acceptodds
Under review as a conference paper at ICLR 2027

Coarse-to-Fine MOF Generation with Periodic U-DiT and Verifier-Guided Linker Editing

13 views

Abstract

Generating novel metal–organic frameworks (MOFs) with application-relevant pore architectures is a vital challenge, requiring both diverse structural exploration and precise geometric control. Small pore apertures are particularly compelling targets given their prevalence in experimental MOFs and critical role in selective molecular separation. To pursue this objective, we introduce CoFi-MOF, a coarse-to-fine framework combining global MOF generation with target-directed linker editing for fine-grained control of small pore apertures, and evaluate its performance in the challenging ultra-microporous regime. At the global scale, a periodic diffusion Transformer generates pore geometry conditioned on the target pore-limiting diameter within a signed-distance-field latent space, which a learned topology-aware Constructor translates into discrete MOF structures. At the local scale, geometric measurements identify promising structures for local refinement by Multi-Edit-Former, which proposes chemically feasible linker modifications to adjust their apertures. Reassembly and verification establish whether these edits improve the realized pore match. Experiments convincingly show that CoFi-MOF generates structurally diverse MOFs with pore distributions concentrated around targets. We further expand PORMAKE’s linker library with carboxylate-based building blocks and introduce MOF-Edit-DB, a large-scale dataset of edited MOFs to train and evaluate target-directed structural editing models. Together, CoFi-MOF enables precise small pore targeting in assembled MOFs and provides a reusable foundation for learning structural refinement.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
metal-organic framework, material generation, AI for science, Periodic Diffusion Transformers, Verifier-Guided Refinement

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_xonuOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the limitations of existing MOF generative models in precisely controlling small pore apertures. To address this challenge, the authors propose CoFi-MOF, a coarse-to-fine generation framework that combines global pore-geometry generation with local structural refinement. Specifically, the framework consists of three main components. First, a Periodic U-DiT generates target-conditioned pore geometries in a signed-distance-field latent space while incorporating periodic spatial relationships. Second, a topology-aware Constructor maps the generated geometries into discrete MOF structures by predicting node and linker assignments. Third, the authors introduce a verifier-guided refinement procedure that uses Zeo++ to select promising candidates and a Multi-Edit-Former (MEF) to propose target-directed linker modifications, followed by reassembly and geometric verification.

Additionally, the authors introduce MOF-Edit-DB, a large-scale dataset of MOF structures and linker-edit responses, to support the training of the editing model. Experimental results demonstrate substantially improved pore-size targeting compared with the evaluated baselines,

Critical strengths
  1. Strong experimental performance. The proposed framework substantially improves pore-size targeting accuracy over existing MOF-generation methods, demonstrating its effectiveness.

  2. Comprehensive ablation studies. The paper provides detailed ablation experiments that clearly illustrate the contributions of different components, particularly geometric verification and MEF-based linker refinement. These experiments clarify where the performance improvements come from and strengthen the empirical evaluation of the proposed framework.

Critical weaknesses
  1. The performance improvements are largely dominated by geometric verification rather than the proposed neural architectures.
  2. The comparisons with existing methods may not sufficiently isolate the proposed methodological contributions.
  3. The generated structures exhibit limited robustness after force-field relaxation
Questions for authors
  1. According to Table 2, introducing Zeo++-based verification increases [email protected] from 17.56% to 83.40%, while MEF provides a further improvement to 88.15%. Given that most of the improvement comes from verification-based candidate selection, could the authors further clarify the contributions of Periodic U-DiT and the topology-aware Constructor to the final PLD-targeting performance? In particular, how much improvement would remain if the proposed architectures were replaced with the original MOFFUSION components under the same verification protocol?

  2. Similarly to the previous question, could the authors provide comparisons under matched candidate counts and Zeo++ evaluation budgets? It would also be helpful to report how [email protected] changes as the number of evaluated candidates increases, to distinguish improvements from learned generation versus inference-time search.

  3. While the experiments demonstrate that MEF outperforms random editing, it remains unclear whether the proposed joint editing architecture is necessary for these improvements. Could the authors provide ablation studies on the unary, pairwise, and global scoring components, or comparisons with simpler learned property-prediction and ranking baselines?

  4. Appendix E.4 reports that CoFi-MOF's [email protected] decreases from 93.6% to 20.9% after force-field relaxation. Since accurate pore-size control is central to the paper's motivation, could the authors discuss whether the framework can directly optimize relaxed PLD rather than unrelaxed geometry? What are the main challenges in incorporating relaxation-aware verification or editing into the proposed pipeline?

6 comments
AuthorsOpenReview ↗

Response to Reviewer xonu

Dear Reviewer xonu, thank you for your constructive feedback. We address your questions with new experiments and quantitative analyses below, revised is highlighted in blue.

W1–W2 / Q1. What do Periodic U-DiT and the Constructor contribute vs. verification?

Verification evaluates candidates, whereas Periodic U-DiT and the topology-aware Constructor determine which candidates exist to be evaluated. Periodic U-DiT generates a target-conditioned continuous SDF, which admits multiple discrete atomistic realizations with distinct pore apertures. Constructor maps this SDF to ranked discrete assignments, and Zeo++ resolves the remaining geometric ambiguity. Controlled ablations confirm each learned module's standalone gain:

  • Periodic U-DiT (App. E.1): Under identical training and evaluation settings vs. a 3D U-Net, assembly success rises from 64.0% to 71.2% and end-to-end validity from 58.0% to 67.4%.
  • Topology/node conditioning (App. A.3): On shared validation pore fields, Constructor's linker Top-1/Top-5 accuracy increases from 36.08%/72.38% to 45.92%/83.32%.

Q2. Can methods be compared under matched candidate budgets and scaled with candidate count KK?

Yes. Under strictly aligned budgets (1,000 upstream attempts/target ×\times 25 slots = 125,000 slots; failed attempts counted as misses without replacement), CoFi-MOF (without MEF) outperforms all baselines at K=5,10,25K=5, 10, 25:

Method / [email protected] (%)K=1K=1K=5K=5K=10K=10K=25K=25
MOFDiff20.4043.4849.5055.48
MOFFlow-217.1639.7247.0454.02
MOFFUSION15.9843.9057.9274.10
CoFi-MOF (w/o MEF)17.5448.8666.4483.34

At K=25K=25, CoFi-MOF beats MOFFUSION by +9.24 pp overall (+8.52 pp at K=10K=10), leading across 3–6 Å (87.40%, 88.90%, 86.90%, 80.20% vs. 72.80%, 72.50%, 76.00%, 74.90%) while slightly trailing at 7 Å (73.30% vs. 74.30%). Candidates derive from one fixed upstream output via each method's native realization (MOFFlow-2: 25 geometric realizations; MOFDiff: 25 annealing seeds; MOFFUSION: top-25 node/linker assignments; CoFi-MOF: Constructor's top-25 assignments). Actual valid Zeo++ counts reflect assembly compatibility: CoFi-MOF records 124,211 measurements (from 124,850 attempts) vs. 105,711 (MOFFUSION), 115,510 (MOFDiff), and 114,374 (MOFFlow-2).

Q3. Does MEF outperform simpler predictors, and how much does it recover in the full pipeline?

Even before Zeo++ verification, MEF achieves 70.59% Top-1 [email protected] across 1,129 shared-source tasks (+31.20 pp over XGBoost, cutting Top-1 MAE from 0.6777 to 0.3268 Å):

Ranking MethodTop-1 [email protected]Top-1 MAE (Å)Best-of-10 [email protected]
MEF70.59%0.326877.41%
XGBoost PLD regressor39.39%0.677774.87%
PLD Prediction MLP34.25%0.824073.31%
Additive Site-wise MLP31.03%0.929376.23%

Under Best-of-10 verification, MEF retains the lowest MAE (0.2119 vs. 0.2501/0.2342 Å) and beats random editing on 937 shared tasks (78.12% vs. 65.21%). Across the full 5,000-attempt pipeline, MEF-guided refinement rescues ~30% of remaining failures without losing any existing hits, boosting [email protected] from 68.24% to 77.46% (+9.22 pp):

ToleranceRemaining MissesRescued HitsRecovery RateHits Lost
±0.5\pm 0.5 Å833235 (4,167→4,4024{,}167\rightarrow4{,}402)28.2%0
±0.2\pm 0.2 Å1,588461 (3,412→3,8733{,}412\rightarrow3{,}873)29.0%0

W3 / Q4. How do structures respond to relaxation, and what obstacles arise in optimizing relaxed PLD?

Under UFF4MOF/LAMMPS relaxation across 1,280 paired structures, [email protected] shifts from 93.6% to 20.9% (MAE 0.140→1.2390.140\rightarrow1.239 Å). The decrease in [email protected] quantifies post-relaxation target mismatch, rather than the fraction of structurally collapsed frameworks. Notably, CoFi-MOF substantially outperforms MOFFUSION in relaxed cohorts: 20.9% vs. 12.3% [email protected], 1.239 vs. 2.999 Å MAE, and 21.6% vs. 9.7% (2.2×2.2\times) initial-hit retention, identifying 276 candidates across 106 topologies with PLD shift ≤0.5\le 0.5 Å (242 still hitting the target).

Our decoupled architecture directly supports selecting by relaxed PLD: erel(C)=∣PLDZeo++(R(C))−d⋆∣e_{\mathrm{rel}}(C)=|\mathrm{PLD}_{\mathrm{Zeo++}}(\mathcal{R}(C))-d^\star|, ranking shortlisted Constructor/MEF candidates after relaxation R\mathcal{R} without differentiating through force fields. The primary bottleneck is compute cost: UFF4MOF three-stage minimization is ~500$\times$ slower than Zeo++ (316 ms/structure, App. B.4). Relaxing all 125,000 candidates would require ~6,250 core-hours (~1 day on 256 cores) vs. 11 core-hours for Zeo++. Fast geometric screening followed by relaxation-based re-ranking of top shortlisted candidates offers an immediately practical solution.

We hope these analyses support a positive reassessment of our paper.

Dear Reviewer xonu,

Thank you again for your thoughtful and constructive feedback, which has helped us substantially strengthen our paper.

As the discussion period approaches its November 18 deadline, we would like to briefly highlight the new evidence addressing your central concerns.

Most importantly, our matched-budget comparison shows that CoFi-MOF retains a 9.24-percentage-point advantage over MOFFUSION (83.34% vs. 74.10% [email protected]) even without MEF, under the same Best-of-25 verification protocol. This provides direct evidence that the learned candidate-generation pipeline contributes beyond Zeo++-based selection alone. We have also added comparisons with simpler learned edit-ranking baselines, where MEF achieves 70.59% versus 39.39% Top-1 [email protected], further clarifying its contribution.

We sincerely appreciate your time and careful consideration. If you have any remaining concerns, we would be grateful for the opportunity to address them before the discussion closes. We hope these additional results and clarifications provide a stronger basis for reconsidering your initial assessment.

Thank you once again for your valuable feedback.

Thank you to the authors for the detailed clarifications and additional experiments. I appreciate the efforts to address my concerns. However, the substantial decrease in [email protected] after relaxation suggests that accurate control of unrelaxed pore geometry does not necessarily translate into reliable control of physically relaxed structures. This discrepancy remains a significant concern and leaves me uncertain about the strength of the paper's conclusions regarding practical pore-size control.

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

The authors propose a coarse-to-fine generative model for MOFs with a specific focus on controlling pore diameter. This is a challenging and important problem. Some of the architectural components are borrowed directly from literature such as the use of signed distance functions (SDFs) to represent pore geometry. A latent diffusion model is introduced to realize a coarser version of the SDF grid with the help of their proposed periodic U-DiT. The resulting latents are fed into a constructor that realizes the MOF topology and predicts molecular fragments for the nodes and edges (linkers). The actual MOF is then constructed from this description by an existing package PORMAKE. Earlier predicted node/linker choices may not be optimal in the resulting structure and authors introduce an additional editor to modify the structure towards desired constraints. The paper is overall clear, well-written but not self-contained. Empirical results are promising.

Critical strengths

Generating property tailored MOFs is a challenging and important problem. The title of the paper accurately captures the key innovative components in the paper, periodic attention in U-DiT and the verifier-guided linker editing. The remaining pieces are largely borrowed from prior literature.

The empirical demonstrations are extensive, helpful and transparent.

The authors also expand the linker library which could further contribute to next generation of MOF design tools.

Critical weaknesses

The paper isn't written in a self-contained manner. Much of the understanding of how the method ultimately builds its generated MOF structure is based on how MOFFUSION and PORMAKE operate.

SDF representation is borrowed from MOFFUSION. The authors propose to train a periodic U-DiT that can generate a coarser version of this grid (8x8x8x3), starting from noise. The constructor then needs to expand this back into a larger SDF grid that is used as an input to decide on the network topology (node placements in fractional coordinates, connectivity) as well as the actual molecular fragment choices for the nodes and linkers. I assume these are then fed into a PORMAKE to construct the structure? Additional edit operations keep the topology fixed but can change the node/linker fragments to sequentially optimize (search for a better) PDL. The whole package is a complex system of routines, some proposed and learned, some borrowed. Ablations and architectural inspection would be more valuable than highlighting specific successes (that can come from many sources). E.g., Figure 3 is compelling but is the benefit primarily coming from MEF (editing layer)?

Questions for authors

how are rotatable linker orientations chosen? Randombly by PORMAKE? Wouldn't these affect PDL?

Since PORMAKE adjusts the unit cell/distances to arrange the chosen blocks and linkers into a structure, the resulting PDL is only computable at that point. How is the proposed model that passes the predicted topology and predicted node/linker choices to PORMAKE able to understand what this fairly complex last step does in relation to PDL?

It would be helpful to characterize how much force-field relaxation changes the predicted structure / PDL. If changes are substantial, the predicted structure itself is largely a sketch, not close to a stable structure.

4 comments
AuthorsOpenReview ↗

Response to Reviewer ckqF

Dear Reviewer ckqF, thank you for your encouraging evaluation and constructive suggestions. We summarize our responses with staged ablations and quantitative analyses below (revised PDF is highlighted in blue).

W1. Pipeline modules and component contributions (Figure 3)?

CoFi-MOF builds on the SDF representation/VQ-VAE from MOFFUSION and atomistic assembly via PORMAKE, introducing Periodic U-DiT, the learned Transformer Constructor, and verifier-guided editing (MEF). The inference chain is: (1) Periodic U-DiT generates a 3×8×8×83\times8\times8\times8 continuous latent; (2) quantization and the frozen VQ-VAE decode a 4-channel 32332^3 SDF; (3) Constructor predicts topology, node pairs, and conditional linkers; (4) PORMAKE assembles atomistic periodic structures and Zeo++ measures realized PLD for selection; (5) MEF proposes linker substitutions from measured PLD and target, followed by reassembly and remeasurement (Secs. 4.1–4.3; Apps. A–B).

Staged results across all 5,000 attempts (failed assemblies counted as misses; MAE over measurable PLDs) show that strong aperture targeting is established before MEF, while MEF provides fine-grained refinement:

Pipeline / Selection[email protected] (%)[email protected] (%)MAE (Å)
MOFFUSION, top-115.986.461.785
CoFi-MOF, Constructor top-117.547.621.734
MOFFUSION, best-of-2574.1058.100.409
CoFi-MOF, best-of-25, no MEF83.3468.240.303
CoFi-MOF, verification + MEF top-186.2071.640.257
CoFi-MOF, verification + MEF top-1088.1077.460.220

Under the same best-of-25 budget, CoFi-MOF without MEF already outperforms MOFFUSION by +9.24 pp ([email protected]) and +10.14 pp ([email protected]). MEF further raises [email protected] from 68.24% to 77.46% and cuts MAE by 27.4% (0.303→0.2200.303\rightarrow0.220 Å). Complementary module diagnostics confirm each component's role:

  • Periodic U-DiT (App. E.1): Under identical data, frozen VQ-VAE, 60k updates, and 500 shared noise seeds vs. a 3D U-Net, assembly success rises from 64.0% to 71.2% and end-to-end validity from 58.0% to 67.4% (+9.4 pp, 95% CI [3.6,15.0][3.6,15.0]) with 6.5×6.5\times fewer parameters (95.7M vs. 620.7M).
  • Constructor (App. A.3): On 13,691 validation pore fields, conditioning on topology/nodes improves linker Top-1/Top-5 accuracy from 36.08%/72.38% to 45.92%/83.32%.
  • MEF (App. E.5): On 937 6 Å tasks, MEF's first edit proposal alone (without selection) hits 68.20% [email protected] vs. 33.62% for random editing (+34.58 pp).

Q1. How are rotatable linker orientations chosen, and do they affect PLD?

Yes. PORMAKE v0.2.3 leaves a rotational degree of freedom around the connection axis for two-connected linkers, which can alter local pore-window geometry. Therefore, CoFi-MOF evaluates every assembled candidate directly from its realized atomic coordinates via Zeo++, ensuring that realized linker orientations and unit-cell adjustments are explicitly captured during candidate selection.

Q2. How does the predictor account for PORMAKE assembly effects on PLD?

During training, the Constructor learns the correspondence between atomistic MOF pore geometries and their topology, node, and linker identities. Its Node and Edge Transformers model these dependencies conditionally, allowing the Constructor to propose chemically compatible realizations of the generated pore field. Because a continuous SDF pore field admits multiple discrete chemical realizations with distinct post-assembly apertures, Constructor proposes up to 25 ranked building-block assignments per pore field. Assembling and measuring all candidates via PORMAKE–Zeo++ systematically explores this one-to-many mapping to identify structures meeting the target PLD.

Q3. How much does force-field relaxation change the structures and their PLDs?

Appendix E.4 quantifies post-relaxation response (UFF4MOF/LAMMPS) across 1,280 selected MEF-edited structures:

MetricBefore →\rightarrow After / Statistic
[email protected] / Target MAE (Å)93.6% →\rightarrow 20.9% / 0.140 →\rightarrow 1.239
Absolute PLD change (mean / median)1.207 Å / 1.104 Å
PLD change ≤0.5\le 0.5 Å276 candidates across 106 topologies

In paired cohorts (App. E.4), CoFi-MOF outperforms MOFFUSION in relaxed [email protected] (20.9% vs. 12.3%), relaxed MAE (1.239 vs. 2.999 Å, 58.7% lower), and initial-hit retention (21.6% vs. 9.7%, 2.2×2.2\times). Crucially, bottleneck PLD shifts from local linker rotation or window adjustment should be distinguished from framework collapse. Importantly, 276 candidates across 106 distinct topologies exhibit PLD changes within 0.5 Å, while 259 candidates satisfy the target tolerance both before and after relaxation.

We thank you again for helping strengthen the paper and hope these additions support your positive reassessment.

Thank you for the clarifications. I'm a bit surprised that linker orientations (initially random) would be determined only through selection of fully assembled structures that contain those random initial choices. This seems somewhat inefficient.

Thank you dear reviewer ckqF ,for this insightful follow-up. We would like to clarify an important distinction regarding linker orientations and the efficiency of our assembly strategy.

Although the axial orientations of two-connected linkers are not explicitly optimized in our PORMAKE-based assembly procedure, this does not mean that our pipeline repeatedly samples random orientations. To verify this in our implementation, we independently assembled 10 fixed topology–node–linker assignments (covering target PLDs of 3–7 Å) three times each. All 30 assemblies succeeded. For each fixed assignment, the unit cell and fractional atomic coordinates were identical at CIF precision, and the measured PLDs agreed to within 10−510^{-5} Å, with an observed standard deviation of 0 Å. Thus, for the tested assignments, assembly is reproducible rather than a stochastic search over linker orientations.

More importantly, CoFi-MOF generates a single SDF pore field per generation attempt, from which the learned Constructor proposes up to 25 distinct chemical realizations. Specifically, for one predicted topology, the Constructor explores up to five node-pair assignments, each associated with up to five conditional linker assignments. These candidates represent alternative chemical structures derived from the same generated SDF, rather than 25 independent SDF generations or different random rotations of an identical MOF.

This one-to-many realization strategy is central to our high-throughput inverse-design objective. A generated pore field does not uniquely determine its underlying chemical building blocks; exploring multiple compatible assignments allows us to identify promising target-matching structures without repeatedly generating new pore fields. Under this strategy, CoFi-MOF achieves 83.34% [email protected] before MEF editing, with at most 25 assembled candidate evaluations per generated SDF.

We agree that explicitly optimizing linker orientations could further improve the atomistic realization of individual structures. Existing orientation-aware approaches, such as the symmetry-guided PORMAKE_v2 extension, provide complementary refinement strategies. Our framework focuses on target-directed chemical structure generation and linker editing, while orientation-aware geometric refinement can naturally follow as a subsequent stage.

We will clarify this distinction and discuss the absence of explicit orientation optimization in the Limitations section. We appreciate the reviewer highlighting this point.

AuthorsOpenReview ↗

Looking forward to your further comment

Dear Reviewer ckqF,

We appreciate your time and insightful comments. With less than 12 hours remaining in the discussion period, we kindly invite you to review our rebuttal and additional experiments if your schedule permits. Please let us know if any points need further clarification before the window closes. We respectfully remind you to update your score should you find our responses satisfactory. We would greatly appreciate your positive consideration.

Thank you for your patience and valuable guidance throughout this process.

Best regards,

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

The paper proposes CoFi-MOF, a framework for generating MOFs conditioned on PLD. Its global representation stage resembles MOFFUSION with VQ-VAE to compress the SDF representations of MOFs and a periodic U-DiT model that learns continuous latent representations. The structure is assembled by categorically predicting topology, node, edge and assembling them with a heuristic algorithm called PORMAKE. It also has a post-processing stage called MEF, which adjusts the linkers based on measured PLD to increase/decrease the pore size to fit the target.

Critical strengths

Precise control over MOF pore sizes is an important problem with practical relevance, particularly for molecular separation.

Critical weaknesses

I have substantial concerns about the experimental design, particularly the fairness of baseline comparisons and the justification of individual components.

  1. Insufficient rigor in experimental comparisons: Inconsistent baseline training protocols, unmatched candidate-selection budgets, and insufficiently documented dataset construction raise concerns about the fairness and reliability of the results.

    • Inconsistent baseline training: The baselines use substantially different training protocols. MOFDiff relies on a pretrained generator trained on a different dataset (BW-DB) and a separately trained PLD predictor, while MOFFlow-2's structure generator, also pretrained on BW-DB, is directly evaluated with the released checkpoint. In contrast, the authors fully train CoFi-MOF from scratch using their own dataset. The overall comparison lacks consistent training conditions, making it difficult to attribute performance differences to the proposed methodology.
    • Unmatched candidate-selection budgets: CoFi-MOF considers up to 25 Constructor assignments and 10 additional MEF edits per generation attempt, whereas the baselines use different candidate-selection budgets. This is particularly concerning because Table 2 shows that verification (i.e., best-of-N selection using Zeo++) alone improves [email protected] from 17.56% to 83.40%. Moreover, without verification or MEF, CoFi-MOF achieves only 17.56% [email protected], below MOFFlow-2's average of 21.88% across targets in Table 1, although their training protocols differ. These results raise concerns about whether the reported improvements stem from the generative model itself or the additional candidate exploration and verification. For a fairer comparison, the authors should evaluate baselines under matched candidate-selection budgets, such as best-of-N sampling with Zeo++ verification, and report the total number of assembled candidates and property evaluations.
    • Dataset construction: The global generation dataset is insufficiently documented. The authors mention combining selected small-pore MOFFUSION structures with self-generated frameworks, but do not clarify how the latter were generated, the proportions of each source, or the selection criteria. Moreover, while Figure 1 motivates the task using experimental CoRE MOFs, the authors do not explain why these structures were not incorporated into training.
  2. Missing novelty evaluation: Although the paper starts its motivation with novel MOFs ("Generating novel metal-organic frameworks ..."), it only reports structural uniqueness within generated batches. Novelty relative to training structures or existing MOFs is not evaluated, leaving it unclear whether the model explores genuinely new structural space.

Questions for authors

My questions are primarily aimed at addressing the weaknesses outlined above:

  1. Could the authors clarify how the training dataset was constructed and why experimental CoRE MOFs were not included?
  2. How would the baselines perform if trained from scratch on the same dataset with comparable training budgets?
  3. Could the authors evaluate the baselines with best-of-N Zeo++ selection under matched candidate-evaluation budgets?
  4. Could the authors evaluate structural novelty relative to the training dataset?
7 comments
AuthorsOpenReview ↗

Response to Reviewer ysap

Dear Reviewer ysap, thank you for your constructive comments. We summarize our responses with new experiments and statistics below (revised PDF is highlighted in blue).

1. Training Fairness and Candidate Budgets

W1.1 / Q2. Would the advantage remain if the baselines used the same training data and comparable training budgets?

PLD-adapted MOFFUSION uses the exact same 109,529-record corpus, 300,000 cumulative updates, and 200-step deterministic DDIM sampling with the same PORMAKE–Zeo++ protocol as CoFi-MOF. Replacing Periodic U-DiT with a 3D U-Net under identical settings (same corpus, frozen VQ-VAE, loss, optimizer, batch 32, 60k updates, 500 shared noise seeds) degrades performance:

MetricPeriodic U-DiT3D U-NetRelative Diff.
Trainable params / Checkpoint95.8M / 1.15GB620.7M / 7.45GB6.5×6.5\times smaller
Peak forward memory (BS=1)0.67 GB3.80 GB5.7×5.7\times less
Assembly / Validity (%)71.2 / 67.464.0 / 58.0+9.4 pp (p=8.2×10−4p=8.2\times10^{-4})

We did not retrain MOFDiff or MOFFlow-2 from scratch as both are tied to BW-DB artifacts (1.99M-entry embeddings [17 GB] and a 59-token vocabulary, requiring 2M steps ×\times batch 128 plus 1.5M BB-encoder steps). These baselines do not disadvantage us in data or compute: they use 1.7–2.7$\times$ more training structures (293,513 and 187,028 vs. 109,529) and up to 27$\times$ more optimization samples (256M vs. 9.6M).

W1.2 / Q3. Unmatched candidate-selection budgets? Compare best-of-NN candidates under aligned budgets.

Even without MEF, CoFi-MOF achieves 83.34% [email protected] (+9.24 pp over MOFFUSION). All methods receive 1,000 upstream attempts across five targets and up to 25 candidates per attempt (5×1,000×25=125,0005\times1{,}000\times25=125{,}000 slots; failed attempts count as misses without replacement). Candidates derive from one fixed upstream output: MOFFlow-2 (25 geometric realizations), MOFDiff (25 annealing seeds 0–24), MOFFUSION (top-25 distinct node/linker assignments under top-1 topology), and CoFi-MOF (Constructor's 25 ranked assignments without MEF).

Method (No MEF)3 Å4 Å5 Å6 Å7 ÅOverall [email protected]
MOFDiff50.00/63.1061.70/74.3045.50/59.6033.60/45.0022.60/35.4055.48%
MOFFlow-242.40/54.3066.20/75.3058.10/72.3033.40/44.9016.70/23.3054.02%
MOFFUSION58.40/72.8057.40/72.5059.30/76.0057.00/74.9058.40/74.3074.10%
CoFi-MOF w/o MEF73.60/87.4077.20/88.9071.30/86.9064.40/80.2054.70/73.3083.34%

([email protected] / [email protected] %). CoFi-MOF leads on 4 of 5 targets, confirming the gain stems from the generation pipeline rather than search budget. One MEF edit further improves [email protected] to 86.20% ([email protected] to 71.64%), yielding one final structure without property-guided selection (the original MEF top-10 setting is retained solely as an editing-budget ablation). Constructor records 124,850 assembly and 124,211 Zeo++ calls, while MOFFUSION, MOFDiff, and MOFFlow-2 yield 105,711, 115,510, and 114,374 finite PLDs (excluding PORMAKE-incompatible outputs).

W1.3 / Q1. How was the training data constructed?

Our 136,911-record corpus (109,529 train) comprises 65.8% (90,130) from the MOFFUSION small-pore subset and 34.2% (46,781: round 1 = 9,774, round 2 = 37,007) newly constructed structures. New structures keep PORMAKE's format and expand 476 carboxylate base edges (in MOF-Edit-DB), sampled from 200k/500k recipes, PLD-proxy binned (0–15 Å), PORMAKE-assembled, UFF4MOF-relaxed, and Zeo++-labeled. Experimental CoRE CIFs contain building blocks outside our library or incompatible with Constructor/PORMAKE decomposition; thus, CoRE serves as an experimental reference rather than training data (which does not claim generalization to CoRE).

W2 / Q4. Are generated outputs novel relative to training and existing MOFs, rather than merely unique within a batch?

At the recipe level, 94.93% (4,741/4,994) of outputs have topology–node–linker combinations absent from the dataset. Among 1,698 MEF-edited outputs, 99.41% (1,688) have novel combinations and 93.64% (1,590) have edited linker graphs non-isomorphic to all 645 base linkers (heavy-atom connectivity). At the crystal level, pymatgen StructureMatcher (tolerances 0.2, 0.3, 5∘5^\circ, primitive reduction & volume scaling; 20/20 verified on self and 0.02–0.03 perturbed matches) finds 0 crystal-structure matches across all 4,994 outputs:

Reference CollectionRecordsComparedSame CompositionCrystal Matches
Full corpus (train/val/test)136,9114,994—0
Reconstructed CoRE (1,034 ASR / 637 FSR / 129 Ion)1,8004,9941550
BW-DB reference308,9624,9946010

We hope these results address your concerns and support reassessing our paper.

I thank the authors for their detailed response and the additional experiments. However, my main concerns about training and evaluation fairness remain, as the new results do not yet fully address them.

W1.1 (Training fairness). I appreciate the MOFFUSION comparison, but my concern was about MOFDiff and MOFFlow-2, which are still trained on different data. The issue is the training distribution, not its size or the training budget. Only 23.56% of BW-DB lies below 7 Å (Fig. 1), whereas the CoFi-MOF corpus is deliberately enriched for small pores, which is the same 3-7 Å range used for evaluation. As a result, the baselines are evaluated off-distribution. Retraining or fine-tuning them on the CoFi-MOF corpus would make the comparison much more convincing.

W1.2 (Candidate budgets). Thank you for the best-of-25 experiment; matching the budget is a clear improvement. The candidates, however, are still not comparable (App. D.4). CoFi-MOF and MOFFUSION try 25 different building-block combinations, whereas MOFFlow-2 and MOFDiff reuse the same combination 25 times and only vary its geometry. Since building blocks largely determine PLD, this setup disadvantages the baselines. MOFFlow-2's sequence generator, for example, could easily sample 25 new combinations instead. This matters because MOFDiff already outperforms CoFi-MOF at K=1 (20.40% vs. 17.54%), so the ranking may change under a fully fair setting.

Overall, I will keep my score for now. Because the baselines differ in training distribution (W1.1) and candidate diversity (W1.2), I can't yet tell how much of the reported gain comes from the method itself, so the main conclusions are not yet reliable. I think the method is promising, and comparisons that resolve these two points would make the paper considerably stronger.

AuthorsOpenReview ↗

Response to Reviewer ysap (1/2)

We thank the reviewer for the continued engagement. Over the past 48 hours, we have made every effort to ensure a fair comparison; however, fundamental architectural differences between methods make perfect alignment of all experimental conditions impossible. We report the actual results below.

W1.1: Training fairness

The reviewer states that "only 23.56% of BW-DB lies below 7 Å" and concludes that the baselines are evaluated off-distribution. This difference in relative training density is real; however, the target range is also substantially represented in the baselines' actual training sets. MOFDiff's training set contains 67,614 structures in the 3–7 Å range, more than our 60,393, and MOFFlow-2's contains 46,636. In absolute terms, the baselines have therefore seen at least as many small-pore structures as we have. Our corpus spans 0–15 Å, not just 3–7 Å.

We attempted conversion of all 109,529 training and 13,691 validation records into MOFDiff/MOFFlow-2 formats. Only 5,793 records, or 5.29%, are compatible with MOFDiff, and 1,664, or 1.52%, with MOFFlow-2; the intersection is 1,595, or 1.46%. The bottleneck is architectural: 81.4% of our structures exceed 20 building-block instances per primitive cell, a cap hard-coded in MOFDiff's CG encoder dimensions and MOFFlow-2's native preprocessing.

We expanded MOFDiff's cap to 64 and fine-tuned both baselines on the expanded compatible sets: 33,710 structures for MOFDiff and 6,682 for MOFFlow-2. Fine-tuning degraded both baselines severely:

Model3 Å4 Å5 Å6 Å7 ÅOverall
CoFi-MOF w/o MEF87.4088.9086.9080.2073.3083.34
MOFDiff + selection, BW-DB63.1074.3059.6045.0035.4055.48
MOFDiff fine-tuned, 64-BB2.5017.5020.0017.502.5012.00
MOFFlow-2, PLD-cond.54.3075.3072.3044.9023.3054.02
MOFFlow-2 fine-tuned2.505.005.0020.007.508.00

MOFDiff dropped from 55.48% to 12.00%; MOFFlow-2 from 54.02% to 8.00%.

Why fine-tuning degraded: failure happens before PLD measurement. The two models fail at different stages:

  • MOFFlow-2: Only the sequence model was fine-tuned; the 3D flow and metal library remain original. Of the 6,682 training sequences, 3,676, or 55.0%, contain metal templates absent from the original library. Among 5,000 generated diverse sequences, 2,675 lack templates, 745 fail organic SMILES parsing, and 95 lack separators or are not strings. The fixed-sequence arm fails admission checks on 147 of 200 groups at the first sequence.
  • MOFDiff: After expanding to 64 building blocks, measurable slots in the fixed arm dropped from 4,775 out of 5,000 to 775 out of 5,000. Native assembly feasibility checks rejected 4,225 slots due to metal/non-metal connection site count mismatch. Validation loss is dominated by the building-block type term. After re-encoding and re-fitting the PLD head, the prescreen ranking degraded across all five targets.

The baselines' building-block space is extremely narrow. BW-DB contains only 6 metal elements: Zn, Cu, V, Ni, Ba, and Cr, all in single-metal structures. MOFDiff's released BB encoder supports only 17 atom types, of which 6 are metals, and MOFFlow-2's vocabulary contains only 59 tokens, of which 6 are metal-related. Our corpus covers 47 metal elements, including 10% multi-metal structures, 572 topologies, 477 inorganic node types, and 1,060 edge types. Conversion flags 46 atom types as OOV for MOFDiff and 115 token types as missing for MOFFlow-2.

Our model achieves better results with less data, fewer parameters, and greater diversity. CoFi-MOF trains on 109,529 structures, versus 293,474 for MOFDiff and 187,028 for MOFFlow-2. It uses 95.8M parameters, versus 620.7M for MOFDiff's 3D U-Net counterpart, and covers 47 metals and 572 topologies, versus 6 metals and a limited vocabulary in the baselines. Our bare generator at 83.34% [email protected] already outperforms their full pipelines at 55.48%, 54.02%, and 74.10%. We consider this is not a data advantage but a model advantage by our coarse-to-fine design.

We acknowledge the limitation and commit to the full comparison. We have not completed full common-corpus retraining of MOFDiff and MOFFlow-2. This is a limitation of our current evaluation. However, our audit demonstrates that this requires fundamental architectural modifications: expanding MOFDiff's element vocabulary from 17 to 50+ atom types, retraining its 1.5M-step BB encoder from scratch, and re-validating its assembly rules against our corpus's metal/non-metal connection site distributions.

Comments on the paper

2 comments

Dear Area Chair,

Thank you for overseeing the review of our submission (CoFi-MOF) and for the time and effort dedicated to this evaluation cycle. As the discussion period concludes, we would like to provide a concise summary of our rebuttal, the additional experiments conducted, and the remaining points of disagreement so that they may be weighed alongside the original reviews.

1. Summary of Rebuttal & Additional Experiments

During the rebuttal and discussion period, we responded to all points raised by the reviewers and added concrete experiments and clarifications (including App. D.4 and App. D.6):

  • Same-corpus & architectural attribution: Against PLD-adapted MOFFUSION on the exact same 109,529-structure corpus, 300,000 updates, and PORMAKE–Zeo++ protocol, CoFi-MOF achieves 83.34% [email protected] without MEF (86.20% with a single MEF edit) versus MOFFUSION's 74.10% (+9.24 pp before MEF). In a same-corpus, same-budget, shared-seed backbone ablation, replacing Periodic U-DiT with a 3D U-Net lowers assembly/validity by 9.4 pp (67.4% vs. 58.0%) while requiring 6.5x more parameters (620.7M vs. 95.8M).
  • Cross-paradigm baselines & compatibility audit (App. D.4): Following field precedent (Park et al., Nat. Commun. 2025), we compared against MOFDiff and MOFFlow-2 using their strongest available configurations, and further tested 64-BB expansion and fine-tuning on compatible subsets (33,710 and 6,682 records), documenting how hard-coded element vocabularies (6 metals in BW-DB vs. 47 metals including 10% multi-metal and 572 topologies in our corpus) and >20-BB cell counts affect compatibility.
  • Candidate budgets & independent resampling (App. D.6): Under one upstream pore-field proposal with up to 25 realizations—matching coarse-to-fine generation from a target pore blueprint—CoFi-MOF leads across baselines (83.34% / 86.20% vs. 74.10%, 55.48%, and 54.02%). Under 25 independent upstream draws, hit rates saturate across models due to repeated sampling (97.0% vs. 99.0% [email protected] and 87.5% vs. 85.5% [email protected] against MOFDiff; 97% for MOFFlow-2), clarifying the distinction between repeated global resampling and realizing a fixed pore geometry.
  • Dataset construction & novelty: We documented our 136,911-record corpus (spanning 0–15 Å, 65.8% from MOFFUSION's small-pore subset and 34.2% newly constructed via PORMAKE + UFF4MOF + Zeo++) and verified novelty (94.93% novel topology–node–linker recipes; 0 crystal-structure matches against the corpus, CoRE, and BW-DB).

2. Current Status of Reviewer Engagement

  • Reviewer ysap raised their score from 1 to 2 after our rebuttal, acknowledging improved transparency in App. D.4 and D.6, while maintaining that dataset differences limit attribution, that independent resampling reduces the apparent gap, and that the paper requires substantial revision. We provided a detailed reply explaining why same-corpus MOFFUSION/U-DiT controls isolate model effects, why repeated independent sampling saturates probabilistically rather than replacing single-blueprint realization efficiency, and how broader chemical diversity (47 metals) supports practical MOF discovery (synthesizability, solubility, and stability considerations).
  • The other reviewer (score 2) did not provide a follow-up response during the discussion period despite our rebuttal addressing their comments.

3. Request for Consideration in Meta-Review

We will continue refining the manuscript and clearly scoping our claims and evaluation protocols in the final version. Because the core methodological comparisons against MOFFUSION, the same-corpus backbone ablation, the novelty verification, and the open 136,911-structure small-pore corpus are already established, we respectfully ask that you weigh these controlled results and clarifications—alongside the lack of further follow-up from the non-responding reviewer—when forming your final recommendation.

Thank you very much for your time, fairness, and consideration.

Best regards, The Authors of Submission 38471

Dear Area Chair,

Thank you for overseeing the review of our submission (CoFi-MOF). As the discussion period closes, we would like to share a concise concern regarding how the review points stand so that our contributions can be evaluated fairly.

First, regarding Reviewer xonu, we have made every effort to address all of their concerns with concrete results and explanations, yet they have not provided any follow-up response or indicated whether their concerns are resolved. Since further back-and-forth is not possible without their engagement, we can only respectfully ask that you evaluate their original comments directly against our rebuttal and experimental evidence.

Second, regarding Reviewer ysap, we believe their continued insistence on penalizing comparisons across different end-to-end generative pipelines is illogical and inconsistent with established practice in this field. Benchmarking across existing generative models with their respective pipelines and training setups is standard in representative works published at NeurIPS and Nature Communications, whose contributions are widely recognized by the community. Moreover, beyond cross-pipeline comparisons, we explicitly ran same-corpus controlled experiments—comparing against PLD-adapted MOFFUSION on the exact same 109,529 structures (+9.24 pp [email protected] without MEF) and conducting a same-corpus Periodic U-DiT vs. 3D U-Net ablation (+9.4 pp assembly/validity with 6.5x fewer parameters)—while also providing a large-scale small-pore/full-range MOF corpus to the community.

We will continue refining our manuscript in the final version, and we sincerely hope that you will assess our full set of contributions—our coarse-to-fine pore-conditioned generation pipeline, chemical/building-block diversity, and curated MOF dataset—fairly and objectively.

Best regard :)

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.