acceptodds
Under review as a conference paper at ICLR 2027

Synthesizability-Guided Discrete Diffusion Model for Molecular Generation

Abstract

Molecular generative models can propose candidates with favorable predicted properties, but are not guaranteed to generate synthesizable molecules. Candidates that are hard to synthesize cannot readily enter wet-lab experiments. Computer-aided synthesis planning (CASP) tools provide route-specific evidence towards synthesizability, but are too slow to integrate into generation. In this paper, we introduce *SynGuide*, a guidance mechanism that distills CASP outcomes into a noise-conditioned guidance signal that guides the generation of discrete diffusion models. We define a fixed-budget retrosynthetic-accessibility (RA) label indicating whether a CASP tool finds a route to purchasable building blocks within a fixed search budget. In order to build guided generative models, we then train a noise-conditioned predictor of the RA-label and use its gradient to guide a frozen discrete diffusion generator. Under RA-label guidance, our method increases the measured planner solve rate by 24%. Across three different protein targets, *SynGuide* achieves *state-of-the-art* performance improving synthesizability. This work provides a practical way to incorporate planner-derived synthesizability objectives into molecular generation without placing expensive retrosynthesis searches in the sampling loop.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Retrosynthetic Accessibility, De novo generation, Molecular design, Discrete diffusion, Guidance

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 3 reviews, above 27% of the others.
Reviewer_DgsEOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary
  • The paper introduces SynGuide a synthesizability guided molecule generation scheme built on GenMol. SynGuide steers GenMol with AiZynthFinder proxy model leading to higher ratio of planner suceeding molecules.
  • The predictor is trained on noisy SAFE strings and while generation the predictor is provided a GenMol generated token probability distribution. The predictor gradient is then used to steer the token logits of SAFE string. Pracitically, the syntehsizability predicotr and docking predictor is balanced while generation.
  • Paper includes experiments on three targets (ACAA1, c-MET, BRD4) and evaluated by planner succes, docking score, and hit ratio. SynGuide improves hit ratio and syntehsiabaility.
Critical strengths

None

Critical weaknesses
  • Limited methodological novelty and insufficient validation of the proposed modifications: The paper primarily combines a frozen GenMol backbone with a planner-supervised predictor and gradient-based guidance. Learning surrogates of retrosynthesis planners has been explored in RAscore [1], while using a learned synthesizability classifier to guide molecular diffusion has been studied in SYNC [2]. Predictor-based guidance for discrete diffusion and flow models is also introduced in prior works. However, the specific design choices in SynGuide (guidance through a soft clean prediction and gradient normalization) are not sufficiently isolated in the experiments. Overall, the paper demonstrates the practical utility of a combinations of existing techniques, but does not sufficiently show a distinct methodological contribution.

[1] Retrosynthetic accessibility score (RAscore) – rapid machine learned synthesizability classification from AI driven retrosynthetic planning (Chemical Science, 2021)

[2] SYNC: Measuring and Advancing Synthesizability in Structure-Based Drug Design (ICLR, 2026)

Questions for authors
  • Regarding evaluation and the docking threshold: The main evaluation uses a relatively mild docking threshold (42.15th percentile; Figure 7). Target-guided molecular generation aims to generate molecules with favorable docking scores, as measured by metrics such as Top-1 and Top-10. Is there a reason the authors chose this threshold? Moreover, under stricter docking thresholds in Figure 7, SAFE-GPT appears to achieve a higher Joint hit rate than the proposed method. Could the authors provide a quantitative comparison under stricter thresholds?
3 comments
AuthorsOpenReview ↗

Response to the Reviewer DgsE [1/2]

We thank the reviewer for the insightful feedback and valuable comments!

W1 — prior art and isolation of the design choices

We agree that both pieces of prior art are relevant. We also agree that the reviewer's concern about isolating the design choices is central.

Prior art. RAscore did learn a planner surrogate, and SYNC did guide molecular diffusion with a learned synthesizability classifier — its Eq. (7) is the same posterior-mean (x̂₀) approximation we use, with the same justification. We had missed SYNC; it is now in the bibliography and will be discussed in the text. Differences: (Table B in the General Comment above.) The axis that matters most is the state space. SYNC operates on continuous-coordinate 3D diffusion, where the intermediate state is a noisy coordinate field. Masked discrete diffusion has no such state: the intermediate object is a partially masked token sequence, so the predictor must be defined on something differentiable that does not exist a priori, and the guidance must act on logits at masked positions only.

RAscore. the comparison that matters is against using such a surrogate the way RAscore is used — as a post-hoc filter. We ran it. Filtering unguided generations by the RA predictor gives Joint rates of 20.9 / 18.5 / 18.7 (ACAA1 / BRD4(2) / c-MET), and filtering docking-only generations gives 39.4 / 23.2 / 25.5, against 52.2 / 45.6 / 47.4 for guidance at λ_syn=0.3. The reason is arithmetic: filtered yield is keep-rate × purity, so to match our unfiltered yield while keeping a quarter of the samples a filter would need purity above 209%. A surrogate used as a filter cannot create molecules the backbone never proposed; used as guidance it changes what gets proposed.

Isolation. Contribution of each design choice: (Table H in the General Comment above.) The last two ablations are the ones raised in the review. Both were run at the paper's operating point (λ_syn = 0.3, γ = 2.0, guidance scale 0.028) as single-factor changes from the main arm, paired over the same nine (target, seed) groups, with 100% AiZynthFinder label coverage on all 7,348 newly generated molecules, so Solve and Joint are comparable to Table 1. Mean ± sd over 3 seeds: (Table I in the General Comment above.) They fail in two different ways, separated by the probability-shift column.

Normalization controls whether guidance acts at all. Removing normalization leaves the per-step probability shift at 0.0014 versus 0.260 — a 186× collapse — and the sampler becomes statistically indistinguishable from the unguided backbone on docking: mean dock −6.93 vs −6.93 (paired p = 0.388), % ≤ τ 33.6 vs 33.3 (p = 0.444). Solve rises by only +0.52 pp over unguided. Guidance is not reweighted but numerically inert.

Two effects contribute to this result: one is a tuning artefact; the other is not. The raw gradients have RMS values of approximately 2–5 × 10⁻³. At the tuned γ and scale, they add a perturbation of only ~1.9 × 10⁻⁴ to backbone logits of order 1–10—298× smaller than the normalized perturbation of 0.056. Increasing the guidance scale could compensate for this effect.

A single global scale cannot correct the second effect: the ratio of the two objectives’ raw gradient RMS values varies by target (syn/dock = 2.45 for ACAA1, 2.40 for BRD4(2) and 1.49 for c-MET). Thus, a nominal λ_syn = 0.3 corresponds to effective values of 0.512, 0.507 and 0.389, respectively—a spread of 0.123. After normalization, the effective value is 0.300 for all three targets by construction. λ is therefore interpretable as a trade-off only after normalization.

The soft relaxation controls the direction of guidance, not its strength. The hard-argmax arm has probability shift 0.254, essentially identical to the main arm's 0.260, yet reallocates that same perturbation: Solve rises to 79.36 (+5.68 pp, p < 0.0001, better in 9/9 groups) while % ≤ τ falls 69.3 → 49.7 and the joint objective drops 9.80 pp (p < 0.0001, worse in 9/9). Straight-through argmax discards the softmax Jacobian, so after renormalisation each term is still unit-RMS but points elsewhere — empirically at the synthesis objective at the expense of binding. This isolates the soft relaxation specifically, as distinct from conditioning on a clean-sequence estimate, the choice SYNC shares.

Revised contribution. We do not claim a new guidance family. We claim a fixed-budget, declared-configuration planner label as the supervision target; a predictor trained to read partially masked sequences so it is usable mid-trajectory; and a normalised multi-objective rule for masked discrete diffusion — each isolated by a paired ablation, with the joint objective shown necessary (both single-objective endpoints are worse) and the label obtainable 29.7× cheaper (AUROC 0.865 at B=1 vs 0.958 at B=500). We will rewrite the contributions and the abstract to claim exactly this and no more.

AuthorsOpenReview ↗

Response to the Reviewer DgsE [2/2]

Q1— the threshold sweep

The quantitative comparison confirms that SAFE-GPT overtakes SynGuide at stricter thresholds. Joint hit rate (%) by threshold percentile:

(Table J in the General Comment above.)

The crossover is between the 25th and 10th percentile: we lead at the operating point and at p25 on all three targets, and below about p10 SAFE-GPT overtakes us on c-MET and BRD4(2), with ToolMol also ahead in the extreme tail. We will therefore delete the §5.3 sentence claiming that method ordering is stable across the threshold sweep and revise the main text to match.

On why 42.15: it is the 42.15th percentile of our 22,792-molecule label pool's docking distribution for each target, applied identically to all methods, chosen as a percentile rather than a fixed energy because AutoDock4 scores are not comparable across pockets. We agree it is mild, and the revision will report this full sweep in the main text instead of a single operating point.

The conclusion is regime-dependent. SynGuide is preferable when the goal is many candidates that are both routable and acceptable binders; SAFE-GPT is preferable for the extreme-affinity tail, because its docking distribution has a much heavier tail (BRD4(2) mean −8.54 vs our −8.12, Top-1 −12.48 vs −10.47) at a lower solve rate (58.0% vs 70.4%). We will state this trade-off.

Thank you for the detailed experiments. The additional ablations clarify the effects of the design choices and provide stronger support for the practical usefulness of the proposed method. However, my concern about the limited methodological contribution remains. Considering the additional experimental evidence, I have raised my score.

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

The paper proposes SynGuide, a guidance method that steers a frozen masked discrete diffusion model (GenMol) toward molecules which can be solved by a retrosynthesis planner. The authors define a binary retrosynthetic-accessibility (RA) label, determined by whether AiZynthFinder finds a route to purchasable building blocks within 500 MCTS iterations. They then train a noise-conditioned classifier on partially masked SAFE sequences to predict the binary target. At sampling time, the classifier scores the denoiser's predicted token distribution at masked positions. The gradient of that score with respect to the denoiser's logits is added to those logits at masked positions to upweight the probability of sampling transitions which result in molecules the planner is predicted to solve. A second classifier is trained to predict whether states are likely to result in molecules with docking score lower (better) than a pre-specified threshold. The gradients of this classifier can also be added to the denoiser's logits to guide sampling toward molecules with better binding affinity, where gradients from the two classifiers are normalized to be on the same scale, and a hyperparameter controls the relative weights of the two gradients.

The method is evaluated against several generative baselines on three protein targets, evaluating retrosynthesis solve rate and docking scores. Further comparison evaluates the same masked discrete diffusion backbone with different guidance mechanisms. A cross-planner evaluation is performed with Syntheseus to evaluate solve according to an alternative retrosynthesis tool besides the one used to train the method. The sensitivity of the method to different hyperparameter choices is also evaluated.

Critical strengths

The paper is careful about framing the evaluation as measuring "a planner outcome under a fixed budget and stock, rather than experimental evidence of synthesis". This is important since their method of evaluating synthetic accessibility (retrosynthesis solve rate) depends on the specific retrosynthesis tool and configuration (search budget, building blocks, etc). They concede in the main comparison against baseline methods (Section 5.3) that the results do not establish that routes proposed by other methods are experimentally infeasible. However, this framing does not appear to be consistent throughout the paper: for example, the abstract claims "state-of-the-art performance improving synthesizability", but considering the concession in Section 5.3, this evaluation cannot support any cross-method ranking.

Critical weaknesses

(W1) The synthesizability claim is evaluated almost entirely under the retrosynthesis tool the method optimizes against.

In Table 1, Synth. measures the AiZynthFinder solve rate at a budget of B=500 MCTS iterations under one fixed configuration (USPTO + RingBreaker expansion policies, USPTO filter, 17.4M ZINC stock). This is precisely the quantity the RA classifier is trained to predict and that guidance optimizes during sampling. The "Hit" metric is defined as a molecule that is both solved and below the pre-specified docking threshold, but SynGuide ranks last or near-last on Dock across all three targets, indicating that the "Hit" result is carried entirely by the Synth. scores. The headline comparison therefore measures agreement with a single planner configuration, which is a narrower result than what is claimed in the abstract.

Furthermore, the comparison is unfair to RxnFlow, which constructs molecules from reaction templates and a pre-defined set of purchasable building blocks. Thus, its low Synth. score primarily reflects AiZynthFinder's coverage rather than the synthetic accessibility of the molecules. The authors acknowledge this discrepancy in Section 5.3, yet the paper still uses "Hit" as the primary metric in the abstract to support the claims of "state-of-the-art performance".

The cross-planner evaluation in Section 5.5 is an important experiment, but it cannot demonstrate what it is meant to. Syntheseus solves 96.0% of unguided top-50 molecules versus 99.8% of guided ones, and 93.3% versus 100.0% among the top-10, where the reported standard deviations correspond to roughly one or two unsolved molecules per group. The configuration is permissive enough that nearly any set of molecules seems to pass, so the authors' conclusion that "the results nevertheless show that the RA-guided samples retain high route-found rates under a different planning configuration" is true under this setting but uninformative since it also holds for unguided sampling and thus provides no evidence of performance gains transferred to this alternative evaluation. Moreover, none of the methods from the Table 1 or Table 2 baselines is evaluated here, which is arguably the most important comparison considering the bias in the primary metric.

Finally, no other measure of synthetic accessibility is evaluated - RAscore, SAscore, SCscore, SYBA, or other CASP tools such as ASKCOS. The only planner setting that was varied is the search budget, but this doesn't address the biased evaluation. The paper's central claim is an improved hit rate and state-of-the-art performance improvement in synthetic accessibility, but the evaluation as constructed cannot support these claims in general. However, this could largely be addressed by re-scoring the existing generated molecules.

(W2) Narrow scope and limited novelty.

The methodological claim is that guiding through the denoiser's predicted clean sequence outperforms guiding on the current masked state (Table 2). However, correcting clean-prediction logits with a reward gradient appears to be the mechanism of existing work (GILC) which is not cited [1].

The method also only applies to masked discrete diffusion, and its components (backbone, guidance mechanism, planner) all exist. Thus, it seems like the primary contribution of the paper is the application.

  1. GILC; Dou et al., ICML 2026, Plug-and-Play Guidance for Discrete Diffusion Models via Gradient-Informed Logit Correction
Questions for authors

Along with addressing the weaknesses outlined above, can the authors please clarify the following:

  1. Table 4 reports "Ours" at c-MET Solve 77.0 and BRD4(2) 74.3, while Table 1 reports 72.7 and 70.4 for SynGuide (λsyn=0.3\lambda_\mathrm{syn} = 0.3); the GenMol and SAFE-GPT rows are identical between the two tables. Table 4's caption states that all rows are reported at 100% planner-label coverage. Can the authors elaborate on what is meant by "planner-label coverage", and why the reported results in Table 4 differ from those in Table 1?

  2. In Table 3, solve rates are reported over top-10 and top-50 docked molecules. For λsyn=1\lambda_\mathrm{syn} = 1, docking does not guide generation, so this ranking selects a subset by a criterion that the method did not optimize. Can the authors report the solve rates over all unique molecules for all three conditions?

  3. Under λsyn=1\lambda_\mathrm{syn} = 1, no docking guidance is applied, so the same molecules appear to be evaluated against all three targets, which is supported by the agreement across targets up to a decimal point in Table 1. Is this correct? If so, evaluating "on three protein targets" is slightly misleading for the RA result. It might be helpful if the authors more clearly separated synthesizability tasks from molecular optimization tasks.

  4. Can the authors elaborate on the scope / applicability? The update modifies logits at masked positions, which relies on the masking structure of the backbone. Can the gradient-based sampling update be applied to other types of generative models?

  5. The RA classifier is trained on 9,321 molecules sampled from the GenMol-V2 backbone, so its supervision covers only the backbone's own distribution. Since guidance steers toward regions where the classifier is confident, it is unclear whether the method is finding novel synthesizable chemistry or just sampling from the training data manifold. Table 5 is consistent with the latter, where SynGuide has the lowest diversity of any sampling mechanism. Could the authors report novelty and distributional statistics (e.g. molecular weight, ring count, similarity to the classifier's training set) for guided versus unguided samples? I would also note that the diversity and uniqueness columns reported in Table 5 are omitted from Table 2 in the main text; given that diversity is the column on which the method ranks last, these would be more appropriately reported alongside the main results rather than in the appendix.

4 comments
AuthorsOpenReview ↗

Response to the Reviewer ivnJ [1/3]

We thank the reviewer for the insightful feedback and valuable comments!

W1 — reliance on one planner

We agree that the current evaluation relies heavily on one planner. Following the reviewer's suggestion, we rescored the generated molecules using four additional accessibility measures (SAscore, SCscore, RAscore, SYBA, 19,749 molecules).

(i) The "ranks last on Dock" premise rests on a mislabeled column. The Dock. column reports Top-1, not the mean —corrected table as Table A in the General Comment above. On the distribution, combined guidance improves mean docking over the frozen backbone on all three targets (−6.22→−6.76, −7.04→−7.61, −7.53→−8.12), and on the fraction clearing τ — the binding term actually inside Hit — it is first on ACAA1 (70.2%) and within 1–3 pp elsewhere. So Hit is not determined by Synth. alone.

(ii) Independence varies across the four measures. SAscore and SYBA use no planner, template library or stock; SCscore is trained on Reaxys reaction counts; RAscore is trained on AiZynthFinder output (its own documentation says so), so its agreement partly reflects shared supervision. (Table C in the General Comment above.)

At λ_syn=1 the result holds across the independent measures: that pool ranks first of seven methods on SAscore, SCscore and RAscore and second on SYBA ES rate, at the lowest mean molecular weight in the table (266 Da).

The panel does not support a surrogate-measured gain at λ_syn=0.3: SAscore favours us (2.32 vs 2.45) but SCscore (3.01 vs 2.95), RAscore (0.941 vs 0.971) and SYBA ES (97.3 vs 98.4) put the unguided backbone marginally ahead, and the MW-matched comparison agrees (Table D in the General Comment above; 16 shared 20 Da bins, n = 7,923) — λ_syn=1 wins on all four, λ_syn=0.3 only on SAscore. Surrogate agreement tracks effect size: all four see the 17 pp move at λ_syn=1 (70.7 → 87.9 solve) and none resolves the 3 pp move at λ_syn=0.3 (70.7 → 72.7), being saturated in this regime (every backbone-derived pool sits at 94–99% SYBA ES). That is why we keep a fixed-budget planner label rather than a cheaper score.

At the molecule level, against our B=500 label (n = 13,389, 58.9% solved): (Table E in the General Comment above.) Note the ordering: the planner-free SAscore predicts our label better than the surrogate trained on AiZynthFinder output.

We have not run ASKCOS or another CASP tool; deploying ASKCOS requires server-scale resources beyond the discussion period.

(iii) RxnFlow. We agree that AiZynthFinder may undercount routes for RxnFlow and will qualify the comparison. Independent measures indicate a mean SAscore of 4.37 for RxnFlow, with only 32.8% below 4. The structural reason is visible — 42.3% of its molecules contain a 3-membered ring, with 6.52 rings, 2.91 stereocenters and 1.01 bridgehead atoms on average (TAGMol: 42.8% contain a ring of size ≥8). These measures suggest that RxnFlow's molecules are challenging under criteria independent of AiZynthFinder. Its assembly guarantees a route in its own action space, not that the product is easy to make by other means. We will rewrite the §5.3 discussion and remove the implication that its low solve rate is simply planner coverage.

(iv) Limitations. The §5.5 Syntheseus evaluation cannot demonstrate transfer — it solves 96.8% of molecules at a median of one model call, close to its ceiling (detail under QfBE W9). The abstract's "state-of-the-art performance improving synthesizability" is not supportable as a cross-method ranking and is inconsistent with §5.3's own caveat; we will rewrite the abstract around the fixed-budget-planner framing and the corroborating analyses above.

AuthorsOpenReview ↗

Response to the Reviewer ivnJ [2/3]

W2 — novelty of the mechanism

We agree that the mechanism claim needs to be narrowed. We will add GILC (Dou, Chen, Li, Li & Deng, Plug-and-Play Guidance for Discrete Diffusion Models via Gradient-Informed Logit Correction, ICML 2026, PMLR 306:26227–26249) to the bibliography. GILC uses an update structurally similar to ours: it takes the reward gradient with respect to the clean-prediction logits η and shifts the logits by that gradient, η^r = η + g_η/β, while deliberately omitting the model Jacobian ∂η/∂z_t. Our update is logits + γ·s·Σ_k λ_k ∇_logits log p_k. We do not claim clean-prediction logit correction as new, and §5.4 will cite GILC as the closest prior work.

An implementation difference is measurable. GILC uses Gumbel–Softmax samples with a straight-through estimator to make the reward differentiable. We instead construct a mixed soft state: probability-weighted at masked positions and one-hot at positions that are already unmasked (soft_state = masked · x̂₀ + (1 − masked) · x_t). Reviewer DgsE asked us to isolate this choice, so we added a straight-through hard-argmax arm at our operating point (3 targets × 3 seeds, fully labeled). Solve increases (79.36 vs 73.68, +5.68 pp), while the joint hit rate decreases (38.61 vs 48.40, −9.80 pp, p < 0.0001; worse in 9/9 groups), despite nearly identical guidance magnitudes (per-step probability shift 0.254 vs 0.260). The hard-forward variant does not weaken guidance; it redirects it toward one objective at the other’s expense. Our arm uses deterministic argmax, whereas GILC averages gradients over n Gumbel samples, so this is not a faithful reimplementation of GILC and that difference may account for part of the gap. The ablation nevertheless shows that the relaxation matters under a two-objective planner reward.

We now claim three contributions beyond GILC:

  1. Multi-objective combination. GILC guides one reward at a time, reporting α, μ, C_v, Δε, HOMO and LUMO separately. Its only normalization is GILC-PG’s group-relative advantage within a single objective. We combine two objectives under λ with per-objective RMS normalization, which a global step size cannot replace. Without normalization, the raw gradient scales differ by a target-dependent factor: a nominal λ_syn = 0.3 effectively acts as 0.512 / 0.507 / 0.389 on ACAA1 / BRD4(2) / c-MET. See Tables H and I in the General Comment above.
  2. A predictor that reads partially masked sequences. GILC is training-free and reuses the denoiser as a variational proxy. We train a predictor with a 20% clean fraction, increasing AUROC on noised inputs from 0.575 to 0.726 and improving the end metric by 2.31 pp.
  3. The supervision target. GILC uses oracle property predictors for rewards, including HepG2 activity, stability and QM9 properties; it does not use a retrosynthesis planner. Our fixed-budget label from a declared planner configuration—and our finding that it can be obtained 29.7× more cheaply at AUROC 0.865—is orthogonal to GILC and, we believe, is the novel contribution.

GILC is plug-and-play and training-free, whereas we train an external predictor. We do not claim an advantage over GILC on its own terms, and have removed wording that implied otherwise.

Our remaining mechanism-level claim is narrower: under planner supervision in masked discrete diffusion, and at matched sampling cost, guiding on the clean estimate rather than the current masked state improves solve rate. Against D-CBG, solve rate is 87.3% vs 74.4% (+12.9 pp; p = 0.0059; 3/3 seeds), at 30.0k vs 30.5k predictor calls and 69 vs 67 s.

Regarding the claim that “the components all exist”: that is true of the backbone, planner and gradient-guidance family, and we will say so in the contributions. The three items above define our claims, each isolated by an ablation. We also accept the scope restriction and state it precisely in our response to Q4.

AuthorsOpenReview ↗

Response to the Reviewer ivnJ [3/3]

Q1 — Table 1 vs Table 4

The reviewer identified a genuine inconsistency. Table 4's "Ours" rows come from an earlier generation campaign than Table 1's, which is why the GenMol and SAFE-GPT rows match while ours do not — those two were never re-run between the tables. We recomputed both from the archived molecules: the current pools reproduce Table 1 (ACAA1 78.0, c-MET 72.7, BRD4(2) 70.4), so Table 1 is correct and Table 4 was stale. The revision will report Table 4 regenerated from these pools, and we will generate both tables from one script so they cannot drift apart again.

"Planner-label coverage" is the fraction of evaluated unique molecules that have an AiZynthFinder label at all. It matters because a solve rate over a labelled subset is comparable across methods only if the subsets are comparably complete — early in this project some conditions were labelled at ~40% coverage, and a solve rate over an unrepresentative 40% is not a solve rate. Every row in the revised Tables 1 and 4 is at 1.00 coverage, and we report the column rather than leaving it implicit.

Q2 — all molecules rather than top-k

We agree that reporting results over all unique molecules is the more appropriate comparison. Over all unique molecules, at 100% planner-label coverage: (Table F in the General Comment above.) The effect is unchanged in direction, +17.1 pp over unguided, while docking-only guidance loses 24.4 pp. Note that every condition's top-k rate is below its all-molecule rate, because ranking by docking enriches for hard-to-synthesize molecules (the +0.175 correlation under W2 of the QfBE thread). We have made the all-unique table the primary panel and kept top-k as secondary with the selection effect stated.

Q3 — λ_syn=1 is one molecule set

Yes, this is correct. At λ_syn=1 there is no docking term, so one generated set is scored against all three grids; that is why the Synth. column agrees to a decimal place across targets. Describing that as an evaluation "on three protein targets" overstated it. We will restructure the tables to separate the synthesizability task (one generator, one molecule set, reported once rather than triplicated) from the joint optimisation task (three pockets, genuinely distinct).

Q4 — scope of the rule

The rule needs two things: a differentiable estimate of the clean sequence at every sampling step, and a set of coordinates still free to change, so the gradient has somewhere to act. Both exist in any absorbing-state / masked discrete diffusion model and in discrete flow matching, where the velocity field is parameterised through a clean-data prediction — we expect the rule to transfer there with only the mask definition changed. Neither exists in a standard autoregressive sampler: once a token is emitted it is not revisited and there is no clean-sequence estimate to differentiate. For continuous-coordinate 3D diffusion the analogue already exists (SYNC), acting on coordinates rather than logits. We will replace the paper's broader phrasing with this, since the honest scope is "masked discrete diffusion and discrete flow matching", not generative models in general.

Q5 — memorisation of the predictor's training pool

We measured this and the result runs opposite to the hypothesis. Against the RA predictor's supervision pool (9,752 molecules), Morgan fingerprints at radius 2: (Table G in the General Comment above.) Guidance moves samples away from the training pool — mean max-similarity 0.908 → 0.515 — while raising the planner solve rate from 70.7% to 87.8%.

One caveat: the pool was sampled from the unguided backbone, so the 78.7% overlap in the first row is expected by construction and is not a finding; the load-bearing rows are the guided ones. The predictor's held-out AUROC on noised inputs is 0.958 at B=500.

Distributional statistics as requested (λ_syn=0.3 vs unguided): MW 288 vs 277, ring count 2.62 vs 2.29, atom stereocentres 0.52 vs 0.54, fraction sp³ carbon 0.38 vs 0.47, a ring of size ≥8 in 2.6% vs 0.7%, a 3-membered ring in 2.4% vs 1.8%. The shifts are modest, towards slightly larger, more ring-rich, less aliphatic molecules.

We have also corrected the reporting: Uniqueness and Diversity now appear in the main Table 2. SynGuide has the lowest diversity of the mechanisms compared (0.847 vs 0.868 unguided).

I thank the authors for a thorough and honest response. Q1 is fully answered by the Table 4 update and definition of planner-label coverage. I am raising Soundness from 1 to 2. I also accept three corrections to my review: with the Dock column correctly reported as a mean, my claim that Hit is carried entirely by Synth does not hold, since %≤τ\%\leq\tau is competitive or best on all three targets; Table G contradicts my point about memorization; and the planner-free measures indicate RxnFlow's molecules are genuinely harder to synthesize, so "primarily reflects AiZynthFinder coverage" was too strong. Given the scope of the rewrite still required, I am keeping my rating at 2; I think the paper would be better served by a full revision and a future submission.

Reviewer_QfBEOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents SynGuide, a new guided diffusion method for synthesizable molecule design. This guidance uses a noise-conditioned predictor trained on fixed-budget synthesis planning outcomes. The results show that SynGuide achieves strong performance on synthesizability and binding affinity evaluations across 3 protein targets. Additional analysis explores other perspectives on the SynGuide method, including changes in the guidance mechanism, balancing synthesizability guidance with docking guidance, and adjusting the budget for synthesis planning.

Critical strengths
  1. This problem is of substantial interest and importance. AI-driven design of synthesizable molecules is one of the leading current problems in the field of AI for drug discovery.
  2. The method design relies on discrete diffusion for 2D molecule generation guided by a noise-conditioned predictor, which makes it relatively straightforward to understand at an intuitive level.
  3. The authors show analysis from several perspectives in their experimental results, not just a single overall analysis.
Critical weaknesses
  1. SynGuide was benchmarked on only 3 protein targets, which is far fewer than most molecule design methods. This is consequential to my overall assessment because it limits my confidence in the generalizability of the results and analysis.
  2. Throughout the results, there are several statements of fact that are left uninterpreted. For example, what do the results in lines 400-402 mean? The last sentence in lines 383-384? Even when interpretation is given, it still feels a little confusing as to what the deeper point is that the authors are trying to make. For example, “SynGuide’s mean score changes from −7.05 to −6.86 kcal/mol, so its greater solve rate does not imply improved docking.” Does this actually suggest some tension between the docking and synthesizability? Or is that not something we should conclude? Should docking and synthesis be correlated, anticorrelated, or neither? These sorts of questions are just examples of what the reader is left wondering. This is consequential because it makes it difficult to identify the main conclusions and claims and ensure they are properly supported. It also makes it difficult for the reader to carry away the key takeaways of the paper.
  3. The docking metric is suspiciously high on Table 1 given the low hit rate and specified thresholds. For example, SynGuide (with only synthesis guidance) has an average docking score of -10.25 on the BRD4(2) target, a synthesis solve rate of 87.9% and a hit rate (synthesis solved AND docking is above threshold) of 19.3%. This implies that 68.6% of the data has a synthesis route solved, but does not meet the -6.61 threshold. Under the most generous assumption, where this 68.6% of generated molecules are all docked infinitesimally close to -6.61 and the remaining 31.4% of data all meet the docking criteria, achieving the average docking score of -10.47 would require the remaining molecules to have an average docking score of -18.2, which is incredibly high, especially for a method which is not conditioned or guided by the pocket. It is possible that the score of -10.47 is only averaged across hits; however, this is never clearly stated. The authors should discuss in more detail how they compute these metrics and their rationale for reporting the metrics in this way. This is consequential because it directly affects the validity and interpretation of the results that support their main conclusions.
  4. Guiding diffusion for molecule generation using a noise-based predictor of synthesizability has been explored by several recent papers not yet cited in this work, for example:

a. Liu et al., “SYNC: Measuring and Advancing Synthesizability in Structure-Based Drug Design,” ICLR 2026.

b. Cretu et al., “SynLaD: Latent Diffusion for Generating Synthesizable Molecules Conditioned on 3D Pharmacophore Profiles,” arXiv:2607.01105v1 The authors should consider including these in related work and explaining how their method is distinct or novel compared to these approaches. This is consequential to my assessment because it raises concerns on the novelty of the effort.

  1. As written, the target-specific thresholds in lines 303-305 appear arbitrary and without rationale. This is consequential because it directly affects the validity and interpretation of the Hit rate metric used later in the results.

  2. In addition to the baselines explored here, why has the PrexSyn work by Luo and Coley not been included? This is consequential because it is a state-of-the-art baseline capable of synthesizable molecule generation, but has been excluded.

  3. The related work has several issues. The first is depth: lines 119-121 describe 3 of the major platforms for synthesis planning by stating that they “differ in their reaction models, search procedures, and knowledge sources.” However, the authors do not elaborate any further, leaving the reader wondering how these differ, why the difference matters, and how this relates to the task at hand (generating synthesizable molecules). Furthermore, the retrosynthesis planning work in the related work is missing several influential papers in the space, for example:

a. Chen et al., 2020. Retro*: Learning Retrosynthetic Planning with Neural Guided A* Search

b. Zhao et al., 2024. Efficient retrosynthetic planning with MCTS exploration enhanced A* search. The authors should consider whether they want to limit the scope of their related work or be more inclusive in their citations. This is consequential because it suggests an unfamiliarity with key methods in the field, including some planning methods that may be more efficient or effective than those employed by SynGuide for guidance.

  1. The “autoguidance” baseline method in Section 5.4 and Appendix E is not clearly defined. This is consequential because it makes it difficult to understand the logic of the claims when the autoguidance baseline is not clearly explained.

  2. For Table 3, it appears that the Syntheseus planner is more powerful, finding routes much more often than AIZynthFinder. If Syntheseus finds evidence of synthesizability for >90% of generated molecules with no guidance, then it seems the retrosynthetic accessibility guidance is mostly unnecessary. The authors should consider discussing why this is or is not the case. This is consequential because it currently seems to undermine the severity of the problem their paper is intending to address.

  3. Several minor issues exist in the paper:

a. Line 479 has an extra space before a comma.

b. The section heading for Section 5.5 does not have proper margins

c. The Table 1 caption omits the information about the ± annotation. Including this here would help the clarity of the paper.

d. Line 184 incorrectly defines the ABFE acronym where the SAFE acronym is intended. Furthermore, the SAFE acronym is first used in the introduction, line 055 and should be cited there.

e. The wrong time is bolded in Table 2, assuming that bold is supposed to mark the best (lowest) time.

f. No validity score is bolded in Table 2, and there is a tie that should be clearly marked.

g. The Mean column also has an issue with formatting, both -7.05s should be bolded.

h. The underlining in ACAA1’s docking column in Table 1 is not correct.

While individually, these are not necessarily consequential, together they show a pattern of errors that don’t quite match the rigor of this venue. I believe these can be easily fixed.

Questions for authors
  1. Is there a reason why SynGuide could not be applied to larger benchmark sets (e.g., the test split commonly used for CrossDocked2020)? Also, what is the rationale behind choosing the targets for consideration? Can readers reasonably extrapolate that the results here will generalize to other types of protein targets? Why or why not?
  2. How were the thresholds in line 303-305 chosen?
  3. The second paragraph in the conclusion suggests absolute binding free energy calculations for evaluation, but does not explain why these were not used here or what might be needed to make such calculations a practical part of the evaluation or guidance for similar frameworks in the future. Can the authors provide more insights here?
  4. In the discussion of autoguidance, what is meant by “extrapolation” and “degraded”? Also, the CFG acronym is never defined.
  5. In the budget analysis of Figure 4, what would happen if evaluation was performed on B>500, say B=1000? While there is precedent in the literature for B=500, is there a reason why a larger budget should not be tried in this solve-rate vs budget study? How long does B=500 take, on average?
  6. In appendix F, can the authors provide the proposed routes? Some of the molecules still appear somewhat unconventional for synthesizable drug molecules. For example, the 3-member ring joined to the 6-member aliphatic ring in the first column of the λ=0.4\lambda=0.4 row Figure 12, or the cyclooctane ring in Figure 13. There could be some interesting discussion precipitate from this on the gap between AIZynthFinder results and actual laboratory chemistry.
4 comments
AuthorsOpenReview ↗

Response to the Reviewer QfBE [1/3]

We thank the reviewer for the insightful feedback and valuable comments!

W1 & Q1 — three targets

We agree that three targets are insufficient to establish generalization, and we will add that in the limitations. The target count bears mainly on the joint claim. At λ_syn=1 there is no docking term, so the generated set is target-independent.

Second, regarding the three-target scope, our backbone is not pocket-conditioned, so a pocket enters the method only through its own learned docking predictor, and each predictor is trained on a 22,792-molecule docking campaign against that pocket's grid (≈13.5 GPU-hours of AutoDock-GPU per target at our measured 2.14 s/molecule, before any generation). CrossDocked2020's test split is built for pocket-conditioned generators and would need ~100 such predictors, so we cannot complete it here. For the camera-ready we will add ESR1 plus one further pocket under the same protocol.

W2 — synthesizability versus affinity

The results show a modest tension that was not clear in our original discussion. In our label pool, solve rate and docking score are positively correlated (corr(solve, ds_min) = +0.175; lower ds_min is better), so better-docking molecules are solved less often. Consistent with this, ranking the unguided pool by docking score and retaining the top 50 lowers the solve rate from 70.7% to 45.8%. The single-objective results show the same trade-off: RA-only guidance raises solve rate to 87.8% but worsens mean docking, whereas docking-only guidance lowers solve rate to 46.3%. We will state this anti-correlation up front in §5.4 and 5.6.

W3 — the Dock. column

The reviewer is right that the Dock. column reports the Top-1 docking score, not the mean. We will revise all relevant statements. We recomputed all 18 cells from the raw docking output and Top-1 reproduces every one exactly, for every method, so the data and pipeline are correct and the error is in the label. We apologize for this mislabeling, which made the values appear inconsistent.

One correction to the arithmetic does not change the reviewer's conclusion: the BRD4(2) threshold is τ = −8.06, not −6.61 (−6.61 is ACAA1's). Redone with the correct threshold and the true mean, BRD4(2) at λ_syn=0.3 has mean −8.12, with 68.7% of molecules at or below τ = −8.06, of which 45.6% are also solved by the planner. The table is internally consistent once the column is labeled correctly, and no extreme tail is required.

Synth. and Hit are unaffected. Both were always computed molecule-level over the full distribution, (solved & (ds_min <= tau)).mean(), never from the Dock. column.

Corrected Table 1, with mean and %≤τ: (Table A in the General Comment above.) On the distribution, combined guidance improves mean docking over the frozen backbone on all three targets (−6.22→−6.76, −7.04→−7.61, −7.53→−8.12). The revision reports mean and %≤τ in Table 1, keeps Top-k in Appendix D under a clear label, and defines all three metrics in §5.2.

W4 — SYNC and SynLaD

We will add both citations in the paper. SYNC is close prior art: it guides molecular diffusion with a learned synthesizability classifier, and its Eq. (7) is the same posterior-mean (x̂₀) approximation we use, with the same justification. We will state that in related work. Remaining distinctions: (Table B in the General Comment above.) SYNC does not address masked discrete diffusion, where the rule must act on logits at masked positions; our contribution is specific to that setting and we will narrow the claims accordingly.

W5 & Q2 — the τ thresholds

We thank the reviewer for raising this important question. Each τ is the 42.15th percentile of our 22,792-molecule label pool's docking distribution for that target, applied identically to every method. We used a percentile rather than a literature constant because AutoDock4 energies are not comparable across pockets. 42.15 is uninterpretable on its own, so the revision moves the full sweep into the main text: (Table J in the General Comment above.) At stricter cutoffs the ordering changes.

W6 — PrexSyn

We will add this citation to the paper. We will add this baseline in the camera-ready version.

AuthorsOpenReview ↗

Response to the Reviewer QfBE [2/3]

W7 — planner comparison

We agree with the reviewer that the sentence in the related work was too general. We will replace it with a direct comparison of the planners’ reaction models (template-based expansion policies or template-free single-step models), search procedures (MCTS, Retro*’s neural-guided A*, or MCTS-enhanced A*), stock, and use of a reaction-feasibility filter. This last factor helps explain why planners can disagree about the same molecule, including in our cross-planner result (W9). We will add citations to Retro* (Chen et al., 2020) and MEEA* (Zhao et al., 2024). Both are more sample-efficient than our MCTS labelling configuration. Labels generated with a stronger search procedure at a matched budget would therefore provide better-grounded supervision, which we will explore in future work.

W8 — autoguidance

Appendix E will define the definition of autoguidance concretely. Autoguidance (Karras et al., 2024) forms a guidance direction from the difference between two generative models — the backbone and a deliberately degraded copy — with no external predictor. Here the degraded model is the same GenMol checkpoint sampled at higher temperature, and the guided logits are ℓ_strong + w(ℓ_strong − ℓ_weak); we will state the degradation, w and the schedule.

We will also say explicitly that its role in §5.4 is as a control: it tests whether our outcomes can be had by sharpening the backbone's own distribution instead. They cannot — 65.1% solve, below the 71.3% unguided baseline, and validity down to 99.0%.

W9 — the Syntheseus experiment

The reviewer was right that this experiment does not support a transfer claim, and we withdraw that claim. The planner solves 96.8% of the evaluated molecules (99.9% under any-node acceptance) with a median of one model call and five seconds per molecule. This near-ceiling performance leaves little room to distinguish methods. We also did not evaluate the Table 1 and 2 baselines with this planner, so the experiment provides no cross-method comparison.

The near-ceiling result is not a budget artefact. A median of one call means that most molecules are accepted after a single disconnection into the 17.4-million-molecule ZINC stock. The RootAligned single-step model is template-free, and we applied no reaction-feasibility filter, whereas our AiZynthFinder configuration applies the USPTO filter. The planners therefore differ in strictness, not search effort. Section 5.5 will describe this experiment as inconclusive for transfer. The narrower claim—that our signal is not tied to one planner configuration—is supported instead by the planner-free analyses in Reviewer ivnJ’s W1 thread.

W10 — presentation

We thank the reviewer for pointing out the presentation issues in the paper. We will fix all eight issues.

Q3 — absolute binding free energy

ABFE costs on the order of 10³–10⁴ GPU-hours for a few hundred ligands, against ~25,000 unique molecules here, so it cannot be the evaluation metric for a study of this shape. What would make it practical is to use it as a terminal filter rather than a metric: take ~20 molecules from the Pareto front of the λ_syn sweep, run ABFE on those alone, and report them as case studies. We will rewrite the paragraph to say this, with the order-of-magnitude cost.

Q4 — terminology

We will define CFG at first use as classifier-free guidance. "Degraded" and "extrapolation" are defined in the rewritten Appendix E quoted under W8: the degraded model is a deliberately weakened copy of the backbone (same checkpoint, higher sampling temperature), and extrapolation means stepping from its prediction towards the backbone's and continuing past it, ℓ_strong + w(ℓ_strong − ℓ_weak).

Q5 — planner cost and larger budgets

We ran both measurements. B=500 costs 23.6 s per molecule on one CPU core in our configuration; B=1 costs 0.79 s, 29.7× cheaper, and a predictor trained on B=1 labels still reaches 0.865 AUROC against 0.958 at B=500.

For B=5000 (10× the iterations, wall clock raised to 3600 s so the clock does not bind first) we probed 100 molecules unsolved at B=500: 11 are rescued (11%) at a median 662 s each, roughly 28× the cost for an 11-point recovery on the unsolved mass; all 20 solved controls remain solved. A larger budget therefore shifts every method's solve rate up by a bounded amount at steeply rising cost, and because it applies to all methods it does not reorder them. We will report the B=5000 numbers explicitly.

AuthorsOpenReview ↗

Response to the Reviewer QfBE [3/3]

Q6 — the Appendix F molecules

Below are routes for the two molecules the reviewer pointed out, written forward (synthesis direction), from the same search configuration as Table 1's Solve column (B=500).

Figure 12, λ_syn = 0.4 row, first column — a cyclopropane spiro-fused to a tetrahydropyran, CC1CCN(C(=O)c2ccc([C@H]3CC34CCOCC4)cc2)CC1. Solved, 7 solved routes. Top-scored route, 0.9976, one step:

1. BrCCOCCBr + CC1CCN(C(=O)c2ccc(C3CC3)cc2)CC1 >> target

That is a double alkylation at an unactivated cyclopropane CH2, needing a dianion that cannot be formed. A recognizable route — sulfur-ylide cyclopropanation of an alkylidene tetrahydropyran — is present but scores lower, 0.9866:

1. CC1CCNCC1 + O=C(Cl)c1ccc(Br)cc1 >> CC1CCN(C(=O)c2ccc(Br)cc2)CC1
2. C=C1CCOCC1 + CC1CCN(C(=O)c2ccc(Br)cc2)CC1 >> CC1CCN(C(=O)c2ccc(C=C3CCOCC3)cc2)CC1
3. C[S+](C)(C)=O + CC1CCN(C(=O)c2ccc(C=C3CCOCC3)cc2)CC1 >> target

Figure 13, λ_syn = 0.15 row, first column — O=C(c1cccc(C2CCCCCCC2)c1)N1CCCCC1. Solved, 13 solved routes. Top-scored route, 0.9940, two steps:

1. C1=CCCCCCC1 + O=C(Cl)c1ccccc1 >> O=C(Cl)c1cccc(C2CCCCCCC2)c1
2. C1CCNCC1 + O=C(Cl)c1cccc(C2CCCCCCC2)c1 >> target

Step 1 alkylates benzoyl chloride under Friedel-Crafts conditions, but that arene is deactivated and meta-directing and the alkylation does not work on it. A Suzuki route sits several places down at 0.9866:

1. OB(O)C1=CCCCCCC1 + O=C(Cl)c1cccc(Br)c1 >> O=C(Cl)c1cccc(C2=CCCCCCC2)c1
2. O=C(Cl)c1cccc(C2=CCCCCCC2)c1 >> O=C(Cl)c1cccc(C2CCCCCCC2)c1
3. C1CCNCC1 + O=C(Cl)c1cccc(C2CCCCCCC2)c1 >> target

In both cases, the search reaches catalogue material, but the highest-ranked route is one a chemist would reject, while a workable route appears lower in the ranking. This illustrates the gap raised in the review: the USPTO filter assesses whether a template is plausible in isolation; it does not establish whether a particular substrate can undergo the reaction given factors such as ring strain, activation, and directing effects. A “solved” result means that a template path to catalogue material exists. This is necessary, but not sufficient, evidence of synthesizability. Section 5.1 states this limitation.

We will add routes for some example molecules in Appendix F as a new appendix in the revision.

The authors' efforts to address the concerns raised are appreciated. However, some of the most fundamental concerns, such as using only 3 targets as the benchmark and the reliability of the scores, are not fully addressed. Given the major changes (correcting errors, withdrawing claims) proposed in the rebuttal, together with the fundamental issues, it is not convincing that this manuscript, even with all the changes promised in the rebuttal carefully implemented and presented in the revision (which would need another round of reviews due to the major changes), is ready for publication. I will retain my current scores.

Comments on the paper

4 comments
AuthorsOpenReview ↗

General Comment [1/3]

We thank the three reviewers for their careful assessments, which helped us identify a genuine error in our reported results. Their comments prompted us to correct a mislabelled metric, withdraw four unsupported claims and run additional experiments. We summarise these changes below and respond to each comment in the individual threads.

An error we are correcting. Reviewer QfBE correctly noted that Table 1’s docking values are incompatible with their stated definition. The Dock. column reports the Top-1 score, not the mean, although the caption calls it the “mean AutoDock4 score” and §5.3 (line 764) describes it as “the mean score over evaluated molecules.” We recomputed all 18 cells from the raw docking output. Top-1 reproduces every reported value for every method, confirming that the data and pipeline are correct and the label was wrong. We apologise for this error.

This correction also changes how the results should be interpreted. Reviewer ivnJ reasonably drew a conclusion from the mislabelled column. Across the full distributions, combined guidance improves mean docking over the frozen backbone on all three targets (−6.22→−6.76, −7.04→−7.61 and −7.53→−8.12). It also ranks first on ACAA1 for the fraction clearing τ—the binding term in Hit—and is within 1–3 percentage points of the best result on the other targets. Thus, Hit is not driven by Synth. alone. Top-1 is an extreme order statistic over approximately 900 molecules and can rank methods differently from the full distribution; reporting it as the mean distorted the comparison in favour of the baselines. We also found that the BRD4(2) threshold was quoted as τ = −8.06 but applied as ACAA1’s threshold of −6.61. The revision will report mean and %≤τ, retain Top-k in the appendix with a clear label, and define Dock., Synth. and Hit explicitly.

Claims we withdraw

  1. The abstract's claim of “state-of-the-art performance improving synthesizability” is unsupported as a cross-method ranking and conflicts with §5.3’s caveat.
  2. The Syntheseus experiment in §5.5 cannot demonstrate transfer: it solves 96.8% of molecules at a median of one model call, near its ceiling, and we did not evaluate the Table 1 and 2 baselines with it.
  3. The claim in §5.3 that method rankings are stable across the threshold sweep is incorrect at stricter cutoffs. Below roughly the 10th percentile, SAFE-GPT outperforms us on two targets.
  4. In response to Reviewer ivnJ, we withdraw any claim that the guidance mechanism itself is new. Correcting clean-prediction logits with a reward gradient is GILC’s contribution (Dou et al., ICML 2026, PMLR 306:26227–26249), which we had not cited.

We also identified two additional errors: line 717 gives λ_syn = 0.4 as the default, although the headline configuration is 0.3; and Table 4’s “Ours” rows came from an earlier campaign than Table 1’s.

Evidence added during the discussion period. These analyses use existing molecules or the paper’s operating point, with 100% planner-label coverage for comparison with Table 1.

  • Two design-choice ablations requested by Reviewer DgsE. Each changes one factor and is paired over nine (target, seed) groups on 7,348 newly labelled molecules. Removing RMS gradient normalisation reduces joint hit rate by 28.43 percentage points (p < 0.0001; worse in 9/9 groups). Replacing the soft relaxation with a straight-through hard argmax reduces it by 9.80 percentage points (p < 0.0001; worse in 9/9 groups). The two ablations show that these choices affect performance in different ways.
  • Four accessibility measures requested by Reviewer ivnJ: SAscore, SCscore, RAscore and SYBA, evaluated on 19,749 molecules. At λ_syn = 1, our pool ranks first among seven methods on three measures and is also the lightest, so the gain is not explained by molecular size. We also report where the measures do not favour us and note that RAscore is trained on AiZynthFinder output, making it non-independent of our label.
  • Further analyses: a budget probe at B = 5000, a planner-free agreement analysis, novelty statistics against the predictor’s training pool, and a null result for querying the predictor at t = 1.

The tables cited in our replies are collected in General Comments [2/3] and [3/3] (Tables A–J).

AuthorsOpenReview ↗

General Comment [2/3]

These are the tables referenced from our per-reviewer replies.

Table A — corrected Table 1, distribution statistics. mean is over all unique molecules; %≤τ is the fraction clearing that target's docking threshold, i.e. the binding term inside Hit. Three seeds, 100% planner-label coverage.

TargetMethodTop-1mean%≤τSynth.Hit
ACAA1 (τ=−6.61)GenMol−8.81−6.2234.270.820.9
TAGMol−10.49−7.0364.54.00.5
SAFE-GPT−9.45−6.6955.857.832.0
ToolMol−10.89−7.1869.927.613.5
SynGuide (λ_syn=0.3)−9.01−6.7670.278.052.2
c-MET (τ=−7.58)GenMol−9.85−7.0433.770.718.8
TAGMol−11.82−8.0364.71.30.2
SAFE-GPT−11.31−7.8362.358.036.8
ToolMol−12.96−8.3269.132.318.6
SynGuide (λ_syn=0.3)−9.99−7.6169.072.747.4
BRD4(2) (τ=−8.06)GenMol−10.50−7.5332.170.718.6
TAGMol−12.50−6.6414.619.30.1
SAFE-GPT−12.48−8.5467.358.040.4
ToolMol−13.44−8.9971.734.419.0
SynGuide (λ_syn=0.3)−10.47−8.1268.770.445.6

Table B — the prior art closest to ours, by axis.

AxisRAscore (2021)SYNC (ICLR 2026)SynGuide
Labelplanner surrogate, binary3D classifier (SE(3)-EGNN)planner surrogate at a fixed, declared search budget
Used aspost-hoc filterguidance + DPOguidance only, frozen backbone, no planner call at sampling
Generator—continuous-coordinate 3D diffusionmasked discrete diffusion over SAFE tokens
Predictor inputclean moleculeclean moleculepartially masked sequences (AUROC 0.726 vs 0.575 for a clean-only predictor on noised inputs)
Multi-objective——RMS-normalised two-term guidance, so λ denotes a comparable trade-off

Table C — the four accessibility measures, 19,749 molecules. SAscore and SYBA are planner-free; SCscore is trained on Reaxys reaction counts; RAscore is trained on AiZynthFinder output and so is not independent of our label.

MethodSAscore ↓%SA≤4 ↑SCscore ↓RAscore ↑%SYBA ES ↑MWAiZynth Solve ↑
SynGuide (λ_syn=1)1.9899.82.580.97998.626687.9
SynGuide (λ_syn=0.3)2.3298.93.010.94197.328872.7
GenMol (backbone)2.4594.12.950.97198.427770.7
ToolMol3.0193.34.110.88299.042632.3
SAFE-GPT3.4475.63.760.88697.535858.0
RxnFlow4.3732.84.400.47250.142712.1
TAGMol5.4010.54.200.2496.93711.3

Table D — the same comparison within the backbone family, matched on molecular weight (16 shared 20 Da bins, n = 7,923, each bin weighted equally).

MeasureGenMol (unguided)λ_syn=0.3λ_syn=1
SAscore ↓2.2762.1521.913
SCscore ↓2.5892.6072.315
RAscore ↑0.9740.9540.982
SYBA %ES ↑96.093.897.2
mean MW269.0287.0266.7

Table E — molecule-level agreement of each measure with our B=500 planner label (n = 13,389, 58.9% solved).

MeasureAUROCindependent of our label?
SAscore0.894yes — planner-free
RAscore0.828no — trained on AiZynthFinder labels
SYBA0.757yes — planner-free
SCscore0.755yes — Reaxys reaction counts
AuthorsOpenReview ↗

General Comment [3/3]

These are the tables referenced from our per-reviewer replies.

Table F — planner solve rate over all unique molecules versus the top-k docked subsets, at 100% label coverage.

ConditionAiZynthFinder, all uniqueTop-50Top-10
Unguided70.745.835.6
Docking-only (λ_syn=0)46.328.225.6
SynGuide (λ_syn=1)87.881.674.4

Table G — overlap of each condition with the RA predictor's 9,752-molecule supervision pool, Morgan fingerprints at radius 2.

Conditionexactly in poolmean max-Tanimoto to pool≥0.4≥0.6
Unguided78.7%0.90898.1%86.0%
SynGuide (λ_syn=1)7.4%0.51577.6%20.3%
SynGuide (λ_syn=0.3)1.7%0.54388.0%27.8%

Table H — every design choice and its measured contribution.

Design choiceContrastMeasured effectStatus
Guide on soft clean prediction x̂₀, not the masked state x_tSynGuide vs D-CBG, matched cost (30.0k vs 30.5k predictor calls, 69 vs 67 s)87.3 vs 74.4%, +12.9 pp, p = 0.0059, 3/3 seedsin paper (Table 2)
Noise-conditioned predictor, not clean-onlytrained with 20% clean fraction vs 100%+2.31 pp; predictor AUROC on noised inputs 0.726 vs 0.575in paper
Query the predictor at t=1 vs the true diffusion timex0 vs x0_t, λ_syn=1, RA term only+0.002 pp, p = 0.9977 (Fisher p = 1.000) — a null, so the Appendix A.1 simplification costs nothingadded during discussion
RMS gradient normalisation--no-normalize_grads at λ_syn=0.3, 3 targets × 3 seedsSolve −2.43 pp (p = 0.064, n.s.); Joint −28.43 pp, p < 0.0001, worse in 9/9 groupsadded during discussion
Soft relaxation vs hard argmaxstraight-through one-hot x̂₀ at λ_syn=0.3, 3 targets × 3 seedsSolve +5.68 pp (p < 0.0001, better in 9/9); Joint −9.80 pp, p < 0.0001, worse in 9/9added during discussion

Table I — the two new ablation arms, against the main arm and the unguided backbone. Each is a single-factor change at λ_syn = 0.3, γ = 2.0, guidance scale 0.028, paired over nine (target, seed) groups, with 100% label coverage on all 7,348 newly generated molecules.

ArmSolve ↑Joint ↑mean dock ↓% ≤ τ ↑per-step prob. shift
Main (x̂₀, RMS-normalised)73.68 ± 3.4848.40 ± 3.26−7.4969.30.260
--no-normalize_grads71.26 ± 0.4119.97 ± 1.62−6.9333.60.0014
x0_hard (straight-through)79.36 ± 3.6038.61 ± 1.55−7.3049.70.254
(reference) unguided backbone70.7319.43−6.9333.3—

Table J — joint hit rate (%) by docking-threshold percentile.

TargetMethodp42.15p25p10p5p1
ACAA1SynGuide (λ_syn=0.3)52.238.821.311.81.8
SAFE-GPT32.023.113.89.42.7
ToolMol13.59.96.35.02.2
GenMol20.911.14.32.00.2
c-METSynGuide (λ_syn=0.3)47.433.616.39.11.8
SAFE-GPT36.828.317.412.35.3
ToolMol18.615.210.48.44.9
GenMol18.88.82.91.60.2
BRD4(2)SynGuide (λ_syn=0.3)45.633.516.07.91.5
SAFE-GPT40.432.723.918.18.8
ToolMol19.015.612.110.16.0
GenMol18.69.13.41.80.4
Area_Chair_S1CqOpenReview ↗

Please respond to the authors

Dear reviewers,

Thank you for your service.

The authors have provided their response to your initial reviews. As we approach the end of the author-reviewer discussion period, could you please respond to the authors’ rebuttal (if you haven’t done so already) and communicate the extent to which your concerns have been addressed, and whether you have any further questions and concerns?

Thanks, Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.