acceptodds
Under review as a conference paper at ICLR 2027

Latent Flow Matching for Molecular Graph Generation

Abstract

Modern graph generative models typically operate directly in the discrete graph space, explicitly generating node and edge variables, which can become costly as graphs grow. In this paper, we perform generation directly on latent representations of entire graphs obtained from a pretrained Variational Autoencoder with high reconstruction fidelity. The generated representations, obtained through flow matching, are then decoded only at the final step. Across molecular benchmarks of increasing size, our approach achieves strong validity and FCD while offering a favorable quality-efficiency trade-off compared with state-of-the-art explicit graph generative models. One of the main advantages of this formulation is that the graph representation only needs to be learned once, after which the same representations can be reused across multiple generative objectives without retraining. We demonstrate generation guided by molecular properties and further introduce validity-aware generation though a classifier learned directly in latent space.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Generation, Generative AI, Flow Matching, Molecular Generation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Bt1aOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a latent flow matching framework for molecular graph generation using a pretrained graph-level VAE. It further introduces classifier-based validity guidance and property-controlled generation. While the approach shows promising FCD results and reduced flow-model training costs, its methodological novelty is limited, and several experimental weaknesses substantially undermine its claims of efficiency, scalability, and practical effectiveness.

Critical strengths
  1. The proposed latent generation pipeline is straightforward.
  2. The method achieves competitive FCD scores, particularly on PubChem16 and PubChem32.
Critical weaknesses
  1. Limited novelty. The framework largely combines an existing pretrained VAE, standard flow matching, and conventional classifier guidance. The contributions appear primarily incremental rather than conceptually novel.
  2. Unconvincing scalability claims. On PubChem32, the proposed model achieves only 74.5% validity, substantially below DiGress (99.0%). This directly challenges its claimed effectiveness on larger molecular graphs.
  3. The reported training costs exclude VAE pretraining, potentially overstating computational advantages. Furthermore, the method is slower than competitive explicit graph generators at comparable low NFE on smaller datasets.
  4. Missing important baselines and datasets. Comparisons are largely restricted to DiGress and Categorical Flow Maps. The paper doesn't include and compare important latent graph generative approaches, such as LGD, making the claimed advantages insufficiently established.
Questions for authors

Please see the Weaknesses.

3 comments
AuthorsOpenReview ↗

Rebuttal: Reviewer Bt1a

We thank the reviewer for their very relevant comments. We acknowledge that our rebuttal is coming at the end of the discussion period. We want to precise that we appreciate the reviewer's work and that this delay is in no way meant to minimize that appreciation.

W1

Limited novelty. The framework largely combines an existing pretrained VAE, standard flow matching, and conventional classifier guidance. The contributions appear primarily incremental rather than conceptually novel.

We thank the reviewer for this comment. We agree that the components of our methods existed beforehand, and do not claim novelty regarding the architecture or training scheme of our model. What our method does is address the gap in the litterature that can be found between:

  • one-shot molecule generation (for instance through VAEs), which is known to be efficient but suffers from a lack of validity without any molecule-specific correction or scheme [1,2,3] due to the fact that valid molecules are sparse in the space of (possibly invalid) molecular graphs
  • graph-space multi-step generation, which suffers from low throughput due to working in the complex graph-space.

Our method proposes to address this by replacing the (gaussian) prior of the VAE by the distribution learned by a flow matching model in the latent space, where sampling from that distribution is made efficient from the fact that the space is a single vector rather than a combination of interacting ones. To the best of our knowledge, our approach is the first to address the problem of graph generation in a graph-level latent space, while other approaches remain either in the explicit graph space or in a node-level latent space. We will clarify this motivation and novelty in the revised version.

[1] Matt J. Kusner, Brooks Paige, and Jos´e Miguel Hern´andez-Lobato. Grammar variational autoencoder, 2017

[2] Wengong Jin, Regina Barzilay, and Tommi Jaakkola. Junction tree variational autoencoder for molecular graph generation, 2018

[3] Trieu Nguyen and Aleksandra Karolak. Transformer graph variational autoencoder for generative molecular design, 2025

W2

Unconvincing scalability claims. On PubChem32, the proposed model achieves only 74.5% validity, substantially below DiGress (99.0%). This directly challenges its claimed effectiveness on larger molecular graphs.

First of all, we retrained our PubChem32 model without batch-OT, and noticed that the results were improved. Our main explanation for that phenomenon is that euclidean-distance batch-OT is challenging in dimension 512 and leads to no improvement. We also replaced the original Heun solver with Euler, leading to quicker generation. We report below the new results:

PubChem32

ModelSamplerStepsValidityUniquenessNoveltyFCDTimeVUN/T
DiGressDiffusion2089.2 ± 0.7100.0 ± 0.093.0 ± 0.83.247.922 ± 0.133107.23
DiGressDiffusion50099.0 ± 0.3100.0 ± 0.095.1 ± 0.51.17193.77 ± 0.0544.9754
VAE SamplingN/AN/A11.1 ± 0.1100.0 ± 0.092.61 ± 0.0211.33N/AN/A
Latent FMEuler279.21 ± 1.193.27 ± 0.935.6 ± 2.212.264.262 ± 0.04240.3
Latent FMEuler2078.7 ± 0.6100.0 ± 0.090.4 ± 1.40.5704.175 ± 0.036245.3
Latent FMEuler50078.44 ± 1.4100.0 ± 0.090.32 ± 1.450.5339.188 ± 0.142111.5

First, this confirms the observation that out approach does not significantly benefit from large numbers of steps, since its 500-steps performance is not significantly better than that of the 20-steps method. We therefore consider our 20-steps performance as our standard in what follows.

It is true that the validity of our approach remains substantially below DiGress on PubChem32. Nonetheless, this result should be considered contextualized by the other metrics:

  • the distributional fidelity of our approach is significantly better, with an FCD of 0.570 compared to 1.17 for DiGress (500 steps regime) or 3.24 (20 steps regime), meaning our approach reflects the distribution of the training data in a more reliable way.
  • low validity means some samples must be rejected: at 78.7% validity we need on average 1.27 samples to obtain 1 valid molecule, where DiGress with 500 steps only needs roughly 1 (with 99% accuracy). On the other hand, DiGress with 500 steps takes 194 seconds per batch, while ours (20 steps) takes 4.175, more than 45 times faster. Even with 20 steps, DiGress is still twice slower than our approach, which is reflected by a VUN/T of 107 compared to our 245; in this regime, DiGress' FCD is significantly worse that ours.

Our claimed scalability therefore stems from that tradeoff, where we reach better distributional fidelity on the larger dataset at a fraction of the computational cost.

We will update those results in the revised paper.

AuthorsOpenReview ↗

Rebuttal: Reviewer Bt1a

W3

The reported training costs exclude VAE pretraining, potentially overstating computational advantages. Furthermore, the method is slower than competitive explicit graph generators at comparable low NFE on smaller datasets.

We thank the reviewer for that comment. Regarding the first point, we situate ourselves in applications where a pretrained representation model is readily available. In turn, our training cost therefore focuses any new generation task that could arise, meaning that training any number of generative models upon that latent, guided by any number of properties, would be cheap compare to the need of retraining from scratch a complex explicit generator since the pretrained model is reused as-is (in the same philosophy as BERT-type models for text, whose latent can be used with simple fine tuning for any number of task once trained once).

We understand that this advantage is mitigated in the case where a single generative model is required, and are currently training a VAE from scratch on PubChem16 for 17 hours, which is roughly the training time for DiGress. We will report the results in the revised version.

Considering the second part of the question, it is true that our method suffers from the flat cost of the single-pass through the decoder. This cost is paid once after the last step, and is therefore a large part of the total time in few-steps regimes. This is especially true on QM9, which consists in very small molecules (9 atoms or less), making the explicit space small and cheap to operate in, mitigating the advantage of working in the latent space.

We nonetheless invite the reviewer to observe the phenomenon in larger graphs. We already discussed PubChem32 in the answer to the previous question. On PubChem16, CFM-CSD and CFM-ECLD reach a higher VUN/T than Latent FM up to 20 steps (about 249 vs. 190 at 20 steps), but with an FCD 6 to 11 times worse (1.155 and 2.083 vs. 0.190). This gap in VUN/T closes as the number of steps grows, since our generation time only goes from 3.5 to 4.3 between 2 and 20 steps, which confirms that it is dominated by the flat decoding cost. At 500 steps, Latent FM reaches the best FCD of all methods (0.171) and a VUN/T of 78.3, about 7 times higher than CFM (10.7) and DiGress (11.8).

W4

Missing important baselines and datasets. Comparisons are largely restricted to DiGress and Categorical Flow Maps. The paper doesn't include and compare important latent graph generative approaches, such as LGD, making the claimed advantages insufficiently established.

We thank the reviewer for that remark. It is true that we have focused our comparison to the state of the art explicit methods for graphs, notably one diffusion approach and one flow-based one, as these are the most widely used.

The reason for this is that the other latent graph generative methods evoked are, as far as we observed in the literature, limited to node-level latent representation, which did not provide direct comparison to our approach, while explicit graph generation is the current state of the art in generation and seemed more relevant to compare against.

We will make this clear in the revised version.

After reading your rebuttal, I still have significant concerns about the proposed method, which appears to offer little methodological novelty. Therefore, I have decided to maintain my original score.

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

The paper proposes applying flow matching to whole-graph latent representations, enabling property-guided and validity-aware generation while improving training efficiency.

Critical strengths
  1. The paper applies flow matching to a whole-graph latent representation, enabling more efficient training compared with flow matching directly over graph representations.
  2. Given a pre-trained generative model, flow matching can be applied to its latent space without retraining the underlying generative model. This provides flexibility and can reduce the overall training cost.
  3. The proposed framework supports both validity and property guidance directly in the latent space, enabling controllable molecular graph generation.
Critical weaknesses
  1. The unique contribution of this paper is not sufficiently clear. Prior work has explored flow matching in the latent spaces of graph generative models, including applications to molecular generation and property-guided generation. The main distinction of this work appears to be the use of a recently proposed autoencoder that represents an entire graph in a fixed-dimensional latent space, followed by flow matching in this space. While the paper reports improved training efficiency compared with another flow-matching approach, it does not clearly explain why the proposed representation leads to this efficiency gain. Moreover, the paper states that “autoencoder architectures are known to yield poor generative performance,” but it remains unclear what properties of the referenced GraViti model overcome this limitation and make its latent space particularly suitable for flow matching.
  2. The experimental evaluation is not sufficiently comprehensive. There are several existing flow-matching approaches based on different graph representations that could serve as relevant baselines. Including comparisons with a broader set of flow-matching methods would better demonstrate the advantages of performing flow matching on the proposed whole-graph latent representation.
Questions for authors

What is the rationale for applying property guidance through intermediate latent states, while handling validity through sampling? Since validity could also be formulated as a binary property, could the authors clarify the fundamental difference between validity guidance and property-constrained generation, and why they require different mechanisms?

1 comment
AuthorsOpenReview ↗

Rebuttal: Reviewer xczM

We thank the reviewer for their very relevant comments. We acknowledge that our rebuttal is coming at the end of the discussion period. We want to precise that we appreciate the reviewer's work and that this delay is in no way meant to minimize that appreciation.

W1

The unique contribution of this paper is not sufficiently clear...

We thank the reviewer for this comment. We agree that the components of our methods existed beforehand, and do not claim novelty regarding the architecture or training scheme of our model. What our method does is address the gap in the litterature that can be found between:

  • one-shot molecule generation (for instance through VAEs), which is known to be efficient but suffers from a lack of validity without any molecule-specific correction or scheme [1,2,3] due to the fact that valid molecules are sparse in the space of (possibly invalid) molecular graphs
  • graph-space multi-step generation, which suffers from low throughput due to working in the complex explicit graph-space.

Our method proposes to address this by replacing the (gaussian) prior of the VAE by the distribution learned by a flow matching model in the latent space, where sampling from that distribution is made efficient from the fact that the space is a single vector rather than a combination of interacting ones. To the best of our knowledge, our approach is the first to address the problem of graph generation in a graph-level latent space, while other approaches remain either in the explicit graph space or in a node-level latent space. We will clarify this motivation and novelty in the revised version.

the paper states that “autoencoder architectures are known to yield poor generative performance,” but it remains unclear what properties of the referenced GraViti model overcome this limitation and make its latent space particularly suitable for flow matching.

It is specifically because GraViti does not overcome that limitation that our approach is justified. While latent sampling from the prior of the VAE yields poor results, we propose to sample from a distribution yielded by a flow matching process, therefore leading to significant improvements.

[1] Matt J. Kusner, Brooks Paige, and Jos´e Miguel Hern´andez-Lobato. Grammar variational autoencoder, 2017

[2] Wengong Jin, Regina Barzilay, and Tommi Jaakkola. Junction tree variational autoencoder for molecular graph generation, 2018

[3] Trieu Nguyen and Aleksandra Karolak. Transformer graph variational autoencoder for generative molecular design, 2025

W2

The experimental evaluation is not sufficiently comprehensive. There are several existing flow-matching approaches based on different graph representations that could serve as relevant baselines. Including comparisons with a broader set of flow-matching methods would better demonstrate the advantages of performing flow matching on the proposed whole-graph latent representation.

We thank the reviewer for that remark. It is true that we have focused our comparison to the state of the art explicit methods for graphs, notably one diffusion approach and one flow-based one, as these are the most widely used.

The reason for this is that the other latent graph generative methods evoked are, as far as we observed in the literature, limited to node-level latent representation, which did not provide direct comparison to our approach, while explicit graph generation is the current state of the art in generation and seemed more relevant to compare against.

We will make this clear in the revised version.

W3

What is the rationale for applying property guidance through intermediate latent states, while handling validity through sampling? Since validity could also be formulated as a binary property, could the authors clarify the fundamental difference between validity guidance and property-constrained generation, and why they require different mechanisms?

We thank the reviewer for this question. In fact, validity guidance and property-constrained generation follow the same underlying principle: both guide the latent flow toward a desired objective. The main difference lies in the nature of the target.

For validity guidance, the target is binary (valid or invalid), so the guidance is achieved through the gradient of the cross-entropy between the true validity (binary property) and the output of a classifier. In contrast, molecular properties are continuous, motivating regressor-based guidance toward a specified target value.

Consequently, the evaluation metrics differ. For validity guidance, we primarily measure the proportion of valid molecules, whereas for property-constrained generation, we evaluate how closely the generated molecules match the requested property values. We also report validity, uniqueness, and novelty to assess how property constraints affect overall generation quality.

Reviewer_4dGoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a method for molecule graph generation that is based on latent flow matching. An autoencoder is trained to represent molecule graphs as latent vectors. A flow matching model is then trained on this latent space. To promote molecule validity, the paper also adds either validity guidance (akin to classifier guidance) or validity-aware endpoint generation by use of variational flow matching. Furthermore, to control the properties of the generated molecules it uses the target property as conditioning signal to the flow matching model and during inference it adjusts the sampling process by taking the negative gradient of a prediction loss wrt the target property.

Critical strengths

Positive points:

  • The idea of turning graphs into latent vectors and then using a latent flow matching model together with the various components leads to a reasonable and sound approach for molecule generation.
  • The proposed approach appears to be quite efficient in terms of training time.
  • Strong performance with respect to FCD metrics
  • Good property-control analysis
Critical weaknesses

Negative points:

  • Novelty is limited. The paper essentially combines existing components (e.g. classifier guidance, variational flow matching, and gradient-based adaptation of flow-matching sampling) and applies them to the problem of molecule graph generation.
  • For large and challenging problems (PubChem32) validity metric is much lower compared to DiGress (78.6% vs 99%). Even worse more iterations does not seem to offer any improvement in validity. On the contrary, it actually worsens the score (78.6 -> 74.4)
  • Furthermore, regarding validity guidance and PubChem32, the paper mentions “on PubChem32, validity guidance fails at improving the validity on that dataset”. No numbers are reported with validity guidance.
  • The paper cites and discusses several latent generation approaches for graphs (in related work) but compares to none of them.
Questions for authors

Guidance scale values reported in tables 9 and 10 appear to have unusually large variance. Is there any explanation for this ?

2 comments
AuthorsOpenReview ↗

Rebuttal: Reviewer 4dGo

We thank the reviewer for their very relevant comments. We acknowledge that our rebuttal is coming at the end of the discussion period. We want to precise that we appreciate the reviewer's work and that this delay is in no way meant to minimize that appreciation.

W1

Novelty is limited. The paper essentially combines existing components (e.g. classifier guidance, variational flow matching, and gradient-based adaptation of flow-matching sampling) and applies them to the problem of molecule graph generation.

We thank the reviewer for this comment. We agree that the components of our methods existed beforehand, and do not claim novelty regarding the architecture or training scheme of our model. What our method does is address the gap in the litterature that can be found between:

  • one-shot molecule generation (for instance through VAEs), which is known to be efficient but suffers from a lack of validity without any molecule-specific correction or scheme [1,2,3] due to the fact that valid molecules are sparse in the space of (possibly invalid) molecular graphs
  • graph-space multi-step generation, which suffers from low throughput due to working in the complex graph-space.

Our method proposes to address this by replacing the (gaussian) prior of the VAE by the distribution learned by a flow matching model in the latent space, where sampling from that distribution is made efficient from the fact that the space is a single vector rather than a combination of interacting ones. To the best of our knowledge, our approach is the first to address the problem of graph generation in a graph-level latent space, while other approaches remain either in the explicit graph space or in a node-level latent space. We will clarify this motivation and novelty in the revised version.

[1] Matt J. Kusner, Brooks Paige, and Jos´e Miguel Hern´andez-Lobato. Grammar variational autoencoder, 2017

[2] Wengong Jin, Regina Barzilay, and Tommi Jaakkola. Junction tree variational autoencoder for molecular graph generation, 2018

[3] Trieu Nguyen and Aleksandra Karolak. Transformer graph variational autoencoder for generative molecular design, 2025

W2

For large and challenging problems (PubChem32) validity metric is much lower compared to DiGress (78.6% vs 99%). Even worse more iterations does not seem to offer any improvement in validity. On the contrary, it actually worsens the score (78.6 -> 74.4)

First of all, we retrained our PubChem32 model without batch-OT, and noticed that the results were improved. Our main explanation for that phenomenon is that euclidean-distance batch-OT is challenging in dimension 512 and leads to no improvement. We also replaced the original Heun solver with Euler, leading to quicker generation. We report below the new results:

PubChem32

ModelSamplerStepsValidityUniquenessNoveltyFCDTimeVUN/T
DiGressDiffusion2089.2 ± 0.7100.0 ± 0.093.0 ± 0.83.247.922 ± 0.133107.23
DiGressDiffusion50099.0 ± 0.3100.0 ± 0.095.1 ± 0.51.17193.77 ± 0.0544.9754
VAE SamplingN/AN/A11.1 ± 0.1100.0 ± 0.092.61 ± 0.0211.33N/AN/A
Latent FMEuler279.21 ± 1.193.27 ± 0.935.6 ± 2.212.264.262 ± 0.04240.3
Latent FMEuler2078.7 ± 0.6100.0 ± 0.090.4 ± 1.40.5704.175 ± 0.036245.3
Latent FMEuler50078.44 ± 1.4100.0 ± 0.090.32 ± 1.450.5339.188 ± 0.142111.5

First, this confirms the observation that out approach does not significantly benefit from large numbers of steps, since its 500-steps performance is not significantly better than that of the 20-steps method. This is consistent with the fact that the underlying latent space is well-formed, and easier to work with than explicit spaces. We therefore consider our 20-steps performance as our standard in what follows.

It is true that the validity of our approach remains substantially below DiGress on PubChem32. Nonetheless, this result should be considered contextualized by the other metrics:

  • the distributional fidelity of our approach is significantly better, with an FCD of 0.570 compared to 1.17 for DiGress (500 steps regime) or 3.24 (20 steps regime), meaning our approach reflects the distribution of the training data in a more reliable way.
  • low validity means some samples must be rejected: at 78.7% validity we need on average 1.27 samples to obtain 1 valid molecule, where DiGress with 500 steps only needs roughly 1 (with 99% accuracy). On the other hand, DiGress with 500 steps takes 194 seconds per batch, while ours (20 steps) takes 4.175, more than 45 times faster. Even with 20 steps, DiGress is still twice slower than our approach, which is reflected by a VUN/T of 107 compared to our 245; in this regime, DiGress' FCD is significantly worse that ours.

We will update those results in the revised paper.

AuthorsOpenReview ↗

Rebuttal: Reviewer 4dGo

W3

Furthermore, regarding validity guidance and PubChem32, the paper mentions “on PubChem32, validity guidance fails at improving the validity on that dataset”. No numbers are reported with validity guidance.

We agree with the reviewer that those numbers are not included, and will fix that in the revised version. Overall, validity remained similar as without validity guidance while FCD was degraded.

W4

The paper cites and discusses several latent generation approaches for graphs (in related work) but compares to none of them.

We thank the reviewer for that remark. It is true that we have focused our comparison to the state of the art explicit methods for graphs, notably one diffusion approach and one flow-based one, as these are the most widely used.

The reason for this is that the other latent graph generative methods evoked are, as far as we observed in the literature, limited to node-level latent representation, which did not provide direct comparison to our approach, while explicit graph generation is the current state of the art in generation and seemed more relevant to compare against.

We will make this clear in the revised version.

Q1

Guidance scale values reported in tables 9 and 10 appear to have unusually large variance. Is there any explanation for this ?

We thank the reviewer for this question. The reported guidance scales were selected through a grid search over 50 evenly spaced values within a predefined range. The value yielding the best performance was then retained. We applied the same tuning procedure to DiGress to ensure a fair comparison.

For instance, for MW on QM9, we evaluated 50 guidance scales between 0 and 30,000 including both bounds, corresponding to a step size of approximately 612.24. The optimal value was the 7th point on this grid, yielding a guidance scale of 3673.47. Therefore, the seemingly unusual values simply result from the grid search procedure rather than any instability in the method.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.