VecMol: Vector-Field Representations for 3D Molecule Generation
| Authors | Yuchen Hua et al. |
| Year | 2026 |
| Field | Statistics / ML |
| arXiv | 2603.12734 |
| Download | |
| Categories | stat.ML, cs.LG |
Abstract
Generative modeling of three-dimensional (3D) molecules is a fundamental yet challenging problem in drug discovery and materials science. Existing approaches typically represent molecules as 3D graphs and co-generate discrete atom types with continuous atomic coordinates, leading to intrinsic learning difficulties such as heterogeneous modality entanglement and geometry-chemistry coherence constraints. We propose VecMol, a paradigm-shifting framework that reimagines molecular representation by modeling 3D molecules as continuous vector fields over Euclidean space, where vectors point toward nearby atoms and implicitly encode molecular structure. The vector field is parameterized by a neural field and generated using a latent diffusion model, avoiding explicit graph generation and decoupling structure learning from discrete atom instantiation. Experiments on the QM9 and GEOM-Drugs benchmarks validate the feasibility of this novel approach, suggesting vector-field-based representations as a promising new direction for 3D molecular generation.
Engineering Breakdown
Plain English
VecMol proposes a fundamentally different way to generate 3D molecules for drug discovery by representing them as continuous vector fields rather than discrete graphs. Instead of the traditional approach of co-generating atom types and coordinates separately (which causes learning conflicts), this paper models molecular structure implicitly through vectors in Euclidean space that point toward nearby atoms. The key innovation is using a latent diffusion model to generate these vector fields parameterized by neural fields, which decouples the discrete-continuous generation problem that plagues existing methods. This vector field representation avoids explicit graph generation entirely, promising to reduce the learning difficulties that come from mixing heterogeneous data modalities.
Core Technical Contribution
The core novelty is reimagining 3D molecules as continuous vector fields over space rather than discrete graph structures. Instead of explicitly generating atom types and positions as separate entities (which creates training friction), VecMol encodes molecular structure implicitly through a continuous vector field where the field values implicitly point toward nearby atoms. The framework uses a neural field function parameterized by learned weights and generates these fields via latent diffusion, sidestepping the graph generation bottleneck entirely. This is a genuine paradigm shift because it transforms a discrete-continuous heterogeneous problem into a pure continuous generation problem, making it amenable to standard diffusion-based generation techniques.
How It Works
The system takes the 3D molecular structure as input and learns to represent it as a continuous vector field defined over all of Euclidean space, where at position points toward nearby atoms and encodes local molecular geometry. A neural field network (likely an MLP or coordinate-based network) is trained to parameterize this vector field given learned latent codes. A latent diffusion model operates on these latent codes in a learned embedding space, generating new vector field representations by progressively denoising from Gaussian noise. During inference, the diffusion model samples a latent code, the neural field decodes it into a continuous vector field, and atom positions are recovered by extracting field singularities or using gradient-based optimization to find local maxima in the field magnitude. Atom types are either generated jointly through an auxiliary head on the latent representation or recovered through a separate refinement step after geometry is established.
Production Impact
For drug discovery pipelines, this approach could significantly reduce training time and improve generation quality by removing the discrete-continuous coupling problem that makes existing graph-based methods unstable and sample-inefficient. Production benefits include: (1) cleaner separation of concerns — geometry generation is decoupled from chemistry, allowing independent optimization of each aspect, (2) better use of standard diffusion machinery which has well-understood scaling properties, and (3) potentially faster sampling since you're not doing expensive graph operations. The trade-offs are non-trivial: the continuous vector field representation requires more memory to store and evaluate than sparse graphs, inference requires either dense field evaluation or iterative optimization to extract atom positions (adding latency), and the method needs validation that implicitly-encoded chemical constraints (valence, aromaticity) are actually learned by the diffusion model. Integration requires rewriting molecular featurization pipelines and validation that generated structures satisfy chemical feasibility without post-hoc filtering.
Limitations and When Not to Use This
The paper doesn't explicitly address how chemical constraints (valence rules, forbidden bonding patterns, aromaticity) are enforced — the implicit encoding through continuous fields may not guarantee chemically valid molecules without extensive post-processing. The approach assumes that vector field topology naturally encodes molecular structure, which may fail for unusual bonding patterns or symmetries that don't align with the implicit field representation. Scalability to large molecules (>100 atoms) is unclear since the continuous field representation becomes increasingly memory-intensive and the latent diffusion requires learning more complex field topology. The paper lacks details on how to handle periodic or symmetric molecular structures, and there's no discussion of how the method compares on standard benchmarks (like QM9 or ZINC) against explicit graph-based baselines in terms of validity, uniqueness, and novelty metrics.
Research Context
This work builds on the success of latent diffusion models (like Stable Diffusion for images) applied to molecular generation, extending them from discrete graphs to continuous geometric representations. It advances prior work on neural fields (implicit neural representations like NeRF) by applying them to molecular geometry rather than just visual rendering, and improves on existing 3D molecular diffusion methods (e.g., FrameFlow, DiffDock) by avoiding explicit discrete-continuous conflicts. The research opens a new direction for molecular generation: treating molecules as geometric objects with implicit structure rather than combinatorial objects, which could inspire similar continuous representations for other discrete-geometric problems like protein folding or crystal structure prediction. This aligns with a broader trend toward continuous geometric deep learning as an alternative to graph neural networks for molecular tasks.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::