monotonically.

Evolution of Recall, Precision, and Energy gap during the diffusion process. As the optimization progresses from input (t=1) to output (t=0), coverage metrics improve and the energy gap to the ground truth decreases.

Downstream Impact

Refining conformers with DiSCO significantly enhances the accuracy of molecular property predictions. For example, the Mean Absolute Error (MAE) for predicting molecular energy using RDKit-generated conformers was reduced from 1.27 to 0.79 kcal/mol after DiSCO optimization.

Discussion and Conclusion

DiSCO demonstrates that the Schrödinger bridge is a powerful tool for molecular conformer optimization. By focusing on the residual task of fixing errors in existing models, it effectively narrows the search space and improves both accuracy and diversity.

While DiSCO adds an iterative sampling step (typically 5–15 steps), it remains significantly faster than full MD simulations. Future work may explore relaxing the independence assumption between ground-truth and generated conformer pairs during training to further refine the diffusion paths. Additionally, extending this guided exploration of the energy landscape to molecular docking presents a promising avenue for drug discovery.

Summary updated

Abstract

The generation of energetically optimal 3D molecular conformers is crucial in cheminformatics and drug discovery. While deep generative models have been utilized for direct generation in Euclidean space, this approach encounters challenges, including the complexity of navigating a vast search space. Recent generative models that implement simplifications to circumvent these challenges have achieved state-of-the-art results, but this simplified approach unavoidably creates a gap between the generated conformers and the ground-truth conformational landscape. To bridge this gap, we introduce DiSCO: Diffusion Schrödinger Bridge for Molecular Conformer Optimization, a novel diffusion framework that enables direct learning of nonlinear diffusion processes in prior-constrained Euclidean space for the optimization of 3D molecular conformers. Through the incorporation of an SE(3)-equivariant Schrödinger bridge, we establish the roto-translational equivariance of the generated conformers. Our framework is model-agnostic and offers an easily implementable solution for the post hoc optimization of conformers produced by any generation method. Through comprehensive evaluations and analyses, we establish the strengths of our framework, substantiating the application of the Schrödinger bridge for molecular conformer optimization. First, our approach consistently outperforms four baseline approaches, producing conformers with higher diversity and improved quality. Then, we show that the intermediate conformers generated during our diffusion process exhibit valid and chemically meaningful characteristics. We also demonstrate the robustness of our method when starting from conformers of diverse quality, including those unseen during training. Lastly, we show that the precise generation of low-energy conformers via our framework helps in enhancing the downstream prediction of molecular properties. The code is available at https://github.com/Danyeong-Lee/DiSCO.
BibTeX
@inproceedings{lee2024disco,
  author = {Danyeong Lee and Dohoon Lee and Dongmin Bang and Sun Kim},
  title = {{DiSCO: Diffusion Schrodinger Bridge for Molecular Conformer Optimization}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  url = {https://openreview.net/forum?id=3HrichmG54},
}
Figure