Torsional Diffusion for Molecular Conformer Generation (2206.01729v2)

Abstract: Molecular conformer generation is a fundamental task in computational chemistry. Several machine learning approaches have been developed, but none have outperformed state-of-the-art cheminformatics methods. We propose torsional diffusion, a novel diffusion framework that operates on the space of torsion angles via a diffusion process on the hypertorus and an extrinsic-to-intrinsic score model. On a standard benchmark of drug-like molecules, torsional diffusion generates superior conformer ensembles compared to machine learning and cheminformatics methods in terms of both RMSD and chemical properties, and is orders of magnitude faster than previous diffusion-based models. Moreover, our model provides exact likelihoods, which we employ to build the first generalizable Boltzmann generator. Code is available at https://github.com/gcorso/torsional-diffusion.

• The paper introduces a novel torsional diffusion framework that isolates torsion angles to reduce computational complexity and enhance molecular conformer accuracy.
• It outperforms traditional cheminformatics and earlier machine learning methods by reducing recall RMSD by 30% with significantly fewer denoising steps.
• The method integrates SE(3) invariance and torsion-based scoring, positioning it as a breakthrough for efficient drug discovery and high-throughput molecular simulations.

An Expert Analysis of "Torsional Diffusion for Molecular Conformer Generation"

The paper "Torsional Diffusion for Molecular Conformer Generation" by Jing et al. presents a novel generative framework termed "torsional diffusion" for molecular conformer generation. This research addresses a significant challenge in computational chemistry: generating molecular conformers that accurately represent the energetically favorable 3D structures of molecules. While cheminformatics methods have been dominant in this domain due to their speed, their accuracy remains suboptimal compared to state-of-the-art machine learning approaches, yet the latter often suffer from inefficiency and complexity.

Core Contributions and Methodology

Jing et al. introduce a significant advancement over previous methods by framing conformer generation within a torsional diffusion process that focuses on the torsion angles, which are a primary source of molecular flexibility. This approach contrasts with traditional methods that incorporate all degrees of freedom and require an excessive number of denoising steps. The authors effectively reduce the complexity of the problem by fixing other molecular degrees of freedom that can be determined rapidly and accurately—such as bond lengths and angles—through existing cheminformatics methods.

Torsional diffusion is implemented using a diffusion process on the hypertorus, with an extrinsic-to-intrinsic score model to predict torsional updates on the molecular structure. This implementation maps the torsional space to a toroidal manifold, utilizing $SE(3)$-equivariant networks that account for symmetries and rotational properties in molecular coordinates. The authors also integrate $SE(3)$ invariance, torsion definition invariance, and parity equivariance into their model to address the challenges of modeling conformer variations derived from molecular symmetries and stereochemistry.

Key Results and Performance

The results on the GEOM-DRUGS dataset demonstrate that torsional diffusion surpasses both traditional cheminformatics software and prior machine learning approaches. Specifically, it outperforms the commercial OMEGA software significantly, reducing the recall RMSD by 30% on average. Moreover, the model achieves this using a fraction of the steps required by Euclidean-based diffusion models such as GeoDiff, leading to enhanced efficiency.

From a chemical accuracy standpoint, the torsional diffusion-generated conformers closely align with experimental structures in terms of energy and other computed properties. The rapid generation of conformers—achieved with just a few denoising steps and orders of magnitude less computational effort—positions this method as a breakthrough for high-throughput applications in drug discovery and other fields relying on 3D molecular conformation.

Implications and Future Directions

The implications of torsional diffusion extend beyond conformer generation. The exact likelihoods provided by the model facilitate the creation of torsional Boltzmann generators, capable of sampling the Boltzmann distribution for unseen molecules, representing a tangible step towards efficient molecular simulations that circumvent expensive molecular dynamics methods. This represents a significant step forward compared to existing Boltzmann generators that are restricted to specific chemical systems.

Theoretically, this work lays the groundwork for future investigations into the potential of diffusion-based models on non-Euclidean manifolds, advocating for broader applications that could benefit protein folding and design, where torsional degrees of freedom are vital.

Conclusion

Jing et al.'s development of torsional diffusion offers a compelling new direction in the generation of molecular conformers, combining the speed of cheminformatics techniques with the precision of advanced machine learning frameworks. The alignment of this model with biological and chemical principles underscores its robustness and promises broad applicability, warranting further exploration in diverse molecular systems and computational scenarios. As the field of molecular generative techniques continues to evolve, torsional diffusion stands out as an innovative and efficient method, poised to influence numerous research areas in computational biology and chemistry.