acceptodds
Under review as a conference paper at ICLR 2027

ForgetU: Leaving Undesirable Molecules Behind in De Novo Design

Abstract

Generative models based on flow matching and diffusion have shown strong potential for 3D de novo molecular design. However, models trained on existing molecular datasets may also learn to generate molecules with toxic or other undesired properties, limiting their reliability and practical utility. Retraining a model from scratch using only retained molecules can address this issue but may be computationally expensive for large 3D molecular datasets. Post-training a pretrained model provides an efficient alternative, but faces two key challenges: suppressing undesired generation while preserving the original generation capability, and preventing the forgotten behavior from being easily recovered through subsequent fine-tuning. In this study, we propose a Reinforcement Learning (RL)-based robust forgetting framework for pretrained 3D molecular flow matching models. RL-based forgetting suppresses undesired generation, while selective distillation preserves generation behavior on retained molecules. We further simulate relearning on undesired molecules and use the resulting loss to further update the model, improving the robustness of forgetting against subsequent fine-tuning. Experiments show that our framework achieves effective and robust forgetting while preserving the original generation capability.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Undesired molecule, Generative model, Molecular Design, Forget

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_crVZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes a Reinforcement Learning (RL)-based forgetting framework for pretrained 3D molecular flow matching models, which can hinder undesired generation while preserving the original generation capability and preventing the forgotten behavior from being easily recovered through subsequent fine-tuning. The experimental results suggest the effectiveness of the proposed method.

Critical strengths
  • This method sounds technical and interesting. Avoiding the model generating undesired molecules is also a realistic application problem.
  • Post-training with the sampling intermediate state is also a unique perspective, and the core method is novel.
Critical weaknesses
  1. The presentation is not good. The algorithm 1 and 2 overlap with the Equation.
  2. This post-training uses the data sampled from its own model, which can be viewed as self-distillation. Therefore, the model might heavily rely on the pretrained model's performance. Especially when the pretrained model generate the the samples with unbalance property (like most of them are non-toxic). Further, this model will use the intermediate state G_t in the sampling trajectory, but use G_T as the label only. The potential risk is that before getting the final state G_T, many molecules with different properties might share the same intermediate state. This will also make the fine-tuning unstable and cause the model to be confused.
  3. This method constructs the SDE version of the original ODE version in order to utilize DPO fine-tuning. However, this will make the original flow matching distinct from the fine-tuning model. Therefore, I was wondering if the model has a discrepancy between ODE and SDE sampling after fine-tuning.
  4. The baselines appear to be too weak to support the proposed method. It would be better to consider some SOTA general unlearning methods like [1]. Further, the baselines should include references, since the details of the baselines are missing in the main text.

[1] Kim, Hoki, et al. "Unlearning-aware minimization." Advances in Neural Information Processing Systems 38 (2026): 93806-93829.

Questions for authors

Did the authors try to directly use the known undesirable molecule dataset to construct Flow Matching intermediate states, and then perform PPO-based forgetting on these states?

5 comments
AuthorsOpenReview ↗

Response to Reviewer crVZ (Part 1/3 - W1 & W2Q1)

We thank the reviewer for recognizing the novelty, practical significance, effectiveness and technical soundness of our work.

Below we address the reviewer’s questions about presentation (W1), intermediate-state handling (W2Q1), ODE–SDE consistency (W3), and an additional baseline (W4).

W1:

We sincerely apologize for the presentation issue and appreciate the reviewer pointing it out. We have carefully reformatted Algorithms 1 and 2 and updated the paper. Now, all equations and pseudocode are clearly presented. We also thoroughly checked the entire manuscript to prevent similar issues.

W2Q1:

We first clarify that sampling trajectories from the current model does not make the forgetting objective self-distillation. Self-distillation typically requires the model’s own predictions, or those of a previous/frozen version, to serve as training targets. Here, the current model provides the sampled trajectories, not the training targets. In our forgetting objective, the supervision instead comes from the external undesirability estimator through the terminal reward. PPO then uses this reward to reduce the likelihood of transitions leading to undesired molecules.

Balanced desired and undesired samples are not required by our formulation. In our reward design, desired molecules receive zero forgetting reward and therefore do not provide a forgetting signal, whereas undesired molecules receive a negative reward that directly drives forgetting. Thus, as long as undesired molecules are sampled during rollout, the model receives a valid forgetting signal. Although most generated molecules may be non-toxic, this only affects the sampling efficiency of the forgetting signal. As a practical strategy, we reuse the sampled trajectories for multiple updates to improve their utilization. As forgetting progresses, fewer undesired trajectories are expected as the forgetting objective is being satisfied, rather than indicating instability. Our QM9 results further support this: despite a pretrained undesired generation rate of only 1.62% for fluorine-containing molecules, our method reduces it to 0.00%, showing effective forgetting even with sparse undesired generations.

Independently sampled trajectories almost surely do not share the same intermediate state. Each intermediate state is Gt=(Xt,At,Ct,Et)G_t = (X_t, A_t, C_t, E_t), where XtX_t contains continuous atomic coordinates. Under our stochastic transition, Xs=μψ,t→s+Vt,sϵX_s = \mu_{\psi,t\to s} + \sqrt{V_{t,s}}\epsilon, with ϵ∼N(0,I)\epsilon \sim \mathcal{N}(0,I) and Vt,s>0V_{t,s} > 0. For two independently sampled trajectories, Xs(1)−Xs(2)X_s^{(1)} - X_s^{(2)} therefore follows a non-degenerate continuous Gaussian distribution, implying Pr⁡(Xs(1)=Xs(2))=0\Pr(X_s^{(1)} = X_s^{(2)}) = 0. Since two full molecular states can be identical only if their continuous coordinates are identical, Pr⁡(Gs(1)=Gs(2))=0\Pr(G_s^{(1)} = G_s^{(2)}) = 0 as well. Therefore, except for numerical precision, two independently sampled trajectories will not encounter exactly the same intermediate molecular state.

Even if two trajectories pass through the same intermediate state, different terminal rewards do not mean that PPO receives contradictory supervision. In the standard stochastic policy-gradient framework, Q(Gt,at)=E[R(GT)∣Gt,at]Q(G_t,a_t) = \mathbb{E}[R(G_T) \mid G_t,a_t], which represents the expected terminal reward after taking transition ata_t from intermediate state GtG_t. The policy is optimized to maximize the expected return, rather than assigning a unique terminal property to each GtG_t. The corresponding on-policy policy-gradient identity is ∇ϕJ(ϕ)=Eτ∼pϕ[R(GT)∑t=0T−1∇ϕlog⁡pϕ(Gt+1∣Gt)]\nabla_{\phi}J(\phi) = \mathbb{E}_{\tau \sim p_{\phi}} \left[R(G_T) \sum_{t=0}^{T-1} \nabla_{\phi}\log p_{\phi}(G_{t+1}\mid G_t)\right], while our method uses the corresponding clipped PPO surrogate. Therefore, even if the same or similar GtG_t leads to different terminal molecules GTG_T, the algorithm can still estimate the expected reward through stochastic trajectories. Different terminal rewards do not constitute conflicting labels.

We had considered directly using a known undesirable molecule dataset to construct flow-matching intermediate states for PPO-based forgetting. However, such states follow the data-conditioned interpolation path from known terminal molecules, rather than on-policy trajectories sampled from the behavior policy ϕold\phi_{\mathrm{old}}. Our PPO likelihood ratio is defined on the same transitions actually sampled by ϕold\phi_{\mathrm{old}}. Directly applying it to externally constructed states would therefore no longer follow the on-policy PPO formulation. As an alternative, we include GA as a baseline, which directly uses known undesirable molecules as terminal states, constructs their flow-matching intermediate states through the original interpolation path, and performs gradient ascent on the corresponding flow-matching objective. As shown in Table R1, it reduces the toxicity UR to 3.84%, compared with 1.43% for our method.

AuthorsOpenReview ↗

Response to Reviewer crVZ (Part 2/3 - W3)

Table R1: Direct Use of Known Undesirable Molecules for Forgetting

MethodValidity (%) ↑Uniqueness (%) ↑Novelty (%) ↑FCD ↓UR (%) ↓Recovery@20 (%) ↓
GA96.91 ± 0.5899.95 ± 0.0399.93 ± 0.046.47 ± 0.273.84 ± 0.345.18 ± 0.57
Ours98.92 ± 0.30100.00 ± 0.00100.00 ± 0.005.31 ± 0.191.43 ± 0.051.52 ± 0.15

W3:

The constructed SDE preserves the marginal distributions of the original ODE, establishing their theoretical consistency. During PPO post-training, the deterministic ODE does not define a stochastic transition density for the continuous coordinates required to construct the PPO transition likelihood, whereas the SDE does. In Appendix B.3, we show how the marginal-preserving SDE is constructed and explain why it is consistent with the original ODE. For the reviewer’s convenience, we summarize the key derivation below.

Let pψ,t(G)p_{\psi,t}(G) denote the marginal distribution of the original flow, where G=(X,D)G=(X,D) contains continuous coordinates and discrete molecular features. The coordinate ODE is dXt=vψ(Gt,t),dtdX_t=v_\psi(G_t,t)\\,dt, and the joint distribution satisfies ∂tpψ,t=−∇X⋅(vψpψ,t)+Qψ,t∗pψ,t\partial_t p_{\psi,t}=-\nabla_X\cdot(v_\psi p_{\psi,t})+\mathcal{Q}_{\psi,t}^{\ast}p_{\psi,t}, where Qψ,t∗\mathcal{Q}_{\psi,t}^{\ast} represents the discrete CTMC dynamics. Keeping these discrete dynamics unchanged, we construct the SDE as dXt=[vψ(Gt,t)+12gt2Sψ(Gt,t)],dt+gt,dWtdX_t=[v_\psi(G_t,t)+\tfrac{1}{2}g_t^2S_\psi(G_t,t)]\\,dt+g_t\\,dW_t, with Sψ(Gt,t)=∇Xlog⁡pψ,t(Gt)S_\psi(G_t,t)=\nabla_X\log p_{\psi,t}(G_t).

Let qψ,tq_{\psi,t} denote the distribution induced by this SDE. Its forward equation is ∂tqψ,t=−∇X⋅[(vψ+12gt2Sψ)qψ,t]+12gt2ΔXqψ,t+Qψ,t∗qψ,t\partial_t q_{\psi,t}=-\nabla_X\cdot[(v_\psi+\tfrac{1}{2}g_t^2S_\psi)q_{\psi,t}]+\tfrac{1}{2}g_t^2\Delta_X q_{\psi,t}+\mathcal{Q}_{\psi,t}^{\ast}q_{\psi,t}. Since Sψpψ,t=∇Xpψ,tS_\psi p_{\psi,t}=\nabla_X p_{\psi,t}, we have ∇X⋅(Sψpψ,t)=ΔXpψ,t\nabla_X\cdot(S_\psi p_{\psi,t})=\Delta_X p_{\psi,t}. Consequently, substituting qψ,t=pψ,tq_{\psi,t}=p_{\psi,t} makes the score correction and diffusion terms cancel: −12gt2∇X⋅(Sψpψ,t)+12gt2ΔXpψ,t=0-\tfrac{1}{2}g_t^2\nabla_X\cdot(S_\psi p_{\psi,t})+\tfrac{1}{2}g_t^2\Delta_X p_{\psi,t}=0. The remaining equation is exactly the same as that of the original flow.

Therefore, during fine-tuning, the SDE introduces stochasticity to support exploration while preserving the original flow's marginal distributions. After fine-tuning, this additional stochasticity is no longer needed, so we retain the original ODE for generation. The ODE and SDE are defined by the same flow matching model and share the same parameters ψ\psi. PPO updates these shared parameters through the SDE trajectories, and generation with the original ODE follows the updated velocity field and discrete predictions. In the main paper, the results in Tables 1-3 are all obtained with ODE and show the learned forgetting effect. For example, Table 2 shows that UR decreases from 9.02% to 1.70% for toxicity and from 10.76% to 2.13% for poor permeability.

In addition, to directly assess consistency after fine-tuning, we compare the ODE and SDE results for the same fine-tuned model in Table R2 below. Their similar UR and validity further support consistent forgetting effectiveness and generation quality. Small differences can arise from stochastic variation across a finite number of generated molecules and numerical discretization.

Table R2. Additional SDE results on the same fine-tuned model (toxicity).

UR (%)Validity (%)
SDE1.4698.87
ODE1.4398.92
AuthorsOpenReview ↗

Response to Reviewer crVZ (Part 3/3 - W4)

W4:

We add Unlearning-Aware Minimization (UAM) (Kim et al., 2025) as an additional baseline and report its results under the same protocol, as shown in Table R3. Compared with UAM, our method achieves lower UR and Recovery@20 on both tasks while maintaining comparable generation quality. We also include references for baselines in the revised manuscript, including Retained (Bourtoule et al., 2021), GA and GA+Retain (Maini et al., 2024), REINFORCE (Williams, 1992), Global KL (Lu et al., 2022), Smoothness (Fan et al., 2025), and UAM (Kim et al., 2025).

Table R3: Comparison with UAM.

Task / MethodValidity (%) ↑Uniqueness (%) ↑Novelty (%) ↑FCD ↓UR (%) ↓Recovery@20 (%) ↓
Toxicity
UAM98.60 ± 0.20100.00 ± 0.00100.00 ± 0.005.47 ± 0.211.80 ± 0.223.30 ± 0.34
Ours98.92 ± 0.30100.00 ± 0.00100.00 ± 0.005.31 ± 0.191.43 ± 0.051.52 ± 0.15
Permeability
UAM98.50 ± 0.35100.00 ± 0.00100.00 ± 0.005.45 ± 0.222.30 ± 0.283.80 ± 0.40
Ours98.82 ± 0.3199.99 ± 0.01100.00 ± 0.005.26 ± 0.141.83 ± 0.192.49 ± 0.22

Bourtoule, L., Chandrasekaran, V., Choquette-Choo, C. A., Jia, H., Travers, A., Zhang, B., Lie, D., & Papernot, N. (2021). Machine Unlearning. IEEE Symposium on Security and Privacy (SP), 141-159.

Maini, P., Feng, Z., Schwarzschild, A., Lipton, Z. C., & Kolter, J. Z. (2024). TOFU: A Task of Fictitious Unlearning for LLMs. ICLR 2024 Workshop.

Williams, R. J. (1992). Simple Statistical Gradient-Following Algorithms for Connectionist Reinforcement Learning. Machine Learning, 8, 229-256.

Lu, X., Welleck, S., Hessel, J., Jiang, L., Qin, L., West, P., Ammanabrolu, P., & Choi, Y. (2022). Quark: Controllable Text Generation with Reinforced Unlearning. NeurIPS 35.

Fan, C., Jia, J., Zhang, Y., Ramakrishna, A., Hong, M., & Liu, S. (2025). Towards LLM Unlearning Resilient to Relearning Attacks: A Sharpness-Aware Minimization Perspective and Beyond. ICML 2025.

Kim, H., Kim, K., Chae, S., & Yoon, S. (2025). Unlearning-Aware Minimization. NeurIPS 38.

We once again thank the reviewer for the thoughtful feedback.

Reviewer_dk8XOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents ForgetU, a reinforcement learning framework for suppressing the generation of undesirable molecules by pretrained 3D molecular flow-matching models. The first stage combines reward-based fine-tuning with selective distillation to preserve desirable generation behavior. The second stage simulates relearning on undesirable molecules and trains the model to resist recovery of the suppressed behavior. The approach provides an interesting perspective on steering molecular generative models, with novelty arising primarily from its application and the integration of its components.

Critical strengths
  • The method is sensible and generally clearly described, addressing both suppression of undesirable outputs and preservation of the model’s original generation capabilities.
  • The second stage is particularly compelling. Training against simulated relearning improves resistance to subsequent fine-tuning on undesirable molecules and adds practical value beyond initial suppression.
  • The baselines and ablations are well chosen to assess the effectiveness of the approach and the contributions of its individual components.
  • Evaluation across different forgetting targets, including predicted toxicity, poor permeability, and fluorine-containing molecules, supports the framework’s applicability beyond a single criterion.
Critical weaknesses
  • Figure 2 does not clearly represent the algorithm. It should distinguish frozen and trainable model copies and clarify how the losses contribute to parameter updates. The arrow from (LforgetL_{\mathrm{forget}}) to (LdistillL_{\mathrm{distill}}) is particularly unclear. Additionally, the “Undesired” label in the distillation branch appears inconsistent with the use of retained trajectories.

  • Although the paper includes held-out and external-data relearning evaluations, generalization across chemotypes remains insufficiently characterized. Scaffold-disjoint or similarity-controlled splits would provide stronger evidence that resistance to relearning extends beyond structurally related molecules.

  • Minor typesetting issues, including awkward citation formatting and overlapping equations and algorithm text, should be corrected.

Questions for authors

How structurally similar are the molecules used for simulated relearning to those used for held-out and external-data evaluation? Does the observed robustness persist when these sets are separated by scaffold?

Could Figure 2 explicitly identify which model copies are frozen or updated, show how the losses are combined, and clarify the selection of retained trajectories for distillation?

Can the method be easily extended to non-binary rewards?

3 comments
AuthorsOpenReview ↗

Response to Reviewer dk8X

We sincerely thank the reviewer for recognizing the novelty and practical relevance of our framework, the sensible formulation and effectiveness of our approach, the careful evaluation, and its broader applicability and practical value.

W1Q2W3: We have revised Figure 2 in the manuscript to distinguish frozen and trainable model copies, show how the losses contribute to the model update, correct the “Undesired” label, and clarify the selection of retained trajectories for selective distillation. We have also carefully corrected the minor typesetting issues throughout the manuscript.

W2Q1: We clarify that the simulated-relearning and recovery evaluation sets are already scaffold-disjoint. We have also clarified this in the revised manuscript. The molecules were split based on Bemis-Murcko scaffolds, with no scaffold overlap between DUtrainD_U^{train} and either the held-out GEOM-Drugs or external ZINC15 evaluation set. We further quantify their structural similarity using nearest-neighbor molecular similarity (lower is better). The average similarities to DUtrainD_U^{train} are 0.51 for held-out GEOM-Drugs and 0.52 for ZINC15. For reference, the public MOSES scaffold-split benchmark reports a train-to-scaffold-test nearest-neighbor similarity of approximately 0.586, with zero scaffold overlap by construction.

Q3: The framework can be directly extended to non-binary rewards because the PPO objective only requires a scalar terminal reward. For example, the binary R(GT)=−U(GT)R(G_T) = -U(G_T) can be replaced by a continuous property score R(GT)=−s(GT)R(G_T) = -s(G_T). We use a binary reward in this study to make the forgetting target explicit and avoid introducing additional preferences among retained molecules.

We once again sincerely thank the reviewer for the thoughtful, constructive, and encouraging feedback.

I appreciate the clarification that these splits are scaffold disjoint, although this seems at odds with the appendix, which describes random splitting:

Splitting. We randomly shuffle the datasets at the molecule level and divide them into training, validation, and test sets using an 80%, 10%, and 10% split, respectively. For GEOM-Drugs, all conformations corresponding to the same molecule are assigned to the same subset to prevent information leakage across the training, validation, and test sets.

Can you explain this incongruity?

AuthorsOpenReview ↗

Clarification on Data Splitting Description in Appendix

Dear Reviewer dk8X,

Thank you very much for pointing this out. The random split described in Appendix C.2 and the scaffold split used for simulated relearning and robustness evaluation correspond to two different stages of our experiments.

First, the description in Appendix C.2 that the dataset is randomly divided into training, validation, and test subsets is for flow-matching pretraining, which is consistent with the data-splitting protocol used in our backbone flow-matching model [1].

Then, in Stage 2, when constructing the molecule sets for simulated relearning and robustness evaluation, we further apply a scaffold-based split to the undesired molecules from the GEOM-Drugs pretraining training set to ensure scaffold separation between the simulated-relearning and held-out GEOM-Drugs recovery-evaluation sets. The external ZINC15 evaluation set is independently constructed and also has no scaffold overlap with the simulated-relearning set.

To make this distinction clear, we will revise the wording in Appendix C.2 as follows:

“Splitting. For flow-matching pretraining, we randomly shuffle the datasets at the molecule level and divide them into training, validation, and test sets using an 80%, 10%, and 10% split, respectively. The conformations in GEOM-Drugs corresponding to the same molecule are assigned to the same subset to prevent information leakage across the training, validation, and test sets. For simulated relearning and robustness evaluation in Stage 2, we further apply Bemis-Murcko scaffold splitting to the undesired molecules from the GEOM-Drugs pretraining training set, ensuring that the simulated-relearning and held-out GEOM-Drugs recovery-evaluation sets are scaffold-disjoint. The external ZINC15 evaluation set is independently constructed and also has no scaffold overlap with the simulated-relearning set.”

We once again sincerely thank the reviewer for bringing this to our attention.

[1] Dunn, I., & Koes, D. R. (2026). FlowMol3: Flow matching for 3D de novo small-molecule generation. Digital Discovery, 5(5), 2052-2066.

Reviewer_dq56OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ForgetU, which aims to reduce the undesired outputs of 3D molecular generative models while preserving their normal generation capability. The authors apply reinforcement learning driven by molecular-property feedback, and distill the original model along normal generation trajectories. To prevent undesired generation from re-emerging under subsequent fine-tuning, they further simulate fine-tuning the model on undesired molecules and adjust the original model according to the post-fine-tuning generation behavior. Experiments on GEOM-Drugs and QM9 demonstrate the forgetting effect, and the ablations support the roles of the capability-preservation and anti-recovery designs. Overall, this is a well-motivated and fairly complete methodological study on safe post-training for molecular generation.

Critical strengths
  1. Treating whether the forgetting effect can withstand subsequent training as an optimization objective is practically meaningful. In the anti-recovery ablation, the two models have similar initial undesired-generation rates, but after fine-tuning their undesired-generation rates are 1.52% and 5.48%. This indicates that the design improves how the model responds to subsequent training.
  2. Each design choice targets a clearly defined problem and is supported by a dedicated ablation. Selective distillation mainly preserves generation quality, simulated relearning mainly suppresses recovery, and the joint optimization of the continuous and discrete components improves the forgetting effect. The components form a coherent, mutually reinforcing scheme.
Critical weaknesses
  1. The conceptual increment of the method is limited. Meta-Unlearning has already adopted the idea of simulating fine-tuning and then re-optimizing against the fine-tuned model. This paper applies that idea to molecular-property rewards, and its added contribution mainly lies in the adaptation to molecular generation.
  2. The scope of the core anti-recovery result still needs to be stated more clearly. The configuration of the recovery evaluation is incomplete, and the results across different data sources are not reported separately, making it hard for the reader to judge whether the benefit extends beyond the update setting simulated during training.
Questions for authors

What data scale, learning rate, and optimizer are used in the recovery evaluation, and how do they differ from the fine-tuning setup simulated during training? What are the individual results for the two data sources mentioned in Figure D.1?

3 comments
AuthorsOpenReview ↗

Response to Reviewer dq56

We sincerely thank the reviewer for recognizing the practical significance, problem formulation, methodological design, and empirical evidence of our work.

Below we address the reviewer’s questions about the difference between our method and meta-unlearning (W1) and the experimental configurations & results (W2Q1).

W1:

Robust forgetting (fine-tuning simulation and then re-optimization) is one component of our overall framework. We first clarify the contribution of our study. In this work, we develop an RL-based forgetting framework specifically for continuous-discrete 3D molecular flow matching, including the construction of transition likelihoods for continuous coordinates and discrete molecular states, trajectory-based optimization, and selective distillation for preserving retained generation. Therefore, our contribution is not the general idea of simulating relearning itself, but its distinct RL-based formulation and integration into a unified framework for continuous-discrete 3D molecular generation. To the best of our knowledge at the time of submission, there are no prior explorations of such a framework for molecular flow matching models.

We clarify the difference between our robust forgetting and meta-unlearning. Meta-Unlearning performs meta-gradient optimization through the simulated fine-tuning process and constructs its meta objective using diffusion losses on the forget and retain sets. In contrast, our method uses simulated relearning to obtain a parameter displacement ΔU\Delta_U, stops gradients through this displacement, and optimizes the model based on its generation behavior after applying ΔU\Delta_U. Specifically, we generate post-relearning molecular trajectories and optimize a PPO-style robust objective using transition likelihood ratios and molecular undesirability rewards, without backpropagating through the simulated relearning process. The two methods also target different behaviors after simulated fine-tuning. Meta-Unlearning intentionally disrupts related benign knowledge under malicious fine-tuning by increasing its diffusion loss. In contrast, our method aims to make undesired generation difficult to recover while preserving normal generation behavior.

We summarize their key mathematical differences in Table R1:

Table R1: Key formulations

StepMeta-UnlearningOurs
Fine-tuning simulationθFT=θ−τ∇θLFT(θ;DFT)\theta^{FT}=\theta-\tau\nabla_\theta\mathcal{L}_{FT}(\theta;\mathcal{D}_{FT})ϕold+←ϕold+−ηU∇ϕold+LFM(ϕold+;DU)\phi_{\mathrm{old}}^{+}\leftarrow\phi_{\mathrm{old}}^{+}-\eta_U\nabla_{\phi_{\mathrm{old}}^{+}}\mathcal{L}_{FM}(\phi_{\mathrm{old}}^{+};D_U), followed by ΔU=ϕold+−ϕold\Delta_U=\phi_{\mathrm{old}}^{+}-\phi_{\mathrm{old}}
Re-optimizationθ←θ−β∇θLmeta(θFT(θ))\theta\leftarrow\theta-\beta\nabla_\theta\mathcal{L}_{meta}(\theta^{FT}(\theta)), where gradients propagate through the simulated updateϕ←ϕ−β∇ϕLrobust(ϕ+sg⁡(ΔU))\phi\leftarrow\phi-\beta\nabla_\phi\mathcal{L}_{robust}(\phi+\operatorname{sg}(\Delta_U)); gradients do not propagate through simulated relearning

W2Q1:

We list the data scale, learning rate, and optimizer used for the recovery evaluation in Table R2 and compare them with the simulated relearning setup. We use the same data scale (10,000 molecules), learning rate and optimizer to keep the settings consistent. The key differences are the data sources and fine-tuning steps. Specifically, we select two disjoint sets of 10,000 undesired molecules from the original GEOM-Drugs training set, one for simulated relearning and the other for recovery evaluation, to assess robustness on unseen molecules from the same source. To further evaluate cross-source robustness, we construct another set of 10,000 undesired molecules from the external ZINC15 dataset. We evaluate recovery at different fine-tuning steps to assess robustness under both shorter and longer fine-tuning (40 steps vs. 10 simulated steps).

Table R2: Simulated relearning and recovery evaluation settings

ParameterSimulated RelearningRecovery Evaluation
Data scale10,000 molecules10,000 molecules per source
Learning rate1 × 10⁻⁴1 × 10⁻⁴
OptimizerAdamAdam
Data sourceGEOM-DrugsGEOM-Drugs (Held-out) / ZINC15 (External)
Fine-tuning steps105, 20, 40

To clearly report the individual evaluation results for the two data sources, we put the results in Table R3. The results show that Recovery@40 remains low for both data sources (1.99% and 2.07%), indicating consistent anti-recovery robustness across data sources, even with longer fine-tuning. The table and results have been updated accordingly in Appendix C.5 and Appendix D.1.

Table R3: Individual evaluation results for the two data sources

Data sourceRecovery@5 (%) ↓Recovery@20 (%) ↓Recovery@40 (%) ↓
GEOM-Drugs (Held-out)1.461.521.99
ZINC15 (External)1.451.562.07

Dear Reviewer dq56, Thank you for your time and thoughtful review. As the discussion closes tomorrow, November 18 (AoE), we would appreciate hearing whether our rebuttal addresses your concerns. We would be happy to clarify any remaining points before the deadline.

Best, The authors

I thank the authors for their detailed rebuttal and the additional experiments conducted during the discussion period. My concerns have been adequately addressed, and I appreciate the clarifications. I am raising my rating to 3 (weak acceptance).

Reviewer_6rVgOpenReview ↗
Rating
3(+1)
Confidence
3(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes ForgetU, a post-training method that makes a pretrained 3D molecular flow-matching generator "forget" an undesired property (e.g., predicted toxicity). Stage 1 applies PPO with a binary reward from a property predictor, using a transition likelihood defined for the continuous-discrete flow, plus a distillation term that keeps the model close to the pretrained one on non-undesired samples. Stage 2 simulates relearning on undesired molecules and optimizes the forgetting loss after that update, so that forgetting resists later fine-tuning. On GEOM-Drugs, the undesired rate drops from about 9-11% to under 2% with generation quality largely preserved, and recovery after fine-tuning stays low.

Critical strengths
  1. Clear effects. The undesired rate is reduced several-fold without visible damage to validity or novelty, and removing Stage 2 clearly increases recovery after fine-tuning (about 5.5% vs 1.5% for toxicity).

  2. A careful construction and sensible evaluation design. The mixed continuous-discrete likelihood is derived with appropriate caveats, and the evaluation separates forgetting, preservation and robustness with a held-out toxic split.

Critical weaknesses

1. The practical benefit over generate-then-filter is unclear. Since only about 9-11% of the generated molecules are undesired, filtering with the same predictor would discard a small fraction. The paper does not compare the quality or downstream utility of this accepted set with ForgetU samples. The QM9 fluorine task is also defined by a simple filtering rule, so it adds little evidence of a practical benefit from modifying the model.

2. The two stages make different contributions, and the robustness claim needs a broader test. Stage 1 uses a standard PPO objective but makes a specific technical adaptation by constructing a mixed continuous-discrete transition likelihood; the role of this construction in the reported gains is not isolated. Related work not cited includes AliDiff by Gu et al., which also handles protein-target-conditioned generation, and RLPF by Zhou et al. Stage 2 is the forgetting-specific contribution, and its ablation supports robustness to the tested procedure. However, that procedure is short fine-tuning on toxic molecules, close to the simulated relearning used in training; the key question is whether the gain transfers to different relearning procedures. Also, Stage 2 requires an explicit set of undesired molecules DUD_U, so the scope of the claim that no predefined forget set is needed should be clarified.

Questions for authors
  1. Given the pretrained UR of 8.74% and 10.56%, how do the validity, FCD and downstream drug-design properties of ForgetU samples compare with pretrained samples filtered by the same UU? This comparison would show whether model editing improves the usable molecule set.
  2. Which reported gains depend on the mixed continuous-discrete likelihood construction in Section 3.2, beyond applying PPO with the binary property reward?
  3. Does Stage 2's Recovery@K advantage over the ablation without robust forgetting and over Smoothness persist when relearning differs from the training simulation, for example with longer fine-tuning, a different optimizer, or benign target-specific data?
  4. What are the size and source of DUD_U and the optimizer and learning rate used for recovery evaluation?
5 comments
AuthorsOpenReview ↗

Response to Reviewer 6rVg (Part 1/2 - W1Q1)

We thank the reviewer for recognizing the effectiveness of our method, the technical construction, and our evaluation design.

Below we address the reviewer’s questions about the benefits of our method compared to generate-then-filter (W1Q1), the contributions of the mixed continuous-discrete transition likelihood, robustness under different relearning procedures, and recovery evaluation settings (W2Q2Q3Q4).

W1Q1:

Compared with generate-then-filter, our method benefits from three aspects: a larger number of usable candidates under the same generation budget, better preservation of useful chemical space, and intrinsic protection upon model release.

Our method provides more usable candidates than generate-then-filter under the same generation budget. Table R1 compares the two approaches when generating 10,000 molecules. While their molecular properties remain comparable, generate-then-filter removes a large number of valid molecules predicted as undesired during filtering, reducing the final number of usable candidates from 9,710 to 8,986. Usable candidates here mean molecules that can be used in subsequent stages. In drug discovery tasks, we usually expect to have as many usable candidates as possible at the initial stage, since downstream screening and validation have very high attrition rates.

Table R1: Generate-then-filter vs. Ours

MethodValidity (%) ↑Uniqueness (%) ↑Novelty (%) ↑# Usable Candidates ↑QED ↑SAS ↓Binding Affinity ↓FCD ↓
Generate-then-filter89.9899.87100.008,9860.761.97-8.45.38
Ours98.9099.88100.009,7100.761.98-8.45.35

Additionally, our method better preserves useful chemical space. As shown in Table R2, generate-then-filter causes hundreds of molecular scaffolds and fragments to completely disappear from the final candidate set. In contrast, several new structures appear among the safer molecules generated by our method, providing greater diversity. Preserving diverse chemical scaffolds is important, since identifying novel scaffolds is a major challenge in many molecular discovery tasks.

Table R2: Structural Diversity: Filtering vs. Ours

MetricComplete BM ScaffoldsGeneric ScaffoldsBRICS Fragments
# Structures Disappearing After Filtering624365641
# Structures Newly Appearing in Ours612353483

Our method provides model-level protection when the model itself is released, which is important for reducing risks of misuse and malicious use. Generate-then-filter is effective only when every output is forced through a controlled API or mandatory filtering pipeline, whereas research models are often released with their weights and can be used without such safeguards. In molecular generation, known toxic compounds and their key precursors may already be regulated, whereas newly generated harmful structures may not be explicitly covered by existing control lists and could still be synthesizable from legally available materials, creating potential risks of misuse or malicious use. Therefore, when model weights are released, controlling the generation of such high-risk molecules at the model level, as in our method, is important to reduce potential safety risks.

We also clarify that the QM9 fluorine experiment is intended to evaluate generalization to a different dataset and undesired target, rather than to demonstrate the practical advantage over filtering.

AuthorsOpenReview ↗

Response to Reviewer 6rVg (Part 2/2 - W2Q2Q3Q4)

W2Q2Q3Q4:

The contribution of the mixed continuous-discrete transition likelihood is examined by the “w/o Continuous” and “w/o Discrete” ablations, as shown in Table R3. Under the same setting (e.g., the same PPO framework and reward), we remove the continuous or discrete component of the transition likelihood, respectively. The results indicate that removing either component substantially weakens both forgetting and robustness. For toxicity, removing the continuous/discrete component increases UR from 1.43% to 6.77% / 4.98% and Recovery@20 from 1.52% to 7.76% / 6.01%. Similar trends are observed for permeability. These results show the importance of jointly modeling the continuous and discrete transition likelihoods, since molecular properties depend on both continuous atomic coordinates and discrete chemical identities such as atom and bond types.

Table R3: Ablation Study of Mixed Continuous-Discrete Transition Likelihoods

MethodToxic UR ↓Toxic Recovery@20 ↓Permeability UR ↓Permeability Recovery@20 ↓
w/o Continuous6.777.768.7510.23
w/o Discrete4.986.016.078.94
Ours1.431.521.832.49

We also cite AliDiff and RLPF in the revised manuscript. We integrate them into the Related Work section as follows:

“RL- or preference-based post-training approaches have also been explored in molecular generation. AliDiff uses preference optimization for property alignment in target-aware molecular diffusion, while RLPF applies PPO to improve the physical stability of molecules generated by diffusion models. However, these methods focus on optimizing desirable properties rather than forgetting undesired generation.”

We further evaluate the robustness of Stage 2 under different relearning procedures. In addition to the original recovery evaluation, we consider longer fine-tuning (40 steps instead of 20) and a different optimizer (SGD with momentum instead of Adam). As shown in Table R4, our method consistently achieves lower Recovery than both w/o Robust Forgetting and Smoothness, with 1.99% under longer fine-tuning and 1.70% with the different optimizer. These results show that the gain transfers to different relearning procedures.

Table R4: Recovery (%)↓ under different relearning procedures

Recovery Settingw/o RobustSmoothnessOurs
Original (20 steps, Adam)5.481.871.52
Longer fine-tuning (40 steps, Adam)6.592.161.99
Different optimizer (40 steps, SGD + momentum)5.831.931.70

We clarify the recovery evaluation settings in Table R5. The undesired set DUD_U used in Stage 2 contains 10,000 molecules from GEOM-Drugs, and has no overlap with the set used for evaluation.

Table R5: Recovery evaluation settings

ParameterRecovery Evaluation
Data SourceGEOM-Drugs (Held-out)
Data Size10,000 molecules
OptimizerAdam
Learning Rate1 × 10⁻⁴

Since Stage 1 does not use an explicit undesired set while Stage 2 does, we revise the statement that “no predefined forget set is needed” in the manuscript as follows:

“without relying on a predefined forget set to drive the forgetting process”

We once again thank the reviewer for the constructive feedback.

Dear Reviewer 6rVg, Thank you for your time and thoughtful review. As the discussion closes tomorrow, November 18 (AoE), we would appreciate hearing whether our rebuttal addresses your concerns. We would be happy to clarify any remaining points before the deadline.

Best, 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.