acceptodds
Under review as a conference paper at ICLR 2027

COMPOSE: Molecular Generation and Optimization with a Reusable Stochastic Rewrite Process

Abstract

Controllable molecular design requires models that describe how molecules can change and steer those changes toward desired outcomes. Existing models often specialize in particular endpoint distributions or task objectives, limiting their reuse as design goals change. To address this, we introduce **COMPOSE** (**CO**ntrollable **M**olecular **P**rocess **O**ver **S**tochastic **E**dits), which generates molecules through stochastic sequences of executable transformations between complete, chemically valid structures. We learn a goal-independent reference process over legal, variable-size transitions defined by an exact rewrite system. Structured programs combine these transitions into coordinated molecular transformations, while a separate controller steers them using task constraints and feedback. Keeping this reference process fixed, COMPOSE maintained 100% validity and achieved the highest mean quality among the compared methods across fragment-constrained generation tasks. This reuse extended to black-box optimization, where COMPOSE achieved a mean final top-10 score of 0.563 across 22 objectives, exceeding the strongest compared baseline's 0.542 under a 1,000-evaluation budget. In protein-specific lead optimization, COMPOSE obtained the best reported docking scores in 23 of 30 cases using 250 evaluations per lead, compared with 1,000 for the baselines. Together with similarity-constrained editing results, these findings show that reusable molecular transformations enable a single learned process to support generation and optimization as structural constraints and design objectives change, without retraining the underlying model.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
molecular editing, graph rewriting, controllable generation, controlled Markov processes, multi-objective optimization, molecular optimization

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

COMPOSE combines an exact molecular graph-rewrite executor, a frozen task-agnostic reference law Rθ over canonical molecular successors, structured multi-edit programs, and task-specific controllers. It is evaluated on fragment-constrained generation, ZINC-QED editing, PMO-1K black-box optimization, and protein-specific lead optimization.

Critical strengths
  1. A coherent executable molecular-editing framework.

The exact executor constrains committed edits to a declared valid molecular graph space. Canonical-successor probabilities and multi-step programs form a reusable system

  1. Measurable benefits from learned transition preferences.

Held-out successor NLL decreases from 5.37 to 3.90, and learned edit preferences improve superstructure quality from 25.40% to 39.03%. Locked-panel selection studies offer additional evidence of utility

Critical weaknesses

1. The distinctive novelty and value of the learned process remain unestablished.

Autoregressive molecular graph generation with valence constraints (e.g., GraphAF), sequential editing and search (MIMOSA, MARS), reusable pretrained diffusion generators (GenMol), and goal guidance of frozen models (Value Matching) are established directions.

Canonical successors and structured programs may be a useful combination, but matched mechanism-level comparisons against strong editing and diffusion baselines are missing. Consequently, the paper does not establish what unique capability is enabled by this modeling paradigm.

Table 6 mainly evaluates prediction of constructed teacher transitions; Appendix F.2 states that ring-edit supervision comes entirely from synthetic walks. Free-running distribution and chemical-space diagnostics (e.g., FCD, novelty, scaffold coverage) are absent.

Moreover, although the paper emphasizes generation trajectories, coordinated edits, and retargeting, it provides almost no visualization of actual molecular transformations. This makes it difficult to verify whether the learned process exhibits distinctive behavior beyond conventional molecular search. The claimed 100% validity follows the declared structural guards; it does not establish synthetic feasibility or drug activity.

2. The optimization gains are insufficiently attributed, and benchmark comparability is limited.

The learned-law ablation is largely restricted to superstructure generation; PMO/T4 lack a controlled intervention on Rθ with the executor, programs, and controller fixed. T4 selects archive parents without score ranking, making the practical roles of Rθ and docking feedback unclear.

The structured-proposer ablation has only 14/18 completed pairs and simultaneously changes multiple proposal factors, preventing isolation of dependency scheduling (Appendix H.6).

On PMO, final Top-10 averages 0.563 versus GenMol's 0.542, but AUC-Top10 is 0.482 versus 0.537. Excluding the two isomer tasks leaves an approximately −0.015 average difference over the other 20 objectives.

The QED comparison uses different source panels and K=8 versus 20 candidates. T4 baselines were not rescored using the same docking pipeline, while the paper reports median repeat-docking variation of 0.70 kcal/mol. Eleven of the 23 nominal wins in Table 3 have margins below that scale.

These results do not yet substantiate broad superiority of the proposed modeling paradigm (Secs. 3.3–3.5; Appendices G, I).

Questions for authors

1. Novelty and comparisons.

What concrete capability does COMPOSE provide beyond GraphAF, MIMOSA/MARS, GenMol, and Value Matching? Can the key claim be tested with matched constraints, controllers, and budgets against strong graph-edit and diffusion baselines?

2. Causal attribution of the reference process and structured programs.

How do PMO and T4 change when Rθ is replaced by a uniform or empirical-family law while other components remain fixed? Where exactly do T4 proposals use Rθ and observed docking scores? Can the full 18/18 program ablation be completed with dependency scheduling isolated?

3. Benchmark comparability and optimization efficiency.

Can QED editing be evaluated on the same source panel with matched K, and can T4 baseline molecules be rescored consistently? Please provide PMO anytime Top-10 curves, sensitivity to the two isomer objectives, and proposal/compute cost per useful oracle evaluation.

4. Verifiable evidence of the actual generative process.

Can the authors show real molecule-by-molecule editing trajectories, highlighting changed atoms/bonds and transition probabilities, as well as dependency-aware multi-edit programs?

Matched visual examples comparing continued editing after a goal switch against restarting from the initial molecule would be particularly informative. Representative random successes and failures, together with free-running chemical-space diagnostics, would also help establish whether the claimed reusable process exhibits meaningful and distinctive behavior rather than merely producing competitive endpoint scores.

7 comments

We thank the reviewer for the careful assessment and specific suggestions. To address the concerns, we conducted new experiments on the learned reference, dependent edits, and continued optimization after a goal switch. We also expanded the baseline comparison to eight methods and completed all 18 paired program experiments. We explain the contribution first, then address each experimental question in turn.

Q1. What capability does COMPOSE add?

What concrete capability does COMPOSE provide beyond GraphAF, MIMOSA/MARS, GenMol, and Value Matching? Can the key claim be tested with matched constraints, controllers, and budgets against strong graph-edit and diffusion baselines?

With COMPOSE, we learn probabilities over executable changes between complete molecular graphs. We can compose these changes into programs with explicit dependencies between edits. For example, we can insert an atom and later close a ring through that same atom. During execution, we track atom identities across steps and verify every intermediate molecule. We aggregate equivalent edits into one probability for each distinct molecular outcome.

The reviewer correctly identifies established work on autoregressive graph construction [1], molecular sampling [2,3], and guidance of frozen generators [4,5]. Our contribution combines canonical transition probabilities with executable programs that preserve dependencies between molecular edits. This makes it possible to continue editing an existing molecule under a new objective using the same frozen reference. We test that capability directly in the goal-switch experiments described in our final comment.

To compare optimization performance, we evaluated eight methods on 22 PMO objectives with 1,000 charged calls per run. With COMPOSE, we obtained the highest mean final Top-10 of 0.567, compared with 0.563 for GenMol and 0.555 for GenMol V2. We retained each baseline's optimization procedure under the common budget. For the mechanism comparisons, we separately tested the reference probabilities and program construction within COMPOSE.

W1. Free-running behavior beyond teacher-transition prediction

We agree that examining free-running molecules is important for evaluating the learned process. To address this, we generated four eight-edit trajectories from each of 512 held-out molecules, producing 2,048 endpoints per arm. We used identical starting molecules, execution rules, and stopping criteria for learned and uniform editing. We disabled optimization controllers and oracle feedback in both arms. We then independently replayed all 32,768 executed edits.

MetricLearned referenceUniform legal edits
FCD ↓2.7714.35
Held-out reference scaffold coverage ↑25.11%15.28%
Mean QED ↑0.5310.212
Mean SA score ↓4.216.43

For FCD, we compared both endpoint sets with the same separate panel of 2,048 held-out molecules. We also measured recovery of the panel's 458 distinct scaffolds. Using the learned probabilities reduced FCD from 14.35 to 2.77 and increased scaffold coverage from 15.28% to 25.11%. We also obtained higher QED and lower SA scores under the same validity rules. These improvements demonstrate the value of learned transition preferences during free-running molecular editing.

We found zero endpoint overlap with the available training-source bank in both arms. The starting molecules were also held out. These are source-conditioned walks, with synthetic ring-edit supervision as described in Appendix F.2. We therefore evaluate the resulting molecular distributions beyond teacher-transition prediction. QED, SA, and docking are computational property measurements.

To further examine reuse, we compared continuation after a goal switch with restarting from the original molecule. Continuation performed better in 35/40 potency-first and 29/40 developability-first comparisons under the same controller and three-edit allowance. We provide the molecular examples and paired confidence intervals in the final comment.

[1] Shi et al., GraphAF, ICLR 2020. [2] Fu et al., MIMOSA, AAAI 2021. [3] Xie et al., MARS, ICLR 2021. [4] GenMol and V2. [5] Jensen et al., Value Matching, ICLR 2026.

Q2. Contributions of the reference process and structured programs

How do PMO and T4 change when Rθ is replaced by a uniform or empirical-family law while other components remain fixed? Where exactly do T4 proposals use Rθ and observed docking scores? Can the full 18/18 program ablation be completed with dependency scheduling isolated?

We appreciate the request to examine these mechanisms directly. We completed all 18 paired program comparisons and added an intervention on created-atom dependencies. To broaden the evaluation, we also compared learned, uniform, and empirical editing across all four fragment tasks. We report those results first, followed by the program and PMO experiments. We explain the T4 implementation in the next comment.

1. Learned transition preferences. For the uniform control, we sampled admissible canonical successors uniformly. For the empirical control, we selected edit families using fixed training frequencies. We then sampled uniformly within each selected family. Both controls retained the same executor and task constraints.

Task quality (%) ↑COMPOSEUniformEmpirical
Motif extension42.63 ± 0.383.109.53
Superstructure39.03 ± 3.091.9733.77
Linker design31.53 ± 1.4022.7523.70
Scaffold decoration36.70 ± 1.803.0711.17

With the learned reference, we obtained the highest quality on all four tasks. For motif extension, quality reached 42.63%, compared with 9.53% for empirical editing and 3.10% for uniform editing. We thus observed substantial gains from learned preferences within the legal editing space. We conducted this ablation separately from the expanded GenMol V2 fragment comparison.

2. Structured construction and created-atom dependencies. To examine structured construction, we retained learned primitive editing and removed the programs. We obtained the following fragment-generation results.

Task quality (%) ↑COMPOSELearned primitive editing
Motif extension42.63 ± 0.387.87 ± 0.12
Linker design31.53 ± 1.4015.43 ± 0.91

We then completed the Appendix H.6 experiment across six objectives and three seeds, giving 18/18 paired comparisons.

Program configurationFinal Top-10 ↑AUC-Top10 ↑
Structured programs0.62980.5139
Uniform editing chains0.55810.4682
Created-atom binding intervention0.57990.4941

For the binding intervention, we retained the broader program recipe. We changed how later transformations attach to atoms created earlier. This reduced final Top-10 from 0.6298 to 0.5799, with paired Wilcoxon p = 0.0182. We therefore find direct support for preserving created-atom dependencies during molecular construction. This intervention tests atom binding within programs. The uniform-chain comparison evaluates the broader contribution of structured construction.

3. PMO ablations (see Appendix G.8 for PMO anytime Top-10 ablation curves). We additionally evaluated uniform and empirical legal-edit chains across the PMO objectives. For the learned primitive control, we retained neural preferences and removed structured programs. These comparisons examine the combined contribution of transition preferences and proposal construction. In the free-running experiment, we isolate the reference probabilities under identical execution rules.

ObjectiveCOMPOSEUniformEmpiricalLearned primitive
Albuterol similarity0.7780.5670.6500.494
Amlodipine MPO0.5710.5210.5090.447
Celecoxib rediscovery0.2860.1930.2700.149
Deco hop0.5860.5820.5900.560
DRD20.9690.8280.9870.029
Fexofenadine MPO0.7520.6570.6860.585
GSK3B0.5020.4240.4470.276
Isomers C7H8N2O21.0000.8550.8820.040
Isomers C9H10N2O2PF2Cl0.8740.8360.8330.682
JNK30.3140.1650.2130.060
Median10.2780.2270.2720.182
Median20.1890.1410.1380.127
Mestranol similarity0.5160.4070.4280.270
Osimertinib MPO0.8430.7880.7790.726
Perindopril MPO0.5030.3870.4060.324
QED0.9380.9250.9180.836
Ranolazine MPO0.7830.5740.6650.209
Scaffold hop0.4590.4470.4640.426
Sitagliptin MPO0.2200.3460.1610.231
Thiothixene rediscovery0.3880.2160.2370.194
Troglitazone rediscovery0.3080.2400.2570.192
Zaleplon MPO0.4200.3810.4160.224

We also compared methods on shared objective sets. On the set shared with uniform editing, we obtained 0.570 with COMPOSE versus 0.490. On the set shared with learned primitive editing, we obtained 0.544 versus 0.405. Across these analyses, we consistently observed gains from learned, structured proposals. The completed paired experiment further supports the specific contribution of created-atom dependencies raised in the review.

Q2 continued. Reference guidance and docking in T4

In T4, we construct shallow edits, anchored replacements, and route-complete programs through the executor. For reference-guided selection, we score the executed transitions with the frozen reference. We average their log-probabilities to reweight program selection. We then dock feasible candidates and update the archive. We sample archive parents without docking-score ranking, with lower sampling frequency for exhausted parents. Docking scores determine the best feasible endpoints. The reference-law interventions above quantify effects in free-running walks and fragment generation.

Q3. Evaluation and efficiency

Can QED editing be evaluated on the same source panel with matched K, and can T4 baseline molecules be rescored consistently? Please provide PMO anytime Top-10 curves, sensitivity to the two isomer objectives, and proposal/compute cost per useful oracle evaluation.

1. Candidate allowance on a fixed QED panel. To measure the effect of candidate allowance, we evaluated the same 800 held-out sources using fixed COMPOSE candidate sequences. We scored the first eight or twenty candidates per source and counted missing positions as failures. Success requires QED ≥ 0.9 and source Tanimoto similarity ≥ 0.4.

Candidate allowanceSuccessful sourcesSuccess rate
K = 8409 / 80051.13%
K = 20487 / 80060.88%

With twenty candidates, we improved success by 9.75 percentage points (95% source-bootstrap CI [7.75, 11.88]). This paired comparison quantifies the effect of candidate allowance within COMPOSE. For GrIDDD, the submission reports results from its published source panel.

2. T4 results and new baseline rescoring. We report all 30 settings across five targets, three starting leads, and two similarity floors. Feasible molecules satisfy QED ≥ 0.6, SA ≤ 4, and source Tanimoto similarity ≥ δ. C, I, and G denote COMPOSE, InVirtuoGen, and GenMol. Column labels specify δ. Docking scores are kcal/mol, with lower values preferred. NR denotes an unreported result.

TargetStartC .4I .4G .4C .6I .6G .6
PARP1-7.3-11.9-14.1-10.6-13.6-12.3-10.4
-7.8-12.3-13.4-11.0-12.7-11.7-9.7
-8.2-14.1-9.0-11.3-11.3-10.7-9.2
FA7-6.4-9.3-8.4-8.4-6.4-7.7-7.3
-6.7-9.4-8.9-8.4-7.8-7.5-7.6
-8.5-9.6-8.0NR-8.5-7.4NR
5HT1B-4.5-12.4-13.3-12.9-13.3-12.4-12.1
-7.6-11.9-12.0-12.3-12.2-12.0-12.0
-9.8-12.5-10.9-11.6-10.8-10.6-10.5
BRAF-9.3-12.2-10.1-10.8-9.0-9.7NR
-9.4-10.5-10.8-10.8-11.0-10.4-9.7
-9.8-10.9-10.6-10.6-11.5-10.3-10.5
JAK2-7.7-10.4-10.2-10.2-10.6-9.7-9.3
-8.0-11.3-10.5-10.0-10.8-10.4-9.4
-8.6-10.7-10.2-9.8-10.8-10.3NR

With COMPOSE, we obtained the best reported docking score in 23/30 settings using 250 evaluations per lead, compared with 1,000 for the baselines. We retain the published benchmark scores in this table.

To examine scoring variability directly, we rescored 90 released InVirtuoGen endpoints three times each using the COMPOSE pipeline. This produced 270 new docking measurements.

Within-molecule docking spreadkcal/mol
Median0.10
90th percentile1.30
Maximum6.50

For these endpoints, we measured a median spread of 0.10 kcal/mol. We report the 90th percentile and maximum to describe the larger variations across individual molecules. The earlier repeat-docking analysis in the submission gave a median of 0.70. Here, we characterize the released baseline endpoints under our pipeline. The lead-by-lead table above retains the published scores.

3. Useful progress per oracle call. To assess evaluation efficiency, we compared seven methods across 16 matched PMO objectives. Each method received 1,000 calls. We counted an evaluation as useful when it improved the running Top-10 after the 16 initialization molecules.

MethodCalls/improvement ↓Relative callsCOMPOSE winsp-value
COMPOSE9.031.00x——
GenMol V210.471.16x13/161.0e-2
GenMol10.701.18x14/161.4e-3
Graph-GA15.871.76x15/161.5e-7
MARS21.872.42x16/161.5e-8
MIMOSA22.362.48x16/163.0e-8
GraphAF24.602.73x16/161.5e-8

With COMPOSE, we required a median of 9.03 calls per useful improvement. GenMol required 1.16–1.18 times as many calls, while graph-based comparators required 1.76–2.73 times as many. COMPOSE won on 13–16 of 16 objectives across baselines. This demonstrates more efficient use of objective evaluations. We measure oracle-call cost here, with proposal-generation time and total runtime as separate quantities.

Reviewer_S9pQOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes COMPOSE, a molecular generation framework based on stochastic graph rewriting. It learns a reusable reference process over valid molecular edits and applies task-specific controllers for molecular generation and optimization without retraining the reference model.

Critical strengths
  • The idea of combining validity-preserving molecular rewrites with a reusable, goal-independent generative process is interesting.
Critical weaknesses
  • Lack of clarity and implementation details. The paper is difficult to follow. While the overall framework is described mathematically, the concrete neural architectures, training procedures, and interactions between the learned model and task-specific controllers are not sufficiently explained. Although some information is provided in the appendix, it remains difficult to understand how the complete system works and to assess its reproducibility.

  • Insufficient comparison with strong baselines. Table 1 only compares COMPOSE with InVirtuoGen and the original GenMol, Lacking advanced baselines such as GenMol (ICLR 2025). Moreover, 100% chemical validity is not particularly novel, as earlier graph-based molecular generators such as MiCaM (ICLR 2023) have already achieved this on de novo generation benchmarks.

References:

Questions for authors
  • Could the authors provide a clearer description of the neural architectures, training procedure, and the implementation of each controller?
  • Could the authors compare against GenMol and other more recent baselines under the same evaluation protocol?
6 comments

We thank the reviewer for recognizing the promise of a reusable, goal-independent molecular process and for identifying where the implementation needed a clearer explanation. To address these concerns, we provide the architecture and controller details below and have substantially expanded the experiments to include GenMol V2 and additional graph-editing and flow-based baselines.

Q1. Architecture, training, and controller implementation

Could the authors provide a clearer description of the neural architectures, training procedure, and the implementation of each controller?

COMPOSE learns probabilities over executable changes between complete molecular graphs. An exact executor determines the legal edits, a neural reference model scores them, and programs coordinate dependent edits. Task controllers then use constraints and feedback to select molecules or programs. We train the reference once and keep its parameters fixed across all four downstream applications.

1. Reference architecture. The inputs are the current graph and a continuous progress coordinate, tt. Learned atom embeddings encode element, formal charge, implicit hydrogen count, and ring topology, with separate bond-order embeddings. A two-layer MLP maps the time features tt, sin⁡(πt)\sin(\pi t), and cos⁡(πt)\cos(\pi t) to a 256-dimensional embedding.

A six-layer message-passing network uses 256-dimensional hidden states. Each layer sums neighboring atom projections and bond embeddings, concatenates this message with the current atom state and time embedding, and applies a 768 → 512 → 256 MLP with SiLU, a residual connection, and layer normalization. Mean pooling and global counts of atoms, bonds, cycles, and hydrogens provide graph features. A pairwise encoder combines atom representations with bond and ring features.

A 256 → 256 → 8 MLP predicts the edit family. Family-specific heads score locations and chemical attributes. The eight families cover atom insertion, deletion, and restatement; bond reordering and rerouting; ring closure and opening; and ring-system restatement. A scalar head parameterizes transition intensity in the continuous-time representation.

The executor masks inadmissible operations using the declared valence, connectivity, and structural rules. We sum probabilities of operations that yield the same canonical molecule, then normalize over distinct successors. The resulting kernel, Rθ(y∣x,t)R_\theta(y\mid x,t), gives the probability of successor molecule yy from current molecule xx at reference progress tt, with learned parameters θ\theta. Reapplying this kernel after each edit produces the stochastic molecular process.

2. Reference training. We trained on 1,803,032 executable transitions from 96,094 source molecules, including synthetic reversible edits. Each example supplies the current graph and progress coordinate as inputs and the canonical successor as the supervision target. We minimize successor negative log-likelihood, combining equivalent edit probabilities before evaluating the loss. Training uses AdamW with weight decay 0.01 and checkpoint selection on held-out source molecules.

The training target is the molecular transition. Downstream goals, property scores, docking measurements, and search history enter through the controllers described in the next comment. The same reference checkpoint supports fragment generation, QED editing, PMO, and protein-specific lead optimization. The QED controller's separate value model uses Adam at learning rate 0.001, a source-held-out validation fraction of 0.15, and five QED thresholds crossed with four similarity floors (Appendix F, Table 5).

3. Executable programs. Each program records primitive operations, attachment requirements, and references to original or newly created atoms. The compiler resolves these references, orders dependent operations, and validates every intermediate graph. For example, a program can insert an atom and subsequently close a ring through that atom. Tracking its identity makes the later transformation executable on the structure created earlier.

These details connect the mathematical formulation to its implementation. The executor supplies admissible molecular transitions, the neural model supplies their probabilities, and the program compiler supplies coordinated execution. The next comment explains how each controller uses these components and reports the new baseline comparisons.

Q1 continued. Task-specific controllers

Each controller reuses the reference and executor described in our preceding comment. Controller learning and feedback update task decisions while the reference parameters remain fixed.

TaskController implementation
Fragment generationPreserve specified fragments and restrict attachment sites. Motif extension and linker design use structured programs with learned transition preferences; superstructure generation and scaffold decoration use constrained edits.
QED editingA 1055 → 512 → 256 → 128 → 1 goal-value MLP guides 32 editing particles. It uses ReLU, dropout 0.1, a sigmoid output, goal-attainment labels, and a Bellman consistency term weighted by 0.3. The reference encoder stays frozen.
PMOMaintain an evaluated population archive. Combine local refinement, structured exploration, and larger dependent programs. Online ridge predictors use molecular and program features with oracle feedback to guide parent selection and candidate allocation.
T4Construct shallow edits, anchored replacements, and route-complete programs. In the reference-guided variant, average the executed transitions' log reference probabilities to reweight program selection. Sample archive parents without docking-score ranking; reduce sampling of exhausted parents. Docking updates the archive and identifies the best feasible molecules.

Q2. Stronger baselines

Could the authors compare against GenMol and other more recent baselines under the same evaluation protocol?

Yes. We added GenMol V2 and expanded PMO to eight methods. Table 1 of the submission already includes the original GenMol [1]; the new evaluation extends that comparison to the V2 release specifically suggested by the reviewer.

For fragment generation, we ran COMPOSE on ten prompts per task, 100 output slots per prompt, and three independent runs. The accounting includes invalid and repeated outputs. GenMol V2 entries below are the published repository results [1], including its one-step linker protocol.

TaskMethodQuality (%) ↑Uniqueness (%) ↑Diversity ↑Validity (%) ↑
Motif extensionCOMPOSE44.77 ± 0.3587.57 ± 0.750.625 ± 0.001100.00 ± 0.00
GenMol V249.0084.500.62699.40
SuperstructureCOMPOSE52.20 ± 1.6672.63 ± 1.700.623 ± 0.004100.00 ± 0.00
GenMol V239.0089.800.55199.70
Linker designCOMPOSE25.87 ± 0.7170.40 ± 0.520.559 ± 0.002100.00 ± 0.00
GenMol V228.6087.100.56681.80
Scaffold decorationCOMPOSE45.13 ± 0.6479.37 ± 1.250.482 ± 0.001100.00 ± 0.00
GenMol V239.7090.500.57199.20

COMPOSE improves quality by 13.20 percentage points on superstructure generation and 5.43 points on scaffold decoration, while maintaining 100% validity across all four tasks. GenMol V2 leads on motif-extension and one-step linker quality. These results identify the tasks where each method performs best and show the strength of COMPOSE for extending protected molecular structures.

We appreciate the reference to MiCaM [2], which generates molecules through connection-aware motif assembly. COMPOSE contributes a learned law over complete molecular transitions and executable programs that track edit dependencies. In our new free-running experiment, the learned law improves FCD from 14.35 to 2.77 relative to uniform editing under identical validity rules, directly demonstrating its distributional benefit.

For PMO, we evaluated all eight methods on 22 objectives with 1,000 charged oracle calls per run, including initialization. Each objective receives equal weight. Final Top-10 averages the ten best evaluated scores; AUC-Top10 uses checkpoints every 100 calls.

MethodFinal Top-10 ↑AUC-Top10 ↑
COMPOSE0.5670.482
GenMol0.5630.483
GenMol V20.5550.474
Graph-GA0.4890.415
MARS0.4590.383
GraphAF0.4220.358
MIMOSA0.3490.297
Value Matching (adapted)0.2950.266

Both GenMol versions use three runs per objective and two warmup settings, with the stronger setting reported. Value Matching uses our PMO adaptation of its published value-training procedure with frozen FlowMol [3]. We retain each method's optimization procedure under the common oracle budget. The next comment provides every objective-level result.

[1] GenMol and V2, official repository. [2] Geng et al., MiCaM, ICLR 2023. [3] Jensen et al., Value Matching, ICLR 2026.

Q2 continued. All 22 PMO objectives

The tables report every objective under the 1,000-call protocol described above. Higher is better. VM denotes our Value Matching adaptation with frozen FlowMol.

Final Top-10.

ObjectiveCOMPOSEGenMolV2Graph-GAMARSGraphAFMIMOSAVM
Albuterol similarity0.7780.7010.6680.4810.4860.4640.4700.382
Amlodipine MPO0.5710.5460.5420.5160.4800.4510.3920.436
Celecoxib rediscovery0.2860.4590.4630.2760.2800.2730.2300.176
Deco hop0.5860.6120.6110.5820.5820.5600.5340.555
DRD20.9690.9940.9990.9850.7100.7160.0540.027
Fexofenadine MPO0.7520.7780.7180.6900.6800.6390.5850.555
GSK3B0.5020.5960.5730.3900.3910.2310.1170.153
Isomers C7H8N2O21.0000.8610.8620.6680.6910.6600.7380.011
Isomers C9H10N2O2PF2Cl0.8740.7630.6350.5990.5340.5650.5970.157
JNK30.3140.3390.4560.2050.2640.1050.0730.092
Median10.2780.2410.2400.2320.1860.1790.2530.164
Median20.1890.1930.2120.1670.1600.1510.1300.139
Mestranol similarity0.5160.4500.4840.3900.3480.3470.3300.253
Osimertinib MPO0.8430.7920.7840.7780.7430.6730.7220.725
Perindopril MPO0.5030.5140.4760.4820.4530.4000.1900.435
QED0.9380.9410.9430.9360.9210.9150.9150.880
Ranolazine MPO0.7830.7420.6610.6900.7330.5830.2120.173
Scaffold hop0.4590.4990.5010.4650.4570.4250.3900.419
Sitagliptin MPO0.2200.2660.2480.3090.1660.1390.1350.112
Thiothixene rediscovery0.3880.3940.4030.2850.2580.2740.1960.224
Troglitazone rediscovery0.3080.2930.2990.2220.2280.2060.2000.174
Zaleplon MPO0.4200.4130.4380.4090.3400.3370.2070.244
Mean0.5670.5630.5550.4890.4590.4220.3490.295

AUC-Top10.

ObjectiveCOMPOSEGenMolV2Graph-GAMARSGraphAFMIMOSAVM
Albuterol similarity0.5670.5610.5250.3930.4350.4140.3910.354
Amlodipine MPO0.5110.4770.4820.4670.4330.4100.3540.409
Celecoxib rediscovery0.2440.3690.3700.2300.2410.2350.1910.165
Deco hop0.5470.5670.5650.5370.5420.5260.5010.516
DRD20.7740.8540.8370.7310.4590.4660.0290.023
Fexofenadine MPO0.6830.6980.6460.6250.6040.5690.5410.513
GSK3B0.4230.4970.4650.2560.2650.1840.0790.123
Isomers C7H8N2O20.8550.7000.7160.5560.5650.5360.5230.007
Isomers C9H10N2O2PF2Cl0.7370.6310.5420.5120.4350.4500.4580.060
JNK30.2020.2740.3280.1340.1580.0800.0510.072
Median10.2380.2050.2060.1840.1650.1560.2070.150
Median20.1680.1670.1800.1470.1450.1360.1210.129
Mestranol similarity0.4410.3730.4070.3180.3080.2960.2700.232
Osimertinib MPO0.7580.7330.7250.7130.6210.5650.6670.670
Perindopril MPO0.4420.4530.4300.4380.4030.3530.1770.383
QED0.8740.8810.8810.8770.8590.8520.8550.815
Ranolazine MPO0.6630.6380.5660.5710.5870.4760.1540.161
Scaffold hop0.4200.4580.4570.4240.4220.3950.3660.395
Sitagliptin MPO0.1710.1900.1910.2270.0860.0810.0800.085
Thiothixene rediscovery0.3150.3140.3210.2530.2290.2410.1840.207
Troglitazone rediscovery0.2510.2430.2470.1870.2000.1830.1670.161
Zaleplon MPO0.3260.3370.3500.3390.2770.2790.1680.220
Mean0.4820.4830.4740.4150.3830.3580.2970.266

COMPOSE achieves the highest mean final Top-10, 0.567, and the highest final score on 9 of 22 objectives. Its AUC of 0.482 is closely matched to GenMol's 0.483, with both exceeding the other methods. Thus, the expanded comparison demonstrates strong endpoint quality together with competitive performance throughout the oracle budget.

We thank the reviewer again! We have provided the requested architecture and controller details, added GenMol V2, and reported complete results against seven baselines. These experiments, in addition to the results on fragment-constrained generation and the T4 docking task at fourfold less oracle calls help to show the practical value of COMPOSE's reusable molecular process and strengthen the case for acceptance. We respectfully ask the reviewer to increase the score and reassess confidence in the submission to support acceptance of our work! Thank you so much!

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

This paper introduces COMPOSE, which first generates stochastic sequences of molecular edits between different valid molecules and then normalizes the edits distribution based on the number of canonical successors. These edits are then put together to form a structured program, which can be combined with different controllers to implement specific constraints for tasks like oracle reward based generation. The main strength of COMPOSE is that the generated edits / program is fixed after training, which provides a shared foundation for different downstream generation tasks. Benchmarks on fragment-constrained generation, similarity-constrained molecular editing, general black-box optimization, and protein-specific lead optimization show that COMPOSE can consistently deliver better or competitive performance compared with the baseline models.

Critical strengths
  • Interesting idea: most of the existing molecular editing frameworks either define a library of legal edits or allow edits that pass basic chemical validity checks and add additional filter post generation. The COMPOSE method proposed here sits in the middle of these two cases and defines legal edits through a stochastic sampling process. This appears to be one step further compared with the reaction based models like SynNet, as here the legal edits and their probability distributions are data-driven.
  • Solid experiment design with strong empirical results: the benchmark and ablation experiments are well designed and present strong model performance under different conditions. Fragment-based design clearly demonstrates the contribution of the learned reference process over uniform distribution.
Critical weaknesses
  • Implicit assumption with bias: It seems that the authors have an implicit assumption that the number of existing molecules / successors represents whether an edit is more favorable or not, as this is the way used to normalize the distribution. This appears to be a too strong assumption as we can have an easy edit for a rarely existing family of molecules, which from the editing perspective should be favorable, but would get a low score in the COMPOSE framework. It would be great if the authors can discuss this point in the paper.
Questions for authors
  1. The author argues that ‘the reference process learns state-dependent transitions’. Is the state-dependent referring to the current molecule or some other states?
  2. When calculating the edits probability in the reference process, do authors calculate each single edit or a series of edits? I think the manuscript indicates the former one, but if that is the case, why different executable edits can produce the same molecule? (see main text line 134)
  3. How large is the legal edits pool that COMPOSE constructed?
3 comments

We thank the reviewer for the thoughtful assessment and for recognizing the value of a reusable molecular editing process. The questions identify an important distinction between learning edit preferences and representing equivalent molecular outcomes. We clarify both below and provide new measurements of the available editing space.

W1. How COMPOSE assigns probability to rare edits

It seems that the authors have an implicit assumption that the number of existing molecules / successors represents whether an edit is more favorable or not, as this is the way used to normalize the distribution. This appears to be a too strong assumption as we can have an easy edit for a rarely existing family of molecules, which from the editing perspective should be favorable, but would get a low score in the COMPOSE framework. It would be great if the authors can discuss this point in the paper.

We appreciate the opportunity to clarify this. COMPOSE learns an edit family's probability from the current molecule, then distributes that probability among the family's legal operations. Thus, a family containing few admissible edits can receive high probability when the molecular context favors those transformations. This directly addresses the example raised by the reviewer.

Canonical aggregation then combines operations that produce the same molecular outcome. Proposition 3 establishes that redistributing probability among equivalent edit representations preserves the reference kernel and successor-level control whenever the total probability of each canonical successor is preserved. The learned model therefore determines transformation preferences, while aggregation represents those preferences consistently over distinct molecules.

Q1. Meaning of state dependence

The author argues that ‘the reference process learns state-dependent transitions’. Is the state-dependent referring to the current molecule or some other states?

Yes. The state is the current complete molecular graph, including its atoms, bonds, and connectivity. The model also receives the reference process's progress coordinate. After each transformation, COMPOSE evaluates the resulting graph and recomputes its legal edits and transition probabilities.

The progress coordinate belongs to the reference process. The downstream controller separately receives the optimization objective, oracle scores, search history, and remaining search budget. This separation allows us to reuse the same frozen reference across design tasks.

Q2. Individual edits and shared molecular outcomes

When calculating the edits probability in the reference process, do authors calculate each single edit or a series of edits? I think the manuscript indicates the former one, but if that is the case, why different executable edits can produce the same molecule? (see main text line 134)

The model assigns probabilities to individual legal edits. Structured programs compose those edits, with the current molecular state determining the available transformations at each step.

For a concrete example, isobutane, CC(C)C, has three equivalent terminal methyl groups. Deleting any one produces propane, CCC. These three atom-indexed operations therefore share one canonical molecular successor. Summing their probabilities gives the probability of producing propane, which is the quantity used for successor-level control.

Q3. Size of the legal editing space

How large is the legal edits pool that COMPOSE constructed?

To quantify this directly, we analyzed 44,458 recorded reference panels spanning 20,336 distinct parent molecules from our PMO campaigns. These molecules contained 14–30 heavy atoms, with a median of 22.

Quantity per molecular stateMedian
Legal edit operations587
Distinct canonical successors511
Edit families with legal support8 of 8

The executor instantiates the eight primitive families on each current graph. COMPOSE therefore typically chooses among hundreds of legal transformations, and each committed edit creates a new editing neighborhood. Programs extend this process by coordinating dependent transformations across successive neighborhoods.

We thank the reviewer again for helping us clarify these mechanisms. The hierarchical probabilities address the concern about rare edit families, the symmetry example explains canonical aggregation, and the new measurements quantify the editing space. We would be grateful if the reviewer would consider increasing the score and continue to support acceptance of our work!

I would like to thank the authors for the clear responses, which directly addressed my questions about this work. I also want to suggest the author to incorporate the detailed explanation of how COMPOSE assigns probability to rare edits into the future version of this paper. I am happy to maintain my previous score of 3.

Dear Reviewer YTQh,

We sincerely appreciate your continued positive assessment and thoughtful feedback. We are glad our responses addressed your questions, and we will incorporate your suggested clarification on how COMPOSE assigns probabilities to rare edits in the revised manuscript. Thank you again for the time and care you have devoted to reviewing our work.

Best regards,
The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response

We thank the reviewers for the thoughtful evaluation of COMPOSE. During rebuttal, we compared eight methods, completed all 18 paired program experiments, and tested continued optimization after a goal change. Reviewer YTQh confirmed that we addressed the questions and maintained support for acceptance. Reviewer S9pQ also announced a score increase following our responses. We appreciate this recognition of our additional work.

Importance of a reusable molecular editing process

During molecular optimization, we often need to preserve part of a lead, coordinate several changes, and revise the objective as design priorities change. With COMPOSE, we learn probabilities over executable transformations between complete molecular structures. We combine those transformations into programs that track dependencies, such as closing a ring through an atom inserted earlier. We then reuse the same frozen reference model for fragment generation, similarity-constrained editing, black-box optimization, and protein-specific lead optimization.

To test what we learned independently of optimization, we generated 2,048 endpoints from 512 held-out molecules with controllers and oracle feedback disabled. Under identical execution rules, we reduced Fréchet ChemNet Distance from 14.35 with uniform legal editing to 2.77 with learned transitions and increased held-out scaffold coverage from 15.28% to 25.11%. We therefore generated molecules closer to the held-out reference distribution through learned transition preferences under the same validity constraints.

New experiments for each reviewer

  • Reviewer YTQh asked whether our probability assignment disadvantages rare edit families and how large the editing space is. We clarified that we predict each edit family's probability from the current molecule before distributing it across legal operations. A family with few available operations can therefore receive high probability when appropriate for that molecule. We also analyzed 44,458 reference panels across 20,336 distinct molecules, finding medians of 587 legal operations and 511 distinct molecular successors per state. After each edit, we recompute the available transformations on the resulting molecule.

  • Reviewer S9pQ requested implementation details and stronger baselines. We explained the neural architecture, training on 1.8 million executable transitions, and each task controller. We then compared eight methods across 22 objectives with 1,000 charged oracle calls per run, including initialization. COMPOSE achieved the highest mean final Top-10 score of 0.567, compared with 0.563 for GenMol and 0.555 for GenMol V2. Performance throughout the budget was closely matched to GenMol, with AUC-Top10 of 0.482 versus 0.483. Across 16 matched objectives, we also required a median of 9.03 calls per improvement to the running Top-10, compared with 10.47 for GenMol V2 and 10.70 for GenMol. These comparisons quantify both final quality and how effectively we used objective evaluations.

  • Reviewer Tdwi asked us to isolate the learned process, test dependent edits, and demonstrate reuse after a goal change. In addition to the free-running experiment above, we compared learned, uniform, and empirical edit preferences under the same fragment constraints. Learned preferences achieved the highest quality on all four tasks. We completed all 18 paired program comparisons and directly changed how later edits attach to atoms created earlier. Final Top-10 fell from 0.6298 to 0.5799 under this intervention, with paired Wilcoxon p = 0.0182, supporting the contribution of these dependencies. We then conducted 80 matched goal-switch comparisons, keeping the reference frozen and allowing the same three additional edits in both arms. Continuing from the current molecule outperformed restarting in 35/40 comparisons after predicted-potency optimization and 29/40 after developability optimization. We also added recorded molecular trajectories and independently replayed their transformations so readers could inspect the actual edits.

Case for acceptance

We have now directly tested the central contribution of COMPOSE. We learned useful molecular transition preferences under fixed execution rules, demonstrated the benefit of preserving dependencies between edits, and continued optimization successfully after changing the goal. Across 22 PMO objectives, we achieved the highest mean final Top-10 among eight methods. We also achieved the highest mean reported quality across fragment-constrained tasks and the best reported docking score in 23 of 30 T4 settings using 250 evaluations per lead, compared with 1,000 for the baselines. With continued support for acceptance and direct evidence addressing the methodological questions, we respectfully ask the Area Chair and reviewers to reconsider the scores and support acceptance of COMPOSE.

Thank you,
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.