acceptodds
Withdrawn from ICLR 2027

Chemistry Before Complexity: Representation-First Molecular Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Molecular machine learning has increasingly focused on expressive architectures, yet classical fingerprints remain highly competitive with modern graph, language, and geometric models. We argue that *molecular representation remains a central bottleneck*: if chemically relevant information is not effectively exposed, additional model capacity may not recover it. We introduce two chemically informed molecular fingerprints. **AtomGrid** encodes atomic-scale covalent, electrostatic, and dispersion interactions in a fixed coordinate system of chemical environments, while **GroupGrid** captures functional-scale organization through group multiplicity, topological separation, and charge complementarity. We study them together with ECFP, AtomPair, and MACCS as five complementary molecular views, evaluating each independently and combining only their prediction scores through a lightweight gated fusion model. On ligand-based virtual screening with hard negatives, fixed fingerprints outperform substantially more complex learned encoders, with AtomGrid among the strongest standalone methods and the five-fingerprint fusion achieving the best overall performance. On molecular property prediction, the same fusion achieves the best aggregate performance across six OGB benchmarks, despite the proposed fingerprints being weaker standalone. These results suggest that future molecular learning should place greater emphasis on *chemistry-aware representation learning*, not only on increasing downstream model complexity.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Molecular Representation Learning, Molecular Fingerprints, AtomGrid, GroupGrid, Molecular Property Prediction, Ligand-Based Virtual Screening, Drug Discovery

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_SjmqOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces two new molecular representations, AtomGrid and GroupGrid. These are combined with three standard fingerprints, ECFP, AtomPair, and MACCS, and their prediction scores are fused through a lightweight gated model for downstream ligand-based virtual screening and molecular property prediction. I largely agree with the paper's high-level message that chemistry-aware representation learning deserves more attention than simply increasing downstream model complexity. However, I view the score-level fusion component as largely incremental, so I primarily evaluate the paper based on the soundness of the proposed algorithms used to construct AtomGrid and GroupGrid.

On this point, I have major concerns. The construction of both representations is described too vaguely to reproduce confidently from the manuscript alone, the provided code repository is currently inaccessible, and the paper does not provide sufficient representation-level ablation studies to validate the key design choices behind AtomGrid and GroupGrid.

Critical strengths

I really like and agree with the message delivered by the paper: "future molecular learning should place greater emphasis on chemistry-aware representation learning, not only on increasing downstream model complexity."

Critical weaknesses
  • AtomGrid reproducibility and soundness are difficult to assess from the paper alone. The main text gives only a high-level description of the chemical environment types and coordinate construction, while the appendix still does not provide the exact atom-typing rules needed to reproduce the representation. The paper also does not include representation-level ablations for key design choices such as the covalent/electrostatic/dispersion channels, conformer dependence, atom typing, or E-state ordering. Since the provided code link is currently inaccessible, it is unclear whether AtomGrid can be faithfully reproduced or whether its reported gains are robust to these choices. This makes it difficult for me to assess the soundness of the proposed representation. A dedicated figure showing the full construction pipeline, or pseudocode specifying the exact steps and rules, would be very helpful.
  • GroupGrid has a similar reproducibility and soundness issue.
  • The paper appears to rely heavily on LLM assistance without sufficiently careful auditing. For example, the bibliography includes the entry “Johannes Klicpera, Janek Groß, and Stephan Günnemann… 2020. Author later published as Johannes Gasteiger.” The note “Author later published as Johannes Gasteiger” is highly unusual for a bibliography and suggests that AI-assisted text may not have been carefully reviewed.
  • The abstract is presented in multiple paragraphs, making it read more like a mini introduction than a conventional conference abstract.
  • The provided code repository link does not work.
  • There appears to be a discrepancy between Table 2 and Table 11, which I understand to be the full version of Table 2. The discrepancies involve both the reported numbers and the methods/configurations shown. In addition, Table 2 in the main paper should report the full results with standard deviations.
  • The paper lists GCN, GIN, GraphSAGE, GAT, GPS, SGFormer, MolBERT, ChemBERTa, MolFormer, Uni-Mol, and GraphMVP, but the sentence introducing them only names the models without individual references. The reference list contains some related papers, but from the manuscript text it is not clear which exact source or implementation corresponds to several of these baselines, especially GPS, SGFormer, MolBERT, ChemBERTa, and GraphMVP.

Minors

  • The average AUROC of 5FP-Fusion in Table 2 is reported as 77.4, but the listed per-dataset values average to approximately 77.6.
Questions for authors

Were the baseline models re-implemented and evaluated on the same data splits as the proposed methods?

Reviewer_k3sSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a "representation-first" perspective on molecular machine learning, arguing that molecular representation can be as important as model capacity. To this end, the authors introduce two fingerprints: AtomGrid, which encodes physical interactions, and GroupGrid, which encodes relationships between functional groups. A separate predictor is trained on each of five fingerprints (the two proposed ones plus ECFP, AtomPair, and MACCS), and their predictions are combined through gated fusion. Experimental results show that this approach outperforms or performs comparably to GNN-based and pretrained models on ligand-based virtual screening (TopU-LBVS) and molecular property prediction (six OGB datasets).

Critical strengths

1. Meaningful Representation-first Perspective

The paper presents a coherent argument that chemically informed representations can be as important as model complexity. Revisiting the effectiveness of fingerprints offers a valuable perspective in a field dominated by increasingly complex learned encoders.

2. Strong Empirical Results

On TopU-LBVS hard negatives, fixed fingerprints and their fusion outperform the evaluated learned models in average rank. Adding AtomGrid and GroupGrid improves average EF@1% from 6.1 (3FP-Gated) to 8.8 (5FP-Gated), demonstrating the benefit of complementary molecular representations.

Critical weaknesses

1. Limited Methodological Novelty

The proposed AtomGrid and GroupGrid appear closely related to existing physicochemical and pharmacophore-based descriptors, particularly E-state features and ErG fingerprints. Moreover, combining multiple fingerprints through gated fusion is a well-established approach. The paper does not sufficiently clarify its methodological novelty or provide direct comparisons with closely related descriptors.

2. Insufficient Justification for CNN-based AtomGrid Processing

AtomGrid sorts atoms by E-state values within each chemical type before constructing the interaction grid. However, spatial proximity in this grid does not necessarily reflect molecular connectivity or physically meaningful locality. The rationale for applying a CNN is therefore unclear, and a comparison with an MLP using the same representation would help justify this architectural choice.

3. Experimental Inconsistencies and Baseline Comparisons

Several results are inconsistent across Tables 2, 7, and 11, raising concerns about experimental reliability. Some reproduced baselines also perform noticeably worse than originally reported. Moreover, the improvement over pretrained models on MPP is marginal (77.4 vs. 77.2 mean AUROC), making the claimed superiority less convincing without further statistical analysis.

Questions for authors

Please address the concerns raised in the Critical Weaknesses section, particularly regarding methodological novelty, representation contributions, and experimental reliability.

Reviewer_KEoXOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a representation-first approach to molecular learning, emphasizing the importance of chemically informative fingerprints over increasingly complex model architectures. The authors introduce two molecular representations: AtomGrid, which captures atomic-level covalent, electrostatic, and dispersion interactions, and GroupGrid, which describes functional-group relationships through multiplicity, topological distances, and charge complementarity. These representations are evaluated individually and combined with ECFP, AtomPair, and MACCS through a lightweight score-level gated fusion framework. Experiments on ligand-based virtual screening and six OGB molecular property prediction benchmarks show competitive results against classical fingerprints and learned molecular models.

Critical strengths
  • The paper investigates an interesting and relevant problem in molecular machine learning. The emphasis on chemically informed representations is well motivated, especially given the increasing complexity of recent molecular learning models.

  • AtomGrid and GroupGrid provide two complementary perspectives on molecular structure, focusing on atomic-level physical interactions and functional-group organization respectively. Their chemical motivations are reasonable, and the resulting representations can be used with relatively lightweight prediction models.

Critical weaknesses
  • The improvements from molecular representations and model ensembling do not seem fully separated. A central argument of this paper is that better molecular representations can be more important than increasing model complexity. However, the best-performing approach combines five independently trained predictors through an additional gated fusion model. This makes it difficult to determine how much of the improvement comes from the proposed chemical representations and how much comes from combining multiple prediction models. The authors have included useful comparisons between 3FP and 5FP fusion under different combination strategies, which partially addresses this concern. Still, a stronger control would be an ensemble of multiple predictors based on the same classical fingerprint, such as ECFP, under a comparable training and fusion setup. This could better isolate the benefit of representation diversity from the general advantage of model ensembling.

  • The paper mainly compares the proposed fingerprints with ECFP, AtomPair, and MACCS as classical representations. Yet AtomGrid and GroupGrid encode information that is also related to several established families of molecular descriptors. For example, GroupGrid is closely connected to pharmacophore-based representations that describe functional-group types and their spatial or topological relationships. AtomGrid also shares some motivation with physical interaction descriptors, such as Coulomb matrices and related atom-pair representations. Although these methods are mentioned in the related work, they are not included in the main experimental comparisons. I think this could be important, since it remains unclear whether the proposed fingerprints capture genuinely additional chemical information or mainly reorganize information already available in existing descriptors. Comparisons with representative fingerprints would be helpful.

  • The computational requirements of AtomGrid are not sufficiently discussed. Unlike conventional 2D fingerprints, AtomGrid requires molecular geometry, partial charges, Wiberg bond orders, and dispersion-related quantities. Obtaining these features can involve substantial preprocessing, depending on the computational methods used. The authors mention that the descriptors are precomputed and cached, but this does not remove the cost of generating them in the first place. This matters especially for virtual screening, where the candidate library may contain millions of molecules. I would like to see a more detailed analysis of feature generation time and computational cost, as well as the sensitivity of AtomGrid to different molecular conformers. Without this information, it is difficult to assess the practical efficiency of the proposed representation relative to conventional fingerprints or learned molecular encoders.

  • The incremental gains are relatively limited in some settings. Although overall results are promising, the advantages are not equally convincing across the two evaluation tasks. Like, on the six OGB property prediction datasets, 5FP-Fusion achieves an average AUROC of 77.4, only slightly higher than MolFormer at 77.2. Meanwhile, AtomGrid and GroupGrid perform considerably worse as standalone predictors on these datasets. This makes their added value in the final fusion model worth examining more carefully.

  • The virtual screening ablations also reveal an interesting pattern. On TopU hard negatives, adding GroupGrid to the three classical fingerprints increases average EF@1% from 6.1 to 8.7, while adding AtomGrid on top of that only brings it to 8.8. This suggests that the contributions of the two proposed representations may be quite uneven in the full model. More detailed per-dataset comparisons and statistical analysis would help determine how consistent these improvements are.

Questions for authors

See weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.