acceptodds
Under review as a conference paper at ICLR 2027

GX-TGVAE: Constraint-Masked Graph Generation for Controllable Molecular Exploration

Abstract

Distributional fidelity in molecular generation alone establishes neither molecule-specific information retention nor controllable out-of-distribution (OOD) exploration. GX-TGVAE connects these capabilities through continuous latent conditioning and constraint-masked atom- and bond-level graph construction. It learns deterministic canonical reconstruction and steers exploration using locally estimated tangent/normal geometry and property gradients. On MOSES, it achieves the highest exact reconstruction among the compared latent baselines and broader structural reach, although distributional fidelity and reconstruction rankings differ across architecture variants. On ZINC, property-guided normal exploration exceeds a random-normal control in Novel Top 5% docking score at the two deeper tested exploration depths. Against the OOD generation baseline MOOD, GX-TGVAE shows high conditional utility within training-similarity bands, but its base sampler lacks sampling mass for fixed-attempt discovery and deep-OOD reach. This distinguishes exploration breadth from reached-region utility. Harvest reallocates sampling budget to selected neighborhoods, improving this deficit and exceeding MOOD's hit yield under a matched evaluation-time budget, while MOOD retains higher upper-tail docking scores. Together, these results identify faithful representation, directional control, and sampling-mass allocation as distinct but complementary design axes for molecular OOD generation.

Primary area:
generative models
Keywords:
molecular generation, graph generative models, variational autoencoders, latent representation, out-of-distribution exploration

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_F9SeOpenReview ↗
Rating
2
Confidence
2(−1)
Informativeness
2
Presentation
2
Soundness
2(−1)
Read review
Paper summary

This paper proposes GX-TGVAE, a latent-space molecular generation framework that aims to faithfully represent existing molecules, enable controllable exploration directions, and ultimately improve the discovery of high-value molecules. Specifically, the framework combines constraint-masked graph generation with Geometry-Guided Exploration and an ES + Massive Harvest extension. The paper demonstrates strong reconstruction and structural reach performance on MOSES. It also provides extensive mechanistic and empirical analyses, showing the effectiveness of property-guided exploration and identifying a sampling-mass bottleneck in OOD molecular discovery. The proposed ES + Massive Harvest extension further improves novel-hit yield under the specified evaluation-time budget. However, although the experiments support the practical usefulness of the proposed components, they do not sufficiently establish the distinct methodological contributions or underlying mechanisms of each component. Consequently, despite the promising results, the framework appears to rely heavily on combining multiple existing techniques, without sufficiently clear evidence isolating the contributions of its key innovations.

Critical strengths

The paper provides an interesting perspective on molecular OOD exploration by distinguishing representation fidelity, directional control, and sampling-mass allocation as complementary factors. In particular, the identification of the sampling-mass bottleneck offers a useful insight into the limitations of the original exploration strategy. The experimental evaluation is extensive, covering reconstruction, structural reach, directional controls, and OOD discovery. The authors also provide meaningful mechanistic and empirical analyses, with promising results on several metrics, while carefully acknowledging the limitations of their methods and findings.

Critical weaknesses
  1. Unclear narrative and insufficient methodological positioning.

Although the paper provides detailed analyses, the Abstract and Introduction do not clearly connect its main contributions. The Related Work section also insufficiently discusses existing latent-space molecular optimization methods, making it difficult to assess the novelty of the proposed geometric exploration.

  1. Limited novelty and insufficient isolation of key contributions.

The framework combines several established techniques, and the existing ablations do not fully establish their independent contributions. In particular, the geometric experiments do not sufficiently distinguish the benefits of tangent/normal geometry from general gradient projection effects. Moreover, while ES + Massive Harvest improves upon the original sampler, which underperforms MOOD, the effects of ES and the Harvest selector are not adequately isolated. The comparison also lacks matched end-to-end computational budgets, limiting claims of practical efficiency. For example, TanNorm and TanNormES use different neighborhood sizes (50 vs. 256), making it difficult to isolate the effect of ES itself (Appendix I.1).

Questions for authors
  1. What is the unique contribution of the proposed tangent/normal geometry to OOD molecular discovery? Although the directional controls show promising results, could the authors further clarify whether the improvements arise from meaningful local latent geometry rather than general gradient projection? Would similar gains be achieved using random or fixed subspaces?

  2. Which components are primarily responsible for the final improvements in novel-hit discovery? Could the authors provide more controlled ablations to isolate the contributions of ES, the Harvest selector, and sampling-budget reallocation? In particular, is geometry-guided exploration still necessary when ES and Massive Harvest are applied?

  3. What is the central methodological contribution that unifies the three proposed design axes? Could the authors clarify how these components form a coherent research contribution rather than a collection of separate improvements?

8 comments
AuthorsOpenReview ↗

Author Response to F9Se (1/5): Q3 — Generative Foundation

Response to Reviewer F9Se

We thank the reviewer for the careful review and for recognizing our analysis of OOD exploration, particularly the treatment of representation fidelity, directional control, and sampling-mass allocation as distinct elements.

We understand the central concerns to be clarifying the methodological contribution that unifies the three research axes and experimentally separating the contributions of Geometry, ES, and Harvest.

In response, we clarified the overall research logic and conducted additional direct controls involving equal-rank subspaces, matched ES, the Learned selector, and Geometry-after-Harvest.

Below, we address the unifying contribution in Q3, Geometry's specific role in Q1, and the contributions of ES, Harvest, and evaluation-budget allocation in Q2.

Q3. What is the central methodological contribution that unifies the three research axes?

The central objective unifying this work is to establish a generative foundation for directionally controlled exploration of unobserved structures using existing molecular knowledge, and scientifically evaluate its exploration capabilities and discovery efficiency.

The three research axes are organized as successive stages of design and validation toward this objective.

First, we developed GX-TGVAE as a generative foundation for exploring unknown structures.

Existing molecular VAEs face the problem that decoder generation can become unstable in latent regions far from the training distribution, producing invalid structures.

Constrained primitive graph generators such as GraphXForm can maintain chemical legality during construction. On their own, however, they do not constitute a Graph-to-Graph VAE that represents entire molecules in continuous latent coordinates and explores by manipulating those coordinates.

We connect whole-molecule continuous latent representations with chemically constrained primitive construction, separating the freedom of continuous latent exploration from the legality of discrete construction.

To do so, we integrate latent conditioning from an empty graph, information delivery to each construction stage through latent tokens and a persistent global construction token, and reconstruction learning through canonical construction traces into one coherent generative architecture.

The important point is not merely assembling existing techniques, but establishing an information interface and training configuration that connects whole-molecule latent information to constrained sequential construction, enabling molecule-specific reconstruction and latent-space manipulation.

We validate this foundation through latent-intervention controls, high reconstruction fidelity, observed Validity and Connected Validity, and reach into unseen structures.

AuthorsOpenReview ↗

Author Response to F9Se (2/5): Q3 — Unifying Contribution

Second, we used this generative foundation to analyze the capabilities and limitations of geometric OOD exploration.

We examined how directional manipulation in latent space affects the structural OOD of generated molecules and external docking utility.

We observed that local direction selection affects generated structures and external utility, and that, in deep OOD regions, the extrapolative accuracy of surrogate predictions does not necessarily align with the usefulness of the directional information supplied by their gradients.

In other words, we experimentally confirmed that useful directional information can persist even in extrapolative regions where predicted values break down.

We also analyzed reaching deep OOD regions, supplying candidates to those regions at sufficient frequency, and obtaining useful candidates there as distinct exploration capabilities.

Third, we analyzed the bottleneck limiting discovery efficiency and intervened to improve it.

In the comparison with MOOD, we decomposed Novel Hit Ratio into screening-pass yield and conditional hit efficiency.

GX-TGVAE's base sampler had relatively high hit efficiency after screening, but a low rate of supplying candidates to the screening-pass set, limiting the number of hits discovered under a finite evaluation budget.

This screening-pass sampling-mass problem is distinct from the sampling density and conditional utility within structural OOD bands analyzed earlier.

Based on this diagnosis, we introduced ES to select candidate exploration directions and Harvest to select promising neighborhoods and allocate evaluation resources, and evaluated their respective contributions through additional controls.

The three axes therefore do not form a collection of independent technical improvements. They constitute a connected study of building a generative foundation that enables exploration of unknown structures, scientifically analyzing its capabilities and limitations, and intervening on the identified bottleneck in discovery efficiency.

The central architectural contribution is the information interface connecting molecule-level continuous latent representations to constrained primitive construction. Geometry, ES, and Harvest are exploration methods for progressively examining and improving directional control and discovery efficiency on this foundation.

AuthorsOpenReview ↗

Author Response to F9Se (3/5): Q3 — Related Work

We also clarified the positioning relative to existing VAE latent optimization and geometric exploration in the revised manuscript.

Bayesian optimization and optimization using property-predictor gradients have already been studied in molecular VAE latent spaces. To address unstable generation far from the training distribution, prior work has proposed validity-constrained search, decoder-uncertainty guidance, and regularization that discourages departure from the latent distribution.

Nevertheless, an important bottleneck in conventional VAE latent optimization is that decoder generation can become unstable in latent regions far from the training distribution, making invalid molecular structures more likely.

To address this problem, prior studies have proposed constraining latent exploration to regions near the training data or where the decoder is reliable, or using decoder uncertainty to avoid uncertain regions.

Thus, there is a tension between advancing exploration and maintaining the validity of generated structures; one research direction has addressed this by restricting exploration to reliable latent regions.

To clarify this positioning, we expanded Section 2 (Related Work) of the revised manuscript, with relevant references, to discuss Bayesian optimization in VAE latent spaces, gradient-based optimization, validity-constrained search, decoder-uncertainty guidance, and regularization that discourages deviation from the latent distribution.

In contrast, we sought to address this problem through the generative architecture so that extrapolation into unknown regions could be pursued actively.

Specifically, we integrated whole-molecule continuous latent representations and chemically constrained primitive graph construction into a single Graph-to-Graph VAE.

This allows exploration to move freely beyond the training distribution in continuous latent space while applying known chemical constraints to the current partial graph at each step of discrete molecular construction.

Rather than avoiding structural breakdown by restricting the range of latent exploration, we separate continuous exploration from discrete construction rules to establish a foundation for generating legal structures even at extrapolated latent points.

An important advance is that establishing this generative foundation allows active latent extrapolation beyond the training distribution to be treated as an explicit exploration strategy for molecular generation.

We then examined Geometry-Guided Exploration, which uses the local geometry of learned molecular representations to control exploration directions rather than leaving extrapolation to random vector movements.

Specifically, we estimate local subspaces of the latent distribution and project property gradients into those subspaces to construct exploration directions toward unobserved regions using learned information.

We further evaluated the effects of geometric direction selection on structural OOD reach and external utility through controls using random directions, gradients in the full active space, and equal-rank random subspaces.

The distinction of our work is therefore not simply vector manipulation or property optimization in a VAE latent space.

It lies in addressing unstable structure generation in OOD regions—a problem in conventional latent optimization—by separating continuous latent exploration from chemically constrained discrete construction, enabling active geometric extrapolative exploration on that foundation, and systematically evaluating its usefulness.

Reviewer_tjPvOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on molecular generation and starts with an idea that connects molecule-specific continuous latent conditioning to constraint-masked primitive graph construction. The results show the advantage of the proposed method.

Critical strengths
  1. The idea that connects molecule-specific continuous latent conditioning to constraint-masked primitive graph construction is clear and reasonable.
Critical weaknesses
  1. Why is discrete graph generation with structural constraints necessary?
  2. Combining molecule-specific continuous latent conditioning and graph construction is not new. The difference from existing work should be highlighted.
  3. Figure 1 is unacceptable. I can not get the key contribution of the proposed method from this figure.
  4. Limited comparison results with existing works.
Questions for authors
  1. Why is discrete graph generation with structural constraints necessary?
  2. Combining molecule-specific continuous latent conditioning and graph construction is not new. The difference from existing work should be highlighted.
  3. Figure 1 is unacceptable. I can not get the key contribution of the proposed method from this figure.
  4. Limited comparison results with existing works.
4 comments
AuthorsOpenReview ↗

Author Response to tjPv (1/3): Q1

Response to Reviewer tjPv

We thank the reviewer for the careful review and for recognizing the idea of connecting molecule-specific continuous latent conditioning with constrained primitive graph construction.

In response, we clarified in the Introduction and Method the design requirements of GX-TGVAE, the specific methodology used to meet them, its differences from prior work, and the capabilities enabled by this integration.

Q1. Why is chemically constrained discrete graph construction needed?

Our objective is to manipulate continuous latent representations learned from known molecules and extend exploration to structures absent from the training data.

In latent regions far from the training distribution, however, the decoder may not have adequately learned the mapping required to generate appropriate structures, and discrete structure generation can become unstable.

We therefore adopted a design that separates freedom of exploration in continuous latent space from chemical legality in discrete molecular construction.

In GX-TGVAE, a whole-molecule latent variable conditions the structure to be generated, while illegal actions, such as valence violations, self-loops, and duplicate bonds, are restricted for each partial graph during construction.

Furthermore, the graph is built through atom- and bond-level primitive operations without requiring a fixed dictionary of high-level motifs or fragments as the generation units.

This design maintains chemical validity by excluding illegal construction transitions while preserving a broad primitive construction space, rather than excessively restricting the explorable structures in advance to high-level structural building blocks.

Chemically constrained discrete construction is therefore an important design requirement for enabling continuous exploration of unknown structures.

AuthorsOpenReview ↗

Author Response to tjPv (2/3): Q2Part 1/2

Q2. Differences from prior work and the central methodological contribution

We sought to meet three requirements simultaneously: (1) map an entire molecule to manipulable continuous latent coordinates, (2) generate a broad range of structures from atoms and bonds without requiring a high-level fragment or motif dictionary, and (3) apply chemical constraints during construction even under latent extrapolation.

Prior work has addressed these requirements individually or combined some of them.

CGVAE combines node-level latent representations with constrained graph construction, and MoLeR uses molecule-level continuous latent representations and hybrid motif-and-atom construction. GraphXForm implements Transformer-based constrained primitive graph generation.

Although MoLeR also has a mechanism for adding individual atoms, the use of frequent motifs as generation units is an important architectural feature. GX-TGVAE, by contrast, requires no high-level motif dictionary and implements the entire construction process through atom- and bond-level operations.

Furthermore, simply using GraphXForm's constrained primitive construction does not map whole molecules to continuous latent coordinates or enable molecular exploration through geometric manipulation of those coordinates.

To establish this connection, GX-TGVAE addresses several design challenges.

Specifically, we addressed three challenges in making this connection.

First, latent conditioning from an empty graph.

GraphXForm implements action selection based on a partial graph, but normally starts construction from a given initial molecule.

When generation is intended to faithfully reflect the latent variable, a randomly placed initial atom can interfere with that generation and introduce inconsistencies.

Building on the virtual global node mechanism inherited from GraphXForm, we therefore constructed a persistent global construction token (a0a_0) that receives latent information even when no chemical atoms exist yet, i.e., in an empty graph. We introduced AddAtomFirst and FirstAtomHead to select the first atom.

This allows us to connect whole-molecule latent information to discrete construction actions from the very first atom-selection step.

Second, use of latent information throughout construction.

Introducing a whole-molecule latent variable is not sufficient if it acts only at the start of generation: subsequent construction decisions may then fail to use molecule-specific information adequately.

We expand the molecular latent variable into multiple latent tokens and place cross-attention in every decoder layer.

In the adopted configuration, the persistent global construction token directly attends to the latent tokens and passes that information to the atom states through graph self-attention.

At each construction step, the decoder states are recomputed from the current partial graph. This repeatedly connects the changing discrete construction state with the continuous latent information representing the same molecule.

We thereby establish an information interface that continuously supplies the whole-molecule latent representation to the sequential graph-construction process.

As functional evidence, the EncoderDrop0 latent-intervention experiment achieved reconstruction Tanimoto of approximately 0.976 with the correct latent variable, whereas shuffling latent variables across molecules reduced it to approximately 0.135.

This shows that the resulting generative system actually uses latent information specific to the input molecule.

AuthorsOpenReview ↗

Author Response to tjPv (2/3): Q2Part 2/2

Third, the formulation of molecular-graph reconstruction learning.

Training an autoregressive action decoder for reconstruction with teacher forcing requires a target action for each construction state. Yet the same molecular graph corresponds to multiple construction sequences, differing, for example, in the order of atom additions or ring-closing bond additions.

GraphXForm also derives action sequences from known molecules for pretraining. To connect this approach to a VAE reconstruction objective conditioned on a molecule-level latent variable, we explicitly define the correspondence between graphs and target sequences.

Specifically, we map each molecule to one canonical construction trace and use that sequence as the teacher-forcing target.

This permits computation of the reconstruction loss from a deterministic target sequence without enumerating and marginalizing over every construction sequence that yields the same molecule.

Moreover, the probability of the canonical sequence is a lower bound on the graph likelihood, and the gap narrows as probability concentrates on that sequence. This explicitly relates the single-sequence reconstruction objective to the probability of the graph as a whole.

These three elements do not merely describe a juxtaposition of existing components.

They implement three requirements—latent conditioning from an empty graph, continued provision of latent information to a changing partial graph, and graph-level reconstruction learning—as one coherent information interface and training formulation.

This simultaneously realizes whole-molecule continuous latent representations, broad structural generation through primitives, and chemical constraints during construction within a single Graph-to-Graph VAE.

The effectiveness of this integration is supported by Correct-z versus Shuffled-z latent-intervention controls, high Exact Reconstruction across four independent training runs, observed Validity and Connected Validity of 100%, reach into unseen generic topologies and ring skeletons, and OOD exploration through latent-direction manipulation.

Our methodological contribution therefore lies not in the first introduction of individual components, but in establishing an information interface and training configuration between molecule-specific continuous latent representations and constrained primitive construction that enables both faithful reconstruction and extrapolative exploration.

AuthorsOpenReview ↗

Author Response to tjPv (3/3): Q3–Q4

Q3. Figure 1

We improved the display size and readability of Figure 1 in the revised manuscript. The figure explicitly identifies four regions—encoder, latent interface, decoder, and action selection—and includes a sentence stating its central message.

We also clarify the information pathways that are particularly important in the figure.

The graph encoder on the left produces a continuous latent representation of the entire input molecule. This representation is expanded into multiple latent tokens and supplied to the central Transformer decoder through cross-attention at each layer.

Crucially, the persistent global construction token (a0a_0) is available even before the partial graph contains any chemical atoms.

This allows molecular latent information to inform the first atom selection through FirstAtomHead, while the same Transformer backbone is used for subsequent sequential atom- and bond-level construction.

At later construction steps, the global token connects information from the current partial graph and the latent representation. After applying masks according to the current graph state, the action heads select the next construction operation.

The central contribution to focus on in Figure 1 is the information interface connecting the whole-molecule latent representation to constrained sequential construction starting from an empty graph.

Further details are provided in Section 3.1 and Appendix B.

Q4. Comparisons with existing methods

Our comparison is not simply a comparison of a few similar models on a single metric.

We compare multiple generative architectures to examine distinct capabilities relevant to our research objective, including reconstruction fidelity, chemical validity of generation, distributional fidelity, and Structural Reach.

At submission, our comparisons already included MoLeR, MAGNet-P/R, TGVAE, and GraphXForm.

MoLeR provides a comparison with hybrid motif-and-atom generation using molecule-level latent representations; MAGNet-P/R with different graph-generation architectures; TGVAE with a related graph-encoder/SMILES-decoder configuration; and GraphXForm with the constrained primitive generation mechanism from which ours is derived.

For each model, we evaluate reconstruction, basic generation performance, distributional fidelity, and Structural Reach where applicable.

In particular, because GraphXForm has no molecule-level VAE posterior, we do not assign it the same Exact Reconstruction comparison as molecule-level VAEs.

Thus, the original comparisons served a clear purpose even before adding FRATTVAE: evaluating different generation mechanisms across distinct capability axes.

Reconstruction fidelity is nevertheless one of GX-TGVAE's central claims, making a direct comparison with a fragment-based molecular VAE with particularly strong reconstruction performance important.

We therefore added the publicly available FRATTVAE model in the revision and compared reconstruction using the same 10,000 molecules from each of the official MOSES test and scaffold test sets under an identical evaluation procedure.

GX-TGVAE achieved comparably high reconstruction fidelity with approximately 29% of FRATTVAE's parameter count while constructing molecules entirely from atom- and bond-level primitives.

We also evaluated Structural Reach through unseen generic topologies and ring skeletons.

FRATTVAE performs better on FCD/Test, so we report reconstruction, distributional agreement, and Structural Reach separately.

Adding FRATTVAE is therefore not simply an increase in the number of comparison models; it strengthens the original multifaceted comparison with a direct control against a molecular VAE with strong reconstruction performance.

Our study shows that atom-level Graph-to-Graph generation without a required high-level fragment dictionary can simultaneously achieve faithful molecule-specific reconstruction, chemical validity, broad Structural Reach, and latent exploration of unknown structures within a single architecture.

The comparisons and evaluation protocols are presented in Table 1, Figure 2, and Appendices C, D, and J.

Reviewer_hNCJOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes GX-TGVAE, which combines a molecule-level continuous VAE representation with constraint-masked autoregressive graph construction. The model reconstructs canonical construction traces and supports molecular exploration via locally estimated tangent/normal spaces and property-surrogate gradients. A Harvest extension learns to select neighborhoods for additional sampling. On MOSES, GX-TGVAE reports stronger exact reconstruction and structural reach than the compared latent baselines. On ZINC, property-guided normal exploration improves docking-based utility over a random-normal control at deeper exploration settings.

Critical strengths
  1. GX-TGVAE improves exact reconstruction over the compared latent baselines. Appendix C.8 reproduces the reconstruction ranking on independent official-test inputs, strengthening this result beyond the internal validation comparison.
Critical weaknesses
  1. The major weakness is that the contribution is not yet established against the most relevant alternatives. The authors appropriately acknowledge that constrained graph generation, latent property optimization, and GraphXForm’s primitive action formulation have precedents. My concern is therefore not missing attribution, but whether the resulting combination provides a sufficiently significant advance. In particular, FRATTVAE is a directly relevant molecular VAE with strong reported reconstruction performance, yet it is absent from the reconstruction comparison. Because reconstruction fidelity is a central contribution, this is a critical weakness.
  2. It is unclear whether learning a predictor for neighborhood productivity selects better centers than a simple baseline of choosing centers with high surrogate scores and chemical quality. It seems like Table 3 does not isolate the benefit of learning neighborhood productivity from the benefit of selecting promising centers and allocating more samples around them.
Questions for authors
  1. How does FRATTVAE perform on the same reconstruction inputs under matched evaluation protocols?
  2. Does learning neighborhood productivity select better centers than simply choosing centers with high surrogate scores and acceptable chemical quality, when everything else is held fixed?
  3. Could the authors clarify which empirical findings most directly establish the added value of the proposed latent-conditioning design beyond the inherited constrained graph generator?
4 comments
AuthorsOpenReview ↗

Author Response to hNCJ (1/2): Q1–Q2

Response to Reviewer hNCJ

We thank the reviewer for the careful review and specific suggestions for clarifying the methodological contribution of GX-TGVAE.

In response, we added a reconstruction comparison with FRATTVAE using common inputs and a direct comparison between a selector that learns neighborhood productivity and a simple center-selection policy. We also clarified the added value of introducing molecule-level continuous latent conditioning into the inherited constrained graph generator, both in terms of the design challenges and the experimental evidence.

Q1. Reconstruction comparison with FRATTVAE

We evaluated Exact Reconstruction for the publicly distributed FRATTVAE model and GX-TGVAE EncoderDrop0 using the same 10,000 molecules from each of the official MOSES test and test_scaffolds sets, under an identical evaluation procedure.

On the official test set, GX-TGVAE achieved 94.715 ± 0.242%, compared with 94.650% for FRATTVAE. On the scaffold test set, the corresponding values were 93.518 ± 0.194% and 92.730%. The GX-TGVAE results are the mean and standard deviation across four independent training runs, whereas the FRATTVAE results use a single publicly distributed checkpoint.

The important point is that GX-TGVAE achieves comparably high reconstruction fidelity with approximately 29% of FRATTVAE's parameter count (19.30M vs. 66.09M), despite constructing molecular graphs sequentially from atoms and bonds rather than relying on a high-level fragment tree as FRATTVAE does.

Furthermore, the same GX-TGVAE EncoderDrop0 model series achieved observed Validity and Connected Validity of 100%, an average of approximately 3,911 unseen generic topologies per 30,000 generated molecules, and an unseen ring-skeleton rate of approximately 1.578%.

These results support the ability of primitive construction without a high-level fragment dictionary to combine high-fidelity reconstruction with broad structural reach.

FRATTVAE, however, performs better on FCD/Test. We therefore evaluate distributional fidelity, reconstruction fidelity, and Structural Reach as distinct capability axes.

We also confirmed that the public FRATTVAE fragment dictionary and frequencies use preprocessing information derived from inputs that include the MOSES evaluation sets. This does not imply that evaluation molecules were included in weight training; we describe this difference in preprocessing conditions between the models in Appendix J.4.

The comparisons are reported in Table 1 and Appendices C.8 and J.

Q2. Added value of the Learned neighborhood-productivity selector

We consider it important to distinguish the contribution of learning neighborhood productivity from that of allocating additional sampling resources to promising centers; the improvement from Massive Harvest alone does not sufficiently separate these effects.

We therefore added a direct control with 128 final evaluation slots: 32 centers were selected from the same 9,000 candidate centers, and the saved K = 4 candidates for each selected center were evaluated.

Compared with the Simple selector, which selects centers using their own surrogate scores and chemical quality, the Learned selector, which predicts neighborhood hit productivity, achieved Novel Hit Ratios of 8.4375% vs. 4.8438% for TanNorm and 10.4688% vs. 6.0417% for TanNormES.

We further conducted the same type of direct comparison in the separate BPE3000-M1 model series. The Learned selector outperformed the Simple selector in both cases: 11.67% vs. 3.125% for TanNorm and 11.67% vs. 4.427% for TanNormES.

Each comparison matched the candidate-center pool, number of selected centers, number of neighborhood candidates, and final evaluation slots.

Thus, a center having a favorable surrogate value and chemical quality is not equivalent to its neighborhood yielding useful molecules at a high frequency.

These direct controls support the conclusion that learning neighborhood productivity itself improves hit-discovery efficiency under a fixed number of final evaluation slots.

The results are reported in Section 7.3 and Appendices L.1 and L.5.

AuthorsOpenReview ↗

Author Response to hNCJ (2/2): Q3Part 1/2

Q3. Added value of the latent-conditioning architecture

The purpose of introducing molecule-level continuous latent conditioning into the constrained primitive construction inherited from GraphXForm is not merely to generate chemically valid molecules. It is to map known molecules into a molecule-level continuous latent space and enable exploration of unobserved structures by manipulating the direction and distance of movement in that space.

GraphXForm implements the selection of legal construction actions based on the current partial graph. That construction mechanism alone, however, does not provide a Graph-to-Graph VAE that acquires a continuous latent variable for an input molecule and uses its molecule-specific information for faithful reconstruction and latent manipulation.

Our technical challenge was to connect continuous latent information representing an entire molecule to sequential discrete construction starting from an empty graph, and establish a generative system usable for both reconstruction learning and latent exploration.

To achieve this, we addressed three design requirements.

First, to faithfully reflect the information in the latent variable, starting generation from an empty graph is more appropriate than placing the first atom at random. Construction from an empty graph, however, requires selecting the first atom based on the molecular latent variable before any actual chemical atom exists.

GraphXForm also uses a virtual node for a global state representation. In our work, we connect this mechanism to latent-conditioned construction from an empty graph and use a persistent global construction token (a0a_0), AddAtomFirst, and FirstAtomHead to design a configuration that can use latent information from the very first atom selection.

Second, the same whole-molecule latent information must remain available at every stage as the partial graph grows, rather than being used only to choose the first atom.

We expand the latent variable into multiple latent tokens and incorporate cross-attention into every decoder layer. In the adopted configuration, the global token directly attends to the latent tokens, and the retrieved information propagates to the construction state through graph self-attention.

This establishes an information pathway connecting the changing partial graph with the fixed molecule-level latent information at each stage of sequential construction.

Third, training an autoregressive action decoder for reconstruction with teacher forcing requires a target action for each construction state. Yet a molecular graph can be constructed in multiple orders that yield the same final structure. Reconstruction learning with a molecule-level latent variable therefore requires a training formulation that handles this multiplicity of construction orders.

We map each molecular graph to one canonical construction trace, defining a deterministic target sequence for teacher forcing. This allows us to compute the Graph-to-Graph VAE reconstruction loss without enumerating and marginalizing over all possible construction sequences.

AuthorsOpenReview ↗

Author Response to hNCJ (2/2): Q3Part 2/2

These three elements are not a juxtaposition of independent improvements. They address specific challenges that arise when connecting molecule-level continuous latent representations to constrained primitive construction; leaving any one unresolved would prevent this connection from functioning. Our central methodological contribution is not the first introduction of individual components, but the simultaneous realization, within a single Graph-to-Graph VAE, of three capabilities: representing an entire molecule in manipulable continuous latent coordinates, generating a broad range of structures from atoms and bonds without requiring a high-level fragment dictionary, and applying chemical constraints during construction even under latent extrapolation.

Three pieces of experimental evidence support the added value of this design.

First, in the EncoderDrop0 latent-replacement control, reconstruction Tanimoto was approximately 0.976 with the correct latent variable, but fell to approximately 0.135 when latent variables were shuffled across molecules. This directly shows that the decoder actually uses molecule-specific latent information.

Second, the same EncoderDrop0 model series achieved Exact Reconstruction of 94.715 ± 0.242% across four independent training runs, observed Validity and Connected Validity of 100%, and reach into unseen generic topologies and ring skeletons. Thus, faithful use of molecular information and structural exploration through constrained primitive construction coexist in the same generative system.

Third, latent-direction manipulation on ZINC generated molecules structurally distant from the training distribution and improved on the anchors in external docking evaluation. This shows that the introduced continuous latent representation is not merely a reconstruction code: it can serve as exploration coordinates for directional control of molecular structure and external utility.

The added value of our work is therefore to extend the constrained graph generator inherited from GraphXForm into a Graph-to-Graph VAE capable of faithful reconstruction and geometric exploration of unknown structures based on molecule-level continuous representations.

Our central methodological contribution is not the first invention of individual components, but their integration into a coherent information interface and training configuration that establishes a functioning generative architecture.

I thank the authors for their detailed response and new experiments. The FRATTVAE comparison and the matched control experiments address my original comparison and selector questions. However, I still find the overall methodological advance modest, which was my original concern. I recognize that GX-TGVAE is effective but remain unconvinced that the reported performance gains and empirical findings demonstrate that its particular architectural choices and combination of components provide a sufficiently significant advance. Therefore, I maintain my rating.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions and Author Responses

We have uploaded a revised manuscript addressing the reviewers' concerns.

The revision clarifies the central methodological contribution of GX-TGVAE, improves Figure 1, and adds controlled comparisons for reconstruction fidelity, latent geometry, ES, and Harvest.

We provide detailed responses to hNCJ, tjPv, and F9Se in their respective review threads, with references to the relevant sections, tables, and appendices.

We thank the reviewers and Area Chair for their consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.