Structure-based drug design by denoising voxel grids

The evolution of drug discovery has transitioned from traditional high-throughput screening to computational approaches leveraging structural biology and deep learning. Virtual screening methods, such as docking algorithms, rely on scoring functions to evaluate ligand-protein interactions but are limited by the vastness of chemical space and the complexity of protein structures. Recent advances include generative models like variational autoencoders (Kingma & Welling, 2014) and diffusion models (Hoogeboom et al., 2022), which can produce diverse molecules but often require complex architectures and extensive training. Body-centered voxel representations have gained attention for their ability to encode spatial features effectively, enabling models to learn shape complementarity crucial for binding affinity. However, integrating these representations into efficient, scalable generative frameworks remains a challenge. This work addresses this gap by proposing a score-based model on voxel grids, combining the advantages of spatial discretization with probabilistic sampling techniques, aiming to improve both diversity and binding affinity in structure-based drug design.

Despite progress, current models face significant hurdles: point cloud diffusion approaches are computationally intensive and struggle with spatial conflicts, while voxel-based methods lack efficient conditional generative frameworks. Achieving high diversity, accurate binding affinity, and computational efficiency simultaneously remains unresolved. Existing methods often involve complex equivariant architectures, which are costly to train and slow to sample from, limiting their practical deployment. Moreover, ensuring molecules do not clash sterically with protein pockets and maintaining chemical validity are persistent challenges. The core problem is to develop a model that can generate high-quality, diverse molecules conditioned on complex 3D protein structures efficiently, with a simplified training process and rapid sampling, suitable for large-scale drug discovery pipelines.

VoxBind introduces several key innovations:

1) Extending neural empirical Bayes to the conditional 3D setting, enabling efficient learning of the score function conditioned on protein pockets.

2) Utilizing a fixed-noise denoising network that simplifies training and accelerates sampling compared to diffusion models.

3) Combining Langevin MCMC with a single-step Bayesian estimator for fast, diverse molecule generation.

4) Employing 3D convolutional U-Net architectures to directly model spatial features, shape complementarity, and avoid steric clashes.

5) Demonstrating superior performance over existing point cloud diffusion models in both speed and quality, with better binding affinity and diversity metrics. This approach bridges the gap between spatial discretization and probabilistic generative modeling, offering a scalable solution for structure-based drug design.

• �� Convert protein pockets and ligands into 3D voxel grids with multiple channels representing atom types.
• �� Train a conditional denoising network (U-Net) to predict the clean ligand from noisy inputs, minimizing mean squared error.
• �� Learn the conditional score function g(y|ξ) via the network, which guides the sampling process.
• �� During sampling, initialize with Gaussian noise, then perform Langevin dynamics steps using the learned score to generate noisy molecules conditioned on the pocket.
• �� After a predefined number of steps, apply the Bayesian estimator to recover the clean molecule.
• �� Use peak detection to convert voxel grids back into atomic coordinates, completing the molecule reconstruction.

The model was evaluated on CrossDocked2020, a benchmark dataset with protein-ligand pairs. Data preprocessing involved voxelizing molecules into 64×64×64 grids, with augmentation through random rotations and translations. Metrics included VinaScore, binding affinity, drug-likeness (QED), synthetic accessibility (SA), diversity, and steric clash count. Baseline comparisons involved diffusion models and variational autoencoders. Hyperparameters such as noise level σ, step size δ, and Langevin steps were tuned for optimal performance. Multiple ablation studies validated the importance of condition encoding, noise level, and network architecture. Results consistently showed superior binding affinity, diversity, and computational efficiency over existing methods.

VoxBind achieved a VinaScore of -6.94, outperforming previous models like DecompDiff (-5.67). The generated molecules exhibited 71.3% high-affinity binders, with higher diversity and fewer clashes. The model’s sampling speed was several times faster than diffusion-based approaches, enabling rapid candidate generation. Ablation studies confirmed that the fixed-noise denoising network and condition encoding significantly contributed to performance. The molecules also scored better on drug-likeness and synthetic accessibility, indicating practical relevance for drug discovery pipelines.

This model can be directly applied to virtual screening, lead optimization, and personalized drug design. It allows rapid generation of diverse, high-affinity molecules conditioned on specific protein structures, reducing experimental costs and time. Its efficiency makes it suitable for high-throughput screening in pharmaceutical industry settings. Additionally, integrating biological activity and toxicity predictions could further enhance its utility for real-world drug development, enabling more targeted and effective therapeutics.

The voxelization process depends on high-quality structural data; inaccuracies or flexibility in protein pockets may reduce effectiveness. The current approach mainly focuses on static structures, lacking dynamic interaction modeling. Computational costs, while reduced, still pose challenges for large-scale applications. The biological activity and toxicity of generated molecules require experimental validation. Future work should address these limitations by incorporating multi-scale modeling, dynamic simulations, and experimental feedback to improve robustness and applicability.