acceptodds
Under review as a conference paper at ICLR 2027

TopoCompile: Molecular Generation with Exact Global Count Constraints

Abstract

Molecular generation designs novel molecules with desired structures and properties. Global count constraints specify molecular size, elemental composition, and ring content without prescribing detailed structure. Property-conditioned generators steer generation toward such targets but do not guarantee exact satisfaction. We introduce TopoCompile, which combines an autoregressive component-tree generator with a dynamic-programming compiler for supported component-additive integer counts. TopoCompile represents molecules as trees of intact ring systems and individual atoms, making each supported count additive over components. At each step, an autoregressive Transformer scores component and attachment actions, while the compiler masks actions with no target-satisfying completion. Within the fixed component vocabulary and typed attachment rules, constrained decoding guarantees exact equality for supported non-negative, component-additive integer counts. Applied at inference time, these checks let the same trained generator support different targets and count combinations without retraining. Across four benchmarks, TopoCompile achieves 100% Exact-valid yield for Common and Rare four-count target groups; the strongest baseline within each dataset reaches 44.9% or less on Rare targets. In unconditional generation, it achieves the highest validity or valid-unique-novel (V.U.N.) among the compared methods across all four benchmark datasets.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Molecular Generation, Dynamic Programming, Constrained Generation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_qWczOpenReview ↗
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

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)(N, \beta, N_O, R_{\text{sys}}), 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)(N, \beta, N_O, R_{\text{sys}}): 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-MM compilation and decoding:

Method (Frozen Checkpoint)Total Online Time (s) ↓Mean Time / Request (s) ↓Exact Outputs Delivered ↑Exact-Valid Yield (%) ↑
STGG+94.0s2.94s93829.3%
STGG+ Count Mask118.8s3.71s1,13235.4%
TopoCompile (ours, all-MM)27.2s0.85s3,200100.0%

(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)r(\tau) without retraining pθp_\theta, 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-MM 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-MM timings and practical requests with unspecified MM, 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 (D0D_0), slot-pair diversity (PP, accounting for all requested output slots), and novel unique exact (NUE) yield. Even without graph editing, TopoCompile (Closed) substantially improves PP and NUE yield over STGG+ Count Mask; count-preserving editing (Edit) further improves both metrics.

DatasetMethodD0D_0 ↑PP ↑NUE Yield (%) ↑
MOSESSTGG+ Count Mask0.8720.11934.7
TopoCompile (Closed)0.8680.86895.3
TopoCompile (Edit)0.9210.92199.8
ZINC250KSTGG+ Count Mask0.8820.20846.8
TopoCompile (Closed)0.8350.83585.2
TopoCompile (Edit)0.9320.93299.6
GuacaMolSTGG+ Count Mask0.8170.03214.7
TopoCompile (Closed)0.8020.80265.5
TopoCompile (Edit)0.9250.92592.7
NPGenSTGG+ Count Mask0.7320.03211.2
TopoCompile (Closed)0.7250.72546.6
TopoCompile (Edit)0.8770.87781.3

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-MM 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).

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.

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

TopoCompile combines an autoregressive component-tree generator with dynamic programming to enforce exact global molecular count constraints. It reports 100% exact-valid yield on tested four-count targets across four datasets.

Critical strengths

The exactness guarantee is clearly formulated and supported by implementation checks. Main baseline comparisons use shared requests, attempt budgets, and evaluators.

Critical weaknesses
  1. Generalization to unseen four-count combinations is not systematically evaluated. The main targets are observed combinations grouped by training frequency. The paper includes a high-rank four-count example, but systematic unseen-target tests vary only cycle rank.
  2. Comparison with other hard-constrained generators is limited. Main baselines use conditioning, guidance, filtering, or local count masks. The paper discusses other hard-constrained approaches, but does not quantitatively compare them on shared requests.
  3. Quality under difficult targets is unclear. Novelty and diversity are reported for Common targets, and targets are excluded if any compared method produces fewer than two exact outputs. The quality comparison therefore does not establish results on Rare targets.
Questions for authors
  1. How does TopoCompile perform on unseen four-count combinations?
  2. What are the novelty and diversity results on Rare targets?
  3. How does it compare with another hard-constrained generator on shared requests?
3 comments

We thank the reviewer EgYb for the constructive comments. We have revised the paper accordingly; changes are marked in blue.


Response to W1 & Q1

We have added a systematic evaluation on unseen four-count combinations across all four benchmarks, where TopoCompile achieves 100% exact-valid yield across all targets compared to <6% for conditional/guided baselines and 0%–31% for hard-constrained baselines.

1. Why TopoCompile generalizes zero-shot to unseen combinations. Conditional and guided generators rely on learned embeddings of training count tuples and degrade when queried on out-of-distribution combinations. In contrast, TopoCompile's DP compiler checks exact suffix reachability over additive component contributions r(τ)r(\tau) at inference time. As long as a requested count tuple can be assembled from the component vocabulary, the compiler guarantees exact satisfaction regardless of whether the joint tuple appeared in training.

2. Results on unseen four-count benchmarks. We evaluated 64 held-out unseen four-count combinations (N,β,NO,Rsys)(N, \beta, N_O, R_{\text{sys}}) on GuacaMol and NPGen, where each individual count appeared in training but the joint tuple was absent from all training sets. With our single-bond port-forest reachability certificate—which checks completion directly via degree-sum and count-sum identities and compiles the hardest unseen NPGen queries in 1.5–3.2s—fixed-vocabulary TopoCompile solves every target, whereas STGG+ and STGG+ Count Mask fail to complete any target:

MethodNPGen Exact Yield (%) ↑GuacaMol Exact Yield (%) ↑Fully Solved Targets (%) ↑
STGG+1.4%2.3%0.0%
STGG+ Count Mask3.1%5.5%0.0%
TopoCompile (ours)100.0%100.0%100.0%

On an additional four-dataset unseen panel across MOSES, ZINC250K, GuacaMol, and NPGen, FragFM+Guidance (0%–0.8%), GenMol+Filter (0%), and Graph DiT (0.4%–3.9%) rarely succeed, while TopoCompile maintains 100% exact delivery across all datasets. Similarly, on unseen multi-count targets in our hard-constraint benchmark, TopoCompile achieves 100% exact delivery (98%–100% unique), whereas CP-BP times out (0%) and NSPSG reaches 0%–31.3%. We summarize these findings in Section 4.2 and report the complete four-dataset unseen yields and tier-level hard-constraint breakdowns in Appendix I.2 (Table I.2) and Appendix K.2 (Table K.2).

Revision: Added unseen four-count generalization experiments in Section 4.2 and Appendix I.2 (Table I.2), the port-forest reachability certificate and compilation timings in Appendix J.1 (Table J.1), and unseen hard-constraint comparisons in Appendix K (Table K.2).


Response to W2 & Q3

We have added two representative hard-constrained baselines, CP-BP (Saikali & Pesant, AAAI 2026) and NSPSG (Zhang et al., ICML 2026), to evaluate exact constraint satisfaction, efficiency, and diversity under shared requests.

1. Exact constraint satisfaction and efficiency.

On Rare four-count targets across all four datasets, TopoCompile achieves 100% exact-valid yield, compared with 0%–8.3% for CP-BP and 17.1%–41.7% for NSPSG. On a separate 100-request shared three-count benchmark, TopoCompile achieves 100% exact yield with a median runtime of 1.0s, substantially outperforming both baselines:

MethodExact Yield (%) ↑Unique Exact (%) ↑Requests Solved (%) ↑Median Time (s) ↓
Graph DiT2.92.942.019.7
CP-BP14.614.436.0300.2
NSPSG (adapted)44.344.394.0243.6
TopoCompile (ours)100.099.7100.01.0

2. Diversity on difficult Rare targets.

TopoCompile+Edit achieves higher equal-survivor Morgan diversity than both hard-constrained baselines across all Rare benchmark groups while preserving 100% exact delivery. For example, on GuacaMol, paired diversity reaches 0.936, compared with 0.883 for NSPSG and 0.670 for CP-BP.

Revision: Added CP-BP and NSPSG comparisons in Section 4.2 (Table 1), and the 100-request shared benchmark and Rare-target diversity results in Appendix K (Tables K.1–K.2).

Response to W3 & Q2

We have added Rare-target evaluations across all four benchmarks, including unfiltered all-target metrics to eliminate survivor selection bias and equal-survivor comparisons to assess structural diversity. TopoCompile achieves 100% exact-valid yield across all Rare targets while substantially improving diversity and novel unique exact (NUE) candidate supply.

1. Unfiltered all-target evaluation.

Excluding targets with fewer than two exact outputs introduces survivor bias, particularly on difficult Rare requests (e.g., STGG+ fails on 40.6% of NPGen Rare targets). We therefore evaluate all 32 Rare targets per dataset (100 attempts each), reporting within-target diversity (D0D_0), slot-pair diversity (PP, accounting for all requested output slots), and NUE yield. Without graph editing, TopoCompile (Closed) substantially improves PP and NUE yield over STGG+ Count Mask. For example, on NPGen, PP increases from 0.032 to 0.725 and NUE yield from 11.2% to 46.6%; count-preserving editing (Edit) further improves them to 0.877 and 81.3%, respectively.

DatasetMethodD0D_0 ↑PP ↑NUE Yield (%) ↑
MOSESSTGG+ Count Mask0.8720.11934.7
TopoCompile (Closed)0.8680.86895.3
TopoCompile (Edit)0.9210.92199.8
ZINC250KSTGG+ Count Mask0.8820.20846.8
TopoCompile (Closed)0.8350.83585.2
TopoCompile (Edit)0.9320.93299.6
GuacaMolSTGG+ Count Mask0.8170.03214.7
TopoCompile (Closed)0.8020.80265.5
TopoCompile (Edit)0.9250.92592.7
NPGenSTGG+ Count Mask0.7320.03211.2
TopoCompile (Closed)0.7250.72546.6
TopoCompile (Edit)0.8770.87781.3

2. Equal-survivor structural diversity.

We also 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), alongside higher Bemis–Murcko scaffold and atom-pair diversity. Applying the same editing procedure to baseline outputs confirms that TopoCompile retains its diversity advantage while preserving all four target counts.

Revision: Added Rare-target evaluations and equal-survivor comparisons in Section 4.2 (Figure 3) and Appendix L (Tables L.1–L.2), with additional vocabulary-support analysis in Appendices J.2–J.3.

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.

Reviewer_pJ69OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The work deals with development of the TOPOCOMPILE approach that allows exact constrained molecule generation. The approach uses an autoregressive Transformer as a generator with a dynamic programming compiler. The molecules are represented as component trees through bridge decomposition of the molecular graph. The DP compiler checks for constraints satisfactions. The proposed approach is demonstrated using four datasets for small molecule design.

Critical strengths
  1. The superior exact-valid yield across all datasets for different targets demonstrates the effectiveness of the dynamic programming compiler.

  2. The ability to handle different target values and combinations of counts without retraining helps in increasing the flexibility and usability of the generator.

  3. The TOPOCOMPILE is benchmarked against different baselines effectively.

Critical weaknesses
  1. The fixed component vocabulary may limit exploration of novel chemical space, particularly for rare components and also impact diversity.
  2. The important point of the dynamic programming compiler's computational complexity with respect to vocabulary size and constraints is not addressed in the work.
  3. The work only discusses the perfectly specified targets. No discussion how the proposed approach handles noisy or incomplete specified targets
  4. The work does not provide a detailed comparison of the diversity of generated molecules against baselines beyond internal diversity metrics, such as structural diversity or novelty relative to existing databases.
  5. The constraint satisfaction does not lead to better drug-like properties or synthesizability. The experiment lacks the analysis for these aspects.
Questions for authors
  1. How will the out of vocabulary and bridge compositions be handled in the proposed method? How sensitive is the performance of the proposed method to the quality of the bridge decomposition?
  2. For material design, how will the dynamic programming compiler scale with constraints and vocabulary size? How does the performance scale with an increasing number of constraints? Is there a point at which the dynamic programming compiler becomes computationally prohibitive?
  3. How will your method handle noisy or incomplet specified targets?
  4. What are the criteria used for grouping actions? What is the trade-offs between compilation time and the potential loss of precision or diversity?
7 comments

We thank the reviewer pJ69 for the constructive comments. We have revised the paper accordingly; changes are marked in blue.


Response to W1:

1. Chemical space coverage and combinatorial recombination. We clarify that TopoCompile fixes the vocabulary of component-level building blocks (intact ring systems and single non-ring atoms), rather than whole-molecule scaffolds. Compatible components recombine freely whenever the partial tree admits a valid DP completion. Across all four benchmarks (MOSES, ZINC250K, GuacaMol, and NPGen), training vocabularies cover 99.6–99.9% of test component occurrences (93.3–99.3% of whole test molecules), and fixed-vocabulary TopoCompile achieves the highest internal diversity on matched Common targets.

2. Utilization and diversity of rare components. We added an analysis of rare ring templates (seen ≤10\le 10 times in training) across equal-sized Rare exact-output streams. Because the DP compiler masks count-incompatible motifs and renormalizes probabilities strictly over feasible completions, rare templates are never crowded out by frequent ones. Across the four benchmarks, fixed-vocabulary TopoCompile samples 1.7$\times$–2.8$\times$ more distinct rare ring templates than STGG+ Count Mask and 2.2$\times$–2.8$\times$ more than STGG+, while maintaining comparable or higher Morgan diversity on rare-ring outputs:

DatasetTopoCompile (ours)STGG+ Count MaskMultiple
MOSES132532.5$\times$
ZINC250K236862.7$\times$
GuacaMol81292.8$\times$
NPGen94561.7$\times$

3. How OOV limitations are handled. We have expanded the paper to detail and evaluate three mechanisms addressing OOV boundaries. Specifically, direct generation cannot create ring interiors or port signatures absent from the vocabulary. We address this boundary in three ways: (i) the root DP check detects vocabulary-infeasible targets before sampling; (ii) terminal ring closure (Closed) converts single-port rings into zero-port templates, increasing NPGen Rare slot-pair diversity from 0.622 to 0.725; and (iii) count-preserving local graph edits (TopoCompile+Edit) explore OOV ring structures while retaining all four target counts, raising this diversity to 0.877. The edits extend generation after decoding and are distinct from the fixed-vocabulary DP guarantee.

Revision: Clarified component recombination and test coverage in Section 3.1, Table 1B, and Appendix F (Table S8a), and added rare-ring utilization in Appendix J.3 (Table J.2), vocabulary-truncation infeasibility experiments in Appendix I.5 (Table I.6), exhaustive singleton support proofs and terminal ring closure in Appendix J.2, and OOV count-preserving local graph edits in Section 3.3, Section 4.2 (Figure 3), and Appendix L (Tables L.1–L.3).


Response to Q1:

1. Out-of-vocabulary (OOV) handling. Please see W1 (Point 3) above.

2. Handling unseen bridge compositions. We clarify that TopoCompile places no training-set restriction on component pairs or bridge connections: any parent port ptp_t and child port qtq_t of matching bond type can connect whenever the successor state Next(zt,γ)\text{Next}(z_t, \gamma) admits a valid completion (Λc⋆=1\Lambda_{c^\star}=1). Without vocabulary modification or retraining, the generator produces 92.7% novel molecules, 76.6% unseen component multisets, and 98.0% unseen attachment-bag signatures on MOSES.

3. Sensitivity to bridge decomposition. We larify that bridge decomposition is uniquely determined by graph connectivity rather than heuristic choices, resulting in zero sensitivity to decomposition order. By Tarjan's theorem (Tarjan, 1974), bridge edges are precisely the non-cycle edges of a connected graph, partitioning any molecule into maximal ring systems and acyclic single atoms. Verification across all 4.24M retained dataset molecules and 4,096 random atom-order permutations confirms 100% decomposition and reconstruction consistency.

Revision: Clarified bridge-decomposition uniqueness and unseen bridge recombination in Section 3.1, Proposition 1 (Appendix A.2), Appendix B.5, and Appendix F.

Response to W2 & Q2:

1. Theoretical complexity.
Compiling across candidate component counts MM takes O(dGmax⁡∑m=1Mmax⁡Sm)O(dG_{\max}\sum_{m=1}^{M_{\max}}S_m) time, with worst-case state space SM≤1+(M+45)∏j=1d(bj⋆+1)S_M \leq 1+\binom{M+4}{5}\prod_{j=1}^{d}(b_j^\star+1). Although Gmax⁡≤4∣V∣G_{\max}\leq 4|\mathcal{V}|, templates sharing identical port and count signatures are grouped into equivalent transitions, reducing practical branching to ∣Γ∣≪4∣V∣|\Gamma|\ll 4|\mathcal{V}| and evaluated transitions by 15×–18×.

2. Empirical scaling across vocabulary sizes and constraint counts.
Cold all-MM compilation completes in sub-second to few-second median time across vocabularies of approximately 11k–71k templates. Increasing the constraint count does not necessarily increase runtime monotonically because additional constraints can improve feasibility pruning. The hardest unseen four-count NPGen queries are further accelerated to 1.5–3.2s using port-forest reachability optimization.

DatasetVocabulary (∣V∣\lvert\mathcal{V}\rvert)Groups (∣Γ∣\lvert\Gamma\rvert)1 constraint (s)2 constraints (s)3 constraints (s)4 constraints (s)
ZINC250K~11k~1.1k0.050.060.180.16
MOSES~15k~1.1k0.040.040.090.09
GuacaMol~58k~6.9k0.160.300.900.74
NPGen~71k~12.7k0.130.250.672.64

3. Scaling to materials design and tractability boundaries.
Representing intact ring systems as components reduces the generation horizon MM. Although the worst-case state bound grows multiplicatively with the number and magnitudes of constraints, correlated chemical counts and reachability pruning can substantially reduce explored states. Compilation is also cached per target, with amortized cost O(dSGmax⁡/B)O(dSG_{\max}/B) over BB samples. Thus, tractability depends primarily on the reachable DP states rather than vocabulary size alone.

Revision: Updated Section 3.3 and Appendix A.3 with the all-MM complexity bound; added grouping and caching results in Figure 5(c,d) and Appendix E (Table S6), and cold-compilation scaling and port-forest optimization results in Appendices I.3 (Table I.3) and J.1 (Table J.1).


Response to W3 & Q3:

TopoCompile supports incomplete and interval-tolerant specifications at inference time without retraining, while detecting infeasible requests before sampling.

1. Incomplete specifications.
When only a subset of counts is specified, the compiler checks feasibility over the active coordinates. Across single-, two-, and three-count schemas on all four benchmarks, TopoCompile achieves 100% exact-valid yield without retraining.

2. Noisy and contradictory specifications.
For tolerance intervals (e.g., NO±1,±2N_O\pm1,\pm2 or β±1,±2\beta\pm1,\pm2), the compiler accepts any completion satisfying an admissible target vector. On MOSES and ZINC250K, it achieves 100% admissible-set satisfaction, 0.85–0.89 internal diversity, and 0.24–1.07s median compilation time on compatible requests. All tested contradictory specifications are detected before sampling (100% rejection).

Revision: Clarified incomplete and interval-tolerant targets in Section 3.2, with subset-schema results in Appendix C (Tables S2a–S2c) and additional evaluations in Appendix I.5 (Table I.7).

Response to W4:

TopoCompile achieves high scaffold and fingerprint diversity alongside 97%–100% external novelty against ChEMBL 37 (~2.9M compounds), delivering up to 25$\times$ more novel unique exact candidates than baselines on Rare targets.

1. Structural diversity. We expanded our Rare-target evaluation with Bemis–Murcko scaffold ratios, atom-pair fingerprint diversity, and training nearest-neighbor similarity. TopoCompile generates novel Murcko scaffolds on the majority of Rare outputs (42.3%–80.6% across the four datasets, with all ranges below spanning MOSES, ZINC250K, GuacaMol, and NPGen) and maintains low training nearest-neighbor similarity (~0.35 mean Tanimoto). Adding count-preserving local edits (TopoCompile+Edit) outperforms both STGG+ and STGG+ Count Mask on every benchmark across matched Murcko scaffold diversity (up to +7.2% higher), atom-pair diversity (+0.07 to +0.13 higher), and unfiltered slot-pair diversity PP (0.88–0.93 vs. 0.03–0.21). Section 4.2 (Figure 3) illustrates the unfiltered slot-pair diversity PP and equal-survivor scaffold comparisons across all four benchmarks, while Appendix I.4 (Table I.4) and Appendix L (Tables L.1–L.2) report the complete per-dataset metrics.

2. External database novelty. We added an external database novelty evaluation comparing all Rare exact outputs against ChEMBL 37 (~2.9M compounds). Across the four benchmarks, 97%–100% of our generated molecules are absent from ChEMBL 37, and our 100% exact yield delivers 2.1$\times$–25.5$\times$ as many novel, unique exact candidates as baselines under the same sampling budget (full per-dataset results in Appendix I.4, Table I.4 and Appendix L.1, Table L.1).

Revision: Added Bemis–Murcko scaffold diversity, atom-pair diversity, nearest-neighbor Tanimoto similarity, and ChEMBL 37 novelty analyses in Section 4.2 (Figure 3), Appendix I.4 (Table I.4), and Appendix L (Tables L.1–L.2).


Response to W5:

We have added QED (drug-likeness) and SA (synthetic accessibility) evaluations across all four benchmarks. The results demonstrate that TopoCompile maintains competitive molecular quality while guaranteeing exact constraint satisfaction. For example, TopoCompile achieves better SA than STGG+ Count Mask on GuacaMol (4.62 vs. 4.95) and NPGen (4.36 vs. 4.51). Comparisons with target-matched reference molecules further show that the generated molecules exhibit physicochemical properties consistent with their specified structural regimes.

Revision: Added QED and SA evaluations against baselines and target-matched reference molecules in Appendix I.4 (Tables I.4–I.5).


Response to Q4:

We clarify that action grouping is exact: it accelerates DP compilation by 12$\times$–14$\times$ without altering the feasible action mask or sampling distribution. At each non-root step, concrete actions a=(p,σ,τ,q)a = (p, \sigma, \tau, q) are grouped by γ(a)=(σ(τ),r(τ),l(q))\gamma(a) = (\sigma(\tau), r(\tau), l(q))—the component's port vector, count contribution, and attaching bond type. Because the state transition Next(z,γ)=(k−1,f+σγ−2elγ,b−rγ)\text{Next}(z, \gamma) = (k-1, f + \sigma_\gamma - 2e_{l_\gamma}, b - r_\gamma) depends only on γ(a)\gamma(a), all actions in γ\gamma reach the same successor state and share the same Boolean feasibility Λc⋆(z,γ)\Lambda_{c^\star}(z, \gamma). Grouping merges only this Boolean DP lookup, while the Action Transformer still conditions on the full prefix hth_t to score every concrete component τ\tau and attachment site (p,q)(p, q) individually. This yields 13.9$\times$ (MOSES) and 12.2$\times$ (ZINC250K) compilation speedups with zero mask or probability discrepancies against ungrouped enumeration.

Revision: Clarified the exactness of action grouping in Section 3.2 (Eq. 2) and Section 3.3 (Eq. 4), reported compilation speedups in Appendix E (Table S6), and added equivalence verification in Appendix I.6.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.