acceptodds
Under review as a conference paper at ICLR 2027

FIRM: From Interfaces to Realizations in Fragment-Based Molecular Generation

Abstract

Individually valid fragments still require compatible attachments to form valid molecules. Yet compatible assembly alone does not ensure that generated molecules follow the reference distribution. This motivates learning connection requirements separately from fragment identities, allowing the model to select among compatible fragment structures according to the learned molecular distribution. We introduce FIRM, the Fragment Interface Realization Model, which generates typed connection requirements before selecting compatible fragment structures. FIRM uses two discrete flows. Interface Flow jointly models typed connections with a latent probabilistic circuit, while Conditional Realization Flow jointly refines fragment templates and attachment mappings through discrete denoising within the interface-defined compatible space. On MOSES and NPGen, conditioning on interfaces rather than untyped fragment topology reduces the empirical training-set conditional entropy of fragment templates and attachment mappings by approximately 61% and 49.5%, respectively. Generation-order comparisons show that interface-first generation improves success without structural correction by 3.25 and 2.67 percentage points over mixed-order generation. Additionally, FIRM more closely matches reference molecular feature distributions than the evaluated fragment-graph baselines. This formulation provides a foundation for controllable molecular design, allowing future methods to explore alternative fragment structures under specified connection requirements.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Fragment-Based Molecular Generation, Molecular Graph Generation, Discrete Flow Matching, Typed Molecular Interfaces

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.67 over 3 reviews, above 99% of the others.
Reviewer_6mRAOpenReview ↗
Rating
4
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary
  • The paper proposes FIRM, a fragment-based molecular generative model that decomposes molecular generation into molecular interface (coarse structure) design followed by fragment identity selection. Interface Flow first generates a typed interface, and Conditional Realization Flow then generates a molecule from a fragment bank constructed from the training set.
  • With the modeling strategies described above, FIRM circumvents the practical difficulty of determining fragment-fragment connections and selecting fragments by factorizing the whole process into two steps, which improves performance and aligns well with recent work.
  • For this two-step generative model, the paper analyzes whether the first-stage generation of the interface reduces uncertainty in the second-stage realization, showing each component’s contribution to molecular generation performance. The paper also analyzes the generation order of the components.
  • The paper reports molecular generation results on multiple benchmarks (MOSES, NPGen, GuacaMol, and ZINC250K).
Critical strengths
  • Novelty: The paper proposes a distinctive approach to fragment-based molecular generation. FIRM uses a typed interface that leaves internal fragment adjacency and atom-level attachment mappings unspecified, and separates interface generation from compatible realization rather than generating both jointly in a single process. This factorization is supported by improvements in realization prediction and generation success.
  • Solid evaluation and supporting experiments are well aligned with the main claims: For this two-step generative model, the paper disentangles the informational contributions of different inputs and analyzes generation order, supporting the design choices of FIRM through interface-first, realization-first, and mixed-order comparisons (Section 5.4 and Appendix C.2). Moreover, the paper shows particularly strong improvements on NPGen, which contains large and complex molecules, achieving the best distribution-matching results among the evaluated graph-based molecular generative models. I recommend acceptance because the interface-realization decomposition offers a modeling contribution, supported by controlled analyses and strong benchmark results.
Critical weaknesses

None

Questions for authors
  • New interfaces in NPGen: The proportion of molecules with interfaces not observed during training is considerably higher in the generated samples than in the test set (33.1% versus 6.2%; Figure 8). Could the authors further explain on this difference and briefly show several molecules produced from new interfaces? (Clarification request)
  • Regarding the extensibility of FIRM: I agree that FIRM provides a chemically sound and effective approach to molecular generation. At the same time, an important practical consideration is whether the model can be conditioned (or rather can be optimized) on a target property, through either conditioning or guidance, while maintaining generation quality. Could the authors provide such experiments using FIRM as a base model, in comparison with several baselines when conditioned on a certain property? My question focuses on whether the FIRM framework can serve as a base model when property conditioning is required, so experiments with simple molecular properties would be sufficient. It would also be useful to explore steering the interface generation also stage or analyze how strongly the interface itself influences the chosen molecular properties. (Experiment suggestion)
6 comments
AuthorsOpenReview ↗

Author Response (1/2)

We thank the reviewer for recognizing the interface-realization decomposition and the controlled analyses supporting it, and for the helpful questions about unseen interfaces and property-conditioned generation. Revisions to the manuscript are highlighted in cyan. We address your comments below.

Comment/Question:

"... The proportion of molecules with interfaces not observed during training is considerably higher in the generated samples than in the test set (33.1% versus 6.2%; Figure 8).... Could the authors further explain on this difference and briefly show several molecules produced from new interfaces? ..."

Answer:

Thank you for reading our results carefully and noting the higher unseen-interface rate on NPGen (33.1% versus 6.2%). This observation prompted a closer look at how FIRM forms new interfaces.

We found that 95.4% of generated unseen interfaces use connection structures observed in training, and their local attachment patterns are also familiar. To test whether recombining these patterns could explain the higher rate, we exchanged branches between test interfaces within five common connection structures. This changes how distant parts are combined while preserving the frequencies of local attachment patterns and neighboring pairs.

Unseen-interface rates within five common NPGen connection structures, weighted equally in the mean. The three-node structure allows no branch exchange.

NodesTestAfter exchangeFIRM (direct)
30.47%0.47%0.91%
42.10%3.47%8.09%
54.00%13.51%22.41%
66.87%33.61%44.26%
70.93%2.49%5.11%
Mean2.87%10.71%16.16%

The increase after exchange shows that familiar local patterns can form many more interfaces absent from training. Together with the observed overlap in connection structures, this supports recombination as a contributor to FIRM's higher rate.

To assess the resulting molecules, we compared them with NPGen test molecules whose interfaces were also unseen during training. Their distributions were similar across six properties, including NP-score, LogP, and TPSA (KS distances: 0.036 to 0.083). This supports FIRM's ability to generate new combinations with molecular properties resembling those of real, unseen combinations. The same property analysis appears in Appendix C.4, Figure 16 of the revised manuscript.

Following the reviewer’s suggestion, we randomly selected three molecules containing unseen interfaces from the molecules generated by FIRM trained on NPGen. Their local attachment patterns, connection rules, and fragment templates occur in training, while the complete interfaces and molecules do not. These examples are also included in Appendix D.2, Figure 28 of the revised manuscript.

AuthorsOpenReview ↗

Author Response (2/2)

Comment/Question:

"... Could the authors provide such experiments using FIRM as a base model, in comparison with several baselines when conditioned on a certain property?... My question focuses on whether the FIRM framework can serve as a base model when property conditioning is required, so experiments with simple molecular properties would be sufficient.... It would also be useful to explore steering the interface generation also stage or analyze how strongly the interface itself influences the chosen molecular properties. ..."

Answer:

Thank you for suggesting property conditioning as a practical test of FIRM beyond unconditional generation. Following your suggestion, we added experiments on MOSES for logP, QED, and ring count, with comparisons against FragFM and DiGress using results taken from the original papers. We use the same target values as FragFM, selected based on its analysis of molecular property distributions in MOSES, to compare the models on the same property targets.

To test whether fragment selection alone provides property control, we condition only the realization stage and leave interface generation unconditional. The table below reports three representative targets. For each, realization conditioning reduces the mean absolute deviation (MAD) from the target while maintaining 100% validity and high uniqueness.

Conditional results use second-stage conditioning only. FCD, validity, and uniqueness refer to conditional generation; FCD uses MOSES TestSF.

TargetUnconditioned MAD ↓Conditional MAD ↓FCD ↓Validity ↑Uniqueness ↑
logP = 30.7430.3201.091100%99.96%
QED = 0.60.2040.1202.530100%99.94%
Ring count = 30.5520.0741.484100%99.93%

We compared property error and FCD with FragFM and DiGress under realization-only conditioning. Together with the results above, this comparison supports using FIRM for property-conditioned generation through fragment and attachment selection. The comparison curves are included in Appendix B.6, Figure 9 of the revised manuscript.

To assess the interface's contribution, Table 5 of the revised manuscript compares conditioning neither stage, Interface Flow only, Realization Flow only, or both. Realization-only conditioning is close to conditioning both stages for logP = 3 and QED = 0.8, whereas one-ring generation benefits substantially from conditioning both stages. These findings show that connection requirements and fragment choices can provide complementary control, with the useful combination depending on the target. Section 5.5, Figure 6 of the revised manuscript provides the property distributions, and Appendix B.6 gives the conditional implementation and training details.

Thank you for the detailed experiments.

I appreciate the authors’ additional analysis on NPGen. It strengthens the claim that FIRM generates novel molecules with property distributions similar to the test set, beyond reproducing molecules seen during training.

However, I have some concerns regarding the conditioning experiments. To the best of my knowledge, both DiGress and FragFM used classifier guidance (CG), and FragFM claims that a fragment-based model improves property control under the same guidance scheme. In that regard, I think an important point in the comparison is how well each framework can be conditioned when using the same guidance method.

The current experiments are sufficient to show that FIRM can be used for property conditioned generation, but I do not think they confirm that FIRM’s architecture itself provides better property control. I still think the first claim is also valuable in practice. However, if the paper includes the second claim, the authors should compare FIRM with other diffusion or flow models using CFG, or evaluate FIRM with CG for a more controlled comparison.

Reviewer_UaAWOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces FIRM (Fragment Interface Realization Model), a fragment-based molecular generative framework that separates inter-fragment connection requirements from the selection of concrete fragment structures. FIRM first generates an anonymous typed molecular interface specifying attachment-port types, pairings, and bond types, and then realizes this interface by selecting compatible fragment templates and atom-level attachment mappings. The framework combines an Interface Flow, built upon a latent probabilistic circuit, with a Conditional Realization Flow for discrete denoising within the interface-constrained solution space, followed by deterministic typed assembly. Experiments on MOSES, NPGen, GuacaMol, and ZINC250K show that interface-first generation reduces uncertainty in fragment realization, improves valid assembly, and reproduces reference molecular distributions.

Critical strengths
  • Technical approach. FIRM separates connection requirements from fragment structures using two discrete flows. This design reduces uncertainty in fragment realization and improves realization prediction and molecular assembly when interfaces are generated first.

  • Article presentation. The paper includes informative figures that illustrate the research motivation and methodological details, as well as a range of statistical plots and tables that present the experimental results clearly and comprehensively.

  • Theoretical formulation. The paper provides a clear problem formulation, supported by mathematical expressions that specify the research task and the implementation of the flow models. This presentation helps readers understand the proposed approach and its objectives.

Critical weaknesses
  • The paper’s central motivation is that one-shot joint generation requires the model to navigate a large combinatorial space; FIRM therefore adopts a two-stage paradigm that first fits the interface and then designs the fragments. However, coarse-to-fine, scaffold-first, and fragment-first strategies are not new to molecular generation. JT-VAE first generates a junction tree of fragments and then assembles an atom-level molecule. FragFM first generates a fragment graph and then recovers atom-level connections through candidate-edge scoring and matching. MAGNet first generates an abstract scaffold and its connectivity, then specifies chemical features and attachment positions. Thus, the rationale for a two-stage approach is relatively straightforward, and the paper does not clearly explain how its core contribution differs from these prior methods.

  • In its first-stage interface-fitting flow model, FIRM deliberately discards the detailed internal structures of the fragments. In practical drug discovery, however, some molecular properties depend not only on how fragments are connected but also strongly on specific functional groups, ring structures, stereochemistry, and electronic effects between fragments. If the interface representation is too coarse, the resulting information loss may prevent the flow model from capturing these global dependencies.

  • The experiments show that, compared with conditioning only on untyped fragment topology, adding typed interface information reduces the empirical conditional entropy of fragment templates and connection mappings by approximately 61% on MOSES and 49.5% on NPGen. This demonstrates that interface fitting provides useful information for subsequent fragment design. However, in comparisons with general-purpose joint molecular generation models, FIRM does not show a clear or substantial overall performance advantage. Moreover, its manually imposed two-stage design appears to reduce its ability to generate novel molecules. Taken together, the results do not establish that future molecular generation methods should adopt a two-stage paradigm.

Questions for authors
  • In light of the weaknesses noted above, I encourage the authors to clarify the central conceptual differences between FIRM and prior coarse-to-fine, scaffold-first, and fragment-first approaches, and to articulate more explicitly where the paper’s novelty difference lies.

  • FIRM requires training two discrete flows, constructing resources for compatible fragments and attachment mappings, resolving mapping symmetries, repairing infeasible interfaces, and performing typed assembly. Does this added pipeline complexity affect generation efficiency relative to joint molecular generation models? An empirical analysis of training cost, sampling time, and generation throughput would help clarify this trade-off.

5 comments
AuthorsOpenReview ↗

Author Response (1/4)

We thank the reviewer for the careful assessment, the positive comments on our technical approach, figures, and problem formulation, and the constructive questions about novelty, chemical dependencies, and computational cost. Revisions to the manuscript are highlighted in cyan. We address your comments below.

Comment/Question:

"... However, coarse-to-fine, scaffold-first, and fragment-first strategies are not new to molecular generation.... Thus, the rationale for a two-stage approach is relatively straightforward, and the paper does not clearly explain how its core contribution differs from these prior methods. ... In light of the weaknesses noted above, I encourage the authors to clarify the central conceptual differences between FIRM and prior coarse-to-fine, scaffold-first, and fragment-first approaches, and to articulate more explicitly where the paper's novelty difference lies."

Answer:

Thank you for asking us to clarify the conceptual difference from existing hierarchical methods. JT-VAE and FragFM organize generation around identified fragments, while MAGNet's abstract scaffolds retain internal atom connectivity. Their assembly or reconstruction stages must therefore resolve attachment choices for structures already specified. Choosing valid substructures alone does not settle this problem: individually valid fragments may be incompatible, and even fixed fragments can admit multiple attachment arrangements.

Your observation that joint generation faces a large combinatorial space also captures an important motivation for our work. FIRM addresses these connected difficulties by first specifying connection requirements, then jointly selecting compatible fragments and attachment sites. The requirements exclude incompatible choices and narrow the subsequent generation space, while leaving internal fragment structures open. Our interface analysis supports this role: connection requirements reduce uncertainty over fragment and attachment choices while retaining multiple possible molecular structures.

The contribution is therefore a molecular modeling approach that separates how fragments must connect from which structures fulfill those requirements. Its value goes beyond dividing generation into two stages: the interface constrains assembly without committing to particular fragments or scaffolds, allowing structurally different molecules to share the same connection requirements. We have revised the second paragraph of the Introduction, the contribution statements, and Related Work to clarify how our model differs from existing approaches.

AuthorsOpenReview ↗

Author Response (2/4)

Comment/Question:

"... In practical drug discovery, however, some molecular properties depend not only on how fragments are connected but also strongly on specific functional groups, ring structures, stereochemistry, and electronic effects between fragments. If the interface representation is too coarse, the resulting information loss may prevent the flow model from capturing these global dependencies."

Answer:

Thank you for emphasizing these chemical dependencies. FIRM's interface specifies how fragments should connect, while the subsequent joint selection of fragments and attachment sites determines the chemical details. Fragment structures retain functional groups and rings within their boundaries, together with the stereochemistry represented in the templates. Here, stereochemistry is carried by the fragment representation, while the interface specifies connection requirements. The attachment-mapping equivalence in Appendix A.1 of the revised manuscript is defined on a feature graph that includes bond stereo but does not distinguish atom-chirality or isotope labels; stereoisomer-specific control is therefore a separate modelling objective. Each selection uses the current assignments across the molecule, allowing it to depend on both nearby and distant fragments. This molecular context provides a basis for learning structural associations relevant to inter-fragment electronic effects.

Existing methods incorporate these chemical factors at different levels. Functional groups and rings are represented through the substructures assembled by JT-VAE and FragFM, or through the atom-and-bond patterns learned by DiGress and DeFoG. Stereochemistry calls for additional treatment, such as JT-VAE's enumeration and scoring of stereoisomers after assembly. Electronic effects are generally approached indirectly through learned structural relationships. More specialized approaches introduce targeted priors: D3FG [1] models functional groups as rigid units; D2L-OMP [2] preserves ring substructures and uses electronic substituent constants in fragmentation; and Tom et al. [3] study stereochemistry-aware adaptations of REINVENT and JANUS using SMILES and SELFIES, respectively, together with a GroupSELFIES-based JANUS variant named GroupJANUS. These adaptations explicitly represent stereochemical information during molecular generation.

Each of these four aspects is itself a focus of specialized research, with distinct representations, constraints, and evaluation objectives. A unified model must also account for their interactions, since fragment composition and connectivity influence stereochemical arrangements and electronic behavior. Integrating all four chemical priors therefore remains an important challenge and direction for future work in molecular generation, including extensions of FIRM's joint fragment and attachment selection.

References

[1] Functional-Group-Based Diffusion for Pocket-Specific Molecule Generation and Elaboration.

[2] Diffusing on Two Levels and Optimizing for Multiple Properties: A Novel Approach to Generating Molecules with Desirable Properties.

[3] Stereochemistry-aware string-based molecular generation.

AuthorsOpenReview ↗

Author Response (3/4)

Comment/Question:

"... However, in comparisons with general-purpose joint molecular generation models, FIRM does not show a clear or substantial overall performance advantage.... Moreover, its manually imposed two-stage design appears to reduce its ability to generate novel molecules.... Taken together, the results do not establish that future molecular generation methods should adopt a two-stage paradigm."

Answer:

Thank you for asking how FIRM's performance gains relate to molecular novelty. FIRM was designed from the outset to better capture molecular property distributions. We therefore expanded the comparison to strong autoregressive generators. On NPGen, ChemGPT, nach0, and G2PT achieve FCD values of 0.43, 0.88, and 1.13, respectively, compared with FragFM's 1.34. FIRM further improves FCD to 0.32 while maintaining 100% validity, and better matches all three natural-product class distributions than these added baselines (Section 5.2, Table 2 of the revised manuscript).

Prior studies have reported a trade-off between distributional fidelity and novelty, showing that higher novelty does not necessarily imply better generation quality [1, 2, 3]. Novelty should therefore be considered alongside distribution matching, as models strike different balances between the two. The radar plots in Appendix C.6, Figure 23 of the revised manuscript illustrate this balance, with FIRM achieving closer distribution matching than the graph baselines and comparable internal diversity. To examine whether this quality extends to novel molecules, we removed all exact training matches: FCD remains 0.385, versus 0.319 before filtering, with logP and TPSA distributions still close to the reference (Appendix C.6, Figures 21–22 of the revised manuscript). We further examined whether property distributions remain faithful when the connection patterns themselves are unseen. Molecules generated from unseen interfaces closely match test molecules with unseen interfaces across six property distributions (Appendix C.4, Figure 16 of the revised manuscript).

To understand how fragment dependencies help achieve this goal, Table 4(b) in Section 5.4 of the revised manuscript compares fragment-selection models within the same two-stage framework on MOSES. Incorporating other fragments' states improves FCD from 0.638 to 0.537 and scaffold similarity from 11.8 to 13.4, while novelty changes from 91.3% to 88.6%. These results support jointly learning fragment combinations and illustrate the same trade-off: accounting for fragment dependencies favors distributional fidelity rather than maximizing novelty alone. Because MOSES TestSF contains scaffolds absent from training, its scaffold metric also tests generalization. FIRM achieves the highest scaffold similarity among the compared fragment-based models, improving both scaffold similarity and molecular novelty over FragFM (Section 5.2, Table 1 of the revised manuscript).

Together, the results on unseen connection patterns and held-out scaffolds show that FIRM's original goal of better property-distribution learning extends beyond combinations observed during training.

References

[1] Masked graph modeling for molecule generation.

[2] De Novo Molecular Generation via Connection-aware Motif Mining.

[3] Data-Efficient Molecular Generation with Hierarchical Textual Inversion.

Reviewer_9QBTOpenReview ↗
Rating
4(+2)
Confidence
4
Informativeness
2
Presentation
2(+1)
Soundness
2(+1)
Read review
Paper summary

The authors propose a FIRM - a hierarchical molecular generation approach, that decomposes the molecular graph generation into two steps - (i) drafting fragment interfaces and (ii) consequent fragments generation that satisfy the drafted interfaces. Authors implement this idea as a two-staged diffusion generation model FIRM and test it on the molecular distribution learning task across several datasets. The whole idea of hierarchical molecular generation is not novel and well explored in the field, including the baselines from 2018 (JTVAE) and 2021 (MolGrow). Still I find the particular implementation of this hierarchical approach as an intermediate generation of interfaces quite original and elegant.

The paper is written quite overcomplicated for this simple and elegant idea. I would recommend authors to simplify method description and make it more intuitive and straightforward. Also, I would recommend to use less models' metrics values in the Experiment sections and provide the findings, analysis and conclusions in more intuitive way without mentioning low-level details like metrics values.

The main weakness of the paper is experiments. Given the current state of the molecular generation field where the majority of modern approaches allow conditional molecular generation and work with 3D instead of molecular graph, the distribution learning task on molecular graph is just a proof-of-the-concept. From my understanding of the approach, nothing from a theoretical side prevents this approach to being adapted for both conditional setup and moving from 2d graph to 3d structure.

What actually prevents this is the chosen diffusion model approach for implementing this idea. Conditioning is not that simple task in diffusion models, it requires non-trivial technical implementation including sometimes training, balancing and guidance from property predictors. In this case the diffusion approach is separated into 2 stages and it requires correspondingly more efforts to make the whole generation framework conditioned, like 2 sets of predictors for each stage of generation and non-trivial joint fragment generation since fragments can no more be independently generated when there is a condition.

Since the whole idea is not limited to a diffusion approach I would recommend authors to consider in the future research testing this interface drafting in other generative approaches more suitable for scaling molecular generation to more complicated data and conditions. It seems that interface drafting and fragment implementation fits perfectly to language models reasoning before giving the final composed molecule as an answer. Speaking of language models, the experiment focus mainly on graph generators and miss huge field of autoregressive molecular generation including ones working directly with molecular string representation, like some language models [a].

Given the authors stated training resources [Appendix B.1], I'm not asking the approach to be scaled in terms of model size and computational resources. The mentioned conditional setup requires engineering efforts, not more computational resources. The mentioned proposal on testing the approach in language models also doesn’t require drastically more resources, several millions parameters language models show good results on molecular string generation [a].

Overall, I find the proposed hierarchical approach quite original and having high potential. But in the current state of work, the experiment doesn’t prove its wide applicability for more advanced molecular generation setups and the current implementation as a 2-stage diffusion model also makes this application more difficult to be implemented.

a. nach0: Multimodal Natural and Chemical Languages Foundation Model

Critical strengths

The (i) drafting fragment interfaces and (ii) consequent fragments generation looks pretty original and having high potential.

Critical weaknesses

The current experiments don't support the wide applicability of the approach.

Questions for authors

The answers for stated weaknesses, including new conditional setup results, could improve my score.

6 comments
AuthorsOpenReview ↗

Author Response (1/4)

We thank the reviewer for recognizing the originality and potential of interface-first generation, and for the suggestions to improve clarity, test conditional generation, and explore autoregressive models. Revisions to the manuscript are highlighted in cyan. We address your comments below.

Comment/Question:

"... I would recommend authors to simplify method description and make it more intuitive and straightforward.... Also, I would recommend to use less models' metrics values in the Experiment sections and provide the findings, analysis and conclusions in more intuitive way without mentioning low-level details like metrics values."

Answer:

Thank you for recognizing the simplicity and elegance of the idea and for suggesting how to communicate it more clearly. We have revised Sections 3 and 4 to explain FIRM more clearly with Figure 2: the first stage specifies how fragments should connect, and the second jointly selects compatible fragments and attachment sites to assemble a molecule. We have moved the more detailed mathematical derivations and implementation details to Appendices A.1–A.7, keeping the main text focused on the formulation and generation process.

We have also organized Section 5 of the revised manuscript around the questions being tested, explaining what each comparison shows and how it supports the conclusions. For example, the generation-order comparison in Section 5.3, Figure 3 addresses why connection requirements are generated before fragment selection, with the detailed protocol in Appendix C.2. The text retains the numerical comparisons needed to support these findings, with the full results available in tables and figures.

AuthorsOpenReview ↗

Author Response (2/4)

Comment/Question:

"... In this case the diffusion approach is separated into 2 stages and it requires correspondingly more efforts to make the whole generation framework conditioned, like 2 sets of predictors for each stage of generation and non-trivial joint fragment generation since fragments can no more be independently generated when there is a condition."

Answer:

Thank you for raising the practical challenges of conditional generation in a two-stage model. You note that conditioning the full framework could require a separate set of predictors for each stage, together with joint fragment generation because fragments become dependent under a molecular property condition. We address these concerns with a simple approach: provide the property target directly to the generation networks, using classifier-free guidance [1] with discrete guidance informed by DeFoG [2]. No separate property classifier or regressor is needed. The fragment generator already considers all current fragment assignments jointly, allowing their choices to depend on one another and on the shared target. Section 4.2 of the revised manuscript describes the shared molecular context, and Appendix B.6 specifies property encoding, training, stage control, and guidance.

The conditioning-stage results and remaining discussion for Response 2 follow in the next comment.

References

[1] Classifier-Free Diffusion Guidance.

[2] DeFoG: Discrete Flow Matching for Graph Generation.

AuthorsOpenReview ↗

Author Response (2/4)

Comment/Question:

"... Given the current state of the molecular generation field where the majority of modern approaches allow conditional molecular generation and work with 3D instead of molecular graph, the distribution learning task on molecular graph is just a proof-of-the-concept. ... ... The answers for stated weaknesses, including new conditional setup results, could improve my score."

Answer:

Your example of separate predictors for the two stages also raises the question of whether both stages need to be conditioned. We therefore conducted a full ablation over all four combinations of conditioning the two stages on MOSES. We use the same target values as FragFM, selected based on its analysis of molecular property distributions in MOSES, to compare the models on the same property targets: logP = 1, 3, and 5; QED = 0.6, 0.8, and 1; and ring count = 1, 3, and 5. The table below summarizes three representative targets: logP = 3 illustrates control through fragment selection, one ring illustrates the complementary contribution of connection structure, and QED = 0.8 illustrates a case where conditioning fragment selection alone performs comparably to conditioning both stages.

MAD measures the mean absolute error from the target; FCD measures the difference from the MOSES TestSF reference distribution. Validity and Unique are percentages; IntDiv measures internal diversity. Stage 1 generates the connection structure; Stage 2 selects fragments.

TargetConditioningMAD ↓FCD ↓Validity ↑Unique ↑IntDiv ↑
logP = 3None0.7430.629100.099.930.8565
logP = 3Stage 1 only0.6360.742100.099.970.8549
logP = 3Stage 2 only0.3201.091100.099.960.8557
logP = 3Both stages0.3041.176100.099.910.8547
Ring count = 1None1.6270.526100.099.990.8556
Ring count = 1Stage 1 only0.5639.766100.099.730.8373
Ring count = 1Stage 2 only1.0493.618100.099.600.8569
Ring count = 1Both stages0.05515.895100.099.430.8300
QED = 0.8None0.0620.675100.099.950.8569
QED = 0.8Stage 1 only0.0620.708100.099.950.8566
QED = 0.8Stage 2 only0.0550.756100.099.960.8572
QED = 0.8Both stages0.0560.775100.099.910.8572

For the three representative targets in the table above, conditioning only fragment selection reduces the mean absolute error relative to the same model with property conditioning disabled. The ablation further shows that the two-stage design provides flexibility in where to introduce control: fragment selection already provides effective control for logP = 3 and QED = 0.8, while conditioning both stages substantially improves one-ring generation. Thus, FIRM supports property-directed generation without separate property predictors, and its two stages can contribute complementary control according to the target.

The property-distribution analysis complements the ablation by showing how the generated distributions respond to different targets. With both stages conditioned, logP and ring-count distributions shift toward the requested values, while higher QED targets shift the generated molecules toward higher QED. These distributional responses show that FIRM can steer molecular properties across multiple targets within the same two-stage generation framework. The distributions also appear in Section 5.5, Figure 6 of the revised manuscript.

We appreciate your suggestion to consider 3D generation. Given the scope of the present work, extending this design to 3D would involve additional modelling of fragment conformations and their relative positions and orientations, together with geometric compatibility at attachment sites and steric interactions between fragments. Jointly modelling these geometric factors and fragment connectivity is a natural direction for conditional 3D generation and one of the future applications we envision for FIRM.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Updates

We thank all three reviewers for their careful reading and constructive feedback. We also appreciate the positive comments on FIRM's design and potential, the experiments examining its design choices, and its performance on complex molecules.

In response to the reviewers' questions, we added experiments and clarified the method and interpretation of the results. We summarize these updates below to make our responses easier to revisit.

Additional experiments and analyses

  • Unseen connection patterns (6mRA). To explain their higher frequency in generated samples, we examined their overlap with familiar connection structures and introduced a branch-exchange experiment. The results support recombination of familiar local patterns as a contributor. We also compared the resulting molecules with test molecules across six properties and added molecular examples (Appendix C.4; Appendix D.2).

  • Broader comparisons and the fidelity–novelty trade-off (UaAW, 9QBT). We added MolGPT, ChemGPT, nach0, and G2PT, including adapting autoregressive baselines to NPGen. We examined distribution matching alongside molecular and scaffold novelty and internal diversity, clarifying how models balance fidelity and novelty and how joint fragment selection affects this balance. Removing exact training matches from NPGen samples preserves close distribution matching, supporting generation quality beyond training-set reproduction (Sections 5.2 and 5.4; Appendix C.6).

  • Property-conditioned generation (6mRA, 9QBT). We added experiments on logP, QED, and ring count, including all four combinations of conditioning neither stage, either stage, or both. The results show effective control through fragment selection and complementary contributions from connection generation for some targets. Further CG and CFG experiments demonstrate compatibility with both guidance methods (Section 5.5; Appendix B.6).

  • Computational efficiency (UaAW). We compared training cost, sampling time, and valid-molecule throughput with FragFM, DiGress, and DeFoG. FIRM combines lower training cost and faster sampling at the settings used for the reported generation results, including the cost of structural correction (Section 5.4, Figure 5).

  • Transfer to autoregressive generation (9QBT). We implemented the design in MolGPT. Generating connection patterns before selecting fragments improves validity and distribution matching, while analysis of branching patterns shows closer agreement with reference molecular structures. This tests the design beyond its original flow-based implementation (Section 5.6; Appendix B.7).

Methodological clarifications and presentation

  • Contribution and chemical dependencies (UaAW). We clarified the distinction between specifying connection requirements and committing to particular fragments or scaffolds. We also explained how joint fragment and attachment selection uses molecular context, discussed functional groups, rings, stereochemistry, and electronic effects, and identified directions for incorporating additional chemical priors (Introduction, Related Work, Section 4.2, and Conclusion).

  • Interpretation of the results (6mRA, UaAW, 9QBT). We clarified the balance between distribution matching and novelty and the role of fragment dependencies. For property conditioning, our focus is the first claim identified by Reviewer 6mRA: FIRM can be used for effective property-conditioned generation. We explained how CFG supplies targets directly to the generation networks without separate property predictors, and discussed conditional 3D generation as a future extension (Sections 5.4 and 5.5; Appendix B.6; Conclusion).

  • Readability (9QBT). We revised the method description around the generation process and Figure 2, moved detailed derivations and implementation material to the appendix, and organized the experimental discussion around the questions tested and the conclusions supported (Sections 3 to 5; Appendix A).

Changes are highlighted in cyan in the revised manuscript. Our individual responses provide the detailed results and explanations. We would appreciate your feedback on these additions and welcome any remaining questions.

Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.