acceptodds
Under review as a conference paper at ICLR 2027

MIRRA: Fast, High-Fidelity Molecular Generation with Chemistry-Informed Discrete Flow Matching

Abstract

Fast, high-fidelity molecular generation requires accurate chemical choices and efficient sampling. Chemical building blocks reduce graph size, but make each node a choice among tens of thousands of related identities. We introduce MIRRA (Molecular Identity Representation, Refresh and Assembly), a discrete flow-matching generator that learns from chemical knowledge and evaluates identity probabilities only when needed, while keeping every identity in its vocabulary available. Chemical features share information across related blocks, and learned identity corrections distinguish them in molecular context. By deciding which positions will refresh before computing their probabilities, the sampler avoids unused predictions while preserving each sampling transition. Against the strongest baselines in each task, MIRRA reduces chemical and graph distribution errors on ZINC250k by 29–31% and 53–60%, respectively, with generation up to 8.6x faster. Property targeting yields an estimated 7.8–20.8x more distinct novel hits per second. On a reserved polymer test, MIRRA achieves 100% periodic validity with known chain topology, versus 57.0% for a polymer variational autoencoder. These results demonstrate fast, high-fidelity generation with the full chemical vocabulary available throughout sampling.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Flow Matching, Molecular Generation, Generative Models, AI for Science, Drug Discovery

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_NnGKOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents MIRRA, a fragment-level discrete flow-matching model that combines chemistry-informed scoring, learned identity corrections, and constrained assembly. Its refresh scheduler retains the full vocabulary while computing fragment scores only at positions selected for resampling. The ZINC250k experiments use 37,821 building blocks.

Critical strengths

The distinction between preserving the full vocabulary and restricting the candidate set is interesting. The candidate-restriction analysis provides a useful warning that naive subset renormalization can alter the sampling distribution.

Critical weaknesses
  1. The novelty is incremental. The chemistry-informed output head is closely related to FragFM's structure-aware fragment representation and the broader structured-output/label-embedding literature, while the identity-specific correction resembles a standard per-class residual. Prediction-independent scheduling also has related precedents.

  2. The role of constrained assembly is insufficiently disentangled from the learned flow model. MIRRA generates fragment identities X and fragment-level adjacency E, but molecular metrics are measured only after a decoder that applies attachment preferences, feasibility constraints, identity repair, and possible rejection. Since fragment identities and connectivity are strongly coupled but are refreshed separately, terminal graphs may be chemically incompatible before assembly. This makes it difficult to determine how much of the improvement comes from the learned flow versus post-hoc assembly.

  3. The main unconditional comparison is concentrated on ZINC250k. The MOSES experiments mainly compare internal output-head variants, and the gains are smaller and vary across runs. Extended evaluation on MOSES and GuacaMol, two larger and more complex benchmarks, would better establish whether the quality and runtime advantages generalize.

  4. Fig. 2 appears redundant, as most of its information is reported in Tables 1, 4, and 5. Similarly, Figs 4 and 5 also feel unnecessary. A compact table would communicate the same information more clearly. It’s unclear what additional insight these figures provide.

Questions for authors

See weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Reviewer NnGK (Part 1/2)

We thank the reviewer for the careful examination of MIRRA's contribution and for suggesting additional experiments that have strengthened the evaluation. MIRRA makes full-vocabulary generation practical through chemically informed learning and refresh scheduling. The revision expands §2 to distinguish these contributions, adds the requested external MOSES and GuacaMol comparisons, and further tests NPGen against seven baselines. The new evidence establishes both broad fidelity gains and the mechanisms behind them.

W1. Contribution relative to prior work.

The expanded Related Work (§2) makes three distinctions explicit:

  • Fragment models: MIRRA learns chemically structured probabilities normalized over the entire vocabulary, with contextual identity corrections. FragFM uses learned fragment embeddings and density-ratio training over stochastic fragment bags. MIRRA changes both how chemical information is shared and which choices are evaluated.
  • Structured-output and residual heads: the contribution is chemical correspondence within full-vocabulary molecular generation. Shuffling only output-feature assignments within attachment/frequency groups holds architecture, capacity and training fixed; correct assignment improves both metrics in both training pairs and all six paired draws (Table 3a).
  • Earlier schedulers: MIRRA decides repeated identity refreshes before full-vocabulary scoring, preserves the finite-step transition and retains full graph context. First-hitting samplers use one-way unmasking, MDLM/RADD exploit absorbing states, and learned planners predict selection policies (§3.2).

To our knowledge, MIRRA is among the first molecular generators to combine chemically informed full-vocabulary prediction with transition-preserving refresh scheduling. It gives 2.2–3.0× faster execution at fixed weights and steps, retaining all 37,821 ZINC identities. Restricting candidates instead raises FCD from 0.3665 to 0.9469 while being slightly slower (Figure 3).

MIRRA’s leading distribution-matching results across ZINC, MOSES, GuacaMol, and NPGen show that these innovations translate into substantial, consistent gains across diverse molecular datasets.

New Figure 7 (p. 20) makes the mechanism concrete: an actual eight-block trajectory requires 44 full-vocabulary evaluations instead of 2,000, and 81 backbone evaluations instead of 250. These operation counts show why scheduling saves work; Table 11 measures the end-to-end benefit.

W3. Added MOSES, GuacaMol and NPGen benchmarks.

On MOSES, MIRRA has the lowest FCD on both splits. We also include MolGPT (Bagal et al., 2022), PCF-VAE (Bhadwal et al., 2025), and VEGA (Delre et al. 2025):

MOSES generatorTest FCD ↓Scaffold FCD ↓
MIRRA (ours)0.08790.5392
FragFM0.11970.5967
MolGPT0.46451.0359
PCF-VAE (D1)0.29620.7467
VeGA0.180.68
Training reference-0.518 ± 0.012

MIRRA closes 73% of the strongest baseline’s scaffold FCD gap to the empirical training reference. The remaining gap is 3.7× smaller (Table 6).

On GuacaMol, MIRRA reduces FCD by 35.1% and improves both official distribution-learning scores:

GuacaMol generatorRaw FCD ↓FCD score ↑KL score ↑
MIRRA (ours)0.48270.9080.987
FragFM0.74340.8620.975
DiGress1.930.6800.929
DeFoG1.520.7380.977

MOSES uses three draws per checkpoint; GuacaMol uses three sampling streams per method (Table 2a–b; Appendix C.1).

We additionally evaluated NPGen: MIRRA leads all seven baselines on all five metrics, reducing FCD 40.9% and NP-score KLD 85.3% relative to the strongest baseline (Table 2c).

NPGen generatorFCD ↓NP-score KLD ↓Pathway divergence ↓Superclass divergence ↓Class divergence ↓
MIRRA (ours)0.8040.00550.00700.05260.2018
FragFM1.3600.03730.02120.15390.3825
DiGress2.0500.19000.02180.36831.1108
JT-VAE4.1000.51590.10481.21232.3343
DeFoG4.4750.15310.12480.41541.4546
HierVAE9.0770.30550.43280.60631.4185
GraphAF25.0080.86650.96623.32676.5374
MolHF31.3890.86281.82978.445910.4227

At fixed weights, 250 steps and full vocabulary, refresh also accelerates MOSES by 2.41× and GuacaMol by 2.17×, including assembly (Table 11). Thus the revision establishes improved quality and speed on the requested benchmarks, with further evidence from natural products.

AuthorsOpenReview ↗

Response to Reviewer NnGK (Part 2/2)

W2. Learned predictions drive fidelity under the same assembly rules.

The revised §3.3 makes the decoder's operation explicit. It reuses the final node and edge probabilities without another model evaluation. Assembly selects a connection tree whose degrees match the blocks' attachment counts, then pairs compatible sites. Edge probabilities guide the tree, while training-derived attachment preferences rank the pairings. Node probabilities guide identity correction if attachment counts are infeasible; a bounded repair search is tried if assembly fails. The decoder thus converts learned preferences into feasible molecular connections.

Node and edge refresh decisions are separate, but their predictions condition on the same incoming graph. On ZINC, 98.58% of terminal states already satisfy attachment-count feasibility, and projection changes only 0.26% of nodes (Appendix E.2).

The table holds tree and attachment constraints active in every row. Full MIRRA has the lowest FCD; assembly alone cannot recover its fidelity:

MIRRA interventionZINC FCD ↓MOSES scaffold FCD ↓
Full MIRRA0.37120.5494
Remove identity projection and repair0.38040.5495
Flatten terminal edge scores0.46480.7952
Replace learned predictions with prior; keep full assembly4.27725.8734

The prior uses training-frequency identities and flat edge scores. Removing learned predictions makes FCD 11.5× worse on ZINC and 10.7× worse on MOSES. Removing identity edits changes FCD by only 0.0092 and 0.0001, with validity still 96.47% and 97.58%. Flattening edge scores also worsens fidelity, establishing the contribution of learned connection preferences.

ZINC averages three draws; MOSES uses one. Failed attempts count against validity, and FCD evaluates accepted outputs in every row (Table 3b; Appendix E.2, Table 13). These controls identify learning as essential within the constrained representation.

W4. Figures 2, 4 and 5 make different findings immediately visible.

We appreciate the suggestion to streamline the presentation. We believe these plots are the most direct way to see the following relationships, complemented by exact values in the main text and supplement:

  • Figure 2 shows fidelity against sampling cost. Its complementary views connect chemical fidelity, graph fidelity, sampling steps and elapsed time. The plotted empirical training reference makes MIRRA's small remaining distributional gap visible alongside its speed advantage. Tables 1, 7 and 8 give the values.
  • Figure 4 separates validity from fidelity. Assembly-only identity correction improves validity but leaves FCD high; generation-time correction lowers FCD approximately 20×, approaching the plotted training reference. The side-by-side views make the different effects immediately apparent. Table 15 gives exact values; new Table 3 adds the stronger controls above.
  • Figure 5 shows the practical return from faster sampling: 7.8–20.8× more useful hits per second across six targets, alongside target-window fidelity. Readers can see both the size and consistency of the benefit. Tables 19–20 provide exact values and conditioning controls.

These figures expose trade-offs and intervention effects that are harder to see from isolated table entries. Alongside the expanded related work and MOSES, GuacaMol and NPGen results in our first reply, we hope the clarified method and new experiments address the remaining concerns and strengthen the case for MIRRA's contribution. We would welcome the reviewer's reassessment of soundness and the overall recommendation.

AuthorsOpenReview ↗

Follow-up before the discussion deadline

Dear Reviewer NnGK, as the discussion period closes on November 18, we wanted to follow up on whether our responses and revised manuscript address your concerns, particularly the clarified contribution relative to prior work (§2), additional benchmark comparisons (Table 2), and controls separating the contributions of learned predictions and assembly constraints (Table 3b; Appendix E.2). We would be glad to clarify any remaining questions before the deadline. Thank you again for your detailed review and constructive suggestions.

Thanks for the detailed reply and new experiments. I’m confused about why MolGPT (2022) was included on MOSES but omitted on GuacaMol. It seems that MolGPT shows comparable results. Several recent methods even report stronger results. Hence, I conservatively keep my original score.

Reviewer_pm8dOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This manuscript proposes a fragment-based molecular generation framework using discrete flow matching, called MIRRA. The method combines chemical features with learned identity corrections to model a large vocabulary of molecular building blocks. It also introduces a refresh scheduling strategy that avoids computing identity probabilities for positions that do not need to be updated. The experimental results on ZINC250k show improvements in FCD and NSPDK over several existing molecular generation methods, including FragFM, while reducing sampling time. The authors also evaluate property-conditioned generation and polymer repeat-unit generation. Ablation studies examine the contributions of chemical features, identity corrections, and refresh scheduling.

Critical strengths

Firstly, I think this is the most convincing part of the paper. By deciding which positions need to be refreshed before computing their probabilities, the method avoids unnecessary computation while preserving the original sampling transition. The fixed-checkpoint comparison in Figure 3 shows a 3.0× speedup with almost unchanged FCD. This is a useful improvement for molecular generation with large fragment vocabularies.

In addition, the proposed MIRRA achieves better FCD and NSPDK than the compared methods on ZINC250k, even with relatively few sampling steps. I also appreciate the experiments comparing correctly assigned and shuffled chemical features, which provide evidence that chemical information contributes beyond model capacity. The property-conditioned experiments further show the potential benefit of faster sampling.

Critical weaknesses
  • FragFM is the strongest baseline, but the two methods use different architectures, representations, and training procedures. Appendix B.2 also indicates differences in the number of training updates and software environments. Although the comparison is useful for evaluating complete generators, it is difficult to determine how much of the improvement comes specifically from the proposed chemical representation and refresh scheduling. A more controlled comparison would make the main experimental claims stronger.

  • This manuscript reports strong performance on ZINC250k, but the MOSES improvements are smaller and vary across seeds and sampling steps. In addition, the polymer experiment gives MIRRA a known linear chain topology, while the baselines must generate complete repeat-unit encodings. I do not think this is a fully comparable setting. The 100% periodic validity result is encouraging, but it is difficult to attribute the large difference entirely to the proposed generation method.

Questions for authors
  • In the comparison with FragFM, how much of the improvement can be attributed to the chemical representation and refresh scheduling rather than differences in architecture and training? Could the authors provide a more controlled comparison, for example by applying the refresh scheduling strategy to FragFM or comparing the methods under a more closely matched computational budget?

  • In the polymer experiments, MIRRA is provided with known linear chain topology, whereas the baselines generate complete repeat-unit encodings. Could the authors clarify how much this difference contributes to the reported validity improvement? Would the advantage remain if the methods were evaluated under comparable topology constraints?

3 comments
AuthorsOpenReview ↗

Response to Reviewer pm8d (Part 1/2)

We sincerely thank the reviewer for the thoughtful assessment of refresh scheduling and for proposing matched-budget and topology controls that have strengthened the evaluation. The new experiments show that MIRRA's fidelity advantage persists under matched training, chemical alignment improves learning, and scheduling accelerates the same trained model. We also report external comparisons on MOSES, GuacaMol and NPGen, alongside the matched-topology polymer control.

W1 / Q1. Equal-budget training.

FragFM has the best baseline FCD and NSPDK in the full ZINC comparison (Table 1). We therefore use this strongest comparator for retraining on the same 208,302 molecules, batch size 128 and 244,050 total optimizer updates, including both FragFM training stages.

Equal-budget retrainingFCD ↓NSPDK × 10³ ↓
MIRRA0.3667 ± 0.00320.1636 ± 0.0009
FragFM0.5206 ± 0.00140.2876 ± 0.0201

Values are mean ± SD across two fresh training runs, with three draws per run; MIRRA also wins at the final checkpoints (Table 5). Data and updates are matched; architecture and compute per update remain method-specific.

MIRRA retains 29.6% lower FCD and 43.1% lower NSPDK. These matched-budget runs also closely reproduce MIRRA’s original performance (mean FCD 0.3659; mean NSPDK ×10³ 0.1610). The 29.6% FCD reduction closely matches the original submission’s 30.6%, confirming that MIRRA’s advantage persists across fresh runs under matched training budgets.

Chemical alignment improves MIRRA at fixed architecture and training.

We compare MIRRA with an otherwise identical version whose output chemical features are shuffled among fragments with matched attachment types and frequencies. Correct chemical assignment lowers FCD by 4.4% and NSPDK by 5.0%:

Alignment-control comparisonFCD ↓NSPDK × 10³ ↓
MIRRA: correct chemical features0.1554 ± 0.00180.1118 ± 0.0050
MIRRA: shuffled chemical features0.1625 ± 0.00010.1177 ± 0.0006

Values are mean ± SD across two training runs per variant, each averaged over three sampling draws. The advantage holds in both training pairs and all six paired draws. Both rows use the same alignment-control evaluation (Table 3a; Appendix E.1). This isolates the value of chemical correspondence beyond architecture, capacity, attachment type and fragment frequency.

Refresh scheduling accelerates the same MIRRA model.

Model and execution comparisonZINC250k speedupMOSES speedupGuacaMol speedup
MIRRA refresh vs. dense (end-to-end)3.01×2.41×2.17×

Each pair fixes architecture, weights, steps, vocabulary, hardware and software; timings include assembly (Figure 3; Table 11). ZINC peak memory falls 37%. The finite-step transition is preserved by construction; paired decoded molecules agree in 99.56% of MOSES and 99.98% of GuacaMol attempts, with numerical discrepancies examined in Appendix D.4. New Figure 7 makes the avoided computations visible along an actual trajectory.

We followed the reviewer's suggested matched-budget alternative; transferring refresh to FragFM's different bag-based sampler remains a separate experiment. Together, equal-budget training, feature shuffling and fixed-model timing isolate the overall fidelity gain, the chemical contribution and the execution benefit.

AuthorsOpenReview ↗

Response to Reviewer pm8d (Part 2/2)

W2. MOSES results across training runs, plus GuacaMol and NPGen.

Every MIRRA training-run mean has lower FCD than all the baselines on both MOSES splits. We also include PCF-VAE (Bhadwal et al 2025) and VeGA (Delre et al 2025).

MOSES generatorTest FCD ↓Scaffold FCD ↓
MIRRA (ours)0.0879 (0.0874–0.0886)0.5392 (0.5329–0.5485)
FragFM0.11970.5967
MolGPT0.46451.0359
PCF-VAE (D1)0.29620.7467
VeGA0.180.68
Training reference-0.518 ± 0.012

Table 2a uses three 25,000-attempt draws; MIRRA uses 250 steps (Appendix C.1).

MIRRA closes 73% of FragFM's scaffold FCD gap to the empirical training reference. The remaining gap is 3.7× smaller (Table 6). The reference is an empirical comparison, not a lower bound. The internal head-ablation benefit varies across runs and steps; the new external comparison establishes the complete generator's consistent advantage.

The additional benchmarks strengthen this finding (Table 2b–c):

  1. GuacaMol FCD falls 35.1%, with both official distribution-learning scores improved;
GuacaMolRaw FCD ↓FCD score ↑KL score ↑
MIRRA (ours)0.48270.9080.987
FragFM0.74340.8620.975
DiGress1.930.6800.929
DeFoG1.520.7380.977
  1. on NPGen, MIRRA leads all seven baselines on all five fidelity metrics, including 40.9% lower FCD than the strongest baseline.
NPGenFCD ↓NP-score KLD ↓Pathway divergence ↓Superclass divergence ↓Class divergence ↓
MIRRA (ours)0.8040.00550.00700.05260.2018
FragFM1.3600.03730.02120.15390.3825
DiGress2.0500.19000.02180.36831.1108
JT-VAE4.1000.51590.10481.21232.3343
DeFoG4.4750.15310.12480.41541.4546
HierVAE9.0770.30550.43280.60631.4185
GraphAF25.0080.86650.96623.32676.5374
MolHF31.3890.86281.82978.445910.4227

W2 / Q2. Learning under identical polymer topology.

Holding chains, lengths, vocabulary and assembly identical gives 100% periodic validity in every row, but learning dramatically improves structural fidelity:

Same polymer topology and assemblyPeriodic validity (%)Test MMD² × 10⁵ ↓
Untrained empirical block prior100.0373.560
Trained conventional head100.04.462
MIRRA aligned head100.02.683

The representation ensures feasibility; learning reduces structural discrepancy 84× with the conventional head and 139× with MIRRA. Chemical alignment lowers the mean error a further 40%, although this additional advantage reverses in one training pair. Both learned heads retain the large improvement over the prior. Results use two training pairs and two draws per model on 9,949 reserved identities (Table 4b; Appendix H.1).

We also provided endpoint constraints to external checkpoints: periodic validity rises from 57.0% to 58.4% for OMG-VAE and from 36.5% to 46.5% for Graph DiT (Appendix H.2). The stronger control above matches the full representation. Accordingly, §6 attributes 100% validity to feasible topology and the substantial distributional improvement to learning.

These controls directly address robustness and attribution, while the additional benchmarks establish broader empirical strength. We hope they support a stronger recommendation.

AuthorsOpenReview ↗

Follow-up before the discussion deadline

Dear Reviewer pm8d, as the discussion period closes on November 18, we wanted to follow up on whether our responses and revised manuscript address your concerns, particularly the matched-budget comparison (Table 5), chemical-alignment control (Table 3a), expanded MOSES evaluation (Table 2a), and matched-topology polymer experiment (Table 4b; Appendix H.1). We would be glad to clarify any remaining questions before the deadline. Thank you again for your detailed review and constructive suggestions.

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

This paper studies the problem of molecular generation. The authors propose MIRRA, which generates the building-block graphs with discrete flow matching and then decodes the block graph into a molecule. MIRRA achieves state-of-the-art performance in ZINC250k and polymer.

Critical strengths
  • The paper is easy to follow.

  • MIRRA can achieve high fidelity and keep efficient.

Critical weaknesses
  • Some important baselines are missing, such as VAE-based methods like JT-VAE and flow-based models such as MoFlow and MolHF. These non-diffusion methods are efficient.

  • The evaluation is limited to old datasets and lacks more recent benchmarks, such as MOSES and NPGen.

  • The proposed method could be evaluated on more practical tasks, such as pocket-conditioned molecular generation.

  • The choice of evaluation metrics could be improved. In particular, Validity and Uniqueness may no longer be sufficiently informative, as most current models achieve nearly 100% on these metrics. More meaningful metrics should be considered to better characterize the actual diversity of the generated molecules.

Questions for authors

See Weaknesses

4 comments
AuthorsOpenReview ↗

Response to Reviewer VAqQ (Part 1/2)

We thank the reviewer for recognizing MIRRA's combination of high fidelity and efficient generation, and for highlighting the importance of broader evaluation. We added JT-VAE, MoFlow and MolHF on ZINC, external MOSES comparisons, and seven baselines on NPGen. We also evaluated GuacaMol, extending the revision across all three additional benchmarks requested across the reviews. MIRRA leads the evaluated methods on both MOSES splits and every NPGen fidelity metric, and improves both GuacaMol distribution-learning scores (Tables 1–2).

W1. Added VAE and normalizing-flow baselines.

MIRRA achieves 9.8× lower FCD and 26.7× lower NSPDK than the strongest of the three requested baselines, with 99.82% validity. Bold marks the best value in each metric; MIRRA's second-best validity is underlined.

GeneratorFCD ↓NSPDK × 10³ ↓Validity (%) ↑
MIRRA (ours)0.371 ± 0.0080.157 ± 0.01299.82‾\underline{99.82}
JT-VAE3.653 ± 0.0424.199 ± 0.11499.76
MoFlow19.896 ± 0.03439.065 ± 0.49499.99
MolHF22.368 ± 0.05137.386 ± 0.28294.84

Values are mean ± sampling SD over three 10,000-attempt draws, using the same references and scoring (Table 1a; Appendix C.1). The gains therefore remain substantial against methods with near-perfect validity.

FragFM is the strongest baseline on both fidelity metrics in the full ZINC comparison (Table 1), making it the most demanding comparator. Its own study also reports the best FCD among its compared methods on ZINC, MOSES, GuacaMol and NPGen (Lee et al., 2026). Retraining MIRRA and FragFM [1] on identical data and update budgets retains 29.6% lower FCD and 43.1% lower NSPDK across two fresh runs per method (Table 5).

W2. MOSES, NPGen, and the additional GuacaMol evaluation.

MIRRA has the lowest FCD on both MOSES splits, and every one of its three training-run means outperforms other baselines, including new tested ones, PCF-VAE [2], and VeGA [3]:

MOSES GeneratorTest FCD ↓Scaffold FCD ↓
MIRRA (ours)0.08790.5392
FragFM0.11970.5967
VeGA0.18000.6800
PCF-VAE (D1)0.29620.7467
MolGPT0.46451.0359
Training reference—0.518 ± 0.012

Each checkpoint uses three 25,000-attempt draws (Table 2a; Appendix C.1). MIRRA closes 73% of FragFM’s scaffold FCD gap to the empirical training reference. The remaining gap is 3.7× smaller (Table 6).

On NPGen, MIRRA ranks first against all seven baselines on all five fidelity metrics:

NPGen generatorFCD ↓NP-score KLD ↓Pathway divergence ↓Superclass divergence ↓Class divergence ↓
MIRRA (ours)0.8040.00550.00700.05260.2018
FragFM1.3600.03730.02120.15390.3825
DiGress2.0500.19000.02180.36831.1108
JT-VAE4.1000.51590.10481.21232.3343
DeFoG4.4750.15310.12480.41541.4546
HierVAE9.0770.30550.43280.60631.4185
GraphAF25.0080.86650.96623.32676.5374
MolHF31.3890.86281.82978.445910.4227

MIRRA lowers FCD by 40.9% and NP-score KLD by 85.3%; pathway, superclass and class divergences fall by 67.0%, 65.8% and 47.2%, respectively.

All methods are scored on 25,000 raw attempts; MIRRA averages three draws. Validity, uniqueness and novelty remain 99.17%, 97.50% and 94.32%. Table 2c reports all eight methods.

The additional GuacaMol evaluation reduces FCD by 35.1% and improves the official FCD score from 0.862 to 0.908 and KL score from 0.975 to 0.987 (Table 2b). Thus the expanded comparison establishes substantial fidelity gains across generative families, drug-like molecules and natural products.


[1] Lee et al. FragFM: Hierarchical Framework for Efficient Molecule Generation via Fragment-Level Discrete Flow Matching. ICLR, 2026.

[2] Bhadwal et al. PCF-VAE: posterior collapse free variational autoencoder for de novo drug design. Scientific Reports, 2025.

[3] Delre and Lavecchia. VeGA: A Versatile Generative Architecture for Bioactive Molecules across Multiple Therapeutic Targets. Journal of Chemical Information and Modeling, 2025.

AuthorsOpenReview ↗

Response to Reviewer VAqQ (Part 2/2)

W3. Practical molecular generation.

We agree that pocket-conditioned molecular generation is a valuable next application, and we thank the reviewer for suggesting it. We will consider extending MIRRA to this setting in future work.

The current property-targeting study evaluates low, median and high QED and logP targets. A useful hit must be valid, within the target window, distinct and absent from training. Across all six targets, MIRRA produces an estimated 7.8–20.8× more useful hits per second than the strongest tested baseline at each target, with the lowest target-window FCD in five of six settings and validity above 99.5% throughout (Figure 5; Table 19).

All methods use 500 sampling steps and 4,096 attempts per target. Rates combine measured hit yields with separately measured generation times; true property labels outperform shuffled-label and unconditional controls at every target (Table 20). The speed advantage therefore translates into more novel candidates meeting the desired property.

W4. Diversity beyond validity and uniqueness.

The expanded evaluation adds fragment, scaffold and nearest-neighbor similarity to FCD and NSPDK, comparing the generated population with the test distribution across complementary levels of chemical structure. MIRRA improves all five measures under equal training budgets; bold marks the better value.

Equal-budget ZINC generatorFCD ↓NSPDK × 10³ ↓Fragment similarity ↑Scaffold similarity ↑Nearest-neighbor similarity ↑
MIRRA0.36670.16360.99930.60250.4490
FragFM0.52060.28760.99610.53680.4215

Both methods use two training runs and three draws per run, with every metric computed on the same generated populations (Table 5; Appendix C.2).

The advantage extends to fragment and scaffold distributions and molecular similarity to held-out data, providing a broader assessment than validity and uniqueness alone.

Table 5 also reports additional metrics showing MIRRA’s competitive performance against the baseline: internal diversity of 0.8712 (FragFM: 0.8732) and 7,898 distinct scaffolds (FragFM: 8,140), alongside its stronger distribution matching.

Together with the new MOSES, GuacaMol and NPGen comparisons in our first reply, these results strengthen the evidence for both broad molecular fidelity and useful property-targeted generation. We hope they support a more positive assessment of the contribution and empirical strength.

Thanks for the detailed response and clarifications. I appreciate the authors' efforts in addressing my concerns. However, I still think the method has limited applicability and offers only minor algorithmic improvements over FragFM. Therefore, my overall assessment remains unchanged, and I will maintain my original score.

We thank the reviewer for the follow-up. The items bearing most directly on the two concerns are §3.2-3.3, Table 6 and Figure 7.

  1. MIRRA is not a variant of FragFM/other baselines. Fragment-based generation has been a shared foundation for decades. But MIRRA designs different sampling and reconstruction algorithms, making it materially distinct from FragFM. Specifically, FragFM uses stochastic fragment bags and a trained coarse-to-fine autoencoder. MIRRA uses refresh scheduling (§3.2) to enable the full chemistry vocabulary, and assembly without a separately trained neural decoder (§3.3).

  2. The improvements are not minor. MIRRA is much closer to the empirical training reference than FragFM (Figure 2). Relative to this reference, MIRRA’s remaining FCD gap is 6.0× smaller than FragFM’s on equal-budget ZINC and 3.7× smaller on the MOSES scaffold split (Table 6). MIRRA robustly leads across ZINC, MOSES, GuacaMol and NPGen. MIRRA also generates 8.6× faster than FragFM at better fidelity.

    Thus, the gains are substantial in both fidelity and speed: Refresh scheduling is not a quality–speed trade-off (canonically identical molecules in >99.5% of attempts, ~3.0× faster; §4.3, Table 11).

  3. On applicability: The scheduling principle applies to a broader class of discrete generators. Refresh gating is applicable to compatible discrete flow and diffusion models for sequences and graphs, with particular value for large vocabularies. The new Figure 7 makes the mechanism concrete. Given its advantage, we really believe this could benefit a wide range of models and applications in the community.

    We also report a result that may be another useful contribution: restricting the candidate set, the standard remedy for large fragment vocabularies, gave no speedup and raised FCD 2.6×, through a proposal bias that finer time discretization does not remove (Appendix D.3). That is a finding about a technique the field uses, not an increment over one baseline.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to the Area Chair and All Reviewers

We thank the Area Chair and the three reviewers for their careful and constructive reports. Their feedback has helped us strengthen the evaluation and clarify the presentation. We have uploaded the revised manuscript and a supplementary comparison with the original submission. The new experiments appear in §4–§6 and the corresponding appendices. We also edited the text throughout for clarity and brevity; the underlying method and theoretical results are unchanged. Added and revised text is marked in the supplementary comparison.

We are encouraged that the reviews recognize:

  • Strong molecular fidelity with efficient generation, even at few sampling steps (Reviewers VAqQ and pm8d).
  • The value of retaining every fragment in the vocabulary throughout sampling (Reviewer NnGK).
  • Refresh scheduling that avoids unused predictions while preserving the sampling transition (Reviewer pm8d).

The reviews raise three main questions about the empirical evidence. We summarize the short answers below. Detailed numerical results, comparison tables and experimental controls appear in the individual replies.

  1. How does MIRRA perform on MOSES, GuacaMol and NPGen against additional baselines? (Reviewers VAqQ, pm8d and NnGK)

    • MOSES: lowest FCD against FragFM, MolGPT and other baselines on both splits; all three MIRRA training runs outperform the strongest baseline, FragFM, with 26.6% lower mean Test FCD and a 3.7× smaller scaffold FCD gap to the training reference (Table 6).
    • GuacaMol: 35.1% lower FCD than FragFM and improvements in both official distribution-learning scores.
    • NPGen: best on all five fidelity metrics against all seven baselines (Table 2), including 40.9% lower FCD and 85.3% lower NP-score KLD than the strongest baseline. Table 2 gives all three benchmark comparisons; Table 1 adds JT-VAE, MoFlow and MolHF on ZINC.
  2. Do the gains persist under matched training, and what produces them? (Reviewers pm8d and NnGK)

    • Matched-budget retraining closely reproduces MIRRA’s original fidelity while retaining 29.6% lower FCD and 43.1% lower NSPDK on ZINC (Table 5). Shuffling chemical features at fixed architecture and training confirms the benefit of correct chemical assignment (Table 3a).
    • Keeping full assembly but removing learned predictions makes FCD 11.5× worse on ZINC and 10.7× worse on MOSES (Table 3b). With polymer topology held identical, learning reduces structural discrepancy 139× (Table 4b).
    • §3.3 now makes assembly explicit: reuse the final predictions to choose a connection tree and pair compatible attachment sites, without another model evaluation. This compact decoder enforces feasibility; the new controls distinguish its role from learned fidelity.
  3. Does faster sampling produce useful, diverse molecules? (Reviewer VAqQ; Reviewers pm8d and NnGK on efficiency)

    • Fixed-model controls establish 2.2–3.0× end-to-end scheduling speedups across three datasets (Figure 3; Table 11). Property targeting yields an estimated 7.8–20.8× more distinct novel hits per second (Figure 5).
    • The expanded diversity evaluation shows stronger distribution matching while retaining high internal diversity, 97.4% novelty and thousands of distinct scaffolds (Table 5).

Together, the new experiments establish leading fidelity across ZINC, MOSES, GuacaMol and NPGen and identify the sources of the gains. We have substantially expanded §2 to distinguish MIRRA from prior molecular generators, structured-output models and efficient samplers. To our knowledge, MIRRA is among the first molecular generators to combine chemically informed full-vocabulary prediction with transition-preserving refresh scheduling. The broader contribution is a reusable approach to large-vocabulary discrete generation: share information across related identities and evaluate their probabilities only when needed. New Figure 7 illustrates this principle along an actual molecular trajectory.

We welcome further discussion and would be happy to conduct additional experiments or provide clarification as needed.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.