acceptodds
Under review as a conference paper at ICLR 2027

MOSA: Molecular Subspace Assembly for Unified Molecular Generation and Editing

Abstract

Molecular design often requires both generating new molecules from multiple property specifications and editing existing molecules while preserving their underlying structures. Although these tasks are closely related, conventional joint adaptation can benefit one task while degrading the other. We introduce , a unified adapter for a frozen language model that organizes low-rank updates into shared, generation-private, and editing-private subspaces. A permutation-invariant property controller uses individual conditions and their pairwise interactions to modulate the shared update. For editing, a source controller combines the input molecule with the requested properties to scale and shift the editing-specific update. The same frozen backbone and text interface serve both tasks, while each request changes the adaptation applied to hidden states. At 39,500 updates, MOSA achieves strict de novo success on 6,000 common requests versus SketchMol's 54.5% under native training budgets, and strict editing success on 5,000 requests. At matched one-pass exposure, MOSA improves both tasks over a conventional joint adapter and avoids the editing-side negative transfer observed relative to the editing specialist.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
molecular generation, molecular editing, large language models, multi-property optimization, parameter-efficient fine-tuning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_AC8vOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MOSA, a unified LoRA-based adaptation method for molecular generation and editing. It uses shared and task-specific low-rank branches, with property-conditioned modulation of the shared branch and additional source conditioning for editing. Experiments show encouraging improvements over conventional joint LoRA and task-specific baselines under matched task exposure.

Critical strengths
  • The shared/private LoRA design provides an interesting and well-motivated approach to reducing negative transfer between molecular generation and editing.

  • The experimental results are encouraging. The paper includes several useful ablations and controlled comparisons to support its findings.

Critical weaknesses
  • Comparisons with existing multi-task LoRA methods are missing, making the contribution of the learned controllers less clear.

  • Generalization remains insufficiently evaluated. Without quantifying train/test molecular overlap or testing on OOD molecules and property combinations, it is difficult to assess how well the method generalizes.

Questions for authors
  • Why does the property controller only use a scaling term gcg_c, whereas the editing controller uses both geg_e and beb_e?

  • How does MOSA compare with existing multi-task LoRA approaches under comparable settings?

  • Could the authors provide train/test molecular and scaffold overlap statistics, and evaluate generalization to unseen molecular scaffolds or property combinations?

7 comments
AuthorsOpenReview ↗

Response to Reviewer AC8v — Part 1/3: Summary and W1/Q2 — MTL-LoRA

Dear Reviewer AC8v,

Thank you for the careful reading and for highlighting the need for a multi-task LoRA comparison and a train/test overlap audit. These requests helped us address gaps in the original evaluation. We have added the matched MTL-LoRA comparison (Tables R1–R2), audited overlap and evaluated connectivity-disjoint panels (Tables R3–R4), and clarified the editing controller’s additive shift. In the matched single-run comparison, MOSA has higher editing success, with the largest reduction in source-retention failures. We hope the new evidence and explanations address your concerns. Our responses are organized into three parts covering W1/Q2, W2/Q3, and Q1.

W1/Q2. Comparison with MTL-LoRA

The matched comparison favors MOSA on editing. Tables R1 and R2 report overall success and editing failure types; Appendix I.1 and manuscript Table 32 give the full comparison.

(1) MOSA gains 2.78 editing points over MTL-LoRA in the matched comparison. Both models use the same backbone, data, and training updates [1]. The paired 95% interval for the editing gain is [1.72, 3.87] percentage points. The generation difference is 1.43 points. Trainable parameter counts are 71.76M for MOSA and 75.39M for MTL-LoRA.

Table R1. Matched MTL-LoRA and MOSA success (%).

MethodGeneration Accstrictgen\mathrm{Acc}^{\mathrm{gen}}_{\mathrm{strict}} ↑Generation Accrelaxedgen\mathrm{Acc}^{\mathrm{gen}}_{\mathrm{relaxed}} ↑Editing Acc0.65edit\mathrm{Acc}^{\mathrm{edit}}_{0.65} ↑Editing Acc0.15edit\mathrm{Acc}^{\mathrm{edit}}_{0.15} ↑
MTL-LoRA51.2783.9368.7485.30
MOSA52.7084.7871.5286.86

Source: manuscript Table 32 (Appendix I.1). Metrics follow Section 4.3. Generation uses 6,000 requests with strict or original SketchMol tolerances. Editing uses 5,000 requests with Tanimoto similarity ≥ 0.65 (strict) or ≥ 0.15 (relaxed). Both criteria score the same saved outputs. Both models use the same backbone, data, and 16,283 updates. Data exposure is matched; rank and compute differ. All entries come from one training run per method.


(2) MOSA produces 139 more successful edits, with all three failure types reduced. Structural failures show the largest reduction, at 1.20 percentage points (Table R2).

Table R2. Editing failure rates (%).

MethodFailstructedit\mathrm{Fail}^{\mathrm{edit}}_{\mathrm{struct}} ↓Failpropertyedit\mathrm{Fail}^{\mathrm{edit}}_{\mathrm{property}} ↓Failinvalidedit\mathrm{Fail}^{\mathrm{edit}}_{\mathrm{invalid}} ↓
MTL-LoRA16.6210.983.66
MOSA15.4210.262.80

Source: the failure breakdown in Appendix I.1 accompanying manuscript Table 32. Rates use the full 5,000-request panel under strict editing criteria (similarity ≥ 0.65). The three categories are mutually exclusive. Lower is better; bold marks the better result.


(3) MTL-LoRA learns task-specific coefficients; MOSA also adapts coefficients to each request. MOSA uses property goals to adjust its shared update, and source and goal features to adjust its editing update.

The clearest benefit over MTL-LoRA is in editing, with the largest failure reduction in source retention. Section 5.1 and Appendix I.1 discuss this finding. This comparison uses one training run per method, so the 1.43-point generation difference remains descriptive. Differences in rank and architecture also limit attribution to individual components or parameter efficiency.

References

[1] Yang et al. MTL-LoRA: Low-Rank Adaptation for Multi-Task Learning. AAAI, 2025. Paper.

AuthorsOpenReview ↗

Response to Reviewer AC8v — Part 2/3: W2/Q3 — Overlap and scaffold generalization

W2/Q3. Overlap and scaffold generalization

MOSA reaches 56.0% strict generation success on unseen-reference scaffold requests, exceeding the generation specialist by 11.67 percentage points. The paired 95% interval is [6.97, 16.28] points. We audited molecular and scaffold overlap and evaluated separate scaffold panels to address your generalization concerns. Section 5.3 summarizes the results; Appendix I.2 provides the overlap audit and scaffold evaluation in manuscript Tables 33–34.


(1) We quantified molecular and scaffold overlap with the fine-tuning data. Exact molecular overlap is 0.17% for generation and 0.04% for editing (Table R3). The corresponding scaffold overlap is 50.28% and 40.54%. To evaluate generalization beyond these shared scaffolds, we additionally constructed separate seen- and unseen-scaffold panels with no molecular connectivity overlap with the fine-tuning data (Table R4).

Table R3. Molecular and scaffold overlap with fine-tuning data.

Evaluated moleculesMolecule overlap (train/test)Scaffold overlap (train/test)
Feasibility reference molecules for generation10/6,000 (0.17%)3,017/6,000 (50.28%)
Source molecules for editing2/5,000 (0.04%)2,024/4,992 (40.54%)

Source: manuscript Table 33 (Appendix I.2), selected columns. The audited training corpus includes generation targets and both sides of editing pairs. Molecular identity uses canonical isomeric SMILES; scaffolds use nonchiral Bemis–Murcko structures. Eight editing sources with empty scaffolds are excluded from the scaffold denominator.


(2) MOSA improves strict generation success on unseen-reference scaffolds. MOSA achieves 56.0%, compared with 44.3% for the generation specialist. On unseen-source editing, MOSA achieves 67.6%, compared with 64.0% for Joint-24 and 64.2% for the editing specialist. The respective editing gains are 3.60 [−0.63, 7.99] and 3.40 [−0.72, 7.43] percentage points; both intervals cross zero. The editing gain over MOSA-Fixed is 3.20 [0.14, 6.44] points. All intervals are paired 95% cluster intervals.

Table R4. Strict and relaxed success on separate scaffold panels (%).

ModelAccstrictgen\mathrm{Acc}^{\mathrm{gen}}_{\mathrm{strict}} SeenAccstrictgen\mathrm{Acc}^{\mathrm{gen}}_{\mathrm{strict}} UnseenAccrelaxedgen\mathrm{Acc}^{\mathrm{gen}}_{\mathrm{relaxed}} SeenAccrelaxedgen\mathrm{Acc}^{\mathrm{gen}}_{\mathrm{relaxed}} UnseenAcc0.65edit\mathrm{Acc}^{\mathrm{edit}}_{0.65} SeenAcc0.65edit\mathrm{Acc}^{\mathrm{edit}}_{0.65} UnseenAcc0.15edit\mathrm{Acc}^{\mathrm{edit}}_{0.15} SeenAcc0.15edit\mathrm{Acc}^{\mathrm{edit}}_{0.15} Unseen
Specialist42.744.378.280.268.464.289.879.8
Joint-2453.854.786.786.868.264.089.881.0
MOSA-Fixed55.553.385.287.071.064.488.282.0
MOSA58.556.088.385.571.067.690.483.0

Source: manuscript Table 34 (Appendix I.2). Training exposure is matched by task. Each group contains 600 generation or 500 editing requests. Strict and relaxed criteria score the same saved outputs using Section 4.3. Bold marks the highest score in each column. These separate panels have no molecular connectivity overlap with fine-tuning data.


(3) The scaffold groups refer to generation reference molecules and editing source molecules. “Unseen” means that the corresponding scaffold is absent from the audited fine-tuning corpus. For generation, the reference molecule supplies the requested property values; the model receives those values, and the grouping places no scaffold constraint on its output. The seen and unseen panels have different request distributions, so comparisons between models are made within each panel. Generalization to unseen combinations of property goals remains a separate, untested setting, acknowledged in Section 6.

AuthorsOpenReview ↗

Response to Reviewer AC8v — Part 3/3: Q1 — Editing shift

Q1. Editing shift

The editing shift adds a source-conditioned offset beyond multiplicative scaling. Both scale and shift depend on the source molecule and property goals. The shift adds a correction independent of the current token feature. Section 3.3 and Appendix I.4 explain this distinction. The fixed-weight controls evaluate the contribution of scale and shift together (Appendix E.2, manuscript Table 27).


(1) Editing must preserve a source as well as meet property goals. Both tasks use shared property scaling. Editing additionally requires source-specific control, which motivates the scale and shift terms in its private block. Shared property scaling reweights low-rank features according to the requested goals, while the editing controller provides an additional source-conditioned adjustment.

(2) The shift can contribute even when the token feature is small. The editing update takes the form

Be(ge⊙Aeh+be).B_e(g_e \odot A_e h + b_e).

The scale geg_e reweights the projected token feature AehA_e h. The shift contributes the additive term BebeB_e b_e, which can remain nonzero even when AehA_e h is small. This gives the editing block an additional way to incorporate source and goal information.


(3) Removing editing modulation lowers strict success by 9.80 percentage points, mostly through source-retention failures. With trained weights fixed, setting ge=1g_e=1 and be=0b_e=0 raises structural failures by 7.40 points and lowers relaxed success by 2.38 points (Table R5).

Table R5. Inference-only editing intervention (%).

ConfigurationAcc0.65edit\mathrm{Acc}^{\mathrm{edit}}_{0.65} ↑Acc0.15edit\mathrm{Acc}^{\mathrm{edit}}_{0.15} ↑Failstructedit\mathrm{Fail}^{\mathrm{edit}}_{\mathrm{struct}} ↓Failpropertyedit\mathrm{Fail}^{\mathrm{edit}}_{\mathrm{property}} ↓Failinvalidedit\mathrm{Fail}^{\mathrm{edit}}_{\mathrm{invalid}} ↓
Full MOSA71.5286.8615.4210.262.80
Inference-only intervention (ge=1g_e=1, be=0b_e=0)61.7284.4822.8212.383.08

Source: manuscript Table 27 (Appendix E.2) and Appendix I.4. Rates (%) use the same 5,000 editing requests. Both success criteria score the same saved outputs. The three failure rates use the strict criterion and sum to 100 with strict editing success. All trained weights remain fixed. For strict success, intervention minus full MOSA is −9.80 percentage points, with a paired 95% source-cluster interval of [−11.08, −8.55].


These controls establish reliance on the combined editing modulation. Isolating the shift’s contribution would require setting only be=0b_e=0 while retaining the learned scale. Whether shared property control would benefit from an additional shift also requires a separate comparison. Appendix I.4 clarifies the scope of the current evidence.

Reviewer_3xRoOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces MOSA, which is a LoRA-based finetuning framework designed for both de novo generation and molecular editing. The main feature of this framework is a shared adapter with private representations for de novo generation and editing tasks respectively. Evaluation on the PubChem-derived structures (de novo generation) and MolEdit-Instruct dataset shows that MOSA outperforms a series of baseline models in both applications for the tasks tested, with a small gap in generated molecules validity compared to the best baselines. Ablation study suggests that without MOSA, the conventional joint adaption method can lead to an increase in de novo generation tasks but hurt model performance for editing tasks. MOSA performs the best in all the different variants tested including the one with the matched adapter capacity, demonstrating that the advantage not solely originates from the stored rank difference.

Critical strengths
  • Comprehensive experimental design with strong results: All the benchmark and ablation experiments are well designed, with a strong performance compared with the selected baselines. One thing I found particularly interesting is the negative transfer in editing tasks, which seems to persist under different conditions but disappeared at high molecular similarity regime. This could provide some insights for finetuning strategy design in real applications.
  • Well written paper: the manuscript is well organized and very easy to follow. The authors explicitly write out not only the performance improvement but also the limitations when interpreting experimental results.
Critical weaknesses
  • Relative simple target properties: the properties used here are quite simple and many of them can be calculated directly from the molecular SMILES (like MW, HBD, HBA, etc). Therefore, it is difficult to evaluate how effective MOSA is in real world applications for either drug or material design.
Questions for authors
  1. MOSA has three paths for finetuning, where a shared path has requested properties, the private path for editing has both properties and source molecule, and the private path for generation does not have any of them. The editing path design is reasonable, but why do the authors choose the shared path rather than the generation private path to have properties?
  2. For the negative transfer in editing tasks, is it because that de novo task naturally has fewer constraints and therefore is more likely to benefit from joint learning? (serve as implicit regulation)
  3. In Table 6, the gain of MOSA compared with the No-source modulation variant is minor. Can authors provide any potential explanation for it?
5 comments
AuthorsOpenReview ↗

Response to Reviewer 3xRo — Part 1/4: Summary and Q2 — Editing constraints and practical implications

Dear Reviewer 3xRo,

Thank you for your positive assessment of the experimental design and for your thoughtful questions about the architecture and ablations. We particularly appreciate your observation that the threshold-dependent negative-transfer pattern may inform fine-tuning strategies in real applications. Your observation has helped us think more deeply about the practical implications of our findings, especially how a model should share knowledge across molecular design tasks with different structural requirements.

We have clarified the design rationale (Sections 3.1–3.3), analyzed source-retention failures (Table R14), and compared retrained and inference-only controls (Table R15). We also report predicted-activity editing beyond the seven training descriptors (manuscript Table 30). Our responses are organized into four parts, covering Q2, Q1, Q3, and W1, respectively.


Q2. Task constraints and practical implications

We agree that the asymmetry in task constraints is a reasonable explanation for the observed transfer pattern. Both tasks can benefit from learning relationships between molecular structure and properties. Generation can satisfy a request with any suitable molecule, whereas editing must satisfy the property goals while remaining close to a particular source. Joint learning may therefore benefit the common property objective while making the additional source-retention requirement harder to satisfy. Our failure analysis is consistent with this interpretation.

(1) Most of Joint’s editing deficit arises from outputs that meet the property goals but fail source retention. In the original run, structural failures increase by 3.54 percentage points relative to the editing specialist, accounting for 79% of the editing regression. Success based on validity and property goals alone falls by 0.94 points. This concentration of failures supports your explanation that editing’s additional constraint matters.

Table R14. Strict editing success and failure rates (%).

ModelAcc0.65edit\mathrm{Acc}^{\mathrm{edit}}_{0.65} ↑Failstructedit\mathrm{Fail}^{\mathrm{edit}}_{\mathrm{struct}} ↓Failpropertyedit\mathrm{Fail}^{\mathrm{edit}}_{\mathrm{property}} ↓Failinvalidedit\mathrm{Fail}^{\mathrm{edit}}_{\mathrm{invalid}} ↓
Editing specialist69.017.110.53.4
Joint64.520.611.43.4
MOSA71.515.410.32.8

Source: manuscript Table 7 (Section 5.4). All entries come from the original training run and use the strict scoring criteria in Section 4.3. Rates use all 5,000 editing requests. Strict success and the three mutually exclusive failure categories sum to 100 before rounding. Joint and the specialist receive the same editing examples in the same order, with loss averaged per record. Differences and the 79% failure share use unrounded rates.


(2) The modulation intervention further connects the editing benefit to source retention. With trained weights fixed, setting ge=1g_e=1 and be=0b_e=0 raises structural failures by 7.40 points. Strict editing success falls by 9.80 points, compared with 2.38 points under relaxed scoring. These results indicate that the trained editing modulation is particularly useful when source retention is required. Part 3 presents this intervention alongside the retrained ablation (Table R15).

(3) Your observation suggests a practical consideration for fine-tuning models that serve both de novo design and lead optimization. Improvements in property satisfaction can coexist with reduced ability to retain an existing molecule. Evaluating joint adaptation against the source-retention requirement of the intended application is therefore important. Our results support sharing property knowledge while providing dedicated adaptation for source-conditioned editing: fewer source-retention failures account for 5.22 of MOSA’s 7.02-point editing gain over Joint in the original run.

As you noted, the transfer pattern also changes at high similarity thresholds. At 0.8–0.9, all compared models have low success and the MOSA–Joint advantage no longer persists. This identifies tightly constrained structural editing as a setting that needs further methodological development.

Your suggestion of implicit regularization offers a possible optimization-level explanation for the generation benefit. The current results establish the transfer pattern and locate the main editing deficit in source retention. A targeted analysis of training dynamics would help determine the contribution of implicit regularization to that pattern.

AuthorsOpenReview ↗

Response to Reviewer 3xRo — Part 2/4: Q1 — Shared property control

Q1. Shared property control

Our design organizes conditional control into two levels: a shared level for the requested property goals, and an editing-specific level for the relationship between the source molecule and those goals. This division motivates placing the property controller in the shared block and an additional source-and-property controller in the editing-private block. Sections 3.1–3.3 and Appendix I.4 explain the architecture.


(1) The shared level controls what properties the output should satisfy. Both generation and editing must meet requested molecular property goals. The property controller maps these goals to gcg_c, which modulates the shared low-rank features. Both tasks therefore train and use a common property-conditioned update.

Placing this controller in the shared block allows the requested properties to guide cross-task sharing itself. Placing it only in the generation-private block would make that particular modulation specific to generation.

(2) The editing-specific level controls how to realize those goals from a particular source molecule. Editing must improve the requested properties while retaining the source structure. The same goal can require different modifications for different sources. For example, reducing molecular weight while preserving the starting structure depends on which molecule is being edited.

The editing-private controller therefore receives both source and property information. It tailors the private update to the source–goal combination. This explains why property information appears at both levels: it specifies the common objective in the shared controller and helps determine a source-dependent adjustment in the editing controller.

(3) Generation uses the shared level together with its task-private update. Generation receives the requested properties through both the input prompt and the shared controller. Its private block operates on backbone hidden states already influenced by that request and learns generation-specific behavior. Because generation has no given source molecule, our design uses the common property control together with this private update, without adding the source-dependent level used for editing.


The asymmetry therefore follows the two levels of conditioning. Shared property control addresses the objective common to both tasks; editing-private control addresses the additional source–goal relationship. This is the design rationale for sharing the property controller while retaining a separate controller for editing.

AuthorsOpenReview ↗

Response to Reviewer 3xRo — Part 3/4: Q3 — Modulation ablation

Q3. Modulation ablation

You are right that the gap to the retrained no-source-modulation variant is modest. A possible explanation is that this variant retains the source molecule in the prompt and retains the editing-private block. These components can adapt during training and may compensate for the removed modulation. We compared this retrained variant with inference-only removal to investigate this explanation (Section 5.6; Appendix E.2, manuscript Tables 26–27).


(1) The retrained variant retains other ways to use the source molecule. Fixing the editing scale and shift to ge=1g_e=1 and be=0b_e=0 removes the explicit editing modulation. The backbone still receives the source SMILES through the prompt, and the editing-private parameters remain trainable. The model can therefore learn to perform source-conditioned editing through these remaining pathways.

In the inference-only control, we apply the same change after training while keeping all learned weights fixed. This tests how the trained model responds when its editing modulation is disabled.

Table R15. Retrained and inference-only editing controls (%).

ConfigurationAcc0.65edit\mathrm{Acc}^{\mathrm{edit}}_{0.65} ↑Property success ↑Failstructedit\mathrm{Fail}^{\mathrm{edit}}_{\mathrm{struct}} ↓
Full MOSA71.5286.9415.42
Retrained (ge=1g_e=1, be=0b_e=0)70.0686.6216.56
Inference-only (ge=1g_e=1, be=0b_e=0)61.7284.5422.82

Source: manuscript Table 27 (Appendix E.2); the retrained comparison also appears in manuscript Table 26. Rates use the same 5,000 editing requests. Property success requires validity and satisfaction of all property goals, without imposing a source-similarity threshold. The retrained row corresponds to the no-source-modulation variant. Inference-only removal keeps all trained weights fixed.


(2) The modest retrained gain is concentrated in source retention. Full MOSA improves strict editing success from 70.06% to 71.52%, a gain of 1.46 percentage points, with a paired 95% interval of [0.58, 2.38]. Structural failures decrease from 16.56% to 15.42%, accounting for 1.14 points of that gain. Property success changes by only 0.32 points.

Generation slightly favors the retrained variant, at 53.07% versus 52.70% for MOSA. The MOSA-minus-variant difference is −0.37 [−1.57, 0.82] points, leaving that comparison unresolved.

(3) The larger inference-only effect is consistent with compensation during retraining. Disabling editing modulation after training reduces strict success by 9.80 points, with an intervention-minus-full paired 95% interval of [−11.08, −8.55]. Structural failures increase by 7.40 points. The contrast between this effect and the smaller retrained gap is consistent with the remaining model components adapting when trained without modulation.

A separate scaffold evaluation provides complementary evidence. On unseen-source editing, MOSA exceeds MOSA-Fixed by 3.20 [0.14, 6.44] points (manuscript Table 34). MOSA-Fixed also fixes the shared coefficients, so this comparison evaluates a broader change than the editing-only ablation.


The small retrained gap therefore helps clarify the role of editing modulation. The source prompt and editing-private block already provide substantial editing capability. Learned modulation adds a measurable source-retention benefit in the retrained comparison, and the trained full model relies on it at inference. These results support its contribution while leaving room for compensation through other components during training.

Reviewer_FnM2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper introduces MOSA, an integrated model for molecular de novo generation and editing. The authors find that jointly adapting both tasks within a single transformer model leads to negative transfer that harms molecular editing performance. To address this, LoRA factors are split into shared and private components responsible for generation and editing. The loss function includes an additional separation penalty to discourage alignment between the shared and private blocks. MOSA eliminates negative transfer, increasing the success rate in both tasks relative to single-task specialized models of the same architecture. It also outperforms other generative models trained to generate molecules matching the given set of molecular properties.

Critical strengths
  1. The Authors discover negative transfer when LoRA on top of a frozen transformer is trained to serve two tasks: de novo generation and molecular editing. This is an interesting observation and a clear problem that needs a solution to build a general generative molecular design model. The performance delta is measured against single-task specialist models, showing negative transfer for the molecular editing task.
  2. The Authors find an effective way to decompose LoRA into shared and private components that can be optimized separately for de novo generation and editing. A loss term is added to discourage alignment between the shared and private down-projection rows. The results show that negative transfer is eliminated and performance improves on both tasks.
Critical weaknesses
  1. It is unclear whether the results are reproducible across multiple runs, including the core negative-transfer experiment. The text mentions three runs, but none of the tables besides Table 3 include confidence intervals. Similarly, without confidence intervals, it is difficult to judge which ablations significantly degrade the model performance. Moreover, the threshold of 0.65 Tanimoto similarity seems a little bit arbitrary, especially since the gap between specialists and joint training peaks around that value (Figure 3b), and the gap closes at 0.8-0.9.
  2. The paper is somewhat difficult to follow because it uses non-standard terms without defining them, or uses them interchangeably. For example, shared and private paths are also referred to as paths, blocks, subspaces, or branches. The abstract introduces these terms alongside other undefined terms, such as “property controller” and “text interface,” making the concept difficult to understand on the first read. Moreover, the figures share the same problem. In Figure 1, the label “7p” suggests “7 times the number of requested properties” rather than “7 properties.” Figure 2 contains symbols that are not introduced, such as LSEPL_{SEP}, and Figure 3 does not explain the categories in subplot (a): what are the failures related to structure only, property, or invalid? Many other passages are unclear; I encourage the Authors to proofread the text again and add necessary clarifications.
  3. The discussion of the problem and results is very brief and usually limited to citing the numbers in the tables. It would be valuable to discuss why negative transfer occurs and how this model addresses it by analyzing each experiment in more depth.
Questions for authors
  1. Are all experiments repeated three times? What is the spread of Δ\Delta in Table 2?
  2. How were the Tanimoto similarity threshold and tolerance values selected?
  3. Why is Table 5 limited to only 3 baselines? There are many excellent de novo generation models that could be included, including some already in the editing benchmark, like REINVENT.
  4. Why does negative transfer affect only the molecular editing task? Do you think that is due to training dynamics or an imbalance in the data?
13 comments
AuthorsOpenReview ↗

Response to Reviewer FnM2 — Part 1/5: Summary and W2 — Method clarity

Dear Reviewer FnM2,

We agree that the original presentation made the method and experiments difficult to follow. Thank you for identifying the unclear terminology and figure labels, and for asking for stronger evidence of repeatability. We have clarified the block operations (Section 3.2), Figure 1, and scoring criteria (Section 4.3), repeated the core comparison across three seeds (Table R10), and added REINVENT4 (Table R12). MOSA outperforms Joint on editing in all three runs; the smaller advantage over the editing specialist is less certain. We hope these revisions make the method and evidence easier to assess. Our point-by-point responses are organized into five parts below.


W2. Method clarity

The revisions address the terminology, figure labels, and evaluation criteria that were unclear.

(1) Blocks, controllers, and prompts have distinct roles. Each block is a low-rank update with its own AA and BB matrices. Controllers adjust low-rank features using the property goals and, for editing, the source molecule. The prompt supplies the molecular request to the backbone. Section 3.2 now explains the sequence AA, modulation, BB; Algorithm 1 matches that sequence and the condition aggregation in the code. Figures 1–2 use the same names.

(2) Figure labels and failure categories are defined. Figure 1 now labels the number of properties explicitly (e.g., “7 properties”). We also clarify that LsepL_{\mathrm{sep}} discourages alignment between shared and private blocks. The metric definitions distinguish invalid outputs, unmet property goals, and failed source retention. Section 4.3 and Table R6 (manuscript Table 3) define these metrics; Section 5.4 and manuscript Table 7 report the failure breakdown. Appendix A uses the same percentage metrics and response-format checks.

Table R6. Definitions of evaluation metrics.

TaskMetricDefinition
GenerationAccstrictgen\mathrm{Acc}^{\mathrm{gen}}_{\mathrm{strict}}Valid; all targets within strict tolerances.
GenerationAccrelaxedgen\mathrm{Acc}^{\mathrm{gen}}_{\mathrm{relaxed}}Valid; all targets within relaxed tolerances.
EditingAcc0.65edit\mathrm{Acc}^{\mathrm{edit}}_{0.65}Valid; all property goals met; similarity ≥ 0.65.
EditingAcc0.15edit\mathrm{Acc}^{\mathrm{edit}}_{0.15}Valid; all property goals met; similarity ≥ 0.15.
EditingFailinvalidedit\mathrm{Fail}^{\mathrm{edit}}_{\mathrm{invalid}}Invalid or missing output.
EditingFailpropertyedit\mathrm{Fail}^{\mathrm{edit}}_{\mathrm{property}}Valid; at least one property goal unmet.
EditingFailstructedit\mathrm{Fail}^{\mathrm{edit}}_{\mathrm{struct}}Valid; all property goals met; similarity < 0.65.

Source: manuscript Table 3 (Section 4.3).

(3) Specialist comparisons identify negative transfer, while Joint-24 controls for stored LoRA capacity. Table R7 presents strict and relaxed success side by side; Section 5.1 explains these comparisons.

Table R7. Strict and relaxed success at matched one-pass exposure (%).

ModelGeneration Accstrictgen\mathrm{Acc}^{\mathrm{gen}}_{\mathrm{strict}}Generation Accrelaxedgen\mathrm{Acc}^{\mathrm{gen}}_{\mathrm{relaxed}}Editing Acc0.65edit\mathrm{Acc}^{\mathrm{edit}}_{0.65}Editing Acc0.15edit\mathrm{Acc}^{\mathrm{edit}}_{0.15}
Specialist44.278.069.086.1
Joint48.383.164.585.1
Joint-2452.584.466.886.4
MOSA-Fixed52.384.570.786.8
MOSA52.784.871.586.9

Source: manuscript Table 4 (Section 5.1). For each task, specialists and joint models see the same training examples. All entries use same training seed; scoring criteria follow Section 4.3. Rates use the full 6,000-generation/5,000-editing panels.

AuthorsOpenReview ↗

Supplement to Part 1/5 — Model definitions and the purpose of each comparison

To clarify the model names in Table R7 of Part 1/5, we summarize their configurations and the purpose of each comparison below. All models use the same frozen Qwen2.5-VL-7B-Instruct backbone.


Table. Model definitions and comparison roles.

ModelConfigurationPurpose
SpecialistTwo separately trained rank-16 LoRA models: one trained only on generation and one trained only on editing. The generation and editing columns report the corresponding specialist.Provides the single-task reference for identifying positive or negative transfer.
JointOne rank-16 LoRA adapter trained on both generation and editing, with the same adapter parameters shared across tasks.Tests the effect of joint training with full adapter sharing.
Joint-24One rank-24 LoRA adapter trained on both tasks. Its stored LoRA parameter count matches MOSA’s three rank-8 blocks.Tests whether increasing the stored LoRA capacity of the fully shared baseline accounts for MOSA’s improvement.
MOSA-FixedRetains MOSA’s shared block and two task-private blocks, but fixes controller scales to one and shifts to zero throughout training and evaluation.Tests the shared/private architecture with fixed modulation coefficients.
MOSAUses one rank-8 shared block and two rank-8 task-private blocks. Each request activates the shared block and its corresponding task-private block. Learned controllers condition the shared update on property goals and the editing-private update on the source and property goals.Evaluates the complete method combining selective parameter sharing with conditional modulation.

The comparisons address complementary questions. Joint versus Specialist establishes the transfer pattern. Joint-24 provides a control for stored LoRA capacity. MOSA-Fixed retains the shared/private structure, allowing us to assess the additional contribution of learned conditional modulation.

Two distinctions are relevant when interpreting these controls. First, Joint-24 matches MOSA’s stored LoRA parameters, but MOSA additionally contains controller parameters; Joint-24 activates rank 24 per request, whereas MOSA activates rank 16. Second, MOSA-Fixed still receives the property request and, for editing, the source molecule through its prompt. Fixing its modulation coefficients therefore does not remove this information from the model.

The comparison in Part 1/5 matches the number of training examples seen from each task. Each specialist receives the same own-task examples as the joint models, while the joint models additionally train on the other task.

AuthorsOpenReview ↗

Response to Reviewer FnM2 — Part 2/5: W1/Q2 — Metrics and thresholds

W1/Q2. Metrics and thresholds

We adopted the 0.65 editing cutoff from MolEditRL and the relaxed generation tolerances from SketchMol’s released code. Strict and relaxed results show how requiring closer source retention or more accurate property matching affects the comparison. Section 4.3 defines the metrics; Appendix A (manuscript Table 9) and Appendix I.3 (manuscript Table 35) list the tolerances.


(1) The editing thresholds measure different levels of source retention. Table R8 compares the two editing cutoffs. We adopted 0.65 to match MolEditRL [2], the source of our editing dataset. The previously reported 0.15 setting relaxes that requirement while still requiring valid molecules that meet every property goal. Reporting them together shows whether a difference lies mainly in preserving the source. MOSA exceeds Joint and the editing specialist at both cutoffs.

Table R8. Editing success at both similarity thresholds (%).

MethodAcc0.65edit\mathrm{Acc}^{\mathrm{edit}}_{0.65} ↑Acc0.15edit\mathrm{Acc}^{\mathrm{edit}}_{0.15} ↑
Editing specialist69.086.1
Joint64.585.1
MOSA71.586.9

Source: selected rows of manuscript Table 4 (Section 5.1). Success rates (%) on the same 5,000 outputs per model, as in Table R7. Both criteria require a valid output that meets every property goal; only the similarity cutoff changes.


(2) Generation tolerances specify how close each property must be to its target. Table R9 compares strict and relaxed generation success. The relaxed tolerances come from SketchMol’s released evaluation code [3] and equal 20% of each predefined property range. For example, the molecular-weight range of 150–600 gives an absolute tolerance of ±90 around the requested value. Our strict setting narrows six property tolerances and keeps HBD unchanged, requiring closer property matching. Under the original SketchMol tolerances, MOSA reaches 87.88%, versus 81.92% for SPMM and 79.12% for SketchMol. The numerical tolerances are listed in the revised manuscript.

Table R9. Strict and relaxed generation success (%).

MethodAccstrictgen\mathrm{Acc}^{\mathrm{gen}}_{\mathrm{strict}} ↑Accrelaxedgen\mathrm{Acc}^{\mathrm{gen}}_{\mathrm{relaxed}} ↑
REINVENT442.082.0
SPMM45.681.9
PEIT-LLM19.450.1
SketchMol54.579.1
MOSA60.587.9

Source: the overall strict and relaxed columns of manuscript Table 6 (Section 5.2). Success rates (%) on the same 6,000 outputs per model. MOSA uses the 39,500-update checkpoint. Relaxed scoring uses the original SketchMol tolerances of ±20% of each predefined property range. Both settings use the same validity rules and denominator.


The reported strict and relaxed comparisons favor MOSA. We now include Acc0.15edit\mathrm{Acc}^{\mathrm{edit}}_{0.15} and Accrelaxedgen\mathrm{Acc}^{\mathrm{gen}}_{\mathrm{relaxed}} alongside the strict metrics in the manuscript. At similarity thresholds of 0.8–0.9, however, the MOSA–Joint advantage no longer persists (Figure 3 and Appendix F.1).

References

[2] Zhuang, Shen, and Sun. MolEditRL: Structure-Preserving Molecular Editing via Discrete Diffusion and Reinforcement Learning. ICLR, 2026. Paper.

[3] Wang et al. Image-based Generation for Molecule Design with SketchMol. Nature Machine Intelligence 7, 244–255, 2025. Paper. Released evaluation code.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Rebuttal Summary for the Area Chair — Part 2/2

(3) Summary of Main New Experiments

Table A2. Across three runs, full sharing helped generation but hurt editing. MOSA retained the generation gain and recovered the editing loss.

Run 1 is the original run. Differences are in percentage points (pp). Brackets give paired 95% confidence intervals at fixed checkpoints. Joint-24 and MOSA-Fixed were each trained once.

AnalysisKey resultConcern addressed
Three training runsThe transfer pattern held in every run. Relative to the respective specialists, Joint-16 averaged +4.72 pp on strict generation and −5.28 pp on strict editing. MOSA averaged +9.22 and +2.07 pp. Relaxed scoring gave the same directions (Tables 23, 38).Repeatability (FnM2 W1/Q1)
Adapter capacityMOSA led on editing at matched stored rank. In Run 1, MOSA led Joint-24 by 4.70 pp on editing [3.46, 5.90] and 0.22 pp on generation [−1.25, 1.71]. At matched active rank, its editing gains over Joint-16 were 7.02, 7.04, and 7.98 pp (Tables 4, 37–38).Capacity and uncertainty (FnM2 follow-up)
REINVENT4MOSA met the property targets more often. On 6,000 requests, strict/relaxed success was 60.45%/87.88% for MOSA and 42.00%/82.03% for REINVENT4. Invalid outputs counted as failures (Table 6).Generation baseline (FnM2 Q3)
MTL-LoRAMOSA produced 139 more successful edits on 5,000 matched requests (71.52% vs. 68.74% strict success). Strict generation success was 52.70% vs. 51.27% (Table 32).Multi-task baseline (AC8v W1/Q2)
Editing failuresMost of the editing loss came from changes to the source structure. Joint-16 and the specialist saw the same 569,905 training edits in the same order. Source-retention failures accounted for 79% of Joint-16’s Run 1 deficit and 5.22 pp of MOSA’s 7.02 pp recovery (Table 7).Editing transfer (FnM2 W3/Q4, 3xRo Q2)
Similarity cutoffThe editing gain held on both sides of 0.65. In Run 1, MOSA led Joint-16 by 6.52, 7.02, and 3.18 pp at cutoffs of 0.60, 0.65, and 0.70, respectively. The gap was near zero at 0.80–0.90. Relaxed generation scoring used SketchMol tolerances (Figure 3, Tables 9, 29, 35).Threshold sensitivity (FnM2 W1/Q2, 3xRo S1)
Unseen scaffoldsThe evidence was clearer for generation. Although molecular connectivity overlap was below 0.2%, scaffold overlap was 50.28% for generation and 40.54% for editing. On unseen scaffolds, MOSA led the specialists by 11.67 pp [6.97, 16.28] and 3.40 pp [−0.72, 7.43], respectively (Tables 33–34).Generalization (AC8v W2/Q3)
Controller ablationsThe trained model relied on editing modulation. Disabling it at inference lowered strict editing success from 71.52% to 61.72%. Retraining with modulation fixed yielded 70.06%. Separately, MOSA led MOSA-Fixed by 0.78 pp [−0.11, 1.63] (Tables 26–27, 37).Controller contribution (3xRo Q3, AC8v Q1, FnM2 follow-up)
Predicted activity55.6% of 500 GSK3β requests met both the predicted-activity and source-retention criteria. Activity awaits experimental validation (Table 30).Practical scope (3xRo W1)
Method clarityThe revision defines controller roles and failure categories. Figures 1–2 and Algorithm 1 use consistent block names. Section 4.3 distinguishes invalid outputs, unmet property goals, and failed source retention.Method and metrics (FnM2 W2, 3xRo Q1, AC8v Q1)

Closing Remarks

We thank the Area Chair and all three reviewers for their time and thoughtful feedback.

Reviewer 3xRo reported that our responses addressed most of their questions. In a preliminary follow-up, Reviewer FnM2 acknowledged the clearer presentation and marked Q2 and Q3 as resolved. FnM2 also noted that the editing gap was largest at the benchmark cutoff and that the Joint-24 and MOSA-Fixed comparisons each used a single training run. Reviewer AC8v noted that our responses addressed some of their concerns and maintained their original score.

We hope the Area Chair will consider the revised manuscript and the additional evidence presented during the discussion in making the final assessment.

Sincerely,
The Authors

AuthorsOpenReview ↗

Rebuttal Summary for the Area Chair — Part 1/2

Dear Area Chair,

Thank you for your time and consideration. We also thank the three reviewers for their careful reading and constructive feedback. We summarize the following points across two comments.

  • The motivation for MOSA and its main contributions
  • The reviewers’ assessments and discussion progress
  • The additional experiments, analyses, and clarifications addressing their concerns

(1) Core Motivation and Novelties of Our Paper

MOSA addresses a question in joint molecular generation and editing: when one model handles both tasks, which parameters should be shared and which should remain task-specific? Both tasks require molecules that meet the requested property goals, but editing adds a constraint: the output must also remain structurally similar to the source molecule.

Our work makes three contributions.

(a) A diagnosis of why joint training hurts editing. We compare each specialist with a jointly trained rank-16 adapter (Joint-16) at matched task exposure. Full sharing improves generation but lowers editing success in all three runs. In the original run, edits that reach the property goals but drift too far from the source account for 79% of the net editing loss.

(b) Sharing property control, not task behavior. MOSA splits the adapter into one shared rank-8 block, conditioned on the requested properties by a controller, and one rank-8 block private to each task. Every request activates the shared block plus its own private block, so MOSA matches Joint-16’s active rank while both tasks still train a common property response.

(c) Edits conditioned on the source, not only the goal. A second controller uses both the source molecule and the property goals to scale and shift features in the editing-private block, so the editing update depends on which molecule is being edited even when the goals are the same. Disabling this modulation in the trained model raises source-retention failures by 7.40 percentage points, accounting for most of the 9.80-point drop in strict editing success.

Across all three runs, MOSA kept the generation gains of joint training and recovered the editing performance lost when the two tasks shared the entire adapter.


(2) Reviewer Ratings and Assessments

Table A1. Original ratings, reviewer confidence, and discussion progress.

ReviewerOriginal ratingConfidenceAssessment and discussion progress
FnM223Recognized editing-side negative transfer and the shared/private design. In a preliminary follow-up, reported improved clarity and marked Q2 and Q3 resolved. Kept the original score for now, noting the peak gap at 0.65 and the lack of repeated runs for Joint-24 and MOSA-Fixed.
3xRo34Highlighted the practical implications of the similarity-threshold results and asked about controller placement, ablations, and activity-guided editing. After reviewing the additional analyses, stated that the responses addressed most questions and confirmed the original rating of 3.
AC8v22Valued the shared/private design and controlled comparisons. Asked about multi-task LoRA, generalization, and the editing controller. Our responses report the MTL-LoRA comparison, overlap audit, and scaffold panels. In the follow-up, AC8v acknowledged that the responses addressed some concerns and retained the original rating of 2.

In the discussion, FnM2 marked Q2 and Q3 resolved while retaining concerns about the ablations, and 3xRo confirmed that the responses addressed most questions. AC8v reported that the responses addressed some concerns and maintained the original rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.