acceptodds
Withdrawn from ICLR 2027

MolField: Representing Molecules as Neural Fields via Hypernetworks

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Molecular representations fundamentally shape how machine learning systems reason about molecular structure and physical properties. Most existing approaches adopt a discrete pipeline: molecules are encoded as sequences, graphs, or point clouds, mapped to fixed-dimensional embeddings, and then used for task-specific prediction. This paradigm treats molecules as discrete objects, despite their intrinsically continuous and field-like physical nature. We argue that molecular learning can instead be formulated as learning in function space. Specifically, we model each molecule as a continuous function over three-dimensional (3D) space and treat this molecular field as the primary object of representation. From this perspective, conventional molecular representations arise as particular sampling schemes of an underlying continuous object. We instantiate this formulation with MolField, a hyper-network-based framework that generates molecular fields. To ensure consistency under rigid motions, these functions are defined over canonicalized coordinates, yielding invariance to global SE(3) transformations. To enable learning directly over functions, we introduce a structured weight tokenization and train a sequence-based hyper-network that generates the parameters of each molecular field. We evaluate MolField on dynamic molecular surface modeling, molecular property prediction, and molecular generation. Our results show that treating molecules as neural fields fundamentally changes how molecular representations generalize and yields downstream behavior that is stable to how molecules are discretized or queried.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Molecule, Drug Discovery, Implicit Neural Representation, Hypernetwork

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_eNyXOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MolField, which models molecules as continuous implicit neural fields in 3D space and explores learning with functions and their parameters as molecular representations. The method combines canonical coordinate construction, structured weight tokenization, and a conditional hypernetwork to generate molecular field parameters, and adopts either field querying or parameter-based read-out depending on the task. The authors evaluate the framework on tasks including dynamic molecular surface modeling, molecular property prediction, and generation. The results show that the framework has advantages in dynamic surface modeling and on some molecular property prediction tasks, offering valuable empirical support for applying continuous neural fields to molecular representation learning.

Critical strengths

The paper combines continuous neural field modeling with hypernetworks, offering an inspiring perspective for molecular representation, and its novelty is commendable. The framework design is coherent, performing relatively well on dynamic surface modeling and some property prediction tasks, and the module ablation and pretraining comparison lend some support to the effectiveness of the overall framework.

Critical weaknesses

(1) The research motivation still needs to be further specified. Taking the continuous physical properties of molecules as the starting point is reasonable, but the correspondence between the specific limitations of existing 3D molecular representations and the expected benefits of this work's choice of a geometric field and parameterized representation could be made clearer. The current discussion of the limitations of existing models is fairly general; it would be helpful to specify more concretely in which aspects of information modeling or task settings these models are constrained, and which problems the continuous field representation can specifically alleviate, thereby strengthening the link between the research motivation, method design, and expected benefits.

(2) The attribution of the method's gains still needs to be strengthened. The paper's central claim is that representing molecules as functions is itself a valuable paradigm. Although the authors provide component-wise ablations that lend some support to the effectiveness of the complete framework, all of these ablations retain the function-based representation framework and only remove components within the paradigm. They can therefore only indicate whether individual components matter, but cannot answer the core question of whether the function-based representation itself yields any benefit. We suggest adding targeted comparisons to clarify the specific contribution of field modeling to the downstream gains.

Questions for authors

(1) In Table 1, only DSR is a domain-specific (molecular) method. The paper cites HyDiF (described as "closest to us") and MoE-DSR but does not include them in the comparison. Could the authors explain why? Is it due to the lack of public code or an inability to reproduce them?

(2) In Dynamic Surface Modeling, z is formed by "concatenating the molecular embedding with the time step," but the main text does not explain how this embedding is obtained. "A learned embedding" is ambiguous. Could the authors specify its construction and dimensionality?

Reviewer_H6WVOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes MolField, which represents a molecule as a continuous function over 3D space, implemented as a small coordinate MLP. A transformer hyper-network generates the weights of this MLP in one forward pass, and the paper uses these generated weights as the representation of the molecule. Points are first moved into a canonical frame built from an equivariant encoder, so the function does not change when the molecule is rotated or moved. On six protein MD trajectories, MolField gets the best average surface accuracy among four methods. On QM9, after pretraining on density fields without labels, it predicts properties from the pooled weight tokens and gets the lowest polarizability error in its table and the second-best average rank.

Critical strengths
  1. the idea is new, and it looks like a promising approach to molecular representation
  2. the surface result on the MD trajectories looks good
  3. the paper is honest about the cases where the frame cannot be defined.
Critical weaknesses
  1. The paper says the generated weights are used directly as the molecular representation. But for property prediction the field is never built: the token sequence is simply pooled into one fixed-size vector. Eq. 8 shows that the transformer's input tokens are the same for every molecule, so the only molecule-specific input is z, the vector the encoder produces for each molecule. So Table 2 shows that some trained function of z predicts properties well. It does not show that the generated weights add anything beyond z. The large gain from pretraining, from average rank 10.0 from scratch to 2.8 pretrained, could simply mean that pretraining makes the encoder better.

  2. Several numbers do not agree with each other. One example is the QM9 experiment. Table 2 gives 24 for mu, but the from-scratch curve in Fig. 7 ends near 12. By the table, pretraining cuts the mu error by about 60 percent; by the figure, by about 25 percent.

  3. The paper does not say which tensors are pooled for property prediction, the transformer outputs or the projected weights, or what the projection Proj from tokens to weights looks like.

Questions for authors
  1. With the same encoder and pretraining, what QM9 errors do you get from a regression head on z alone, over two or three seeds?
  2. Which tensors are pooled for property prediction, with which head, and what exactly is Proj? How many of the field weights are specific to each molecule?
  3. Which from-scratch numbers are correct, Table 2 or Fig. 7, and what differs between the runs?
  4. Where do the Equiformer and NequIP values in Table 9 and the FuncMol numbers come from?
  5. How many parameters do the two arms of Table 8 have, and on which frames was the model trained and tested in Table 1?
Reviewer_EeWfOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This manuscript is proposing to represent each molecule as a neural field over 3D space, evaluated in a learned canonical frame for SE(3) invariance, and using the field parameters generated by a transformer hyper-network over structured weight tokens as a molecular representation. The approach improves over per-trajectory INR fitting on dynamic protein surface modeling, gives the best polarisability among the compared models on QM9 after density-field pretraining, extending to unconditional molecule generation.

Critical strengths
  • Using generated field parameters as a general molecular representation is interesting and relevant.
  • Evaluation is relatively broad, covering six protein trajectories, QM9 and MD17 property prediction, and molecule generation on QM9 and Drugs.
  • Rotations are well handled. Canonical frame comes with proofs, a clear characterisation of the degenerate cases, and an empirical check of their effect in the appendix.
  • Some gains are relatively large, e.g. IoU on YiiP, NhaA and adk_equi and polarisability on QM9.
  • The authors are open limitations and on computational overhead.
Critical weaknesses
  • The conceptual novelty rather limited. The authors present the work as bringing “neural function-space learning to molecular representation”, however, DSR and FuncMol already represent molecules as neural fields, and HyDiF already generates them with a hyper-network. There is not comparison run against this prior art. The hyper-network follows Chen & Wang (2022) and the canonical frame follows Kaba et al. (2023).
  • The paper does not cite Lhotka & Probst (arXiv:2502.10848, ICLR 2025 LMRL workshop). They already proposed molecular neural fields as a representation to complement existing ones. Their fields are the same kind of object as the density field here, namely, sums of atom-centred radial basis functions per atom type or atomic property. They also fitted via a shared network with latent modulation. In addition, they also show super-resolution reconstruction, latent interpolation between conformers, and a latent space that, by visual inspection, reflects physicochemical properties. Compared to this work, the new elements here are the canonical frame, generating weights instead of modulating a shared network, and quantitative downstream benchmarks. While a preprint does not affect my rating, the work here should still be positioned against it.
  • I think that the fields can't add information beyond the discrete input. That is, atom-type densities and signed distances are deterministic renderings of atom positions and types that also enter through the GNN encoder as a discrete graph. The motivation via continuity and the electron distribution, as stated, and the claimed behavior that “stable to how molecules are discretized or queried” are not tested directly, such as via super-resolution as in the work mentioned above.
  • The authors state that "we do not explicitly instantiate the neural field", does this mean that for property prediction, the field is never queried? Pooled tokens are a deterministic function of the encoder output. From scratch as shown in Table 2, the model is mid-table (average rank 10.0). This looks like the gains come from the density-field pretraining rather than from the function-space representation.
  • QM9 comparison is not matched. MolField is pretrained, fine-tuned on all targets jointly, and has about 30M parameters versus 3.5M for Equiformer. The baseline numbers seem to come from papers that only train one model per target, partly on other splits. It also seems that not all standard targets are reported.
  • The claim stating "the gains concentrate on targets governed by molecule-wide spatial structure" doesn't match Table 2. MolField ties the best HOMO, while its R2 ranks 5th of 22 rows, that's more than twice the best value.
  • The claim "the best average performance on all three metrics" seems overstated for CD, as the CD advantage comes entirely from adk_equi. Without it, MolField’s mean CD is above that of DSR and CanFields. On YiiP, the IoU gain over DSR comes with twice the CD. This should be stated as a more specific advantage and investigated/explained..
  • What is the generative prior? Without a KL term or a learned prior, I don't understand why/how the samples decode to 91.3% valid and unique molecules.
  • The generation comparison leaves out prior art. EquiFM (Song et al., 2023), cited in Appendix A.2, reports higher values on all three metrics of Table 6, and FuncMol (a direct competitor) is missing. In Table 7, transfer is worse than training from scratch for every method, so I'm not sure about the claim "degrades less than the compared models".
  • No variance or seed information is reported. This is an issue as some ablation gaps are small.
  • Implementation details for QM9 pretraining, fine-tuning, and generation are missing, this is an issue as the code is not currently available.
Questions for authors
  • What conformation enters the canonical map at time t, in training and at test time? Is it X_t or a fixed reference?
  • Where does the canonical frame enter the property read-out?
  • How would the results change for the pretrained encoder with a regression head of similar size, and for the same encoder pretrained with a shared decoder, or latent modulation as FuncMol and Lhotka & Probst?
  • Was the density-field pretraining restricted to the 100K training molecules? And what exactly does "from scratch" mean, given that Fig 5 reports an INR reconstruction loss for that model?
  • Are the MolField results in Table 1 computed on the same training frames as the DSR numbers and with the same evaluation code? Were NDF and CanFields run and tuned (as well)?
  • Each weight token is already a free learnable vector, so I think that the layer and role embeddings in equation 9 do not change what can be represented. So, what explains the gap without positional embeddings your report in Table 3?
  • What are the number of tokens, d_model, and the block sizes? The appendix claims a 30M parameter model, but the other numbers reported in the manuscript, especially the transformer, suggests a higher number.
  • Where do the FuncMol property numbers in Table 2 come from? Was the model rerun?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.