Structure-based drug design by denoising voxel grids (2405.03961v2)

Abstract: We present VoxBind, a new score-based generative model for 3D molecules conditioned on protein structures. Our approach represents molecules as 3D atomic density grids and leverages a 3D voxel-denoising network for learning and generation. We extend the neural empirical Bayes formalism (Saremi & Hyvarinen, 2019) to the conditional setting and generate structure-conditioned molecules with a two-step procedure: (i) sample noisy molecules from the Gaussian-smoothed conditional distribution with underdamped Langevin MCMC using the learned score function and (ii) estimate clean molecules from the noisy samples with single-step denoising. Compared to the current state of the art, our model is simpler to train, significantly faster to sample from, and achieves better results on extensive in silico benchmarks -- the generated molecules are more diverse, exhibit fewer steric clashes, and bind with higher affinity to protein pockets. The code is available at https://github.com/genentech/voxbind/.

• The paper introduces VoxBind, a structure-based method that harnesses 3D voxel grids and score-based denoising to generate binding-efficient drug molecules.
• It employs a two-step process using Langevin MCMC for sampling and a 3D U-Net convolutional neural network for denoising noisy molecular structures.
• VoxBind outperforms state-of-the-art methods by achieving superior binding affinities and faster sampling, offering promising advances in computational drug discovery.

Understanding VoxBind: Voxel-Based Generative Model for Structure-Based Drug Design

Introduction

Structure-based drug design (SBDD) has always been a critical area within computational drug discovery. Essentially, the goal is to generate molecules, or ligands, that bind effectively to specific biomolecular structures, such as protein pockets. Traditional methods often involve virtual screenings, where databases of molecules are searched to identify potential candidates. However, this approach can be inefficient as the complexity of chemical space increases.

VoxBind introduces a novel voxel-based generative model aimed at improving upon the existing SBDD methods, primarily by using a score-based approach. Let's explore the components that make this method intriguing and effective.

Key Components of VoxBind

3D Voxel Representation

One major innovation in VoxBind is its use of a 3D voxel grid to represent molecules. Each molecule is discretized into a three-dimensional grid, where each voxel (a 3D pixel) represents a tiny volume within the molecule. This method contrasts with point-cloud approaches, which treat atoms as discrete points in space.

Here's why voxel grids are a game-changer:

• Expressive Representation: Voxel grids can capture detailed 3D patterns and surfaces, making them particularly suitable for modeling the intricate shapes of protein pockets and molecules.
• Adaptation from Image Generation: The architecture used in VoxBind draws inspiration from image generation techniques, which have already been proven effective in other domains like computer vision.

Neural Empirical Bayes (NEB) Framework

VoxBind extends the NEB framework to handle conditional generation tasks. The two-step generation process involves:

1. Sampling Noisy Molecules: Using a process called Langevin Markov Chain Monte Carlo (MCMC), voxels are sampled from a Gaussian-smoothed distribution.
2. Denoising: A convolutional neural network (CNN) is employed to clean up the sampled noisy molecules to produce more accurate, binding-efficient molecules.

Training and Architecture

The training involves a conditional denoising model that consists of two main parts:

• Encoders: Separate encoders for noisy ligands and protein pockets.
• U-Net Architecture: A 3D U-Net structure processes the combined encoder output to produce the final denoised molecule.

This design is computationally efficient due to the convolutional filters' ability to capture 3D spatial patterns effectively.

Results and Performance

The performance of VoxBind is quite impressive compared to state-of-the-art methods like TargetDiff and DecompDiff. Here are some key findings:

• Binding Affinity: VoxBind produces molecules with higher binding affinities, an essential metric for SBDD. In benchmark comparisons, VoxBind achieved the best VinaScore and VinaDock among competing models.
• Efficiency in Training and Sampling: VoxBind is significantly faster and simpler to train. Sampling is also more efficient, sometimes achieving results an order of magnitude quicker than other methods.

Practical Implications and Future Directions

VoxBind's framework allows for flexibility in practical applications, such as initializing the model with fragments of molecules rather than starting from scratch. This property can be leveraged for scaffold hopping or linking tasks, similar to in-painting tasks in image generation.

However, the method does have some limitations. The use of voxel grids can consume significant memory, which might limit the model's scalability for larger biomolecules like nucleic acids or full proteins. Future work could focus on optimizing data representation and architecture to overcome these limitations.

Conclusion

VoxBind offers a compelling alternative to traditional SBDD methods, showing notable improvements in binding affinity and efficiency. By leveraging voxel representations and advanced generative modeling techniques, this approach promises to unlock new possibilities in drug discovery. Future developments might even lead to scaling up the model to tackle larger and more complex biomolecules, broadening the horizons of computational drug design.