acceptodds
Under review as a conference paper at ICLR 2027

WEECFP-SuRGE: A Position-Aware Substructure Encoding Method for Molecular Property Prediction

Abstract

Computational molecular property prediction requires representations that capture local chemistry, long-range interactions, and molecular topology. Conventional fingerprints provide efficient local substructure features, whereas learned graph and sequence models can represent broader context but often rely on pretraining or three-dimensional conformers. We introduce Wide Encoded Extended Connectivity Fingerprints (WEECFP) with Substructure Rotary Graph-distance Encoding (SuRGE), a tokenized hierarchical Morgan representation in which graph-distance-dependent rotations are applied at the input and within transformer self-attention. Across MoleculeNet and the Therapeutic Data Commons ADMET benchmarks, WEECFP-SuRGE is competitive with recent pretrained and geometry-aware methods without external pretraining or conformer generation. We also show that the rotated token representation remains structurally informative. A guided confirmed-handshake overlap procedure reconstructs the correct constitutional isomer for 92.6% of a 4,200-molecule self-library evaluation. Together, the predictive and reconstruction results indicate that WEECFP-SuRGE preserves local substructure identity while making relative topology available to the model.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
cheminformatics, rotary position encoding, extended connectivity fingerprint, transformer

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_5kV7OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper designed Wide Encoded Extended Connectivity Fingerprints based on atomic neighbor information with Transformer to predict the molecule properties.

Critical strengths

NA

Critical weaknesses

The paper is not well written, very hard to follow:

  1. in formula 1, h_1, b_i, p_i, v_i, f [p_i] is not defined. The selection of number 17, 4 and 8 is not clearly stated.
  2. the representation construction is based on a lot of arbitrary engineering choices without clear motivation.
Questions for authors

NA

1 comment

We thank the reviewer for their consideration of the manuscript and hope that the revised version provides a clearer representation of the presented algorithm. Specifically, we do not intend to present the WEECFP fingerprint method and the associated transformer as the sole contribution of the work. Instead, WEECFP is intended to enable the application of SuRGE, which adds graph-distance information in a format similar to RoPE, and secondarily supports library-free reconstruction in the evaluated settings. Additionally, we would like to point out the use of anchor reference distance encoding of SuRGE in the attention layer as a scalable surrogate for pairwise graph distance (Section 3.2). With respect to the reviewer’s comment on the missing variable definitions, we thank the reviewer for noting this discrepancy and have made corrections throughout the manuscript to make the methods clearer. Section 3.1 now summarizes the token construction, and SI Section K specifies the digest bytes, coordinate indices, and signed updates, with each symbol defined. With respect to the reviewer’s comment on the selection of engineering choices in both the fingerprint construction and model design, we have added further discussion throughout the manuscript to specify the design motivations. Additionally, we have incorporated ablation studies shown in Section 5.4 and Table S14 (SI Section J). These indicate that most engineering choices outside SuRGE have little effect on predictive performance: the token-content, transformer-size, and fingerprint-constant variants shift RMSE by a few percent, and 45 of the 224 comparisons reach p<0.05 without adjustment, against roughly 11 expected by chance. Removing SuRGE, by contrast, is consequential on particular targets (16.6% on FreeSolv and 111.1% and 174.0% on QM9 α and C_v, against under 3% on Lipophilicity and QM9 μ).

Reviewer_z1akOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

WEECFP replaces hashed Morgan bits with dense signed continuous tokens that retain local substructure information, while SuRGE applies rotary rotations based on graph-distance coordinates. A seven-model ensemble combines transformer, MLP, and CNN members, and is evaluated on scaffold-split MoleculeNet tasks and the TDC ADMET leaderboard. The paper also reconstructs constitutional molecular graphs from rotated tokens through a fragment-library and confirmed-handshake procedure.

Critical strengths
  1. The method is simple, deterministic, and does not require pretraining or 3D conformers. The ref3 coordinate construction addresses a real implementation problem: full pairwise graph-distance rotations would lose the usual key-factorization and scale quadratically in sequence length.

  2. The evaluation is broad and reasonably transparent about trade-offs. On MoleculeNet, the ensemble is competitive with 3D and learned baselines across regression and classification, and the paper reports five scaffold-split seeds rather than a single split. It also acknowledges that SuRGE is not uniformly useful and that the reconstruction test does not cover stereochemistry or unseen fragments.

Critical weaknesses
  1. The representation and “reversibility” claims are not sufficiently established.

    Section 3.1 states that identical substructures at different atoms collapse into one entry before aggregation, while Section 3.4 and the supplement reconstruct one token per heavy atom and assign each token a graph position. The paper does not clearly explain how atom multiplicity, token identity, token ordering, and permutation invariance are preserved.

    In addition, the 1024-dimensional signed MD5 encoding, log compression, and aggregation may introduce collisions and erase structural information. The claim that collisions behave as approximately zero- mean interference is asserted without collision-rate analysis or robustness experiments. The 92.6% reconstruction result is also evaluated with a self-library built from the same 4,200 molecules. This measures closed-world reconstruction under complete fragment coverage, rather than invertibility or generalization to unseen fragments and scaffolds. A disjoint-library evaluation and comparisons with unrotated WEECFP, Morgan, and random/hash controls are needed.

  2. The experiments do not cleanly establish that SuRGE itself provides the claimed benefit.

    The headline model is a seven-member ensemble containing four SuRGE transformers, an MLP, and two CNNs without SuRGE; several members also use Avalon+ErG descriptors. Thus, the reported performance reflects tokenization, descriptor augmentation, architecture diversity, and ensembling together. The NoSuRGE comparison changes only the transformer members, and the TDC average-rank difference is small (3.9 versus 4.1) without paired statistical testing.

    The per-task results are mixed, with some conventional fingerprints, 3D models, or NoSuRGE variants performing better. The TDC comparison additionally relies on published leaderboard results obtained under potentially different implementations, tuning budgets, and evaluation conditions. The paper needs component-isolated comparisons, matched ensemble-size and compute controls, per-task significance tests, and an evaluation against a single WEECFP model with and without SuRGE.

Questions for authors

N/A

1 comment

We thank the reviewer for their thorough and constructive feedback on the manuscript. The reviewer has correctly pointed out that the use of the term “reversible” with respect to the encoding method is misleading, and we have revised the language of the claim to reflect an effective reconstruction algorithm utilizing substructures with graph distance and not specifically reversibility inherent to the fingerprint. With respect to the reversibility claims on the use of a reconstruction algorithm that utilizes a pre-generated fragment library, the reconstruction method has been revised to utilize a deterministic, library-free extraction of the central atom and its immediate neighbors from explicit token blocks. The radius two hash is no longer used in the fragment inversion and is instead a component of the token verification process. The method then uses immediate neighbor reconstruction based on valid bond matches from other fragment central atoms with SuRGE-derived distance filtering. This adjustment no longer requires a training fragment library. The revised encoder reduces the complexity of the fragment alignment computations, formally referred to as “handshakes”, and replaces it with simpler radius one bond formations. This simplified format reduces many of the error mechanisms of the formerly used algorithm, such as failed cyclic bond attachments. Instead, the revised method generates a larger candidate fragment connection list and relies more heavily on the SuRGE-based distance filter. The method searches bond arrangements until the first graph matches the complete signed-token representation by re-encoding within tolerance under optimal token-to-atom assignment. Reconstruction without SuRGE has an approximately 23× median paired per-molecule slowdown, shown in Figure 2E. Additionally, with SuRGE, it recovers 4,197/4,200 Lipophilicity molecules (99.9%) and 133,875/133,885 QM9 molecules (99.993%). These results supersede the original 92.6% self-library result, which used a radius two fragment assembly algorithm. Additionally, we have revised the language of the methods section to clarify the unspecified components of the algorithm raised by the reviewer (Sections 3.1–3.5; Algorithm S1 in SI Section C). In particular, one radius-2 token per heavy atom is retained as an unordered multiset, preserving multiplicity. We have also added the collision analysis the reviewer asked for (SI §E), run on ChEMBL 37 rather than on the evaluation collections. Among its 707,914 distinct radius-2 environments, no two whose tokens carry a Morgan hash block fall within the decoder’s tolerance or reach cosine 0.99 in any of 2.5×10^11 pairs, whereas folded 1024-bit ECFP4 places 49% of the same environments in a group with an identical representation. Identical tokens arise only for terminal atoms, which carry no radius-2 hash (1.8% of environments); at the molecule level these collide exactly where the unfolded ECFP4 count multiset does, so the ambiguity is a property of the radius-2 description rather than of the hash, and the SuRGE distances separate every such pair in Lipophilicity. We also measured the separation behind the verification threshold: every valid single-edit neighbor of a Lipophilicity molecule lies at least 0.31 from its fingerprint, about 150 times the acceptance tolerance. With respect to the approximate zero-mean assertion, that feature of the fingerprint is not sufficiently justified and has been removed from the manuscript. We agree with the reviewer’s concern on the demonstration of SuRGE as an enhancing component of the model architecture and that some of the traits are masked by the ensembling methods used. In response, we have included individual ensemble member results (Tables S1 and S10; SI Sections A and H) along with ablation studies (Section 5.4; SI Section J) and appropriate statistical analyses to evaluate the influence of SuRGE (Sections 5.2–5.4; SI Section J). With respect to the TDC benchmark training configuration, the official split framework described in the Leaderboard Guide was followed [11, 38]. Additionally, per our general note to reviewers, we have corrected test-set checkpoint selection for our original TDC regression runs. In the interest of transparency along with appropriate context to the limitations of the leaderboard, we report the corrected results in the main text with reference to the original test-set-stopped results in the supporting information.

References [11] Huang, K., et al. Therapeutics Data Commons: Machine learning datasets and tasks for drug discovery and development. Advances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track, 2021. [38] Therapeutics Data Commons. Leaderboard Guide. https://tdcommons.ai/benchmark/overview/ (accessed November 15, 2026).

Reviewer_RukyOpenReview ↗
Rating
3
Confidence
3(+1)
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a substructure position-aware encoding method without pre-training, which transforms the discrete binary bit flipping in traditional methods into a 1024-dimensional continuous signed Token sequence. Drawing on the rotation position encoding in large language models, for unordered molecular graphs, the topological graph distance between atoms is converted into a three-dimensional reference distance and injected into the self-attention rotation phase. Finally, a large number of experiments have verified its feasibility.

Critical strengths

The author's introduction of RoPE into the disordered topological graph structure is highly insightful. It compresses the expensive pairwise attention bias in the classic Graph Transformer into a linear form.

Critical weaknesses

The author's main model integrates the architecture of 7 models. This integration combines elements such as Transformer, CNN, MLP, as well as traditional physical and chemical descriptors from Avalon and ErG. The vast majority of the comparison baselines are independent single models.

Questions for authors

I would like the author to provide the results of a standalone Transformer model using WEECFP-SuRGE on MoleculeNet and TDC datasets.

2 comments

We thank the reviewer for their consideration of the manuscript and positive feedback. We agree with the reviewer with respect to the comment regarding ensemble reporting instead of individual models. The intent of the ensemble methods is to provide a case study application of the methodology and not necessarily to present a core strategy of the manuscript. This point was not clear in the original manuscript, and as such we have aimed to more clearly present the key findings of the work. First, in response to the reviewer’s request, we have provided all individual ensemble member results for both MoleculeNet and current TDC ADMET regression runs in Tables S1 and S10 (SI Sections A and H). Additionally, the individual model results for each of the WEECFP-SuRGE variants were not clearly present in the original draft, so the central method (WEECFP-SuRGE Input, md) has been added to Table 1 as an isolated entry alongside NoSuRGE; Table S4 contains the other standalone variants, which are distinct from the compact ensemble members. Finally, we have included more extensive ablation studies and statistical analyses on the model architecture in Table S14, Sections 5.2–5.4, and SI Section J.

Thank you for the author's explanation which has solved all my problems. I will maintain a positive attitude and enhance my confidence.

Comments on the paper

1 comment

During our review of the submitted results, we identified test-set early stopping in the WEECFP-SuRGE ensemble experiments: the transformer and CNN members selected checkpoints on the test split, while the shared MLP used a training holdout. We therefore reran all nine TDC ADMET regression tasks over five seeds and six standalone MoleculeNet variants on nine affected targets over five scaffold splits, in both cases selecting checkpoints on a validation partition held out from the training pool. These corrected runs are treated as the current evaluation throughout the manuscript, and historical individual-member results are omitted. Recent work has highlighted limitations of the TDC leaderboard itself. An independent audit of top-ranked TDC ADMET models examined reproducibility, data leakage, and overfitting across leaderboard entries [35]; it validated CaliciBoost, MapLight, and MapLight+GNN, identified leakage in MiniMol, GradientBoost, and XGBoost, and reported that other methods could not be fully reproduced because of code availability or execution limitations. The audit provides important context for interpreting leaderboard comparisons, but it does not validate our historical results or imply that biases across methods are comparable. Separately, TDC issue 394 reports a possible leaderboard score-mapping error for Bioavailability_Ma [37]; that benchmark is excluded from all leaderboard rank calculations in Table 2. The main text now reports only the corrected evaluation. We retain the original TDC ensemble results in the supplement (SI §H), labeled “leaky (test-set stopped)”, so that the earlier leaderboard standing can be read against the held-out-checkpoint evaluation that supersedes it.

References [35] Koleiev, I., et al. An End-User Audit of Reproducibility, Data Leakage, and Overfitting of the Top-Ranked ADMET Prediction Models in TDC Leaderboards. Journal of Chemical Information and Modeling 66(14):8045–8058, 2026. DOI: 10.1021/acs.jcim.6c00819. [37] Taekya. [Bug] Score discrepancy and data contamination in Bioavailability_Ma leaderboard. TherapeuticsDataCommons GitHub issue #394, opened June 8, 2026. https://github.com/mims-harvard/TDC/issues/394 (accessed November 15, 2026).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.