acceptodds
Withdrawn from ICLR 2027

MolGrid: Which Chemistry Do Molecular Models Read?

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Molecular property prediction guides which molecules are made and tested, in drug discovery and beyond. Its models read graphs, SMILES strings or 3D structures, which leave relations between atoms, such as how strongly two atoms couple through bonds, for the network to infer. A good score does not tell us which of these relations a model actually relies on. We introduce , which states this pairwise chemistry explicitly as a symmetric image over atom pairs, with channels for through-bond coupling, partial charge and stereochemistry. Built from SMILES in milliseconds with no 3D input, it is read unchanged by pretrained vision backbones. Its diagonal holds one value per atom and its off-diagonal one per atom pair, the first two terms of a property's expansion over atoms, which predicts that additive properties read the atoms and delocalised ones the pairs. Because every value has a fixed address, we can shuffle either part or remove a channel and retrain a chemically pretrained backbone, against a control that only relabels the atoms. leads reproduced fingerprint, graph and SMILES-transformer baselines on five of seven MoleculeNet scaffold-split tasks each, and 3D Uni-Mol on solubility. Across nine properties, only the HOMO–LUMO gap reads the pairs, percentage points beyond the control, while solvation reads the charges and permeability stereochemistry. Pretraining improves of backbone–task pairs without changing which property reads the pairs.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
molecular property prediction, molecular representation learning, pairwise chemistry, causal interventions, interpretability, vision foundation models, chemical pretraining

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_D6MDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents MOLGRID, an explicit, fixed-size 3×N×N symmetric tensor encoding of a molecule (built from SMILES in milliseconds, no 3D input) with three chemically meaningful channels: through-bond coupling via shortest-path products of bond-type attenuations and normalised bond orders, Gasteiger charges (per-atom on the diagonal, pairwise products off it), and stereochemistry (R/S and E/Z). Feeding this "image" to a chemically pretrained DINOv2-B yields competitive results on seven scaffold-split MoleculeNet tasks (leading reproduced D-MPNN, MolFormer-XL and ECFP4-RF on five of seven each, and 3D Uni-Mol on ESol). The paper's main scientific claim is analytical: because every cell has a fixed chemical address, retraining after scrambling the atom-pair assignments (against an atom-relabelling control) shows which register each property reads. Across nine properties, only the HOMO–LUMO gap depends on which atom pair carries which coupling (+5.7 pp pair excess, positive on all ten seeds), solvation tasks depend most on per-atom charges, and only BBBP reads stereochemistry — consistent with an additivity-based prediction that additive properties read atoms while delocalised ones read pairs. Chemical pretraining improves 25/28 backbone–task pairs without changing the reading pattern.

Critical strengths

Outstanding practicality: SMILES input only, millisecond-level construction, bit-level accuracy, and no need for 3D conformations; as an image, it can directly reuse visual backbones such as DINOv2 and SigLIP; it is compatible with 12 backbones spanning three orders of magnitude; it outperforms 5 out of 7 baselines, and ESol outperforms Uni-Mol.

Critical weaknesses

W1. The intervention analysis is not tied to MOLGRID itself. This process is, in principle, applicable to any explicit molecular representation; the paper does not demonstrate that MOLGRID’s specific encoding produces more reliable or informative results than simpler explicit encodings; when performance is merely “competitive,” the unique contribution of the encoding to the core findings remains unclear.

W2. The comparisons fail to isolate the contribution of the encoding. Table 3 simultaneously alters the architecture, initialization, and pretraining; SI Table 13 shows inconsistencies in training set size and splitting, so improvements cannot be attributed solely to input representations.

W3. No controlled ablation for key components: the path product in Ch0 was not compared to the direct adjacency matrix or shortest-path distance; the three pre-training objectives and courses were not decomposed individually; the claim that “pre-training does not alter reading patterns” was only supported by a before-and-after comparison in SigLIP/ESol.

W4. Unclear boundaries in the argumentation: The path product involves intermediate atoms yet corresponds to a two-body term; the approximation is not explained; the in-depth argumentation does not apply to Transformer/MACE-class architectures; the path product is insensitive to key-type sequences and selects the first path among parallel paths, yet sensitivity has not been evaluated.

W5. The report lacks rigor: The denominators for the number of failures (278) and the success rate (93.05% ⇒ 1,390) in Table 2 are inconsistent; there is no physical explanation for qᵢqⱼ; and terminology is ambiguous in multiple places (“chemistry,” “ImageNet-initialized,” “nearly independent”).

Questions for authors

Q1. Given the same backbone, initialization, and budget, by how much does the Ch0 path product improve distance accuracy compared to the direct adjacency matrix and shortest path distance? How insensitive is it to key-type sequences, and how sensitive is it to selecting the first path among parallel paths?

Q2. What are the respective contributions of the three-objective pretraining and the complexity course? Please provide individual ablation results.

Q3. Before and after DINOv2-B chemical pretraining, are the nine-property intervention plots (Fig. 5a) consistent? Please provide before-and-after comparisons for at least gap, ESol, and BBBP.

Q4. Can the MolSight comparison in Table 13 be replicated under the same molecular partitioning, teacher information, and pretraining conditions?

Q5. Why do the path descriptors correspond to the two-body terms in Equation (2) (even though they include intermediate atoms)? Does the denominator for the 278 failures in Table 2 match the 93.05% success rate?

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

This paper proposes MolGrid, a three-channel symmetric atom-pair tensor computed from SMILES, whose channels hold a through-bond coupling term, Gasteiger partial charges and stereochemistry, with per-atom values on the diagonal and per-pair values off it. The tensor is read by vision backbones that are further pretrained on PCQM4Mv2 with a 3D teacher, and the authors shuffle the diagonal or off-diagonal values, or remove a channel, retrain, and compare against an atom-relabelling control to infer which part of the input each property depends on. Experimental results show that MolGrid outperforms selected baselines on most MoleculeNet scaffold-split tasks, and that among nine properties only the HOMO-LUMO gap depends on the pair values.

Critical strengths
  1. A controlled intervention protocol. Every intervention is retrained with seed-paired arms against a relabelling control and corrected for multiple comparisons.

  2. A cheap input with fixed chemical addresses. The representation is built from SMILES alone, and every cell has a fixed meaning that makes register-level interventions possible.

Critical weaknesses
  1. The experiments do not support the motivation. The depth-reach argument and the register prediction are stated without proof and never tied to MolGrid's own results, and the one property that reads the pairs rests on an unmatched control and on non-significance elsewhere.

  2. The comparison does not isolate the representation, and recent close work is missing. MolGrid is scored on smaller molecule sets than its baselines, its gains track a pretraining recipe that no baseline input receives, and the closest pair-aware and image-based methods the paper itself cites are not compared.

Questions for authors
  1. The depth-reach argument (lines 193 to 200) is never tied to MolGrid's own results. Lipophilicity, BACE, HIV and ClinTox have the largest beyond-reach shares in Figure 3b, yet none shows a significant pair excess in Figure 5a, and QM9, the only pair-reading task, is missing from Figure 3b. An intervention restricted to cells beyond 2L bonds is needed before "MolGrid states it outright" (line 99).

  2. The reach argument has no formal statement or proof. Line 194 treats a depth-L network as a degree-L polynomial in the graph operator, which holds for linear filters but not for nonlinear message passing, and the 2L bound covers a pairwise readout, whereas every task uses a graph-level readout over all atom states. Eq. (1) is likewise an approximation without a validity condition, and Channel 0 does not store it (lines 191 to 192). Which result bounds what a depth-L graph-level model can predict?

  3. Eq. (2) is said to predict which register each property reads (lines 229 to 231), but no derivation links the order of a term to the register a nonlinear backbone uses. The dipole contradicts it: the QM9 target is the dipole magnitude, and for a neutral molecule ∣μ∣2=−12∑i,jqiqj∥ri−rj∥2|\mu|^2 = -\tfrac{1}{2}\sum_{i,j} q_i q_j \lVert r_i - r_j \rVert^2 is second order in the charge products that Channel 1 stores off the diagonal, so its null pair excess is a failed prediction, not the additive control of lines 445 to 447. How was each property labelled additive or delocalised before the map was seen?

  4. The relabelling control is not matched to the content-scramble: the scramble keeps the SMILES-ordered diagonal, whose locality Table 1 documents, while the relabelling destroys it. The significantly negative pooled BBBP pair excess (p = 0.040, Table 6) is the symptom. A relabel-then-scramble arm, in which both arms share one random atom order, is the matched control.

  5. "Only the HOMO-LUMO gap reads the pairs" rests on non-significance. On the reported backbone, the pair excess on ESOL, FreeSolv and ClinTox is about as large as the gap's (Figure 5a), on small test sets. Per-cell intervals and an equivalence test against the gap's effect are needed for every "does not read" cell.

  6. On BBBP, BACE, ClinTox and HIV, MolGrid is scored on fewer test molecules than ECFP4-RF, D-MPNN and MolFormer-XL (lines 1630 to 1632), although Appendix H.1 states that "comparing two models over two different molecule sets is not a comparison" (lines 1611 to 1612). Every Table 3 cell needs scores on aligned molecule sets.

  7. Line 405 credits the standing to the representation, but without chemical pretraining the best of twelve backbones (Table 9) trails Table 3's D-MPNN on five of seven tasks, and in Table 13 the ImageNet SigLIP 2 on MolGrid is worse than MolSight's copied number. The missing control is the same pretraining and finetuning recipe on a baseline input, such as a rendered depiction or a plain bond-order adjacency in the same grid.

  8. The newest baseline in Table 3 is Uni-Mol. The paper calls the dual-graph line [1] "closest of all" (line 143) and cites Transformer-M [2], R-MAT [3], TGT [4], its own teacher Uni-Mol2 [5], the image models ImageMol [6], VideoMol [7] and MoleCLIP [8], and the 3D-distilled GraphMVP [9] and 3D Infomax [10], but compares against none of them. How does MolGrid compare with them on the same split?

  9. The Reproducibility Statement says the builders, the intervention, the configs and a regression suite are released and "reproduce every result in the paper" (lines 488 to 490), and Appendix B says the pseudocode follows the released code line for line, but the submission contains no code and no link. The 12 RDKit descriptors and the 85 functional groups are not listed either.

[1] Yunqing Liu, Yi Zhou, and Wenqi Fan. Enhancing molecular property predictions by learning from bond modelling and interactions. In ICLR, 2026.

[2] Shengjie Luo, Tianlang Chen, Yixian Xu, Shuxin Zheng, Tie-Yan Liu, Liwei Wang, and Di He. One transformer can understand both 2D & 3D molecular data. In ICLR, 2023.

[3] Łukasz Maziarka, Dawid Majchrowski, Tomasz Danel, Piotr Gaiński, Jacek Tabor, Igor Podolak, Paweł Morkisz, and Stanisław Jastrzębski. Relative molecule self-attention transformer. Journal of Cheminformatics, 16:3, 2024.

[4] Md Shamim Hussain, Mohammed J. Zaki, and Dharmashankar Subramanian. Triplet interaction improves graph transformers: Accurate molecular graph learning with triplet graph transformers. In ICML, 2024.

[5] Xiaohong Ji, Zhen Wang, Zhifeng Gao, Hang Zheng, Linfeng Zhang, Guolin Ke, and Weinan E. Uni-Mol2: Exploring molecular pretraining model at scale. In NeurIPS, 2024.

[6] Xiangxiang Zeng, Hongxin Xiang, Linhui Yu, Jianmin Wang, Kenli Li, Ruth Nussinov, and Feixiong Cheng. Accurate prediction of molecular properties and drug targets using a self-supervised image representation learning framework. Nature Machine Intelligence, 4(11):1004-1016, 2022.

[7] Hongxin Xiang, Li Zeng, Linlin Hou, Kenli Li, Zhimin Fu, Yunguang Qiu, Ruth Nussinov, Jianying Hu, Michal Rosen-Zvi, Xiangxiang Zeng, and Feixiong Cheng. A molecular video-derived foundation model for scientific drug discovery. Nature Communications, 15:9696, 2024.

[8] Yonatan Harnik, Hadas Shalit Peleg, Amit H. Bermano, and Anat Milo. Data efficient molecular image representation learning using foundation models. Chemical Science, 16(24):10833-10841, 2025.

[9] Shengchao Liu, Hanchen Wang, Weiyang Liu, Joan Lasenby, Hongyu Guo, and Jian Tang. Pre-training molecular graph representation with 3D geometry. In ICLR, 2022.

[10] Hannes Stärk, Dominique Beaini, Gabriele Corso, Prudencio Tossou, Christian Dallago, Stephan Günnemann, and Pietro Liò. 3D Infomax improves GNNs for molecular property prediction. In ICML, 2022.

Reviewer_7MhvOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The authors present a method, MolGrid, which aims to improve the pairwise representations of molecular properties used as inputs to models. Their findings show that the pairwise grid across all constituent atoms in a molecule expanded to include physically inspired properties as part of a higher rank tensor improves property prediction on the PCQM4Mv2 and QM9 datasets. The use of these inputs for molecular modeling into a DINOv2 model is an interesting alternative angle.

Critical strengths

The strongest area of the paper is the core concept. Expanding the set of properties used as inputs to the model is an interesting, and with the evidence provided fruitful, avenue to improve the representational capacity of the model. The idea of a pair tensor input is a natural extension of current approaches.

Critical weaknesses

This paper has a good concept that I find to be lacking evidence to sufficiently support.

The method section was challenging to follow, and despite working in this exact area I couldn’t follow what exactly was used to construct the additional representations for the molecules.

The key claims that the representation allowed the model to read largely the relevant input register, and that the model was recovering the “chemistry” feel under-developed. Lots of models can be shown to correlate certain areas of molecules with certain properties, what’s unclear is (as in Fig 5) are these properties just strongly correlated with these tasks or is there something about the way they are represented in the input tensor that specifically exposes these properties. For instance, does the removal of charge information affect FeeSolv in the same way for MolGrid as other models?

The comparisons to other models / model families feels somewhat scattered. The comparison to vanilla GNN methods looks quite strong, but given that a key argument is the lack of longer distance message passing, there is no comparison to the impact of the models where they rely on virtual nodes to provide this kind of message passing. Nor to the way the MACE family of models use higher representations to reach a larger receptive window. The other clear area lacking was the comparison to models where they use the entire input, there are many transformer variants that operate on the 2D space of atom locations and the various other pairwise properties included either as inputs, or as attention biases.

I would also like to see a much more detailed analysis on what the new tensor inputs actually represent, and what information they introduce. This is where sections 4 and 5 felt particularly weak. It is unclear exactly what information is added here, the paper reads like it comes directly from the SMILES (which provide a complete 2D representation of the graph) but also are contributing charge and interaction terms.

Questions for authors

I would ask the authors to consider rewriting in particular sections 4 and 5 - they contain the core contribution which might be of strong interest to the field, but is currently lost under both a lack of detail to allow reproduction of the method, and extraneous mathematical / chemical detail that hinders comprehension.

In order to be convinced that the results here are due to the representation being significantly better than the alternatives I would ask them to tighten the analysis, focus on one or two key areas, and the most important thing to show is that the improvement is from the method under controlled settings. As presented I cannot compare the results, the models have different inputs, but also so many different characteristics that it’s impossible to disentangle architecture from method.

Please can the authors from a stylistic POV look at all their figures and tables and try to present them in a clearer way. Figure 2 was particularity difficult, Figure 7 was missing, as was the appendix (which for this paper feels like it might have been important)

Finally - could the authors please consider rewriting more of the text by hand, the LLM assistance in editing and grammar has made a large portion of the paper almost impossible to follow. And the style which LLMs tend to follow seems to me to have introduced a lot of strong claims that editorialize and feel out of step with how an academic paper ought to read.

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

Related papers

Discussion (0)

Sign in to comment.