acceptodds
Under review as a conference paper at ICLR 2027

Inertial Fragment Matching: scalable shape-conditioned molecular generation via descriptor decomposition

Abstract

Prior work has demonstrated that three-dimensional molecular structures can be generated effectively using an equivariant diffusion model conditioned on the ordered principal moments of inertia as a compact and interpretable shape descriptor. While guidance of the generation process by this coarse-grained representation enables high shape similarity for relatively small molecules, its limitations become increasingly apparent for larger molecular structures. We introduce Inertial Fragment Matching (IFM), an inference framework that exploits the additivity and reference-point translatability of the moment-of-inertia (MOI) tensor while using its ordered principal moments as the model-facing conditioning signal. IFM enables fragment-based molecular generation while preserving global shape consistency, substantially improving generation quality and scalability for generation of larger molecules. IFM achieves an average shape similarity of 0.698, corresponding to a 27.8% improvement over the conventional generation workflow. Under fixed-fragment generation conditions, IFM further attains a fragment integration success rate of 0.638, a 73.8% improvement over conventional inpainting-based approaches. Additionally, the impact of IFM on molecular quality, structural fidelity, and inference efficiency is evaluated. We present a proof of concept for context-window-independent molecular generation by combining inertial fragment matching with chain merging, demonstrating the potential to generate linear molecular chains of arbitrary size. Finally we formulate a general transport-additive fragment matching beyond the molecule generation task.

Primary area:
generative models
Keywords:
equivariant diffusion models, molecular generative models, shape-conditioned generation, geometric deep learning

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_SphrOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Inertial Fragment Matching (IFM), an inference framework that generates molecular fragments conditioned on local principal moments of inertia, then assembles and jointly denoises them to match a reference shape. IFM improves average shape similarity and fixed-fragment integration success.

Critical strengths

The core idea is clean and physically motivated. MOI tensors are additive and can be shifted between reference points with the parallel-axis theorem.

Critical weaknesses
  1. The main text is thin and does not provide a self-contained description of the method, leaving essential formulation and implementation details to the appendix. Consequently, the paper reads more like a technical note or workshop submission than a fully developed study.

  2. The underlying model and training data are only briefly described, with substantial reliance on prior work. This makes it difficult to assess reproducibility and distinguish the contribution of IFM from that of fragment-level training augmentation.

  3. The headline shape-similarity gain comes with a large strain cost, while the benchmarks and baselines cover a narrow evaluation setting. The paper lacks a sufficiently broad, fair comparison with other moment-of-inertia-conditioned generation approaches under matched optimization and computational budgets.

Questions for authors
  1. Can the authors expand the main text to include the essential method formulation, inference procedure, and implementation choices? What additional evidence would establish IFM as a broadly applicable framework beyond the current proof of concept?

  2. Can the authors provide fuller details of the model architecture, training procedure, dataset construction, and fragment augmentation? Can an ablation separate the effects of fragment augmentation from those of IFM inference?

  3. How does IFM compare with other moment-of-inertia-conditioned generation approaches under identical geometry optimization and matched computational budgets? Can the evaluation cover more diverse molecular architectures and jointly report chemical validity and inference cost?

2 comments

1. Method formulation and generality

We agree that the original main text was insufficiently self-contained. We have moved the essential IFM formulation into the revised manuscript, including the distinction between the additive full MOI tensor and the principal-moment conditioning vector, rigid-motion transport, fragment recomposition, and the inference procedure.

The revised description covers reference fragmentation and local descriptor computation, storage of fragment rotations and translations, independent fragment generation, rigid placement in the global frame, and late-stage global-MOI-conditioned denoising in relaxed or clamped mode. We also specify the main implementation settings: 100 denoising steps, with merging during the final 10 steps.

We have clarified that the current implementation requires a decomposable reference geometry or externally supplied fragment descriptors and placements; the global principal-moment vector alone does not determine the fragmentation.

Finally, although Section 2.2 formulates a broader transport-additive principle, our experiments establish feasibility only for MOI-conditioned molecular generation. Demonstrating broader applicability would require validation with other generative backbones, descriptors, or object classes. We have therefore revised the Abstract and Conclusion to present generalization as future work rather than an established empirical result.

2. Architecture, training and augmentation

We have expanded Appendix A.3–A.5 and added a summary of the architecture and training procedure to the main text.

All experiments use MLConformerGenerator, an equivariant diffusion model conditioned on ordered principal moments of inertia, with bond prediction performed by AdjMatSeer. The original checkpoint was trained on ChEMBL-derived conformations containing 15–39 heavy atoms. For IFM, we trained an augmented checkpoint using the same architecture and objective, extending the training data with 6–15-heavy-atom fragments obtained by iteratively breaking non-ring single bonds, canonicalizing by SMILES, and removing duplicates. The resulting checkpoint covers 6–39 heavy atoms.

Importantly, all primary comparisons (conventional generation versus relaxed IFM, and inpainting versus clamped IFM) use the same augmented checkpoint. Thus, the observed differences cannot be attributed to unequal fragment augmentation.

We additionally evaluated the effect of augmentation on conventional whole-molecule generation:

MetricOriginal (15–39)Augmented (6–39)Paired Δ (95% CI)
Shape similarity0.5140.517+0.001 [−0.005, +0.007]
Chemical similarity0.1100.108−0.000 [−0.004, +0.003]
Validity0.4590.492+0.033 [+0.001, +0.065]

Shape and chemical similarity remain statistically unchanged, with confidence intervals centred near zero. This suggests that fragment augmentation has little to no effect on conventional generation, while the larger improvements arise from combining the augmented checkpoint with IFM inference.

3. Computational efficiency and molecular architectures

We compare IFM with conventional generation and inpainting using the same MOI-conditioned backbone to isolate the effect of the inference procedure.

On an NVIDIA L4 GPU with 4 CPU cores, generating 100 samples per reference across the full dataset, using 100 diffusion steps (including 10 merging steps for IFM), required:

  • IFM: 31,949.8 seconds.
  • Conventional generation: 30,126.7 seconds.

Thus, IFM introduces a 6.1% increase in total generation time, indicating a relatively modest computational overhead.

Regarding molecular architectures, we consider the diversity of the reference dataset sufficient for the present feasibility study. We also argue that chemical-space coverage is difficult to interpret without a specific application or target distribution. We therefore consider application-driven evaluation of generated chemical space a more meaningful direction for future work.

Thanks for the rebuttal from authors. Although the new clarifications and method introduction are helpful, I think this manuscipt still below the acceptance bar of ICLR. It needs better paper writing with clearer background and related work discussion, more detailed and rigorous introduction of the method, and more comparison with existing baselines. Therefore, I will maintain my score.

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

This paper proposes IFM (Inertial Fragment Matching), an inference framework for shape-conditioned 3D molecular generation that leverages the additivity of the moment of inertia (MOI) tensor and its translation properties across different reference points. Specifically, the framework first decomposes a reference molecule into multiple fragments and generates each fragment independently using a generative model. The generated fragments are then transformed into a global coordinate system and merged under global MOI constraints to form a complete molecule. Experimental results show that IFM achieves certain performance improvements over the selected baselines. In addition, the paper presents a theoretical framework with potential for broader applications, although it has not yet been experimentally validated.

Critical strengths
  1. The method leverages the additivity of the MOI tensor and the parallel axis theorem to decompose global shape constraints into local fragment-level constraints, enabling the generation of complete molecules. This design is supported by clear mathematical motivation and theoretical foundations.
  2. The framework shows potential for extension to different molecular generation tasks and macromolecular design.
Critical weaknesses

1. Insufficient discussion of related work and unclear novelty compared to existing methods. The paper lacks a comprehensive discussion of closely related research, particularly Fragment-Based Molecular Generation, Hierarchical Molecular Generation, and Compositional Diffusion Models. The core workflow of IFM, including fragment decomposition, independent conditional generation, geometric transformation and assembly, and late-stage diffusion merging, shares conceptual similarities with existing approaches. However, the paper mainly discusses whole-molecule MOI-conditioned generation and inpainting methods, without sufficiently clarifying how IFM fundamentally differs from these related approaches. In particular, the unique advantages of leveraging the translational additivity of the MOI tensor over existing fragment generation and assembly strategies remain unclear. The authors are encouraged to provide a more systematic comparison with related work and better clarify the key technical contributions of IFM.

2. Limited geometric validity of generated molecules and substantially reduced performance gains after geometry optimization. Although IFM achieves notable improvements in shape similarity, the geometric validity of its generated molecules remains a significant concern. According to Section 2.1, the average strain energy of IFM-generated molecules reaches 11,452.9 kJ/mol, substantially higher than the 866.9 kJ/mol reported for the conventional method, suggesting that fragment assembly may introduce severe geometric distortions. After MMFF94 geometry optimization, the shape similarity of IFM decreases from 0.698 to 0.575, while its relative improvement over the conventional method drops from 27.8% to only 5.3%. This suggests that the shape-matching advantage of IFM may partially rely on geometrically unrealistic molecular conformations. Moreover, the paper does not report results for the conventional method under the same geometry optimization procedure, making a fair comparison difficult. Therefore, the current experiments do not sufficiently demonstrate that IFM can achieve substantial improvements in shape matching while maintaining reasonable molecular geometry.

3. Lack of key ablation studies to validate the effectiveness of the core design. IFM consists of several key components, including MOI-conditioned fragment generation, rigid-body transformation and assembly, and late-stage joint denoising. However, the paper lacks systematic ablation studies to identify the contribution of each component to the overall performance. In particular, the authors do not experimentally demonstrate the unique benefits of the translational additivity of the MOI tensor compared to general fragment generation and assembly strategies. Furthermore, the paper merges fragments during the final 10% of denoising steps by default, without analyzing the effects of merging timing, fragment size, or fragment count on generation quality. As a result, it remains unclear whether the performance gains primarily arise from the MOI decomposition and recombination mechanism or from other design choices. Additional ablation studies on the core components and key hyperparameters are recommended.

4. Limited baseline coverage and insufficient comparisons with relevant state-of-the-art methods. The experiments mainly compare IFM against conventional whole-molecule MOI-conditioned generation and inpainting-based fixed-fragment generation, resulting in relatively limited baseline coverage. Given the close connections between IFM and existing approaches such as Fragment-Based Molecular Generation, Hierarchical Molecular Generation, and 3D Fragment Linking, comparisons with representative methods from these research directions are missing. In particular, comparing only against inpainting in the fixed-fragment generation task is insufficient to fully establish the competitiveness of IFM. The authors are encouraged to include more representative baselines under comparable experimental settings or explain why direct comparisons are not feasible. The current results demonstrate improvements over the selected baselines but do not sufficiently establish the superiority of IFM over a broader range of existing methods.

Questions for authors

None

2 comments

We thank the reviewer for the constructive assessment and the recognition of IFM's mathematical foundations and potential broader applicability. We address the main concerns below.

1. Related work and novelty

We have expanded the Introduction to discuss related fragment-based and compositional approaches, including HierDiff (Qiang et al., 2023), DecompDiff (Guan et al., 2023), D3FG (Lin et al., 2023), DiffLinker (Igashov et al., 2024), and compositional diffusion models (Du et al., 2023).

We have also clarified our novelty claim: IFM does not introduce fragment decomposition or assembly themselves. Its contribution is exploiting closed-form rigid-motion transport and the additivity of the full MOI tensor to analytically connect local fragment conditions with a global shape condition. Composition therefore occurs in a physically interpretable conditioning space, using the same MOI-conditioned architecture for local generation and global reconciliation.

2. Geometric validity and optimisation

We acknowledge that high pre-optimization strain is an important limitation. The unrelaxed shape similarity of 0.698 demonstrates descriptor-level matching, not physically plausible equilibrium geometries. Fragment merging introduces local distortions, while constrained MMFF94 relaxation reduces strain but partially compromises shape fidelity. The model is trained on ETKDG conformations rather than energy-minimized structures and requires downstream geometry optimization (Sapegin et al., 2025).

We also acknowledge that the reported 5.3% difference is not a matched post-optimization comparison, since the conventional baseline did not undergo equivalent constrained optimization. We have clarified this and no longer claim a definitive post-optimization advantage.

Nevertheless, IFM improves fixed-fragment integration success from 0.367 to 0.638 relative to inpainting. This complementary graph-level improvement demonstrates better preservation of specified molecular substructures, independently of geometric quality. We have revised the manuscript to distinguish shape fidelity, post-relaxation geometry, and fragment preservation, describing unrelaxed outputs as intermediate structures requiring further optimization.

3. Ablation studies

To address fragment-size sensitivity, we performed a post-hoc stratified analysis of the existing benchmark.

Integration and validity are attempt-level metrics, with invalid outputs counted as failures. Shape similarity, chemical similarity and strain are calculated over valid outputs only. Consequently, their denominators differ.

IFM

Fragment sizeNIntegrationShape similarityChemical similarityValidityStrain (kcal/mol)
6–8430.9270.6780.1570.463387
9–111330.8750.6730.2340.481322
12–142010.7620.6830.2840.432346
15–204040.5750.6970.3150.423314

Inpainting

Fragment sizeNIntegrationShape similarityChemical similarityValidityStrain (kcal/mol)
6–8430.5730.5970.1140.401368
9–111330.4000.5540.1440.331506
12–142040.3130.5170.1540.284627
15–204130.3890.4910.1860.238541

Overall, IFM achieves higher integration (+0.18 - 0.47), shape similarity (+0.08 - 0.21), and validity (+0.06–0.18) across all fragment-size bins. Strain is substantially lower for fragments of at least 9 atoms. Chemical similarity increases with fragment size, as expected from greater scaffold overlap.

These findings support IFM's feasibility while highlighting fragment-size sensitivity. Comprehensive hyperparameter analysis remains future work.

4. Baseline selection

We agree that our experiments do not establish superiority over general fragment-based or hierarchical generators.

Our baselines address a controlled question: what is the effect of introducing IFM while keeping the MOI-conditioned backbone, preprocessing, and denoising setup fixed? Whole-molecule generation and inpainting provide corresponding baselines for relaxed and fixed-fragment generation, respectively, isolating the effect of descriptor factorization and fragment merging.

Direct comparison with HierDiff, DecompDiff, D3FG, and DiffLinker is not straightforward because these methods use different conditioning settings and specialized architectures, and generally do not accept global MOI targets. Adapting them would require substantial retraining and method-specific modifications.

We have included these approaches in the related-work discussion and narrowed our claims accordingly. We now present IFM as an extension of an existing MOI-conditioned generator relative to controlled same-backbone baselines, not as generally superior to specialized fragment-generation methods. Broader task-matched benchmarking remains future work.

Reviewer_ShGaOpenReview ↗

Thanks for the Rebuttal

I appreciate the authors’ detailed response and the efforts made to address my concerns. While the additional clarifications and analyses are helpful, I believe that several important issues remain unresolved. Overall, I do not consider the current version of the paper sufficiently convincing for acceptance at ICLR. Therefore, I will maintain my original rating.

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

The paper introduces inertial fragment matching to enable fragment-based molecular generation. In an empirical evaluation, the proposed method is compared to standard inpainting.

Critical strengths

N/A

Critical weaknesses

The paper seems to be in an unfinished state. There is neither a methodology, background, nor a related work section.

  • Many related works are missing.
  • The empirical evaluation is unclear. Neither the setup is explained, nor are the selected baselines sufficient.
  • The methodology section is missing, making it not possible to understand the approach.
Questions for authors

See weaknesses

2 comments

We thank the reviewer for their time to assess our manuscript. While we appreciate the suggestions on clarity and presentation, we respectfully disagree with several points raised, and would like to provide a comment on those.

  1. Background and related work

The Introduction (pp. 1-2) defines the shape-conditioned generation problem, summarizes prior MOI-conditioned molecular generation, explains its limitations for larger molecules, and discusses fragment growth, linker design, and diffusion inpainting. The submitted version cited, among others, Gebauer et al. (2022), Sapegin et al. (2025), Imrie et al. (2020), Igashov et al. (2024), and Schneuing et al. (2022). We nevertheless agree that the positioning relative to fragment-based and hierarchical generation was too brief. We have added a concise discussion of HierDiff, DecompDiff, D3FG, DiffLinker, and compositional diffusion models. We have also narrowed the novelty claim: IFM is not presented as the first fragment-based generation method. Its specific contribution is the use of the closed-form rigid-motion transport and additivity of the full MOI tensor to relate local fragment conditions analytically to a global shape condition.

  1. Methodology

The methodology was present, although too much of it was placed in the appendix. Section 2.1 and Figure 1 (pp. 2-3) describe the complete inference workflow: reference fragmentation, computation of local MOI conditions, storage of fragment rotations and translations, independent fragment generation, rigid placement, and late global-MOI-conditioned denoising in relaxed or clamped mode. Appendix A.1.1-A.1.3 (pp. 8-10) provides the mathematical formulation, the parallel-axis transport equation, the transport-additive descriptor factorization, and the inference procedure. Appendices A.3-A.4 describe the generative model, fragment augmentation, and molecular splitting. To make the paper self-contained, we have moved the essential tensor-recomposition equation and the step-by-step IFM procedure into an explicitly labelled Method section in the main text, while retaining the detailed derivation in the appendix.

  1. Experimental evaluation and baselines

The experimental protocol was specified in Appendix A.5-A.6: benchmark construction, fragmentation criteria, fragment-size ranges, 100-step denoising schedule, final-10-step merging, inpainting resampling, and definitions of shape similarity, chemical similarity, strain, and fragment integration success. We have corrected the omitted benchmark counts: the relaxed and fixed-fragment evaluations contain 993 and 799 reference molecules, respectively. The two comparisons answer controlled same-backbone questions: relaxed IFM versus conventional whole-molecule generation increases average shape similarity from 0.546 to 0.698; clamped IFM versus inpainting increases fragment integration success from 0.367 to 0.638. All compared methods use the same underlying MOI-conditioned architecture, checkpoint, preprocessing pipeline, and denoising setup. This design isolates the effect of IFM inference rather than differences between independently trained architectures. We agree that these experiments do not establish superiority over all fragment-native or hierarchical molecular generators. We have therefore clarified that the paper presents controlled same-backbone evidence, not a broad state-of-the-art claim, and have added the relevant method families to the related-work discussion.

  1. Scientific soundness

The review does not identify a specific mathematical inconsistency, experimental error, or unsupported result that would constitute a soundness problem that is difficult to address. The manuscript explicitly reports the principal limitation of the current implementation: the high strain of unrelaxed assemblies and the reduction in shape similarity after geometry optimization. It also reports failure modes and error accumulation in sequential chain merging.

For these reasons, we respectfully disagree with the characterization of the work as having fundamental soundness concerns. We do agree that the organization of the submitted version made the contribution unnecessarily difficult to assess. We have addressed this by correcting the incomplete presentation elements, moving the essential methodology into the main text, expanding related work, making the experimental setup explicit, and narrowing the claims to what is directly supported by the reported experiments.

I thank the authors for their thorough response. I agree with the other reviewers that the paper still does not meet the acceptance bar for ICLR. Therefore, I keep my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.