acceptodds
Under review as a conference paper at ICLR 2027

Structure-Constrained Molecular Design with Discrete Graph Bridge Diffusion

Abstract

Structure-constrained molecular design, such as decorating a scaffold, extending a motif, or linking fragments, requires generating around a fixed substructure while leaving it untouched. Diffusion models are a natural fit for this: unlike autoregressive language models, they can enforce such constraints directly by construction, simply by clamping the prescribed subgraph during sampling. However, much of the diffusion literature has focused instead on de novo generation or property optimization via guidance, with structural constraints more often imposed at inference time than trained for directly. We revisit this setting along two complementary directions. The first focuses on the role of preprocessing, examining how much of the apparent gap between graph diffusion and chemical language models can be attributed to preprocessing rather than architecture. This alone yields a substantially stronger constrained-generation baseline than prior graph diffusion work. The second, Graph Bridge Diffusion (), goes further by training the model to see the constraint throughout the generative process rather than only at inference, closing a train–inference mismatch that every prior clamping-based approach leaves open. The result is a single model that serves a variety of structure-based generation tasks without task-specific retraining, and that better matches the reference chemical series in composition, not just in satisfying the constraint. Trained on 1.6M molecules with 16M parameters, GBD performs competitively with chemical language models with more parameters and over 1.1B molecular strings for pretraining, while outperforming graph diffusion baselines at matched budgets.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
structure-constrained molecular design, molecular optimization, diffusion models, bridge diffusion, generative models

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_qhhwOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies structure-constrained molecular generation using Graph Bridge Diffusion (GBD), a conditional flow-matching model built on DeFoG. GBD keeps a prescribed subgraph fixed while generating the remaining nodes and edges. The method is evaluated on MolGenBench and SAFE, with additional analysis of the effect of Kekulization on constrained generation.

Critical strengths

The observation that aromatic-bond representation can severely hurt constrained graph generation is interesting and might be useful.

Critical weaknesses
  1. The novelty appears incremental. Keeping the condition clean while noising only the target is already standard in partial noising and conditional diffusion. Diffusion Forcing uses the same high-level principle.

  2. Proposition 1 is largely a direct consequence of standard discrete flow-matching theory rather than a new theoretical result. It assumes an exact conditional posterior, while Algorithm 1 is not shown to induce the same conditional distribution assumed by the theorem.

  3. The “bridge” term is misleading. GBD fixes part of the coordinates and conditionally generates the rest. It is not a conventional endpoint-conditioned diffusion bridge. A different name may be clearer.

  4. Most of the validity gain comes from preprocessing, not the proposed methodology. Validity changes from 0.081 for DeFoG to 0.926 for DeFoG with the preprocessing, and only then to 0.945 for GBD. Thus, the additional validity gain from bridge conditioning is modest. Kekulization is standard preprocessing and should be separated from the methodological contribution.

  5. Table 4 uses only 1k samples and omits standard distributional metrics such as FCD. If de novo generation is part of the generality claim, a standard MOSES evaluation with recent baselines such as GruM, HOG-Diff, and Cometh would be more convincing.

Questions for authors

Atom, Ring, and FG are described as Jensen-Shannon divergences, but Table 1 treats them as higher-is-better. Why?

Reviewer_khm3OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper adapts DeFoG to structure-constrained molecular generation. It combines Kekulé preprocessing with conditional training that keeps prescribed atoms and internal bonds fixed while corrupting and generating the remaining part. A mixture of scaffold, fragment, and unconditional training examples enables one checkpoint to support several generation tasks.

On MolGenBench and SAFE, most of the validity gain comes from preprocessing, with anchored training adding gains in SA and distribution fidelity.

Critical strengths

Experimental gains:

  • The preprocessing is useful: on MolGenBench, validity improves from 0.081 to 0.926 without bridge training.
  • Bridge training helps beyond preprocessing at the same 500-step budget (SA 0.615→0.748, functional-group fidelity 0.718→0.777), though diversity drops somewhat.

Expressing different editing tasks as different anchor sets is a clean, practical interface, and one compact model covers scaffold decoration and several fragment settings.

Critical weaknesses
  • Incremental method: This is essentially partial noising plus standard conditional flow matching applied to DeFoG, and the exactness argument follows the usual conditional DFM line. The real contribution is the empirical diagnosis, and the large validity gain should be credited to preprocessing, not the bridge.
  • Theory and algorithm don’t quite match. Algorithm 1 samples a molecule first and then picks an anchor with a molecule-dependent rule r(S|G). The induced conditional is p_train(G|S) ∝ p_data(G)·r(S|G), which differs from Eq. (3) unless r is constant across compatible molecules. Scaffold extraction and BRICS sampling don’t obviously satisfy this. Proposition 1 holds for the idealized objective, but needs justification or a redefined target to apply to Algorithm 1. This undercuts the “exact conditional” claim.
  • Reference-derived graph sizes. Atom counts are drawn from the reference molecules, which leaks information beyond the anchor. Comparisons with other methods and the fidelity and data-efficiency claims are affected.
  • The tenfold sampling claim isn’t isolated. In Table 2, both models hold up at 50 steps (validity 0.945 vs. 0.926 at 500 steps for GBD and 0.900 vs. 0.885 for no-bridge). That shows robustness at low budgets, but not that bridge training itself yields a 10× reduction in steps for comparable quality.
Questions for authors

Please respond to the weaknesses.

Additional questions:

  • Does Algorithm 1 actually optimize Eq. (11) with the Eq. (3) conditional, or should the theorem refer to the conditional induced by the training pair distribution?
  • What node and edge priors do GBD and the no-bridge baseline use? How does the absorbing “unassigned” initial state fit the full-support prior assumed in Proposition 2?
  • How much performance remains if sizes come only from training data?
Reviewer_JQNzOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The work deals with designing molecules constrained to subgraphs (scaffolds, motifs) using Graph Bridge Diffusion (GBD) approach. The main idea of this approach is to handle the mismatch between the train and the inference gap due to clamping. The work trained a 16M-parameter models using MOSES, 1.9 M drug-like molecules from ZINC databases. The trained model with bridge and without bridge were benchmarked against the chemical language models, and Graph-based approaches on the MolGenBench (scaffold decoration task) and SAFE (fragment-constraint generation tasks). The proposed approach GBD after the preprocessing (restricting bond vaculbulary, bond types) performs better on structural fidelity metrics.

Critical strengths
  1. The preprocessing step (kekulization) demonstrates a substantial improvement in baseline performance, highlighting the importance of molecular representation.
Critical weaknesses
  1. The work lacks robustness and error analysis. The ablation study in Sec 4.4 does not comprehensively evaluate variations in the size or complexity of the constrained subgraph S. Furthermore, p and q are kept constant in this study.
  2. The preprocessing step constraints the molecule design space significantly and hence, the proposed approach (Graph Bridgue Diffusion) may not perform well without any preprocessing step. In that case, the preprocessing is the main reason for the performance of the method.
  3. The proposition 2 assumes the base model recovers the true unconditional distribution. However, the neural network of the denoisers learns approximately distribution and hence, the sampling error is due to the clamping error and model error.
Questions for authors
  1. How does the choice of the preprocessing step affect the performance on datasets with different chemical characteristics? Are there cases where the preprocessing leads to poor results? Table1 notes that graph diffusion without preprocessing fails outright. Could the authors provide more details on the specific failure modes observed and the impact of aromatic bonds on the diffusion process?
  2. The section 3.5 'graph-specific issues' does not elaborate on the specific challenges encountered beyond aromaticity. Is it possible to provide more details on other graph-specific considerations such as acyclic, multiple rings, molecular weight or solubility constraints etc. and how the proposed approach addresses them?
  3. How does the proposed approach's computational efficiency compare to autoregressive models like SAFE-GPT in terms of inference time or resource usage?
  4. How does the function approximation affect the sampling?
  5. How is your work different from the DeFoG and the work in [1]?

[1] Kim, J. H, et al, Discrete Diffusion Schrödinger Bridge Matching for Graph Transformation, ICLR 2025.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.