acceptodds
Withdrawn from ICLR 2027

FragFlex: Fragment-Based Flexible Molecular Generation for Exploring Vast Chemical Space

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Drug discovery requires efficient exploration of an astronomically large chemical space. Molecular graph diffusion models offer a promising approach to this task, but many require graph size to be fixed before generation. Although atom-level insertion and deletion can enable flexible-size generation, jointly modeling molecular size, atom types, and bonds while maintaining chemical validity remains challenging. We introduce FragFlex, a flexible-length discrete diffusion model that performs node insertion and deletion at the fragment level. During denoising, FragFlex jointly learns the size, fragment identities, and inter-fragment connectivity of the fragment graph. A neural attachment-site predictor and valence-aware assembler then convert the generated fragment graph into an atom-level molecule. In large-scale experiments generating 10 million molecules, FragFlex achieves near-perfect chemical validity while maintaining high uniqueness, demonstrating a strong balance between the two.

Primary area:
generative models
Keywords:
molecular generation, discrete diffusion, graph generative models, fragment-based generation, drug discovery

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_ppRgOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents FragFlex, it combines fragment-level discrete diffusion with node insertion and deletion for variable-size molecular generation. A neural attachment-site predictor and valence-aware assembler reconstruct atom-level molecules.

Critical strengths
  1. The design is well motivated. Fragment-level generation reduces structural complexity, while the neural attachment predictor provides a flexible alternative to the attachment lookup table used by FragDiffusion.

  2. The large-scale evaluation is useful for examining repeated generations beyond small sample sets. FragFlex reports impressive validity and uniqueness metrics at 10 million samples, with additional scaffold and retrosynthetic-feasibility analyses.

Critical weaknesses
  1. The baseline comparison is not sufficiently controlled. Experiment section retains each implementation’s default dataset and configuration, without specifying the exact datasets and filtering differences. Coverage of the 100-fragment vocabulary is also unclear. These differences may affect validity, QED, SA and retrosynthetic success, making the improvements difficult to attribute to the method.

  2. The contribution of flexible-size diffusion is not isolated. Table 3 changes the assembly selection rule, insertion schedule and terminal state, but provides no fixed-size counterpart with the same neural assembler. Connectivity repair and beam search also influence the final results. A controlled comparison is needed to separate the benefit of size adaptation from assembly.

Questions for authors
  1. Are FragFlex and FragDiffusion trained on exactly the same molecules? Please specify dataset sizes, filtering rules and how molecules containing fragments outside the vocabulary are handled.

  2. See weaknesses.

Reviewer_5sxqOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces FragFlex, a discrete diffusion model over BRICS fragment graphs that allows the number of fragments to change during denoising through node insertion and deletion, and that converts the generated fragment graph into an atom-level molecule with a learned attachment-site predictor and a valence-aware beam-search assembler. The combination is sensible and clearly described: it joins the fragment-level coarsening of FragDiffusion with the insert/delete mechanism of GrIDDD, and it replaces FragDiffusion's precomputed attachment-mode lookup table with a neural site predictor. The evaluation at 10^7 generated samples is unusually large for this literature and is a welcome contribution in itself: FragFlex reaches 99.67% validity with 95.52% uniqueness and 9,520,600 distinct valid molecules, the highest count in Table 1, together with the best QED (0.764), SA (3.021) and AiZynthFinder success rate (43.2%).

Critical strengths

A smart combination of the fragment-level coarsening of FragDiffusion with the insert/delete mechanism of GrIDDD, replacing FragDiffusion's precomputed attachment-mode lookup table with a neural site predictor.

The evaluation at 10^7 generated samples is unusually large for this literature and is a welcome contribution in itself.

Critical weaknesses

There are many technical weaknesses (suggestions for improvements below), but I believe they can be addressed with some revisions. With those revisions, this paper would be a useful contribution to ICLR and the field.

Questions for authors

My concerns are mainly about what the experiments establish. First, the central claim is flexible size, but the paper never shows that the generated size distribution is better than that of fixed-size models, which is the only direct evidence that flexibility helps. Second, the evaluation is narrow: validity and uniqueness are close to saturation for every fragment-based model, no distributional metric is reported, novelty is asserted but not shown, and Table 5 shows that FragFlex produces fewer unique scaffolds than three of the five baselines. Third, several comparisons are hard to interpret: baselines use their own default datasets and configurations, no compute or parameter budget is matched, and no seeds or error bars are given anywhere. The empirical case for the claims needs substantial strengthening, so I recommend many revisions:

  1. The central claim of flexible size is never tested directly. The motivation is that fixing the graph size before denoising hurts quality, yet no experiment compares the generated size distribution with the training distribution, nor against a fixed-size model whose size is sampled from the training-set histogram. Please report size distributions (fragment count and heavy-atom count) for FragFlex and for all baselines against ZINC, with a distance such as Wasserstein or KL. Without this, flexibility is motivated but not demonstrated.
  2. Validity is close to saturation and may not measure what the paper suggests. For fragment-based models, chemical validity is largely enforced by construction: the assembler only adds bonds that pass the RDKit valence check and returns failure otherwise. The reported validity is therefore close to the assembly success rate, not an intrinsic property of the generative model. Please state this explicitly, report the assembly failure rate separately, and clarify whether failed assemblies are counted as invalid samples.
  3. Scaffold diversity is a weakness, and the discussion understates it. Table 5 shows 3,492,520 unique Bemis-Murcko scaffolds at 10^7 samples, below GrIDDD (5,889,181), IFH (5,402,044) and FragDiffusion (4,734,767). Appendix B.1 frames this as a diversity-quality trade-off, but the claim in the title and introduction is about exploring vast chemical space, which scaffold coverage measures more directly than the count of distinct molecules. Please move this result into the main text and temper the exploration claim accordingly.
  4. No distributional or property-matching metrics. Validity, uniqueness and novelty say little about whether the model learned the data distribution. Please add standard benchmarks such as Frechet Chemnet Distance, the KL divergence of physicochemical descriptors, scaffold similarity to the test set, and ideally the MOSES or GuacaMol suite, so that FragFlex can be compared with the wider literature.
  5. Novelty is asserted but not reported. Line 320 omits novelty because it is "nearly saturated at 100% across all evaluated models", but with a vocabulary of 100 fragments and at most 32 nodes the amount of training-set memorization at 10^7 samples is a genuine question. Please report novelty numbers.
  6. Baseline comparability. Section 4.1 says each baseline keeps "the default dataset and model configuration specified in its original implementation". If the models are trained on different ZINC subsets, differences in validity, uniqueness and synthesizability may partly reflect the training data rather than the method. Please retrain all models on one dataset split, or state clearly which subsets were used and what the effect is likely to be.
  7. No seeds, no error bars. All tables report single numbers, including the ablation in Table 3, where the differences are small: the default configuration has 99.67% validity and 99.94% uniqueness against 99.66/99.22 for Top1 and 99.76/99.32 for Middle. Differences of this size might be within seed-noise. It would be beneficial for the paper if the results of Table 3 were reported with at least three seeds.
  8. The ablation does not isolate the paper's own contributions. Table 3 varies the assembly selection rule, the schedule s_p and the terminal state, but not the two components presented as contributions: the neural attachment-site predictor and the valence-aware assembler. Please add the natural controls, namely FragDiffusion's lookup-table assembly combined with the FragFlex denoiser, and a greedy (beam width 1) or random-feasible assembler, to show how much each part contributes.
  9. The schedule choice is weakly justified. The paper sets s_p = 0, concentrating fragment insertions at the beginning of reverse sampling, and Table 3 shows that moving the peak to the midpoint gives higher validity (99.76 against 99.67) and lower uniqueness (99.32 against 99.94). Only two values of s_p are tested, with one seed. Please report a sweep and explain why early growth is expected to be better, beyond the observed difference.
  10. Uniform selection from the product pool is a design choice that trades quality for diversity. The assembler builds a pool of up to P_max = 16 valid products and then samples uniformly, which the paper justifies by the higher distinct-molecule count in Table 3. But Top1 gives essentially the same validity and the authors do not report whether the uniformly sampled products are chemically comparable to the top-ranked one. Please report QED, SA and the assembler score for uniform versus top-1 selection.
  11. The quality metrics deserve more careful interpretation. The QED advantage over JT-VAE is 0.764 against 0.762, which is within noise without seeds, and SA and AiZynthFinder results are strongly influenced by the fragment vocabulary, since BRICS fragments taken from ZINC are by construction synthetically accessible. Please report the same metrics for molecules sampled from the training set as a reference line, so the reader can see how much of the advantage comes from the vocabulary rather than the model.
  12. Missing fragment-based baselines and a trivial control. FragFM (Lee et al., 2026) and HierVAE (Jin et al., 2020) are discussed in related work but not compared. More importantly, a non-generative control is missing: sample a fragment count from the training distribution, sample fragments from the empirical marginal, connect them at random with the same valence-aware assembler. If this control also reaches high validity and uniqueness, the metrics in Table 1 are not discriminative and the paper needs stronger ones.
  13. Scope limitations should be stated. The model is limited to N_max = 32 fragments and a vocabulary of 100 fragment types, binary fragment edges cannot express multiple connections between the same fragment pair, and only single bonds are formed during assembly, so inter-fragment double or aromatic bonds appear impossible. Generation is also unconditional, while the motivation is drug discovery, where property or target conditioning is what matters. Please discuss the coverage of ZINC achievable under these restrictions and add a limitations paragraph.
Reviewer_yZCROpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper presents FragFlex, a fragment-based molecular generator based on discrete diffusion. During the forward process, FragFlex progressively deletes or corrupts fragments and their connections until only one remains. During the reverse process, it learns to insert or denoise fragments and connections, allowing to generate variable-sized graphs. These graphs are then assembled into valid molecules through through neural attachment-site prediction, followed by a beam search that rejects valence violeations, and final RDKit sanitization. The model is tested on ZINC, evaluating up to 10M generated molecules. FragFlex reports high validity and uniqueness, good QED and SA scores, and improved retrosynthetic planning success.

Critical strengths

Contribution and significance

The paper addresses a relevant problem and factual limitation of many discrete diffusion molecular generators (aside GrIDDD and few others, though they operate at the atom level): generating molecules of variable size without having to choose the number of fragments. It does so at the fragment level, which is clever as BRICS fragments represent chemically meaningful substructures. This modeling choice is also convenient, since it reduces the number of generative decisions compared with atom-level construction.

Clarity

The paper is readable and coherently structured, and the main contribution is clear, simple and well-motivated.

Critical weaknesses

Integration rather than novelty

The paper reads more like a practical integration of flexible-size diffusion with fragment-based molecular generation, rather than a fundamentally new diffusion mechanism. At its core, the main contribution is integrating the grow/shrink diffusion mechanism introduced by GrIDDD to the fragment-level setting of FragDiffusion, plus learned attachment prediction and constrained assembly. The author correctly and honestly do not misrepresent this as plain novelty. While this is not fatal, this contribution stands below ICLR's novelty bar.

Data and evaluation reproducibility/robustness

The paper does not report the precise ZINC dataset used (my guess is they used ZINC250k, but this doesn't appear to be cleared out in the paper), its size and splits, or how many molecules remain after filtering non-frequent fragments. This completely undermines reproducibility. I also have concerns about the experimental comparison, since the authors do not disclose whether baselines were retrained or run from existing checkpoints. Furthermore, the paper reports only point estimates, with no standard deviations or repeated seed results. This is particularly problematic because some differences between methods or ablations are small and may be caused by natural data/training variability.

Expressiveness concerns

The model uses a fixed vocabulary consisting of the top-100 frequent BRICS fragments extracted from the ZINC training data. This implies that FragFlex cannot represent or generate molecules whose BRICS decomposition requires out-of-vocabulary fragments, although it can produce novel combinations of in-vocabulary fragments. By my experience, the ZINC250k dataset contains 15k to 30k fragments depending on how the BRICS algorithm is set up. So, if the dataset is ZINC250k, this model cannot learn its distribution. This clashes even with the paper title, which claims "exploring vast chemical space". The chemical subspace that FragFlex can explore, while vast, is itself arguably much more smaller than, eg. atom-based molecular generators.

Missing ablation

The paper misses a crucial ablation to determine whether the neural attachment site predictor adds anything compared to simple frequency-based lookup (as done in FragDiffusion). Testing frequency-based lookup and the learned predictor through the same downstream assembler would help understand whether learning attachment sites actually improves validity, diversity, or molecular quality in general.

Validity comparison is unfair

FragFlex is claimed to outperform competitors numerically and as having the “strongest balance” of validity and uniqueness. A main role is played by the post-hoc assembly beam search, which can explore more alternatives than single-shot validity checks. I find this unfair with respect to competitors, which use a less "forgiving" assembly machinery.

Questions for authors

In its current form, the paper falls below the bar for ICLR, despite addressing an interesting problem and presenting a potentially valuable contribution. The concerns above are substantial and affect the experimental design and comparability of the results. Therefore, even resolving them would require a whole re-evaluation rather than a limited revision.

More than questions, I have suggestions based on the weaknesses above. I encourage the authors to resubmit once these points have been improved:

  • report dataset details and provenance. document or disclose filtering, preprocessing and data splitting.
  • address coverage of the vocabulary: how many ZINC molecules can/cannot represent? How are out-of-vocabulary fragments handled? At the same time, acknowledge that this model can represent a limited portion of chemical space with respect to other methods.
  • add an ablation of the neural attachment predictor to justify its existence. The most immediate one is to compare it with frequency-based lookup.
  • to make the validity comparison fairer, it would be better to present a comparison without the beam search component, or to add the beam search component to the baselines (although this is arguably harder). This would serve a double purpose: besides making the comparison fairer, it would demonstrating that it's needed to improve validity rates. Using beam search is not a problem by itself; the problem is claiming superior validity/uniqueness balance which is clearly driven by a component that the competitors don't have.
Reviewer_xtGrOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses large-scale molecular generation, focusing on producing many unique and valid molecules. The method applies discrete diffusion to molecular fragment graphs. Starting from a single root fragment, it jointly denoises fragment identities, connectivity, and graph size through insertion and deletion operations. A neural attachment-site predictor and a valence-aware assembler then convert the fragment graph into an atom-level molecule.

Critical strengths

The fragment-based formulation could support incorporating medicinal chemistry knowledge through fragment selection and assembly constraints, although this potential is not explored in the paper.

Critical weaknesses
  • Other fragment based molecular generation methods, such as FragFormer, are neither cited nor included in the comparisons. This makes it difficult to assess the novelty and advantages of the proposed approach.

  • The paper evaluates QED and synthesizability, but unlike models such as GCPN and GraphAF - which generate molecules while simultaneously optimizing for properties of interest - the goal here is simply to generate a large library of molecules. This can also be achieved by manipulating SELFIES-type string representations, which make it easier to rearrange and combine fragments to generate novel and valid molecules. It remains unclear how useful these libraries would be for a specific drug discovery task in practice - it would be more helpful if molecules were created based on pharmacophore, retrosynthesis, or property optimization objectives.

  • The paper does not include any analysis of computational complexity or runtime comparisons.

Questions for authors
  • Did you consider using Bemis-Murcko scaffolds or other medicinally meaningful substructures to define the fragments? How might this choice affect the usefulness of the generated molecules?

  • Can you comment on using this model to preserve certain scaffolds (those that preserve bioactivity) or exclude certain fragments containing toxic groups or similar?

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

Related papers

Discussion (0)

Sign in to comment.