Score-based 3D molecule generation with neural fields (2501.08508v1)

Abstract: We introduce a new representation for 3D molecules based on their continuous atomic density fields. Using this representation, we propose a new model based on walk-jump sampling for unconditional 3D molecule generation in the continuous space using neural fields. Our model, FuncMol, encodes molecular fields into latent codes using a conditional neural field, samples noisy codes from a Gaussian-smoothed distribution with Langevin MCMC (walk), denoises these samples in a single step (jump), and finally decodes them into molecular fields. FuncMol performs all-atom generation of 3D molecules without assumptions on the molecular structure and scales well with the size of molecules, unlike most approaches. Our method achieves competitive results on drug-like molecules and easily scales to macro-cyclic peptides, with at least one order of magnitude faster sampling. The code is available at https://github.com/prescient-design/funcmol.

• The paper introduces a novel continuous atomic density representation for 3D molecule generation using neural fields.
• It employs a combined walk-jump sampling method integrating Langevin MCMC for efficient and scalable molecular generation.
• FuncMol achieves competitive results on drug-like molecules and macrocycles, significantly reducing sampling time while enhancing expressivity.

An Expert Overview of "Score-based 3D Molecule Generation with Neural Fields"

This paper presents a novel approach to 3D molecule generation using a representation based on continuous atomic density fields, titled "FuncMol." The authors introduce a model that diverges from prevalent methods by focusing on neural field representations rather than point clouds or voxel grids, overcoming constraints associated with both scalability and expressivity.

The representation utilized in this work describes molecules through continuous atomic density fields mapped from 3D coordinates to atomic densities. This captures the inherent continuity of molecular structures more naturally than previously used discrete formats. The molecular fields are parameterized using a neural field, which utilizes implicit neural representations or coordinate-based neural networks. This allows the representation to be both more compact and expressive, providing data at arbitrary resolutions.

FuncMol applies a learned generative model using a combination of walk-jump sampling, entailing a Langevin MCMC-based walking to sample noisy modulation codes, followed by a single-step denoising jump to generate the final molecular forms. The introduction of neural field-based modulation results in a reduction of assumptions regarding molecular structures (e.g., bonds or symmetries), while ensuring the model remains scalable with respect to the number of points and molecular size.

Several critical results are highlighted in the paper:

• FuncMol achieves competitive results on drug-like molecules against various baselines.
• The approach scales efficiently to larger molecules like macrocyclic peptides and demonstrates a drastic improvement in sampling time, being at least an order of magnitude faster than existing methods.
• The model effectively captures the complexity of macrocyclic conformations in the CREMP dataset, maintaining molecular validity and realistic features without explicit structural priors.

Theoretical implications of this work include the effectiveness of neural fields in representing complex, continuous molecular data, potentially setting a new standard for 3D molecule modeling presaging more robust all-atom generative models. Practically, the efficiency and scalability of FuncMol open prospective applications in drug discovery and material sciences, where handling large molecules swiftly and accurately is pivotal.

Future directions discussed by the authors entail refining these models further, exploring various neural field architectures, and extending capabilities to conditional generation (e.g., structures aligned with specific biological targets) and more sophisticated representations involving electron densities or pharmacophore features.

Overall, the paper demonstrates a promising advancement in 3D structural modeling, coupling neural field representations with swift generative processes, thereby broadening the practical applicability of generative models in computational chemistry and beyond.