acceptodds
Under review as a conference paper at ICLR 2027

eBOND: An Electron-Guided Fragmentation Dataset for Molecular Learning

Abstract

Molecular fragmentation, which decomposes molecules into smaller substructures, is widely used in fragment-based molecular learning. Existing fragmentation schemes primarily determine cleavage sites using hand-crafted chemical rules or cleavage patterns derived from reaction data. These approaches produce fragment spaces shaped by pre-defined structural assumptions or the coverage of recorded reactions, while largely overlooking molecule-specific electronic structure. Electronic structure underlies molecular properties, motivating its use in molecular fragmentation. However, obtaining such electronic information at scale requires costly quantum-chemical calculations, limiting the development of large-scale electron-guided fragmentation datasets. In this work, we introduce eBOND, a large-scale quantum-chemical dataset for electron-guided molecular fragmentation, covering 452,775 molecules at a computational cost of approximately 1.26 million CPU core-hours. Using three-state density functional theory and wave-function analysis, eBOND provides atom- and bond-level electronic descriptors and translates them into molecule-specific scores for fragmentation. On USPTO-50K, eBOND-derived cuts cover 81.1% of recorded synthetic disconnections and recover 73.4% of those missed by both BRICS and RECAP. We further show that eBOND-derived fragments provide useful learning signals for retrosynthesis and fragment-based molecular optimization. These results demonstrate the value of molecule-specific electronic information for molecular fragmentation and establish eBOND as a reusable resource for fragment-based molecular learning.

Primary area:
datasets and benchmarks
Keywords:
Molecular Fragmentation, Electronic-Guided, Molecular Learning

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_3RcYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces EBOND, a large-scale quantum-chemistry dataset for electron-guided molecular fragmentation. The authors run DFT calculations on 452,775 molecules sampled from USPTO, derive atom-level and bond-level electronic response descriptors (condensed dual descriptor, bond dual descriptor, Parr functions), and map them to bond-level fragmentation scores. Validation shows the resulting fragmentation covers 81.1% of recorded synthetic disconnections on USPTO-50K, recovers 73.4% of sites missed by both BRICS and RECAP, and is evaluated on two downstream tasks: single-step retrosynthesis pretraining and fragment-based molecular optimization (DRD4).

Critical strengths

1. The dataset itself is a substantial, reusable contribution.

Three-state DFT on 452,775 molecules at ~1.26M CPU core-hours is beyond what most labs can compute, and the atom/bond-level electronic descriptors are a genuinely new resource for molecular learning.

2. Complementarity to rule-based fragmentation is convincingly demonstrated.

EBOND recovers 73.4% of disconnection sites missed by both BRICS and RECAP (vs 51.0% for a size-matched random baseline), and combined coverage reaches 96.1%. This is strong evidence that electronic information carries signal beyond local structural patterns.

Critical weaknesses

1. Evidence that electron-guided fragmentation improves molecular learning is insufficient.

(i) Table 5 uses single-step retrosynthesis to show that EBOND fragmentation outperforms other fragmentation schemes, but the None baseline already reaches 48.59% Top-1 while EBOND only raises it to 51.65%, and all fragmentation variants cluster within a narrow band. Moreover, only BRICS is compared among the fragmentation methods, the runs are single, so it is unclear whether the margins are more than chance variation; a harder and more discriminative task is needed to establish the advantage. (ii) The optimization study gives a larger gap, but it also covers a single case (one target, DRD4) and compares only BRICS, so the evidence remains insufficient.

2. No comparison with learned fragmentation methods. The baselines are only random, BRICS and RECAP, all rule-based or heuristic. Section 2.1 acknowledges recent data-driven fragmentation and substructure-learning methods, but none are evaluated. So it is unclear whether the electronic signal offers an advantage over learning disconnections directly from data.

3. Fragmentation depends on a single global threshold.

All boundaries come from one threshold, τ = 0.25, applied to molecule-wise z-scored descriptors and combined by union across the three scores. Sensitivity to τ is discussed only in Appendix E.2, with no analysis in the main text, and it is not shown whether one value holds across molecules of different size and across chemical space. Since τ directly controls how many bonds are cut per molecule, and that count is fixed to match EBOND in the random baseline, the reported gains may partly reflect this choice rather than the electronic signal itself.

Questions for authors
  1. Are the Table 5 improvements statistically significant across random seeds? Please report standard deviations for all fragmentation variants.

  2. Can you evaluate under a harder, more discriminative task (e.g. a subset without high-frequency reaction templates, multi-step retrosynthesis, or low-data fine-tuning) to demonstrate the advantage of EBOND fragmentation ?

  3. What is the sensitivity to the threshold τ = 0.25? How do RDC, the number of fragments, and the downstream metrics change as τ varies from 0.1 to 0.5?

  4. Can you add a comparison against at least one learned fragmentation baseline?

  5. In 5.2 both methods receive the same 5,120 generation attempts, but the fragment vocabularies differ in size (26,168 EBOND fragments vs 14,930 BRICS). Since the vocabulary is the search space, a larger pool alone could improve the hit rate under a fixed attempt budget. Can you repeat the comparison with matched vocabulary sizes, or report search diversity statistics, so the gap is attributable to fragment quality rather than pool size?

1 comment

Dear Reviewer 3RcY,

We have addressed your concerns in our rebuttal and would appreciate your feedback on our responses. Please let us know if any points require further clarification.

Best regards,
The Authors

Reviewer_dtqwOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The paper introduces eBOND, a dataset of 452775 molecules with atomic and bond electronic-response descriptors based on Fukui and Parr functions. Their calculation required approximately 1.26 million CPU core-hours, mainly for single-point DFT calculations at the B3LYP/def2-SVP level. These descriptors are used for molecular fragmentation. The authors compare their method with rule-based baselines such as BRICS and RECAP on recovering disconnections in the USPTO-50k dataset. They also report improvements from using eBOND fragments for retrosynthesis pretraining and molecular optimization.

Critical strengths
  • The core idea behind the paper is worthwhile: fragmentation based on molecule-specific electronic responses seems promising and meaningful.
  • The presented electronic-response method complements baselines based on fixed chemical rules. In Table 2, eBOND recovers 73.4% of in-scope disconnections missed by both BRICS and RECAP, compared with 51.0% for the random control.
Critical weaknesses

The central fragmentation rule is inconsistent with both the threshold analysis and the supplied examples.

  • Section 3.3 selects a bond when any standardized score satisfies z_n >= tau. Therefore, increasing tau should not increase the number of selected bonds, since the selection criterion becomes stricter. Instead, Table 11 (Appendix E.2 related to the threshold value) shows that the number of cuts (selected bonds) increases as tau grows from -0.5 to 0.5. Beyond this contradiction, no quantitative criterion is provided for choosing tau.
  • In example_001 of examples/ebond_examples.jsonl, the bond (1, 2) has a molecule-normalized electrophilic BDD fragmentation score of 0.997, exceeding the selection threshold of tau = 0.25. However, this bond is omitted from selected_bonds. Applying the stated rule to all four stored score channels also gives selections that differ from the supplied ones for 41 of 50 molecules: 65 bonds that meet the criterion are missing, while 18 selected bonds meet none of the criteria. These counts use population standard deviation, exclude null CDD values from normalization, and skip zero-variance channels. These inconsistencies in the supplied examples leave it unclear which selection procedure was used to generate the annotations and experimental results. A minimal executable implementation of the stated rule would help resolve this issue.
  • Simply changing the criterion from z_n >= tau to z_n <= tau does not resolve these contradictions and would also be inconsistent with Figure 5, where all fragmentation scores are high for disconnected bonds.

The usefulness of the dataset is not clear

  • If I understand correctly, to perform fragmentation on a new molecule, one needs 3 DFT calculations. This makes such a fragmentation method impractical compared to BRICKS or RECAP. A possible solution is to train a NN to predict the electronic properties (or bond splits directly) and use it as a proxy.
  • The benefit of using DFT over the cheaper xTB method is not demonstrated. In the current pipeline, DFT is used for descriptor calculations, which take almost 92% of the total computational cost. However, it is unclear whether this expense improves fragmentation. If this pipeline is intended for broad use, it would be helpful to evaluate xTB-based descriptors as well, especially since xTB is already used for geometry optimization. At the same time, it would also be useful to test whether further geometry optimization with DFT improves fragmentation.

Other weaknesses

  • There is no penalty for the number of cuts. Table 1 shows that eBOND obtains 81.1% disconnection coverage with 6.17 cuts per molecule, whereas BRICS obtains 82.3% with 5.07 cuts, and RECAP obtains 58.1% with 1.83 cuts. It is worth testing whether the proposed criteria outperform these baselines while keeping the same number of cuts per molecule (for example, by adjusting the threshold value). It is also important to expand the analysis of descriptor contributions in Appendix E.1. Table 10 shows that using subsets of descriptors also changes the number of cuts. It is therefore unclear whether RDC drops because the omitted criterion helps select relevant bonds or simply because removing it reduces the number of cuts.
  • The retrosynthesis application section does not provide enough details about the experiments. The resulting Top-k accuracy is also below published results for several models evaluated on USPTO-50k, including LocalRetro, RetroSynFlow, and the Chimera ensemble (see Table 1 in Yadav et al., 2025; DOI: 10.52202/085713-1846). It motivates testing whether eBOND provides additional value on stronger baseline models.
  • Figure 5 does not clearly distinguish raw descriptors from the bond-selection scores defined in Section 3.3. Panel (a) displays negative BDD values, whereas pi_BDD_nuc = max(-BDD, 0) is nonnegative. For consistency, panels (a) and (b) could show pi_BDD_nuc and pi_BDD_elec, respectively, alongside pi_CDD and pi_Parr in panels (c) and (d). The axes or caption should specify the exact plotted quantities.
  • Table 6 gives property and solved rates on valid molecules, and sweet-spot solved rates on valid sweet-spot molecules. Without validity rates and counts of unique successful molecules, these percentages do not establish the practical yield under the fixed budget of 5120 attempts.
Questions for authors
  • It is unclear how the eBOND fragments are obtained in the molecular optimization experiment (section 5.2). Were DFT calculations needed?
  • How were the 200 molecules from each method selected for the docking comparison in Section 5.2? Please specify the sampling procedure, any filtering criteria, and whether the same procedure was applied to both methods.

Suggestions

  • Prioritize a minimal executable path from supplied descriptor values to selected bonds and the reported evaluation tables.
  • Does increasing the threshold or requiring agreement between multiple electronic criteria increase the fraction of selected bonds that correspond to recorded USPTO disconnections? This would help show whether stronger electronic signals lead to more selective identification of these bonds.
  • It would be useful to compare eBOND with modern fragmentation methods, such as DigFrag (Yang et al., 2024).
3 comments

Dear Reviewer dtqw,

We have addressed your concerns in our rebuttal and would appreciate your feedback on our responses. Please let us know if any points require further clarification.

Best regards,
The Authors

Thank you for the clarification. The updated examples resolve my concern about consistency, and the same-cut comparison, in my view, better shows the contribution of each descriptor used in this work. I am raising my soundness score from 1 to 2 accordingly.

It is now clear that three single-point DFT calculations are required for each previously unseen molecule, and that replacing them with xTB single-point calculations noticeably degrades the metrics (RDC 70.08% vs 78.83%). This confirms my original concern that the method is impractical for large molecular collections compared with more classical approaches such as BRICS and RECAP. The reported cost is about 3 CPU core-hours per molecule on average, while BRICS and RECAP require only sub-millisecond substructure matching.

At the same time, I do not see a corresponding gain in quality. I acknowledge that eBOND provides complementary disconnections, but the demonstrated benefits appear limited relative to its computational cost. At a matched cutting budget, eBOND remains below BRICS and RECAP in RDC. For example, at 5.07 cuts per molecule, eBOND achieves 73.68% RDC compared with 82.30% for BRICS. On the stronger retrosynthesis backbone, the improvement is +0.73 percentage points in Top-1 while Top-5 and Top-10 fall slightly below the original KnowRetro, and no run-to-run variability is reported for this comparison.

While the rebuttal strengthens the technical soundness of the work, this balance between cost and demonstrated benefit leaves me unconvinced about its practical value, and I will therefore maintain my overall rating.

Dear Reviewer dtqw,

We thank the reviewer for recognizing the clarified examples and descriptor comparisons, and for raising the soundness score.

Our central contribution is a reusable quantum-chemical dataset that connects molecule-specific electronic responses with fragmentation, providing atom- and bond-level descriptors and fragmentation annotations for 452,775 molecules. Making these precomputed data available enables the community to investigate electronic information in molecular learning without repeating the calculations for the included molecules.

The reviewer acknowledges that eBOND provides complementary disconnections. This contribution is supported quantitatively: eBOND recovers 73.4% of recorded disconnections missed by both BRICS and RECAP, compared with 51.0% for matched random additional cuts. At a matched average budget of 1.86 cuts per molecule, it also achieves 40.46% RDC versus 22.67% for DigFrag.

The utility extends to molecular optimization. Under the same model, objectives, and generation budget, eBOND fragments increase the sweet-spot rate from 37.31% to 51.39% and the synthesis-planner solved rate from 46.41% to 59.65% relative to BRICS (Table 6).

We will clarify that eBOND’s contribution lies in its shared electronic-structure data, complementary fragmentation information, and demonstrated utility for molecular learning.

Best, The Authors

Reviewer_KyvEOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces eBOND, a quantum-chemical dataset covering 452,775 molecules derived from USPTO. The authors perform B3LYP/def2-SVP calculations for the neutral N-electron and corresponding electron-added N+1 and electron-removed N-1 states at the optimized molecular geometries. They extract atom- and bond-level electronic descriptors, including CDD, BDD, and Parr functions. Then they evaluates each bond relative to the other candidate bonds in the same molecule and translate these scores into discrete fragmentation decisions. Two downstream applications are demonstrated: fragmentation-based pretraining for single-step retrosynthesis and fragment-based molecular optimization on DRD4.

Critical strengths

1.Constructing three-state electronic descriptors for 452,775 molecules represents a considerable computational contribution. The dataset provides atom- and bond-level information across approximately 12.1 million heavy-atom bonds. Electron-guided fragmentation selects approximately 1.87 million fragmentation boundaries, producing approximately 154,000 unique fragments. 2.On USPTO-50K, EBOND-derived cuts cover 81.1% of recorded synthetic disconnections and recover 73.4% of those missed by both BRICS and RECAP, demonstrating their synthetic relevance and complemenstarity to rule-based methods. The downstream experiments hold the model architecture or optimization framework fixed, providing evidence that changing the fragmentation procedure can improve performance within those settings.

Critical weaknesses

1.In the downstream applications, Table 5 compares four pretraining variants of one Transformer, while Table 6 compares BRICS and eBOND fragments within the existing DeepFMPO framework on one target. These are useful controlled experiments, but they do not establish whether the gains persist on stronger established backbones or against other competitive pretraining and fragmentation approaches. Stronger evidence is needed to substantiate the breadth and robustness of the practical benefit. 2.The fragment evaluation does not yet establish improved fragmentation efficiency or final-fragment quality. In Table 1, eBOND selects more bonds than BRICS while achieving lower recorded-disconnection coverage. Overall building-block coverage is also slightly lower.

Questions for authors

1.The bond is selected with a thresthold τ\tau. With fixed scores, increasing τ\tau must produce a subset of the previously selected bonds. Why the Cuts/mol in Table 11 becomes less when τ\tau gets larger? 2.What is the incremental benefit of three-state electronic response over less expensive electronic features? How do neutral-state charges and bond orders compare with the proposed descriptors in disconnection coverage and downstream utility? 3.Does the calculation pipeline introduce systematic selection effects? Around 86,000 molecules are failed. What’s the structual distribution of these molecules? Is this pipeline not applicable to certain types of molecules? 4.Only one conformer is used for each molecule. How reliable are the electronic annotations and resulting cut sets? Could the authors evaluate a representative subset across multiple low-energy conformers and reasonable electronic-structure settings?

1 comment

Dear Reviewer KyvE,

We have addressed your concerns in our rebuttal and would appreciate your feedback on our responses. Please let us know if any points require further clarification.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.