acceptodds
Under review as a conference paper at ICLR 2027

FlexMOF: Joint Composition and All-Atom Coordinate Generation for Metal Organic Frameworks

10 views

Abstract

Metal–organic frameworks (MOFs) are porous crystals assembled from metal nodes and organic linkers, offering tunable chemistry and pore structures for gas storage, carbon capture, and catalysis. Their vast design space presents an opportunity for AI-assisted discovery, but useful generation requires more than selecting building blocks: these building blocks must assemble into a periodic crystal with appropriate atomic conformations and pore geometry. Existing approaches often restrict the internal geometry of building blocks or learn composition independently of structure, limiting the connection between composition generation and its three-dimensional realization. We introduce FlexMOF, a unified model for joint composition and all-atom structure generation. Its coordinate diffusion process incorporates molecular topology and periodic interactions while allowing every atomic coordinate and the lattice to evolve. Building on this structural representation, joint training uses a composition encoder shared by the composition and structure pathways, together with latent features, to propagate coordinate supervision into the graph representation used for composition generation. This enables composition learning to benefit directly from modeling three-dimensional structures. FlexMOF achieves significantly better single-sample crystal structure recovery than the reported state-of-the-art all-atom baseline. Joint composition and structure generation also yields chemically and structurally valid, novel, and unique MOF candidates. By connecting flexible all-atom modeling with composition learning, FlexMOF represents an important step toward AI-assisted MOF discovery through the unified design of composition and all-atom periodic structure.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Metal–organic frameworks, Crystal structure generation, Diffusion models

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_D4nbOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces FlexMOF, whose main contributions and scientific significance lie in two aspects. First, and most importantly, it integrates building-block generation with all-atom diffusion, substantially increasing the flexibility of MOF generation compared with previous approaches. The joint training strategy is also well-motivated, as the generation of building blocks should account for the constraints imposed by periodic crystal structures. Second, FlexMOF introduces an attention architecture operating at the atom, block, and cell levels. Overall, I believe this work has sufficient scientific merit, as all-atom MOF generation is undoubtedly an important task in materials science. Nevertheless, several closely related works deserve attention. In particular, Mofasa has already demonstrated all-atom MOF generation, while Crystalformer has explored periodic attention mechanisms (which the authors do not appear to cite), with a formulation highly similar to the proposed cell-level attention. Despite these concerns, given the strong experimental results, I still consider this work worthy of acceptance. However, I believe the authors should address several important questions, particularly regarding the practical generalizability of the proposed method and the actual benefits of joint training for generation.

Critical strengths
  1. Unified framework. The paper unifies crystal structure prediction and de novo MOF generation within a coherent framework, providing a clear and well-motivated approach to both tasks.
  2. Joint composition–structure learning. The proposed method integrates building-block generation with all-atom coordinate diffusion through joint training, effectively improving generation quality.
  3. Efficient attention design. The attention architecture incorporates atom-level, building-block-level, and periodic cell-level interactions, enabling efficient modeling of relatively large MOF systems.
  4. Comprehensive experiments. The experimental evaluation is thorough and well-designed, covering not only structure recovery and generation quality but also important MOF properties, particularly pore geometry.
  5. Strong empirical performance. FlexMOF demonstrates substantial improvements over existing baselines across multiple evaluation metrics, providing convincing evidence of its effectiveness.
Critical weaknesses
  1. Unclear benefits of joint training for composition generation. My primary concern is how joint training actually changes the learned representations and generation behavior of the building-block autoencoder. Given the authors' emphasis on periodicity, I would like to understand what limitations arise when building blocks are generated without structural supervision. For example, are independently generated building blocks less likely to assemble into valid periodic MOF structures?
  2. Limited evaluation of generalizability. The experiments are primarily conducted on the hypothetical BW dataset, while more realistic MOF datasets, such as CoRE MOF, are also important. It remains unclear whether FlexMOF can effectively perform all-atom training and inference on these datasets, particularly given the larger size and structural complexity of real MOFs.
  3. Insufficient discussion of related work. The paper does not appear to cite Crystalformer, which introduces a periodic attention mechanism that is mathematically very similar to the proposed cell-level attention. The authors should clarify the methodological differences and the novelty of their design.
  4. Lack of computational efficiency analysis. The paper does not provide a systematic analysis of computational cost or scalability, and the experiments appear to focus on MOFs of limited size. Since MOFs can be substantially more complex than conventional crystals, the computational feasibility of all-atom generation for larger and more complicated frameworks remains an important concern.
Questions for authors

see Weakness. If the authors can provide an interpretable analysis of the actual benefits of joint training (rather than merely reporting improvements in final performance) and demonstrate the efficiency and effectiveness of their model on real-world MOF datasets, I would be willing to raise my score.

7 comments

Response to Reviewer D4nb

We thank the reviewer for the careful assessment and constructive suggestions on joint learning, experimental-MOF applicability, related work, and scalability. We hope the reviewer will carefully consider the additional evidence and clarifications below when reassessing our submission.

Weakness 1: Explain how structural supervision improves building-block chemistry and assembly

We have conducted additional experiments to investigate how structural supervision improves composition generation, including a controlled comparison using a fixed structure predictor and an analysis of generated chemical graphs.

1. Improved ability to form valid periodic MOF structures. We compare independently trained and jointly trained composition generators using the same fixed structure-predictor weights and 10,000 generation requests. As reported in Section 4.5, Table 3 of the revised manuscript, the results are:

\hline \textbf{Composition generator} & \textbf{Validity (\\%) }\uparrow & \textbf{EVNU (\\%) }\uparrow \\\\ \hline \text{Independently trained} & \text{56.79} & \text{49.61} \\\\ \text{Jointly trained (FlexMOF)} & \textbf{79.82} & \textbf{69.73} \\\\ \text{Improvement} & \textbf{+23.03} & \textbf{+20.12} \\\\ \hline \end{array}$$ These substantial improvements indicate that **compositions learned with structural supervision are more suitable for generating valid periodic MOF structures**, rather than the improvements arising solely from changes in the structure-prediction model. **2. Improved chemical connectivity of generated building blocks.** We further analyze the chemical graphs produced by the two composition generators and their resulting structures: $$\begin{array}{lcc} \hline \begin{array}{c}\textbf{Chemical consistency} \\\\ \textbf{issue}\end{array} & \begin{array}{c}\textbf{Separate training} \\\\ \textbf{(\\%) }\downarrow\end{array} & \begin{array}{c}\textbf{Joint training} \\\\ \textbf{(\\%) }\downarrow\end{array} \\\\ \hline \text{Disconnected linker graphs} & \text{3.64} & \textbf{2.60} \\\\ \begin{array}{c}\text{Carbon atoms with more} \\\\ \text{than four graph neighbors}\end{array} & \text{6.65} & \textbf{4.41} \\\\ \begin{array}{c}\text{Structures flagged for} \\\\ \text{carbon undercoordination}\end{array} & \text{22.48} & \textbf{11.33} \\\\ \hline \end{array}$$ Carbon undercoordination refers to **carbon atoms having insufficient bonding neighbors for their expected chemical environment**, potentially indicating missing bonds, missing hydrogen atoms, or incorrect local geometry. Unlike the first two metrics, which directly assess generated chemical graphs, carbon undercoordination is detected by MOFChecker from the resulting 3D structures. These results reveal concrete limitations of composition generation without structural supervision: independently trained generators produce more disconnected or chemically implausible building-block graphs, and their resulting structures exhibit more coordination defects. In FlexMOF, coordinate-denoising supervision influences the shared composition representations and graph-posterior latents during joint training (**Section 3.4 and Appendix E.1**). This connects chemical graph learning to the geometric constraints of periodic structure formation. **Together, these results provide empirical evidence that structural supervision improves both the chemical consistency of generated building blocks and their compatibility with valid periodic MOF structures.**

Weakness 2: Demonstrate training and generation on complex experimental MOFs

We have conducted preliminary experiments to show that FlexMOF can be trained on experimental MOF structures and support de novo generation across a broad range of system sizes.

We constructed an experimental MOF training dataset comprising 25,683 structures from publicly accessible sources, including QMOF, the computation-ready subset of CoRE-MOF-2024 SI, the CSD-modified computation-ready CoRE-MOF structures, and ARC-MOF DB12/DB14 structures. Some additional CSD-derived structures require licensed access and are not included in our current dataset.

We use Mofasa as the primary literature reference for experimental-MOF generation because it provides the closest available comparison in task scope and training-data provenance. Its experimental model generates complete all-atom MOF structures and is trained on a mixture of QMOF, computation-ready CoRE-MOF-2024, and ARC-MOF structures, closely corresponding to the sources used in our experimental dataset. Moreover, its evaluation extends to 500 atoms per unit cell and reports validity rates, directly addressing the reviewer’s concerns about realistic structural complexity and generative effectiveness.

Our experimental training set is considerably smaller than the hypothetical BW dataset and also smaller than the approximately 49,000 experimental structures used to train Mofasa. We jointly train FlexMOF's composition and all-atom structure generation modules on our experimental training set, using the same general modeling framework without introducing experimental-dataset-specific architectural modifications.

We sample 10,000 MOFs from the model trained with the experimental MOF dataset. We evaluate generation quality by grouping outputs according to their actual atom counts, including hydrogen, and reporting the validity rates in each interval. Mofasa reference values are approximate estimates obtained by interpolating its published unconditional-generation curve in Figure 7 (without geometry relaxation) of its paper at the atom counts of our samples and averaging within each interval.

\hline \textbf{Atoms per cell} & \begin{array}{c}\textbf{FlexMOF} \\\\ \textbf{validity (\\%) }\uparrow\end{array} & \begin{array}{c}\textbf{Mofasa} \\\\ \textbf{validity (\\%)}\end{array} \\\\ \hline \text{10–100} & \textbf{59.03 }\text{(2,658/4,503)} & \text{58.29} \\\\ \text{101–200} & \textbf{51.35}\text{ (1,616/3,147)} & \text{49.28} \\\\ \text{201–300} & \textbf{29.82 }\text{(339/1,137)} & \text{29.79} \\\\ \text{301–400} & \textbf{24.09 }\text{(159/660)} & \text{18.20} \\\\ \text{401–500} & \textbf{16.19 }\text{(40/247)} & \text{7.72} \\\\ \text{>500} & \textbf{10.73 }\text{(31/289)} & \text{—} \\\\ \hline \end{array}$$ FlexMOF’s observed validity is comparable to or higher than the approximate Mofasa reference values in most evaluated size intervals. These results are encouraging for larger frameworks and demonstrate that FlexMOF can generate valid structures containing more than 500 atoms per unit cell. However, differences in training subsets (we will collect more experimental MOF data in the future) and conditioning protocols mean that this comparison provides context rather than a controlled ranking. These preliminary experiments support the applicability of FlexMOF to experimental MOFs across a broad size range. Improving generation quality and diversity, particularly for larger frameworks, remains an important direction for further work.

Weakness 3: Clarify the relationship to Crystalformer and the methodological contribution

We agree that the core log-sum-exp formulation of our source-atom aggregation is mathematically equivalent to the periodic-image attention reduction in Crystalformer. We have explicitly acknowledged and discussed this connection in Sections 2 and 3.2 of the revised manuscript, rather than claiming the aggregation formula itself as a new attention mechanism.

However, our methodological contribution lies in extending this aggregation to chemically conditioned, dynamic periodic structure generation, which introduces two important distinctions:

Dynamic periodic geometry modeling. Unlike Crystalformer, which encodes given periodic structures for property prediction, FlexMOF jointly denoises atomic coordinates and unit-cell geometry in Cartesian space. As both evolve, the periodic multigraph and its neighboring periodic images are dynamically reconstructed at each denoising step. While attention contributions from images of the same source atom are aggregated, their individual relative displacement vectors are retained to guide atomic and lattice-coordinate updates. This enables efficient periodic attention within a generative process where the underlying periodic geometry continuously changes (Sections 3.1–3.2 and Appendix D.2).

Chemical graph conditioning. FlexMOF further integrates chemical connectivity information from the molecular graphs of MOF building blocks, including linker bond connectivity, into the attention computation. Specifically, Graphormer-style shortest-path and bond-path biases provide chemical context alongside the geometric interactions from periodic images. This allows the model to jointly consider molecular connectivity and spatial periodic interactions when predicting atomic and lattice-coordinate updates (Sections 3.1–3.2 and Appendix B.3).

These distinctions clarify that our contribution is not a new periodic-attention aggregation formula, but its extension to chemically conditioned, dynamically evolving all-atom periodic structure generation. We have revised the related-work discussion accordingly.

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

This paper proposes FlexMOF for joint composition and all-atom structure generation of metal–organic frameworks. It connects graph reconstruction, composition generation, and coordinate denoising through a shared encoder, and aggregates periodic-image interactions during structure generation. Experiments cover structure recovery, pore geometry, complete MOF generation, and MOF–adsorbate configurations.

Critical strengths
  1. The framework models atomic coordinates and lattice geometry flexibly while incorporating molecular connectivity and periodic interactions.

  2. The experiments cover several relevant tasks and include topology and periodicity ablations. The paired pore-property evaluation provides evidence of improved geometric accuracy.

Critical weaknesses
  1. The novelty of the periodic-image aggregation is limited, as its core formulation is closely related to Crystalformer. The paper does not discuss this connection or sufficiently establish the additional methodological contribution.

  2. Relevant baselines are unevenly covered across experiments, and several key comparisons use different evaluation populations or postprocessing protocols. These differences limit the interpretation of the reported performance gains (Figure 3; Tables 2–3).

  3. Joint training improves complete-generation validity, but the experiments do not isolate the contributions of composition generation and structure prediction. Comparing the composition generators with a fixed structure predictor would more directly test whether structural supervision improves composition learning.

Questions for authors

See above

4 comments

Response to Reviewer zUFm

We thank the reviewer for the careful assessment and constructive suggestions on methodological novelty, evaluation fairness, and the benefits of joint training. We hope the reviewer will carefully consider the additional evidence and clarifications below when reassessing our submission.

Weakness 1: Distinguish the periodic-attention contribution from Crystalformer

We agree that the core log-sum-exp formulation of our source-atom aggregation is mathematically equivalent to the periodic-image attention reduction in Crystalformer. We have explicitly acknowledged and discussed this connection in Sections 2 and 3.2 of the revised manuscript, rather than claiming the aggregation formula itself as a new attention mechanism.

However, our methodological contribution lies in extending this aggregation to chemically conditioned, dynamic periodic structure generation, which introduces two important distinctions:

Dynamic periodic geometry modeling. Unlike Crystalformer, which encodes given periodic structures for property prediction, FlexMOF jointly denoises atomic coordinates and unit-cell geometry in Cartesian space. As both evolve, the periodic multigraph and its neighboring periodic images are dynamically reconstructed at each denoising step. While attention contributions from images of the same source atom are aggregated, their individual relative displacement vectors are retained to guide atomic and lattice-coordinate updates. This enables efficient periodic attention within a generative process where the underlying periodic geometry continuously changes (Sections 3.1–3.2 and Appendix D.2).

Chemical graph conditioning. FlexMOF further integrates chemical connectivity information from the molecular graphs of MOF building blocks, including linker bond connectivity, into the attention computation. Specifically, Graphormer-style shortest-path and bond-path biases provide chemical context alongside the geometric interactions from periodic images. This allows the model to jointly consider molecular connectivity and spatial periodic interactions when predicting atomic and lattice-coordinate updates (Sections 3.1–3.2 and Appendix B.3).

Importantly, we also provide a controlled experiment to quantify the practical contribution of our efficient implementation. We compare source-atom aggregation with explicitly expanded periodic-image attention using identical model weights, periodic neighborhoods, and random seeds. Both approaches achieve nearly identical structural recovery (70.57% versus 70.48% weighted MR@5\mathrm{MR}@5), while our implementation achieves 15.09–30.23$\times$ faster sampling across different MOF sizes, with peak GPU memory reduced from 6.04 to 2.81 GiB for the largest structures (Section 4.6 and Appendix F.5, Table 9).

Therefore, we do not claim novelty in the log-sum-exp aggregation identity itself. Instead, our contribution is extending this principle to dynamically evolving periodic structures with explicit molecular graph conditioning, enabling efficient and chemically informed all-atom MOF generation.

Weakness 2: Align baseline coverage, evaluation populations, and postprocessing

We agree that differences in evaluation populations or postprocessing can mislead comparisons. We revised our experiments to align the populations and evaluation criteria for our key comparisons.

Baseline coverage reflects task scope and available results. The original MOFFlow and AtomMOF-L require supplied compositions and cannot generate them, so they cannot be evaluated on complete de novo design. Table 3 includes Mofasa's published de novo results. Mofasa reports no Boyd–Woo structure-prediction match rate, and we could not locate public inference code or model weights to evaluate it on the same inputs.

1. Structure prediction (revised Figure 4). We identified that, in their original paper, the reported AtomMOF-L match rates were underestimated due to a unit-cell orientation error in its evaluation code, which has recently been acknowledged and corrected by the AtomMOF authors in their Github repository v1.1.0 release two weeks ago (https://github.com/nayoung10/AtomMOF/releases/tag/v1.1.0).

We re-evaluated FlexMOF, AtomMOF-L, and MOFFlow-2 on the same 19,792 Boyd–Woo test structures, with identical matching criteria and the unit-cell orientation correction in AtomMOF-L's evaluation script. MR@1 is 53.49% for FlexMOF, 40.01% for AtomMOF-L, and 28.40% for MOFFlow-2. MR@5 is 73.12% for FlexMOF and 61.46% for AtomMOF-L. Our advantage also persists under a stricter structure-matching criterion, with lattice-length tolerance of 0.3, site tolerance of 0.3, and angle tolerance of 10∘10^\circ. Under this setting, FlexMOF achieves 43.32% MR@5\mathrm{MR}@5, compared with the corrected 28.97% reported for AtomMOF-L and 16.0% reported for MOFFlow-2, corresponding to differences of 14.35 and 27.32 percentage points, respectively. All these results use the same full test population (Figure 4(c), Appendix F.1, Table 5, Figure 6, Table 6).

2. Pore properties and MOF–adsorbate evaluation (revised Tables 1–2). Pore-property errors use each method's first candidate on the same property-specific sets, including matched and unmatched predictions, without relaxation. FlexMOF lowers MAE for all five pore descriptors by 17.5–24.4% relative to AtomMOF-L (Table 1, Appendix F.3). For adsorption configuration coverage, we reproduce AtomMOF-L on the same 683 systems and 3,582 reference configurations, with identical candidate counts, matching criteria, and aggregation. At ten candidates, COV-R is 26.07% for FlexMOF and 22.52% for AtomMOF-L (Table 2, Appendix G.2). Literature-derived adsorption-energy results are explicitly sourced and use aligned postprocessing (Appendix G.3).

3. De novo MOF generation (Appendix F.4). MOFFlow-2 and MOF-BFN use validity criteria different from Table 3, so their published results and protocols are reported in Appendix F.4. MOFFlow-2's 16 MOFChecker checks add porosity to the 15 checks in Table 3. Under this validity criterion, FlexMOF achieves 68.14% validity and 59.19% VNU, versus the reported 38.84% and 31.35% for MOFFlow-2. Both use 10,000 unrelaxed samples, with VNU following each method's own uniqueness convention. MOF-BFN reports 32.3% overall validity on 1,000 unrelaxed samples, requiring both MOFChecker checks and connection-point matching. Using only the 15 checks in its official MOFChecker 0.9.6 script gives 69.29% (6,929/10,000) for FlexMOF and 32.5% for MOF-BFN's public samples. This re-evaluation excludes connection-point matching because FlexMOF lacks equivalently defined endpoint representations.

We evaluated FlexMOF on as many tasks supported by these baselines as possible and provide the corresponding comparisons.

We release evaluation scripts, checkpoints, and sampled structures for FlexMOF and the evaluated baselines, enabling reproduction with the same references and evaluation procedures.

Code: FlexMOF repository

Checkpoints and sampled structures: Zenodo reviewer preview

Weakness 3: Compare composition generators using a fixed structure predictor

We have performed the proposed experiment by comparing independently trained and jointly trained composition generators using the same fixed structure predictor. Both approaches use identical structure-predictor weights and the same 10,000 generation requests, allowing us to directly evaluate the effect of composition generation on the quality of the resulting MOF structures.

As reported in Section 4.5, Table 3 and Appendix F.4 of the revised manuscript, the results are:

\hline \textbf{Composition generator} & \textbf{Validity (\\%) }\uparrow & \textbf{EVNU (\\%) }\uparrow \\\\ \hline \text{Independently trained} & \text{56.79} & \text{49.61} \\\\ \text{Jointly trained (FlexMOF)} & \textbf{79.82} & \textbf{69.73} \\\\ \text{Improvement} & \textbf{+23.03} & \textbf{+20.12} \\\\ \hline \end{array}$$ These substantial improvements demonstrate that **the jointly trained composition generator produces chemical compositions that lead to higher-quality MOF structures even when the structure predictor is fixed**. The benefit therefore cannot be attributed solely to improvements in the structure-prediction model. We further analyze the chemical graphs produced by the two composition generators and their resulting structures: $$\begin{array}{lcc} \hline \begin{array}{c}\textbf{Chemical consistency} \\\\ \textbf{issue}\end{array} & \begin{array}{c}\textbf{Separate training} \\\\ \textbf{(\\%) }\downarrow\end{array} & \begin{array}{c}\textbf{Joint training} \\\\ \textbf{(\\%) }\downarrow\end{array} \\\\ \hline \text{Disconnected linker graphs} & \text{3.64} & \textbf{2.60} \\\\ \begin{array}{c}\text{Carbon atoms with more} \\\\ \text{than four graph neighbors}\end{array} & \text{6.65} & \textbf{4.41} \\\\ \begin{array}{c}\text{Structures flagged for} \\\\ \text{carbon undercoordination}\end{array} & \text{22.48} & \textbf{11.33} \\\\ \hline \end{array}$$ Carbon undercoordination refers to **carbon atoms having insufficient bonding neighbors for their expected chemical environment**, potentially indicating missing bonds, missing hydrogen atoms, or incorrect local geometry. Unlike the first two metrics, which directly assess generated chemical graphs, carbon undercoordination is detected by MOFChecker from the resulting 3D structures. These results reveal concrete limitations of composition generation without structural supervision: independently trained generators produce more disconnected or chemically implausible building-block graphs, and their resulting structures exhibit more coordination defects.

Dear Reviewer,

We thank you again for the constructive suggestions. As the author–reviewer discussion period will end soon, we would appreciate your consideration of our rebuttal and additional experimental results. We would be happy to provide further clarification or respond to any remaining concerns.

Reviewer_QLv9OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces FlexMOF, a generative model for structure prediction and de novo generation for MOFs. FlexMOF represents a MOF with metadata, building block graphs, coordinates, and lattice, and factorizes the problem into p(H)p(G∣H)p(X,L∣G,H)p(H)p(G \vert H)p(X, L \vert G, H). p(H)p(H) is an autoregressive GRU trained separately, and p(G∣H)p(G \vert H) and p(X,L∣G,H)p(X, L \vert G, H) are, respectively, a latent flow model and an ambient/spatial diffusion model trained jointly. The model outperforms baselines in structure prediction (both clean MOF and MOF with adsorbates) and in DNG.

Critical strengths

FlexMOF demonstrates strong empirical gains compared to existing baselines in all tasks (structure prediction, property evaluation, and generation).

Critical weaknesses
  1. Insufficient ablations of the proposed contributions:

    • The PBC ablation demonstrates the importance of modeling periodic interactions but does not evaluate the benefits of the proposed source-atom aggregation mechanism. Comparing the proposed method against conventional attention over expanded periodic images (with equivalent cutoff thresholds) would help validate the claimed efficiency gains and assess any differences in performance.
    • Although the paper provides a joint-versus-"separate-training" ablation, it is unclear how the latter is implemented and how each component contributes to performance. It will be helpful to investigate which aspects of joint training contribute to the improvements: (1) sharing the encoder between the graph and structure prediction models and (2) backpropagating coordinate-loss gradients through the graph posterior latents.
  2. Reproducibility: The proposed methodology is relatively complex, yet neither code nor pretrained checkpoints are provided. Furthermore, important computational details, such as total training time and GPU model, are not reported.

Minor points (non-critical; simply to help improve paper quality):

  • Clarification of joint generation: Although the paper emphasizes "joint composition and structure generation," inference remains sequential. This distinction should be made clearer in the abstract and introduction to avoid confusion with models that jointly sample composition and structure, such as Mofasa.
  • The BW-DB structure prediction results do not include the strict matching criterion evaluated in prior work.
  • Figure 1 contains partially clipped headings and the baseline names are formatted inconsistently.
  • Including the architectural overview (Figure 6) in the main text would help readers understand the methodology.
Questions for authors
  1. How does the proposed source-atom aggregation compare with expanded-image attention in terms of performance and computational efficiency?
  2. Are both the shared encoder and backpropagation through the graph posterior latents necessary to achieve the full benefits of joint training?
  3. How does FlexMOF perform under the stricter structure-matching criterion?
  4. What computational resources were used for training, and how many parameters does the model have?
5 comments

Response to Reviewer QLv9

We thank the reviewer for the careful assessment and constructive suggestions on controlled ablations, joint-learning mechanisms, and clarity of presentation. We hope the reviewer will carefully consider the additional evidence and clarifications below when reassessing our submission.

Weakness 1: Insufficient controlled ablations of the proposed contributions

Part 1: Test aggregation efficiency and accuracy against explicit periodic images

Concern: Compare source-atom aggregation with expanded-image attention under the same cutoff to assess computational gains and recovery accuracy.

We have performed the proposed controlled experiment, comparing our source-atom aggregation with conventional attention over explicitly expanded periodic images.

Both implementations use identical model weights, a 40 Å cutoff, the same periodic-image neighborhoods, and paired random seeds. We evaluate 200 size-stratified MOF structures, generating five candidates per structure.

As reported in Section 4.6 and Appendix F.5 (Table 9) of the revised manuscript, both methods achieve identical population-weighted MR@1\mathrm{MR}@1 of 42.45%, while their weighted MR@5\mathrm{MR}@5 values are 70.57% and 70.48%, respectively, demonstrating nearly identical structural recovery accuracy.

However, our source-atom aggregation achieves 15.09–30.23$\times$ faster sampling than the tested expanded-image attention implementation across different structure sizes. For the largest MOFs (≥385\geq 385 atoms), peak allocated GPU memory decreases from 6.04 GiB to 2.81 GiB, a reduction of approximately 53.5%.

\hline \textbf{Atoms per MOF} & \begin{array}{c}\textbf{Sampling} \\\\ \textbf{speedup}\end{array} & \begin{array}{c}\textbf{Aggregation} \\\\ \textbf{memory (GiB)}\end{array} & \begin{array}{c}\textbf{Expanded-image} \\\\ \textbf{memory (GiB)}\end{array} \\\\ \hline \text{1–96} & \text{15.09}\times & \text{2.30} & \text{2.66} \\\\ \text{97–192} & \text{19.69}\times & \text{2.34} & \text{3.07} \\\\ \text{193–384} & \text{25.27}\times & \text{2.48} & \text{4.16} \\\\ \geq\text{385} & \textbf{30.23}\times & \textbf{2.81} & \textbf{6.04} \\\\ \hline \end{array}$$ These results directly validate the computational benefits of our source-atom aggregation strategy while demonstrating that it preserves the structural prediction performance of explicit periodic-image attention.

Weakness 1: Insufficient controlled ablations of the proposed contributions

Part 2: Define the separate-training baseline and distinguish the two joint-learning mechanisms

Concern: Clarify the separate-training baseline and distinguish the contributions of shared encoder parameters and coordinate-loss gradients through graph posterior latents.

We have clarified the implementation of our separate-training baseline and the gradient propagation mechanisms in Section 3.4 and Appendix E.1 of the revised manuscript. In particular, the separate-training baseline independently trains a chemical-graph autoencoder and then trains a latent-flow generator on its frozen representations, without coordinate-denoising supervision.

(1) Effect of shared encoder parameter between graph and structure prediction models. We have conducted an additional controlled experiment to distinguish improvements in composition generation with shared encoder by composition and structure modules. We compare the separate-training baseline against our jointly trained composition generator using the same fixed structure predictor and the same 10,000 generation requests. In the separate-training baseline, the encoder in structure predictor is different with the encoder in the separately trained composition model. While in the jointly trained version, the encoder is shared.

As reported in Section 4.5, Table 3, the results are:

\hline \textbf{Composition generator} & \textbf{Validity (\\%) }\uparrow & \textbf{EVNU (\\%) }\uparrow \\\\ \hline \text{Independently trained} & \text{56.79} & \text{49.61} \\\\ \text{Jointly trained (FlexMOF)} & \textbf{79.82} & \textbf{69.73} \\\\ \text{Improvement} & \textbf{+23.03} & \textbf{+20.12} \\\\ \hline \end{array}$$ These substantial improvements demonstrate that **jointly trained model with a shared encoder produces better composition representations and generated chemical graphs, rather than merely improving the structure prediction model**. Furthermore, the fraction of disconnected linker graphs decreases from **3.64% to 2.60%**, and the fraction of generated compositions containing carbon atoms with more than four bonds decreases from **6.65% to 4.41%**. These findings provide additional evidence that structural supervision improves the chemical consistency of generated compositions. **(2) Gradient propagation through graph posterior latents.** In FlexMOF, coordinate-denoising supervision influences composition learning through two pathways: direct updates to the shared composition encoder and backpropagation through the graph posterior latents used to condition structure generation. Both mechanisms are enabled during joint training. We agree that further isolating the contribution of gradient propagation through the posterior latents would provide a more fine-grained understanding of the joint learning mechanism. However, this requires **retraining a controlled model variant with gradients through the posterior latents detached while retaining encoder sharing**, which entails substantial additional computational cost. We therefore leave this targeted ablation for future work. Our current fixed-predictor experiment establishes the **combined benefit of these mechanisms**, rather than attributing the observed improvements to either pathway individually.

Weakness 2: Missing code, pretrained checkpoints, and training-resource details

We provide the source code, evaluation scripts, model checkpoints, and sampled structures to support independent reproduction of the reported results.

Code: FlexMOF repository

Checkpoints and sampled structures: Zenodo reviewer preview

We also report the computational resources in Appendix F.6 of the revised manuscript. FlexMOF contains 426.33M parameters. Joint training and subsequent composition fine-tuning used resources equivalent to 32 NVIDIA H100 GPUs for 208.92 and 49.20 hours, respectively. These details complement the training-step runtime and memory measurements reported in Appendix F.5–F.6.

Minor Point 1: Clarify joint training versus sequential generation

We have clarified in the Abstract, Introduction, and Section 3.4 that joint refers to the training process, where coordinate-denoising supervision informs composition learning. At inference, FlexMOF sequentially generates building-block chemical graphs, followed by atomic coordinates and unit-cell geometry. We have also clarified this distinction from Mofasa in Section 2.

Minor Point 2: Report stricter structure-matching results

We have added results under the stricter matching criterion with lattice-length tolerance 0.3, site tolerance 0.3, and angle tolerance 10∘10^\circ. FlexMOF achieves 43.32% MR@5\mathrm{MR}@5, compared with the literature-reported 28.97% for AtomMOF-L and 16.0% for MOFFlow-2 (Appendix F.1, Figure 6).

Minor Point 3: Correct Figure 1 labels and formatting

We have corrected the clipped headings and standardized the baseline names and numerical formatting in Figure 1 of the revised manuscript.

Minor Point 4: Place the architectural overview in the main text

We have moved the architectural overview from the original Appendix Figure 6 to Figure 3 in the main text, providing a clearer illustration of the joint training framework and sequential generation process.

Question 1: Compare source-atom aggregation with explicit periodic-image attention

As detailed in our response to Weakness 1, we have conducted the suggested controlled comparison using identical model weights, a 40 Å cutoff, periodic-image neighborhoods, and random seeds. Source-atom aggregation achieves 15.09–30.23$\times$ sampling speedups with nearly identical structural recovery (70.57% versus 70.48% weighted MR@5\mathrm{MR}@5). Please refer to Section 4.6 and Appendix F.5, Table 9.

Question 2: Separate the contributions of encoder sharing and posterior-latent gradients

As discussed in our response to the Weakness 1, our fixed-structure-predictor experiment demonstrates substantial benefits from joint composition learning, improving validity from 56.79% to 79.82% and EVNU from 49.61% to 69.73% (Section 4.5, Table 3). Both gradient pathways are enabled in FlexMOF. Isolating their individual contributions would require additional controlled retraining at substantial computational cost, which we leave for future work.

Question 3: Evaluate recovery under stricter structure matching

Please refer to Minor Point 2 above and Appendix F.1, Figure 6, where we report results under the stricter matching criterion.

Question 4: Report training resources and parameter counts

FlexMOF contains 426.33M parameters, approximately 13.3% fewer than AtomMOF-L (491.91M). Joint training and subsequent composition fine-tuning used resources equivalent to 32 NVIDIA H100 GPUs for 208.92 and 49.20 hours, respectively.

Reviewer_zu3rOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents FlexMOF, a diffusion model that does all-atom MOF generation. It has a shared composition encoder that models coordinates when denoising, which feeds into a latent space. The model reports promising results on structure prediction and de novo generation compared to strong baselines.

Critical strengths
  • I like the flexbility offered by using building blocks, without relying on RDKit or conformers.
  • I like the ablations provided in the appendix, they show which parts are the most useful in making the final FlexMOF framework
  • Pore property prediction metrics look great, showcasing the design model's ability to also be a regression model. It indicates some sort of successful feature or representation learning
Critical weaknesses
  • I'd like to flag limited novelty in the individual components used to create FlexMOF. It uses artefacts from graphormer, DiT, and some other geometric GNNs. The individual components aren't really ablated well either, which makes it hard to understand how much they're helping with the MOF design task.

  • In figure 3c, The improvement over AtomMOF-L and MOFFlow-2 in the MR@1 metric is not that signifcant. Also, the 46% vs 30% is a comparison over the reported AtomMOF-L numbers. If this is a different population of generated samples, that'd be an unfair metric. I'd urge the authors to rerun and rescore on a new batch of AtomMOF-L samples for fairness.

  • There is no information given on the efficiency gains of the model despite it being a central claim.

Questions for authors
  1. I'd urge the authors to rerun and rescore on a new batch of AtomMOF-L samples for fairness.

  2. Can the authors please explain why MOFFlow-2 is removed from Table 4 even though it also does BW generation?

  3. Can the authors please provide info on compute used, parameter count, memory usage, and wall-clock times and compare them to the baselines?

5 comments

Response to Reviewer zu3r

We thank the reviewer for the careful assessment and constructive suggestions on novelty, baseline comparisons, and computational efficiency. We hope the reviewer will carefully consider the additional evidence and clarifications below when reassessing our submission.

Weakness 1: Limited novelty of individual components and insufficient ablations

We agree that Graphormer-style graph biases, diffusion Transformers, and several other architectural components are built upon established foundations. However, the novelty of FlexMOF lies in two key methodological contributions.

First, we propose a unified Cartesian-space diffusion framework that explicitly models atomic coordinates and periodic cell geometry. Unlike existing crystal generation methods that employ fractional-coordinate diffusion or separately parameterized lattice generation, FlexMOF represents both atoms and the unit cell directly in Cartesian space. Specifically, we introduce virtual nodes at the unit-cell corners, allowing the model to directly perceive and update the cell geometry alongside atomic coordinates. The lattice vectors are derived from four independent corner coordinates, while periodic interactions are explicitly modeled through dynamically reconstructed periodic neighborhoods during denoising. This enables flexible all-atom generation without restricting molecular conformations or requiring a separate lattice-parameter diffusion process.

Second, we introduce joint composition–structure learning that couples latent-space chemical graph generation with explicit Cartesian-space structure diffusion. Unlike approaches that train composition and structure generators independently, or generate both chemistry and geometry entirely through a compressed latent representation, FlexMOF shares a composition encoder across the two pathways and propagates coordinate-denoising supervision into the composition representations. Consequently, the composition model can learn representations informed by the structural constraints of periodic MOFs.

Our experiments demonstrate the benefits of this joint training strategy: compared with separate training, joint training improves MOF validity from 56.79% to 79.82% and EVNU from 49.61% to 69.73% when sharing the same model weights in the structure decoder part. We updated this shared structure decoder improvement in Table 3 in the main text.

Together, these contributions establish FlexMOF as more than an assembly of existing architectural components: they connect flexible all-atom periodic diffusion modeling with structure-supervised composition generation in a unified framework, and proving that the composition generation benefits from the accurate Cartesian space periodic modeling.

Weakness 2: Small recovery gains and fairness of AtomMOF-L comparisons

We identified that, in their original paper, the reported AtomMOF-L match rates were underestimated due to a unit-cell orientation error in its evaluation code, which has recently been acknowledged and corrected by the AtomMOF authors in their Github repository v1.1.0 release (https://github.com/nayoung10/AtomMOF/releases/tag/v1.1.0). This explains why AtomMOF-L achieved only approximately 30% MR@1\mathrm{MR}@1 in our original Figure 3(a), which directly cited the results reported before the correction, whereas Figure 3(c) showed higher match rates, as we had independently evaluated AtomMOF-L using the correct unit-cell orientation on the common comparison subset.

We have therefore updated our evaluation using the corrected AtomMOF-L results. On the same filtered Boyd–Woo test set of 19,792 MOFs, evaluated using identical structure-matching criteria, FlexMOF achieves 53.49% MR@1\mathrm{MR}@1, compared with 40.01% for AtomMOF-L and 28.40% for MOFFlow-2. This represents improvements of 13.49 and 25.09 percentage points, respectively. At MR@5\mathrm{MR}@5, FlexMOF further achieves 73.12%, compared with 61.46% for AtomMOF-L, an improvement of 11.66 percentage points.

Our advantage also persists under a stricter structure-matching criterion, with lattice-length tolerance of 0.3, site tolerance of 0.3, and angle tolerance of 10∘10^\circ. Under this setting, FlexMOF achieves 43.32% MR@5\mathrm{MR}@5, compared with the corrected 28.97% reported for AtomMOF-L and 16.0% reported for MOFFlow-2, corresponding to differences of 14.35 and 27.32 percentage points, respectively.

Importantly, we have also improved our periodic-neighborhood construction by removing the previous global limit on the number of periodic images. We retain all eligible periodic interactions within the 40 Å cutoff in the updated result, allowing more complete modeling of periodic structures. Our source-atom aggregation strategy avoids redundant scalar attention computations over repeated periodic images, helping control computational and memory overhead even when a larger number of periodic neighbors are considered.

The performance advantage becomes particularly pronounced for larger MOFs. For structures containing more than 200 atoms, FlexMOF achieves 33.8% MR@1\mathrm{MR}@1, compared with 12.4% for AtomMOF-L and 0.5% for MOFFlow-2.

Weakness 3: Missing evidence for computational efficiency gains

We have conducted additional experiments to systematically evaluate the computational efficiency gain of our efficient periodic attention. The results demonstrate substantial improvements in both sampling speed and memory efficiency, particularly for larger MOF structures. We have also incorporated these experiments into Appendix F.5 of the revised manuscript.

Experimental setup. We compare our source-atom aggregation against a conventional expanded-image attention implementation, where periodic images are explicitly represented as separate key/value entries and attention is evaluated in chunks. Both implementations use identical model weights, a 40 Å cutoff, the same periodic neighborhoods, and paired random seeds. We evaluate 200 MOF structures stratified into four atom-count groups, with 50 structures per group and five generated candidates per structure. All experiments are conducted on a single NVIDIA H100 80 GB GPU.

Structural accuracy. Both implementations achieve identical population-weighted MR@1\mathrm{MR}@1 of 42.45%. Their population-weighted MR@5\mathrm{MR}@5 values are 70.57% for source-atom aggregation and 70.48% for expanded-image attention, respectively. These nearly identical results confirm that our aggregation strategy preserves the structural prediction quality of explicit periodic-image attention.

Computational efficiency. The table below summarizes the median paired sampling-time speedups and peak allocated GPU memory across different MOF sizes.

\hline \textbf{Atoms per MOF} & \textbf{Speedup} & \begin{array}{c}\textbf{Aggregation} \\\\ \textbf{memory (GiB)}\end{array} & \begin{array}{c}\textbf{Expanded-image} \\\\ \textbf{memory (GiB)}\end{array} \\\\ \hline \text{1–96} & \text{15.09}\times & \text{2.30} & \text{2.66} \\\\ \text{97–192} & \text{19.69}\times & \text{2.34} & \text{3.07} \\\\ \text{193–384} & \text{25.27}\times & \text{2.48} & \text{4.16} \\\\ \geq\text{385} & \textbf{30.23}\times & \textbf{2.81} & \textbf{6.04} \\\\ \hline \end{array}$$ Our source-atom aggregation achieves **15.09–30.23\$\\times\$ faster sampling** than the tested expanded-image implementation, with greater speedups for larger structures. In the largest size group, peak allocated memory decreases from **6.04 GiB to 2.81 GiB**. These efficiency gains arise because periodic images of the same source atom share their scalar keys and values. By aggregating their contributions before scalar attention, FlexMOF avoids redundant computations over repeated images while retaining image-specific geometric information for coordinate updates. **Overall, these results demonstrate that our periodic attention implementation significantly reduces computational cost and memory consumption while maintaining essentially identical structure prediction accuracy.** This efficiency is particularly valuable for all-atom MOF generation, where large unit cells and numerous periodic neighbors would otherwise introduce substantial computational overhead.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.