- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 2
Read review
Paper summary
This paper presents TopoCompile which generates molecules by assembling intact ring systems and individual atoms, using a Transformer and dynamic programming to reject additions that cannot lead to a valid completion. Its main contribution is a compact state representation that tracks remaining components, typed attachment ports, and residual counts. The paper reports 100% exact validity across four datasets, with ablations on joint versus factorized feasibility and learned policies. The validity guarantee is limited to the defined template vocabulary and attachment rules, rather than the full molecular space.
Critical strengths
- The compact completion oracle is a useful contribution. Prior work has explored motif-tree generation and hard-constrained molecular generation [1–4], so the novelty lies mainly in the sufficient-state reduction. Checking exact completion without enumerating molecular structures is a meaningful improvement.
- The ablations help explain the method's effectiveness. Joint versus factorized checking shows that individually feasible requirements may not admit a common completion. The learned-policy ablation further shows that feasibility alone is insufficient to generate realistic molecular distributions.
References-- [1] Jin et al. Junction Tree Variational Autoencoder for Molecular Graph Generation. ICML, 2018. [2] Saikali and Pesant. Constrained Molecule Generation Modelled Using the Grammar Constraint. AAAI, 2026. [3] Geng et al. Neural Proposals, Symbolic Guarantees: Neuro-Symbolic Graph Generation with Hard Constraints. arXiv:2602.16954, 2026. [4] Manibod et al. Constrained Sequential Inference in Machine Learning Using Constraint Programming. IJCAI, 2025. [5] Jolicoeur-Martineau et al. Any-Property-Conditional Molecule Generation with Self-Criticism using Spanning Trees. arXiv:2407.09357, 2024.
Critical weaknesses
- The baseline setup does not yet support the claimed comparative advantage cleanly. Appendix B.3 supplies requested counts to released Graph DiT and STGG+ checkpoints without explaining how their conditioning interfaces learned those quantities. Original STGG+ ZINC conditioning uses molecular weight, logP, and QED. Figure 4(b) also compares 2.08 hours of STGG+ “adaptation” with 0.67 seconds for TopoCompile, although STGG+ can change trained property values without retraining and Appendix B describes frozen checkpoints. Training for new count definitions must be separated from serving new values. These ambiguities affect both the hit-rate interpretation and the efficiency claim.
- The practical evidence is perhaps weak for difficult requests. Table 1(B) evaluates Common targets after excluding those with too few baseline survivors; NPGen contributes only 45 matched samples across 15 targets. This cannot establish useful diversity on Rare targets. Likewise, fixed-component-count timings in Table S6 do not measure cold compilation over all candidate component counts for the four-count task. The reported increase from 0.034 seconds for one count to 19.011 seconds for three makes that omission consequential.
Questions for authors
Which count labels trained each baseline checkpoint, and what operations constitute “adaptation”?
2 comments
As the discussion period is drawing to a close, we would greatly appreciate it if you could let us know whether our responses and additional experiments have addressed your concerns.
Thank you again for your time and thoughtful consideration.
We thank the reviewer qWcz for the constructive comments. We have revised the paper accordingly; changes are marked in blue.
Response to W1 & Q1
We clarify that all evaluated conditional baseline checkpoints were explicitly trained with supervised labels for the four target counts (N,β,NO,Rsys), and when serving new target values online using frozen checkpoints, TopoCompile is 3.5$\times$–4.4$\times$ faster while delivering 2.8$\times$–3.4$\times$ as many exact molecules.
1. Baseline checkpoint training labels.
Rather than using mismatched default property checkpoints (such as ZINC logP/QED/MW), all evaluated STGG+ and Graph DiT checkpoints were trained directly on the four target counts (N,β,NO,Rsys): STGG+ fits its property scaler on these four count channels during training, and Graph DiT binds its conditioning slots to the same four counts (while FragFM+Guidance uses dataset-specific scalar predictors trained on the active counts). All conditional baselines therefore received direct supervision on the requested counts.
2. Separating offline training from online serving.
We agree that online serving new target values should be distinguished from offline training (2–3 hours). Figure 5(b) compares online serving on 32 shared MOSES Rare four-count requests (100 slots each) with loaded checkpoints; TopoCompile includes cold all-M compilation and decoding:
(Moreover, if a user introduces a new component-additive count definition such as nitrogen or halogen count, TopoCompile supports it zero-shot at inference time via component annotations r(τ) without retraining pθ, whereas conditional baselines require retraining.)
Revision: Clarified baseline count-training details in Appendix B.3 and Appendix I.1, and updated the online serving latency comparison in Section 4.5 (Figure 5b) and Appendix I.1 (Table I.1).
Response to W2
We have added cold all-M compilation experiments and Rare-target diversity evaluations across all four datasets. The results demonstrate practical compilation efficiency and diversity without excluding failed targets.
1. Cold compilation across all component counts.
To address the gap between fixed-M timings and practical requests with unspecified M, we evaluated cold compilation across all candidate component counts with 1–4 constraints. For four-count requests, median compilation times are 0.09s (MOSES), 0.16s (ZINC250K), 0.74s (GuacaMol), and 2.64s (NPGen). Our port-forest reachability optimization further reduces compilation of challenging unseen NPGen requests to 1.5–3.2s. These results demonstrate that full target compilation remains practical even with large component vocabularies.
2. Rare-target diversity without survivor exclusion.
We evaluated all 32 Rare four-count targets per dataset (100 attempts each), including targets where baselines produce insufficient exact outputs. We report within-target diversity (D0), slot-pair diversity (P, accounting for all requested output slots), and novel unique exact (NUE) yield. Even without graph editing, TopoCompile (Closed) substantially improves P and NUE yield over STGG+ Count Mask; count-preserving editing (Edit) further improves both metrics.
We additionally compare equal numbers of exact outputs on surviving Rare targets. TopoCompile+Edit achieves higher paired Morgan diversity than STGG+ and Count Mask across all four datasets (0.916–0.935 vs. 0.816–0.882), demonstrating that the advantage persists under matched-output comparisons.
Revision: Added cold all-M compilation results in Appendix I.3 (Table I.3), reachability optimizations in Appendix J.1 (Table J.1), and Rare-target diversity evaluations in Section 4.2 (Figure 3) and Appendix L (Tables L.1–L.2).