- 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
-
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.
-
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.
-
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.
-
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
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 intervention | ZINC FCD ↓ | MOSES scaffold FCD ↓ |
|---|---|---|
| Full MIRRA | 0.3712 | 0.5494 |
| Remove identity projection and repair | 0.3804 | 0.5495 |
| Flatten terminal edge scores | 0.4648 | 0.7952 |
| Replace learned predictions with prior; keep full assembly | 4.2772 | 5.8734 |
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.
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.
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:
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):
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:
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).
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.