acceptodds
Under review as a conference paper at ICLR 2027

Plan, Guide, and Repair: Decomposed Control for Conditional Molecular Graph Diffusion

Abstract

Conditional molecular graph generation must align requested properties without losing the discrete chemical structure that makes a sample usable. Existing generators typically leave molecular size, property control along the reverse trajectory, and chemical correction to a single denoising process, although these decisions require different information at different stages. We introduce Plan, Guide, and Repair (PGR), a sampling framework that controls them separately while keeping the generative backbone frozen. Plan selects a condition-dependent, training-supported molecular size before denoising. Guide learns the endpoint property implied by sampled categorical states from actual reverse trajectories and, at every reverse step, differentiates the target discrepancy on soft node and edge distributions to steer generation toward the requested condition. Repair then uses the backbone edge posterior to revise late bond choices that violate valence, aromaticity, or connectivity constraints. On joint O/N/CO generation, PGR attains 0.9991 raw validity and 0.497 average property error. On BACE, BBBP, and HIV, it reaches property accuracies of 0.9716, 0.9665, and 0.9912 with raw validities of 0.9248, 0.9549, and 0.7745. Ablations confirm the intended roles of all three components: Plan improves condition-aware size selection, Guide improves condition alignment by actively controlling reverse generation, and Repair restores raw chemical validity while preserving the alignment gain. These results support stage-specific sampling control across continuous and categorical molecular design tasks.

Primary area:
generative models
Keywords:
conditional molecular generation, molecular graph diffusion, discrete diffusion models, property-guided generation, chemical validity, inverse molecular design

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_SRzxOpenReview ↗
Rating
3(+1)
Confidence
1(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Plan, Guide, and Repair (PGR), a sampling-time control framework for conditional molecular graph diffusion that aims to improve property alignment and chemical validity without retraining the pretrained diffusion backbone. PGR decomposes the generation process into three complementary stages: (1) Plan predicts a condition-dependent molecular size before sampling; (2) Guide trains a property predictor on actual reverse diffusion trajectories to estimate endpoint properties and uses its gradients to steer generation toward the desired conditions; and (3) Repair performs chemistry-constrained edge resampling at late diffusion steps to correct valence, aromaticity, and connectivity violations. The method is evaluated on continuous multi-property polymer generation using MELD and categorical molecular generation using Graph DiT on BACE, BBBP, and HIV. Experimental results show improved property alignment and raw molecular validity compared with the underlying backbones. Ablation studies further support the contributions of the three components, showing that Plan improves size selection, Guide enhances property control, and Repair improves chemical validity while preserving property alignment.

Critical strengths
  • The experiments are comprehensive, and the results show the effectiveness of the proposed method in condition control.
  • Figures are helpful for understanding.
Critical weaknesses
  • The main concern is the lack of details, necessary background, and critical explanation. There is no background, qualitative example, or problem setting in this paper, which makes it hard to understand the scope of this paper. Moreover, in the methods section, some critical steps lack explanation or further demonstration. Some methods lack motivation and seem heuristic.
Questions for authors
  • It would be better to give an example of ZtZ_t and the background of your problem.
  • What is the representation of cc? A text prompt?
  • If we can build the training pairs (ci,Ni)(c_i, N_i), does it mean that we can infer NiN_i from cic_i? If so, then why can't the model itself do such an inference? Also, what if we need a smaller or larger count in inference that does not exist in training?
  • I am not clear about the motivation of Eq.5 and 6. If my understanding is correct, we are using the property of the final Z0Z_0 to guide the generation. However, as Z0Z_0 is the final outcome of the generation, it seems that we are emphasizing the generation toward Z0Z_0, but not introducing new signals.
  • Is Sec.3.4 for inference or training?
  • The definition of illegal properties in Eq.10 seems heuristic. I am an expert in molecule generation, so I wonder if the four potential mistakes are well considered by the community. It would be better if the authors could show evidence that supports the construction of the four mistakes.
  • I think the paragraph on backbone-specific instantiation can be moved to the experiment section
8 comments

Response to Reviewer SRzx

We sincerely thank the reviewer for the specific questions concerning the problem formulation, notation, and algorithmic stages. We address each point below.

W1: Insufficient background, problem scope, and qualitative examples

The paper provides insufficient background, problem formulation, and qualitative examples, making the scope of the study difficult to understand.

Response. We sincerely thank the reviewer for identifying the need for additional background. A common molecular-design setting is to generate new chemical structures from requested functional properties, such as synthesizability, gas permeability, biological activity, or blood-brain barrier permeability, rather than merely scoring existing candidates. This work studies precisely this structured-property-conditioned generation of new two-dimensional molecular graphs. On top of a frozen discrete graph-diffusion backbone, PGR separately handles condition-dependent molecular size, property deviation during reverse generation, and local graph validity. We expanded the application context and problem background in Section 1, “Introduction” (lines 33–45), and specified the input, output, and study scope in Section 3.1, “Problem Formulation” (lines 148–161).

W2: Insufficient motivation for key steps, making them appear heuristic

Several methodological steps are not sufficiently explained or motivated and therefore appear heuristic.

Response. We sincerely thank the reviewer for emphasizing the design motivation. The three modules follow the order in which information becomes available during reverse generation: Plan selects molecular size before topology forms, Guide corrects endpoint properties from the current state during sampling, and Repair addresses valence, aromaticity, and connectivity after atom and bond identities begin to resolve. The necessity of Guide and Repair and the corresponding ablation evidence are detailed in our response to Reviewer TmK1, W2.

Q1: What is ZtZ_t? Please provide an example and additional problem context

Response. We sincerely thank the reviewer for requesting an intuitive definition. Zt=(Xt,Et)Z_t=(X_t,E_t) is the partially denoised molecular graph at reverse step tt, containing its current atom and bond categories together with unresolved mask states. Z0Z_0 is the final discrete molecule. We clarified the problem setting and the definition of ZtZ_t in Section 3.1, “Problem Formulation” (lines 148–161), and Section 3.2, “Conditional Graph Diffusion with Decomposed Control” (lines 182–195). We also added authentic Test reverse trajectories for all four tasks in Appendix D.4, “Authentic Reverse Trajectories” (lines 1041–1044), and Figure S2 (lines 1118–1122). For O₂/N₂/CO₂, BACE, BBBP, and HIV, Figure S2 follows one authentic trajectory at t/T=1.00,0.60,0.40,0.20,0.00t/T=1.00,0.60,0.40,0.20,0.00, using a fixed graph layout within each row.

Q2: What does c\mathbf c represent? Is it a text prompt?

Response. We sincerely thank the reviewer for requesting clarification. c\mathbf c is not a text prompt. It is a structured numerical or categorical property condition. For example, the continuous task may request (SA, O₂, N₂, CO₂)=(3.32,1200,610,4700), whereas BACE may request (SA, class)=(3.46, active). Continuous quantities are normalized using training-set statistics before being provided to Plan, the backbone, and Guide.

Q3: Why can the backbone not infer NN itself? What happens for node counts smaller or larger than those seen during training?

Response. We sincerely thank the reviewer for asking about this interface. This is not an additional restriction imposed by PGR, but the native sampling design of MELD and Graph DiT. The node count NN and its corresponding node mask must be specified before reverse diffusion begins. The diffusion process updates atom and bond categories on this fixed node support, while node count itself is not part of the diffusion state. The backbone therefore cannot infer NN during sampling; the count must be selected by Plan or specified by the user in advance. For a node count not observed during training but still within the backbone's capacity, the user can directly set num_nodes=N.

Q4: What motivates Eqs. (5)–(6)? Does using the final Z0Z_0 create a circular dependency?

Response. We sincerely thank the reviewer for raising this important question. There is no circular dependency because Z0Z_0 is used only when constructing the supervision data offline. With Guide disabled, we first complete a reverse trajectory and then use y(Z0)y(Z_0) from that same trajectory to supervise its intermediate state ZthardZ_t^{\mathrm{hard}}. During generation, the future Z0Z_0 does not yet exist. The frozen fϕf_\phi predicts endpoint properties only from the current pX0,pE0,tp_X^0,p_E^0,t, and its gradient modifies the next sampling distribution. Eqs. (5)–(6) therefore learn where the current state is likely to end, rather than reading a future endpoint during generation.

Q5: Is Section 3.4 used for inference or training?

Response. We sincerely thank the reviewer for requesting a direct clarification. Section 3.4 in the original manuscript described inference/generation. In the revised manuscript, Guide supervision is presented in Section 3.4, “Guide Supervision from Reverse Trajectories” (lines 213–263), while the active guidance process corresponding to the original Section 3.4 appears in Section 3.5, “Guide as an Active Reverse-Process Controller” (lines 265–301). The former trains fϕf_\phi using completed trajectories. The latter freezes its parameters, predicts from the current soft posterior at every reverse step, differentiates with respect to the inputs, and then samples the next hard state.

Q6: Are the four violation statistics in Eq. (10) of the original submission supported by established domain practice?

Response. We sincerely thank the reviewer for asking about the basis of these definitions. The four-dimensional violation statistic referenced by the reviewer was Eq. (10) in the original submission. Because the revised manuscript adds an equation, this definition is now Eq. (11), ν(G)=(a(G),o(G),e(G),k(G))\nu(G)=(a(G),o(G),e(G),k(G)); Eq. (10) in the revised manuscript is the categorical guidance loss for Guide. Eq. (11) operationalizes commonly used checks for valence, aromatic-cycle consistency, and single-molecule connectivity as four monotonically comparable statistics. These checks are used or discussed in MOSES, Open Babel, the SMILES specification, and DiGress [1,4–6]:

StatisticExact implementation definitionCorresponding domain check
a(G)a(G)Number of edges in the aromatic-edge-induced subgraph that cannot remain in its degree-two cyclic coreAromatic bonds must belong to a consistent Kekulé/cyclic system
o(G)o(G)Number of atoms whose current total bond order exceeds the capacity of the atom typeIncidence of overvalent atoms
e(G)e(G)∑vmax⁡(valence(v)−capacity(v),0)\sum_v\max(\mathrm{valence}(v)-\mathrm{capacity}(v),0)Total severity of valence excess
k(G)k(G)max⁡{ncomp(G)−1, 0}\max\{\mathrm{ncomp}(G)-1,\,0\}, where ncomp(G)\mathrm{ncomp}(G) is the number of connected componentsA generated molecule should form one connected molecular graph

o(G)o(G) and e(G)e(G) measure the incidence and severity of valence violations, respectively. Repair accepts an edit only when at least one statistic decreases and none increases, preventing an apparent improvement obtained by introducing a more severe local valence violation.

[4] O'Boyle et al., “Open Babel: An open chemical toolbox,” Journal of Cheminformatics, 2011. DOI: 10.1186/1758-2946-3-33.
[5] Weininger, “SMILES, a chemical language and information system. 1. Introduction to methodology and encoding rules,” Journal of Chemical Information and Computer Sciences, 1988. DOI: 10.1021/ci00057a005.
[6] Vignac et al., “DiGress: Discrete Denoising diffusion for graph generation,” ICLR 2023. arXiv: 2209.14734.

Q7: Move the backbone-specific instantiation paragraph to the experimental section

Response. We sincerely thank the reviewer for this structural suggestion. We adopted it and moved “Backbone-specific instantiation” to Section 4, “Experimental Setup,” immediately after “Tasks and backbones” (lines 339–353; the relocated paragraph is in lines 345–353).

Thanks for your careful rebuttal. However, a few concerns still exist.

  • Q3: The revised Eq.4 still requires the model to select N^\hat{N} from the training pairs. Though the users can assign the count themselves, it actually skips the plan step.
  • Q4: Can we use a property predictor instead of an endpoint predictor to guide the generation? Also, the difference of states during training and inference may introduce drift. It is hard to tell the results in Table S8 where hard states are better than the soft ones
Reviewer_uw4vOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PGR, a conditional molecular graph generation method that utilizes the pretrained molecular diffusion model. The proposed method consists of three stages: (1) Plan selects molecular size based on the requested properties, (2) Guide provides property guidance during reverse diffusion, and (3) Repair corrects chemical violations at late steps. Experiments on polymer generation and BACE, BBBP, and HIV show improved condition control and raw validity over the corresponding backbones.

Critical strengths
  • The Plan, Guide, and Repair structure is clear and maps the three controls to important aspects for conditional generation. Selecting molecular size based on the requested properties is intuitive, since this may prevent assigning extreme atom size that is not compatible with desired property. Providing property guidance throughout the reverse trajectory also makes sense.
  • PGR achieves the best reported results on the condition-control metrics in both continuous and categorical properties.
Critical weaknesses
  • The authors train Guide using property labels assigned to generated endpoints by surrogate predictors. However, these labels are predictions rather than measured properties, and their errors may be inherited by Guide. This is particularly concerning if generated molecules differ from the training distribution of the predictors, where their predictions could be less reliable.
  • GraphDiT and MELD use classifier-free guidance, which does not need auxiliary property predictors. Thus, the statement "Auxiliary property predictors may be trained on clean or artificially corrupted molecules even though guidance is applied to imperfect states produced by the reverse sampler." is misleading.
Questions for authors
  • Are the predictors used to label trajectory endpoints different from the evaluation oracles?
  • Why is Guide trained on hard-sampled graphs but evaluated on soft predictions during generation? Would training Guide directly on the soft predictions better match its intended use?
  • What motivates training Plan and Guide separately rather than jointly?
  • In Table S4, Plan accounts for a large fraction of the improvement in condition control, while Guide provides a smaller gain. Could the authors give an intuitive explanation for this? A case study comparing sizes drawn from the empirical prior (adopted in GraphDiT and MELD) with the Plan-selected size for the same requested condition may be useful.
6 comments

Response to Reviewer uw4v

We sincerely thank the reviewer for the detailed questions concerning the supervision source, guidance interface, and module-training procedure. We clarify each design choice below and provide the corresponding experiments.

W1: Endpoint labels come from surrogate predictors and may be unreliable out of distribution

Guide is trained using endpoint labels produced by surrogate predictors. Their errors may be inherited by Guide, particularly when generated molecules deviate from the training distribution.

Response. We sincerely thank the reviewer for pointing out that teacher errors may be inherited. We use an endpoint teacher not because surrogate labels are inherently more accurate than experimental measurements, but because newly generated molecules have no experimental property labels and Guide must predict the endpoint reached by continuing to sample from a current, authentic reverse state. Prescribed forward noising pairs artificially corrupted states with the label of the original clean molecule. In contrast, our method pairs states actually visited by the sampler with the generated endpoint of the same trajectory, making the supervision problem more closely match the control problem encountered at deployment.

We agree that out-of-distribution reliability is an important concern. Guide is not intended to extrapolate across arbitrary chemical space; it predicts reverse states produced by the same frozen backbone under the same task and sampling policy. Any learned predictor faces distribution shift, and we collect authentic reverse trajectories to reduce the mismatch between training and deployment states. The matched comparison below shows better control than forward-noising supervision on all four tasks.

The direct comparison below uses the same Plan, requested conditions, guidance strength, and 10,000 Test samples:

Test 10K primary metricForward noisingReverse-trajectory endpoint
O₂/N₂/CO₂ Avg. MAE ↓0.58240.5050
BACE Accuracy ↑0.95870.9682
BBBP Accuracy ↑0.94430.9539
HIV Accuracy ↑0.98000.9860
Mean Accuracy across three categorical tasks ↑0.96100.9694

Reverse-trajectory endpoint supervision therefore performs better on the continuous task, each categorical task, and their mean Accuracy. The teacher is fitted only on the training split and is used solely to label Guide trajectories offline; it does not participate in Test generation or Test evaluation.

W2: The original text conflated classifier-free guidance with auxiliary-predictor guidance

MELD and Graph DiT use classifier-free guidance without an auxiliary property predictor, making the original discussion of auxiliary predictors misleading.

Response. We sincerely thank the reviewer for identifying this mechanistic distinction. Graph DiT and MELD natively implement classifier-free guidance through conditional and unconditional denoiser predictions, whereas MOOD uses an additional differentiable property objective for predictor guidance. PGR retains the backbone's native conditioning mechanism and adds a trajectory-endpoint Guide. We now distinguish these two mechanisms in Section 1, “Introduction” (lines 46–56; MOOD is described in lines 51–53), and restrict the relevant discussion to auxiliary-predictor guidance.

Q1: Is the endpoint-label teacher different from the evaluation oracle?

Response. We sincerely thank the reviewer for requesting this clarification. Yes. The endpoint teacher and benchmark evaluator are distinct model instances with different checkpoints, parameters, and call stages. The teacher is fitted only on the training split and labels Guide trajectories offline. The benchmark evaluator is called only after generation to compute Test metrics. Test molecules and evaluation results are not used for Guide training, checkpoint selection, or guidance-strength selection.

Q2: Why train on sampled hard states but guide soft predictions? Would soft-state training be better?

Response. We sincerely thank the reviewer for this important question. Three considerations motivate this design:

  1. A hard state is the state actually sampled and propagated by the reverse chain, so it forms an unambiguous one-to-one pair with the endpoint of the same trajectory. It is not the argmax of a stored soft posterior.
  2. The generation-level role of Guide depends both on its endpoint estimate and on whether the gradient direction and scale induce useful categorical graph changes under the subsequent sampling dynamics. Hard-state supervision directly corresponds to the discrete states actually propagated by the reverse chain and to their own trajectory endpoints.
  3. Hard inputs align supervision with authentic reverse trajectories, whereas soft inputs provide gradients during generation. They therefore serve trajectory supervision and differentiable control, respectively.

We also performed a controlled comparison on the continuous MELD task. Each version selects its own guidance strength using validation generations only, while sharing Test conditions, Plan node counts, backbone, and evaluation protocol:

Guide training inputValidation-selected strengthCorrected / raw validityTest Avg. MAE ↓
Sampled hard state (one-hot)800.9988 / 0.98250.5050
Generation-interface soft posterior1300.9971 / 0.96900.5709

After each version selects its guidance strength using validation generations only, the hard-state-trained Guide obtains a Test Avg. MAE of 0.5050, lower than the 0.5709 of the soft-posterior-trained Guide. This result shows that hard-state training provides a more effective control signal in the present MELD generation process.

Q3: Why are Plan and Guide trained separately rather than jointly?

Response. We sincerely thank the reviewer for asking about this design. Plan must first use a discrete arg⁡max⁡\arg\max to determine the node count, node mask, and tensor shape before the complete reverse trajectories required to train Guide can be collected. There is no ordinary backpropagation path through the discrete node-count selection and categorical sampling operations. Plan and Guide therefore cannot be trained end to end in the current architecture.

Q4: Why does Plan contribute so much? Can the empirical node-count prior and Plan be compared under the same conditions?

Response. We sincerely thank the reviewer for requesting this comparison. Node count defines the global graph support and cannot be changed by later modules, so Plan provides a large first-order gain. The original cumulative ablation evaluates Guide only after Plan; it therefore cannot directly compare Plan-only with Guide-only and also gives Guide a stronger baseline with less remaining headroom. We added an independent single-module ablation in which Guide-only is retrained on Plan-off trajectories and selects its guidance strength independently. The results are below; a fuller explanation appears under “Necessity of Guide” in our response to Reviewer TmK1, W2:

Primary metricBackboneGuide-onlyPlanPlan + Guide
O₂/N₂/CO₂ Avg. MAE ↓0.79810.72960.56920.5050
BACE Accuracy ↑0.91350.94890.95620.9682
BBBP Accuracy ↑0.94170.95230.95270.9539
HIV Accuracy ↑0.97770.98320.98460.9860

Guide-only is retrained on Plan-off trajectories and selects its guidance strength independently using validation generations. Before Repair is enabled, Plan + Guide already achieves stronger conditional-control results than the displayed baselines. The following examples further illustrate the difference between an empirical-prior draw and a Plan-selected size for the same requested condition:

TaskRequested conditionEmpirical-prior size and generated resultPlan-selected size and generated result
O₂/N₂/CO₂SA/O₂/N₂/CO₂ = 3.32/1200/610/4700N=36N=36; output 7.41/271.79/59.98/1087.75N=21N=21; output 3.09/962.06/854.22/4188.70
BACEactive; target SA = 3.46N=20N=20; activity score 0.11; SA error 0.984N=29N=29; activity score 0.95; SA error 0.060
BBBPpermeable; target SA = 3.02N=17N=17; permeability score 0.22; SA error 1.197N=28N=28; permeability score 0.96; SA error 0.347
HIVactive; target SA = 2.80N=18N=18; activity score 0.15; SA error 0.174N=26N=26; activity score 1.00; SA error 0.021
Reviewer_TmK1OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
3
Soundness
2(+1)
Read review
Paper summary

The authors propose a PGR (Plan, Guide and Repair) approach that decomposes molecular graph generation into 3 steps aimed to stabilize molecular diffusion models to provide more valid structures while satisfying generation constraints and conditions.

The idea is quite straightforward and not that novel. First, sample the number of nodes for generation with trained conditioned NN (Plan), then perform diffusion sampling with guidance from property predictor (Guide) and finally fix object inconsistencies (disjoint fragments, broken geometry, misplaced aromatic bonds) with final Repair step. Both learnable priors (Plan) and guidance from property prediction are not novel in the area of diffusion models and well adapted in molecular generation. The repair step is quite an interesting modification of the diffusion process. At the same time this last Repair step looks the most useless in the whole pipeline. Table S4 with ablation studies shows minor improvement over just Plan+Guide steps. Even if diffusion models can produce broken molecular graphs, it’s not that big a problem (typically the validity of generated structures is more than 98% in Table 1, the only exception is Graph DiT) and these objects could be just resampled instead of repaired.

The experiments show good property constraints satisfaction abilities of the PGR model in Tables 1, 2, but given the ablation studies it looks like the main contributor is surprisingly just a simple backbone+Plan step, not Guide and Repair. So the whole message of the work is just “Smartly sample the number of nodes before running diffusion sampling” which makes the part of the work less significant. The Guide and Repair steps just complicate the whole framework training and inference and do not provide any significant improvement.

Another weak point of Experiments is almost full absence of baselines that generate molecular structure as just molecular string (SMILES/SELFIES). Molecular strings are a very good prior, the fragments, including aromatic rings, are localised in the string; the next generated atom is attached to some of previous ones, so it’s almost impossible to generate disjoint fragments; and the overall format syntax is easily learnable by language models [a, b]. Also the language models don't need explicit predictive guidance, the condition can be just textually represented in input prompts, the language models directly control the size of the structure since it’s just basically the number of generated tokens, and the whole training of language models is straightforward.

The paper text is written in an easy-to-follow style and main parts accompanied with corresponding figures.

The main problem of the paper is very limited novelty and that the necessity of some proposed parts not supported by experiments. The approach have good results in Tables 1,2, but it's not clear is it winning over straightforward language models trained on molecular strings.

a.Galactica: A Large Language Model for Science

b.MMAI Gym for Science: Training Liquid Foundation Models for Drug Discovery

Critical strengths

None

Critical weaknesses

As stated in Summary, (i) the lack of simple and straightforward molecular string generator baselines and (ii) the lack of experimental evidence to support the necessity of Guide+Repair steps.

Questions for authors

The answers for stated Critical Weaknesses and experimental proof could improve my score.

7 comments

Response to Reviewer TmK1

W0: Limited novelty and apparently dominant contribution from Plan

Learned size priors and property-predictor guidance have precedents, making the overall framework appear straightforward. The existing ablation also seems to indicate that most of the improvement comes from Plan.

Response. We sincerely thank the reviewer for raising this central concern. We respectfully disagree that the larger contribution of Plan makes Guide redundant or limits the novelty of the framework, for three reasons.

First, the principal contribution of PGR is to formulate global size planning, reverse-trajectory property control, and local validity repair as three interfaces decoupled from the generative backbone. We evaluate the full PGR framework with two conditional backbones, MELD and Graph DiT, and further evaluate the Plan–Guide controller on a frozen unconditional DiGress backbone. These experiments demonstrate that the sampling-time control components can operate across different graph-diffusion settings without modifying the backbone parameters.

Second, Guide is not a generic property predictor trained with conventional forward-noising supervision. Our reverse-trajectory endpoint supervision takes reverse states actually sampled by the backbone as inputs and uses the properties of the final molecule from the same trajectory as supervision. Guide therefore learns the relationship between a current state and its actual generated endpoint. The complete comparison is provided in our response to Reviewer uw4v, W1. With Plan, requested conditions, and guidance strength held fixed, reverse-trajectory endpoint supervision reduces the continuous-task Avg. MAE from 0.5824 to 0.5050 and increases the mean Accuracy across the three categorical tasks from 0.9610 to 0.9694.

Third, our original experiments used strong conditional backbones, and Guide appears after Plan in the cumulative ablation. It therefore faces a stronger baseline and less remaining headroom. The independent single-module ablation and unconditional-backbone experiment in W2 further isolate the roles of the modules and directly test the necessity of Guide and Repair.

W1: Missing SMILES/SELFIES and molecular-language-model baselines

The experiments do not include methods that directly generate molecular strings, so it remains unclear whether PGR outperforms language models based on SMILES or SELFIES.

Response. We sincerely thank the reviewer for this suggestion. Using the same training splits, 2,000 validation generations, 10,000 Test conditions, and property evaluators, we added Conditional SMILES LSTM, MolGPT, Conditional SELFIES Transformer, and task-specific LoRA adaptation of Galactica-6.7B. Decoding settings are selected using validation generations only. Each Test condition is sampled once, without rejection sampling or validity retries. Values in parentheses are raw validity.

MethodO₂/N₂/CO₂ Avg. MAE ↓BACE Accuracy ↑BBBP Accuracy ↑HIV Accuracy ↑
Conditional SMILES LSTM0.7288 (0.5301)0.7991 (0.7244)0.8041 (0.2997)0.7036 (0.4636)
MolGPT0.8920 (0.2717)0.7112 (0.4120)0.8206 (0.1057)0.6004 (0.1438)
Conditional SELFIES TransformerN/A0.7146 (0.9998)0.5749 (0.9916)0.5736 (1.0000)
Galactica-6.7B LoRA1.2646 (0.9891)0.5175 (0.9703)0.5203 (0.8637)0.5152 (0.8994)
PGR0.4966 (0.9991)0.9716 (0.9248)0.9665 (0.9549)0.9912 (0.7745)

PGR achieves the best primary metric on all four tasks. Although SELFIES and Galactica attain high validity on several tasks, their Accuracy is substantially lower, showing that string validity does not substitute for property control. The continuous-task SELFIES result is N/A because the standard representation cannot encode 547 of 553 polymer repeat units containing * connection sites. Molecular properties often depend on specific functional groups and their graph context. Language models must infer this structure–property mapping indirectly from token sequences, whereas graph models operate directly on atom–bond topology and can associate requested properties with the relevant substructures more directly.

We did not include MMAI Gym for Science as a numerical baseline because it is a multitask SFT+RFT system rather than a single string generator. Its molecular protocol starts from an input molecule, optimizes BBBP/DRD2/logP/QED, and aggregates 20 samples. This differs from our protocol of generating one new molecule from a requested target-property vector. A directly reproducible complete checkpoint and environment are also not currently available.

W2: Insufficient evidence for the necessity of Guide and Repair

Guide and Repair increase training and inference complexity, but the existing ablation does not sufficiently establish that these two stages are necessary.

Response. We sincerely thank the reviewer for asking us to test the roles of Guide and Repair more directly. The original cumulative Backbone→Plan→Guide→Repair ablation measures sequential gains but cannot fully separate the independent contributions of Guide and Repair. We therefore added an independent single-module ablation with 10,000 Test samples on all four tasks and used unconditional DiGress to further test the gain from Guide after Plan.

Necessity of Guide

We completed an independent single-module ablation with 10,000 Test samples on all four tasks. Backbone + Guide uses an independently trained Guide based on trajectories recollected with Plan disabled. Backbone + Repair disables both Plan and Guide. This experiment therefore directly tests whether each module remains effective when separated from the other modules.

The continuous-task results are:

MethodAvg. MAE ↓Raw validity ↑
Backbone0.79810.9651
Backbone + Plan0.56920.9820
Backbone + Guide0.72960.9667
Backbone + Repair0.79750.9960
Full PGR0.49660.9991

For the categorical tasks, each cell reports Accuracy / raw validity:

MethodBACEBBBPHIV
Backbone0.9135 / 0.84950.9417 / 0.85050.9777 / 0.7415
Backbone + Plan0.9562 / 0.90980.9527 / 0.91340.9846 / 0.7192
Backbone + Guide0.9489 / 0.85190.9523 / 0.85340.9832 / 0.7418
Backbone + Repair0.9157 / 0.86070.9492 / 0.87170.9848 / 0.7763
Full PGR0.9716 / 0.92480.9665 / 0.95490.9912 / 0.7745

The independent and cumulative ablations answer complementary questions. The independent ablation tests whether a module improves the original backbone when used alone, whereas a later module in the cumulative ablation faces a stronger baseline and less remaining headroom. Even under the latter, stricter comparison, Guide reduces the continuous-task Avg. MAE from 0.5692 to 0.5050, an 11.3% relative reduction that accounts for approximately 21.3% of the total error reduction from Backbone to full PGR. With Plan, requested conditions, and guidance strength held fixed, the reverse-trajectory Guide also clearly outperforms forward-noising supervision, 0.5050 versus 0.5824. This improvement therefore cannot be attributed to Plan.

To further separate Guide from conditioning already learned by the backbone, we attached Plan and Guide to the public unconditional DiGress model. The DiGress denoiser never receives the target properties. Conditions enter only through the external Plan and Guide interfaces, and the backbone remains frozen. The new paired Test results on 2,000 conditions are:

TaskUnconditional DiGress+ Plan+ Plan + GuideGuide vs. Plan
O₂/N₂/CO₂ Avg. MAE ↓1.97651.96611.9132−2.69%
BACE Accuracy ↑0.51790.53130.5851+5.38 pp
BBBP Accuracy ↑0.50960.54200.6292+8.72 pp
HIV Accuracy ↑0.46710.47040.5296+5.92 pp

Under identical Plan node counts, requested conditions, and base random streams, Guide reduces the continuous-task Avg. MAE from 1.9661 to 1.9132, a 2.69% relative reduction, and increases BACE, BBBP, and HIV Accuracy by 5.38, 8.72, and 5.92 percentage points, respectively. The improvement on all four tasks shows that Guide provides effective property control after Plan determines molecular size, even when the backbone receives no property condition.

Necessity of Repair

Repair addresses a problem that neither Plan nor Guide directly optimizes. Guide can steer a molecule toward the requested properties, but this intervention does not directly constrain whether the local atom-bond structure lies in the valid-molecule region. A separate mechanism is therefore still required to address structural validity after property control.

The validity rate of generated structures is a core molecular-generation metric rather than a secondary measure, because a graph with correct properties but invalid structure cannot be used as a successful output. MOSES and GuacaMol both treat validity as a principal evaluation metric, and GraphAF likewise frames the simultaneous satisfaction of chemical validity and target properties as a fundamental molecular-generation problem [1–3].

Repair neither reselects the node count nor replans the requested properties. Instead, it directly improves the raw graph along the current conditioned trajectory:

TaskBackbone raw validityBackbone + Repair raw validityPlan + Guide raw validityFull PGR raw validity
O₂/N₂/CO₂0.96510.99600.98250.9991
BACE0.84950.86070.91400.9248
BBBP0.85050.87170.91410.9549
HIV0.74150.77630.72150.7745

Repair-only improves raw validity on all four tasks. After Plan and Guide, Repair further improves raw validity on all four tasks without reducing the continuous-task Avg. MAE or Accuracy on any of the three categorical tasks. Direct resampling discards the current conditioned trajectory and invokes the full backbone again, without guaranteeing that the new trajectory will preserve the same property alignment. Repair instead preserves the alignment already achieved within the same reverse chain while correcting local structural violations. Guide and Repair therefore address two problems that Plan cannot replace: property deviation can remain after the node count has been selected, and structural invalidity can remain after the generated properties approach the request.

References

[1] Polykovskiy et al., “Molecular Sets (MOSES): A Benchmarking Platform for Molecular Generation Models,” Frontiers in Pharmacology, 2020.
[2] Brown et al., “GuacaMol: Benchmarking Models for de Novo Molecular Design,” Journal of Chemical Information and Modeling, 2019.
[3] Shi et al., “GraphAF: a Flow-based Autoregressive Model for Molecular Graph Generation,” ICLR 2020.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

Dear Area Chair and Reviewers,

We hope this message finds you well. We sincerely thank you and the reviewers for the time and effort devoted to evaluating our submission. Our paper proposes PGR, a sampling-time framework that decomposes conditional molecular graph generation into condition-aware size planning, reverse-process property guidance, and late-stage graph repair. During the rebuttal, we provided point-by-point responses to all three reviews, conducted additional experiments and analyses, revised the manuscript, and prepared this consolidated summary to facilitate the final assessment.

The follow-up discussion can be summarized as follows.

Reviewer SRzx explicitly confirmed: “My concerns are addressed. I have raised my score.” We greatly appreciate the reviewer’s careful follow-up questions, which helped us clarify the problem formulation, the intended scope of Plan, endpoint prediction, hard-state supervision, and the chemical constraints used by Repair.

Reviewer uw4v stated that “Most of my concerns and questions have been resolved” and maintained the original weak-acceptance rating. The reviewer’s remaining question concerned the reliability of the endpoint teacher on generated molecules. In response, we completed an additional reliability audit using real held-out labels, independent evaluator agreement, and paired Plan/Plan+Guide outputs. The results show strong teacher–evaluator agreement and consistent Guide improvements on the subsets where teacher and evaluator disagree most. We sincerely appreciate this clarification, which motivated a more direct analysis.

Reviewer TmK1 did not provide follow-up feedback during the discussion period despite our detailed response and subsequent reminders. Nevertheless, we directly addressed the reviewer’s central concerns. For baseline coverage, we added Conditional SMILES LSTM, MolGPT, SELFIES Transformer, and Galactica-6.7B LoRA. For the necessity of Guide and Repair, we conducted independent single-module ablations across all four tasks, compared reverse-trajectory endpoint supervision with forward noising, evaluated Plan–Guide control on an unconditional DiGress backbone, and analyzed Repair’s raw-validity gains and graph transitions. Because no follow-up response was received, we remain uncertain whether this new evidence resolved the reviewer’s concerns or affected the original weak-rejection assessment.

In summary, we made every effort to address all concerns carefully, transparently, and within the discussion period. Two reviewers engaged in follow-up discussion: one confirmed that the concerns were resolved and raised the score, while the other stated that most concerns were resolved and retained a weak-acceptance rating. We respectfully ask the Area Chair to consider the complete rebuttal and the substantial additional evidence when evaluating the remaining review for which no discussion-phase feedback was available.

We deeply respect the ICLR review process and sincerely appreciate the constructive feedback, which has materially improved our paper. Thank you again for your valuable time and careful consideration. We remain available should any further clarification be helpful.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.