acceptodds
Under review as a conference paper at ICLR 2027

Fraglingo: Molecular Design via Attachment-Aware Autoregressive Fragment Generation

Abstract

Molecular design is most effective when generation mirrors the edits chemists actually make: extending a scaffold, replacing a substituent, or decorating a scaffold at a specified attachment site while optimizing molecular properties. Fragment-based molecular design naturally supports this workflow, yet existing approaches often separate fragment selection from attachment prediction, first choosing a fragment from a fixed vocabulary and then predicting how it should be connected. This decoupling restricts generation to a closed fragment vocabulary and treats attachment as a separate prediction problem. We introduce Fraglingo, an autoregressive fragment-based molecular generator that jointly models fragment identity and attachment in a continuous latent space. Fraglingo predicts an attachment-aware fragment embedding and retrieves the next fragment through latent-space nearest-neighbor search. To encode attachment context, we introduce a wildcard-anchored readout that represents the growing molecule from the perspective of its active attachment site, enabling the predicted embedding to capture both the molecular context and the required attachment. Because generation operates in a continuous embedding space rather than over fixed fragment identifiers, new fragments can be added to the inference-time vocabulary without retraining, provided their embeddings can be computed by the trained fragment encoder. This retrieval-based formulation provides a unified generation primitive for molecule generation, scaffold generation, scaffold decoration, and molecular optimization. On controlled property-conditional benchmarks, Fraglingo achieves stronger joint property control than comparably trained baselines while maintaining competitive validity, uniqueness, and novelty. Furthermore, Fraglingo generalizes to fragment libraries up to larger than those used during training without retraining. Code is available at: https://anonymous.4open.science/r/FragLingo-3551.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Molecular Generation, Fragment-Based Generation, Attachment-Aware Representation, Property-Conditioned 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_v2ncOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Fraglingo, a molecular generation framework that performs autoregressive fragment assembly via continuous representation retrieval. The model uses a shared graph encoder and performs attention readout with active attachment points as queries to predict attachment-point-aware representations of the next fragment. By retrieving candidate "fragment–attachment point" representations, the model simultaneously determines the new fragment and how it is attached. Training combines InfoNCE, auxiliary fragment classification, and RVQ regularization; at inference time, the model allows encoding of new fragments and expansion of the candidate library without retraining the generator. The authors apply the framework to molecular generation, molecular optimization, scaffold generation, and scaffold decoration, and report experiments on joint property control, component ablations, data and model scale, and inference-time vocabulary expansion. The paper has a reasonable technical motivation and potential practical value. However, the current manuscript does not adequately define its added contribution relative to existing attachment-point-aware generation and retrieval methods; there remains an unresolved confounding factor between the strongest performance and the fragment partitioning scheme; and the experimental coverage and several reporting issues also limit the credibility of the core conclusions.

Critical strengths
  1. Clear correspondence between problem and design. Different attachment points of the same fragment may lead to different assembly outcomes, so establishing conditioned representations for fragments and their attachment positions is a reasonable modeling choice.
  2. Decoupling the candidate library from fixed output classes has practical significance. Table 4 shows that the model can select fragments outside the training vocabulary, indicating that expanding the candidate library at inference time is operationally feasible.
  3. Broad experimental coverage. The appendix includes experiments on loss components, data scale, vocabulary size, model scale, and other molecular design tasks, providing a certain basis for analysis.
  4. Some results are attractive. The BFE version demonstrates good joint property control, and the molecular optimization results also show good target matching and structure retention.
Critical weaknesses

W1. The core novelty has not been adequately established, especially the lack of positioning and comparison with directly related work such as MiCaM. Attachment-point-aware fragment generation already has clear precedents. FREED explicitly models the current molecular attachment point, candidate fragments, and candidate fragment attachment points, and makes subsequent decisions depend on prior selections; MoLeR also includes an attachment point selection mechanism. Therefore, "considering attachment point context" itself cannot be regarded as a new contribution. More directly relevant is MiCaM (ICLR 2023). Its §2.2 uses "*" to retain attachment information, starts from the current attachment point, constructs queries from graph and node representations, and scores candidate attachment point representations, thereby simultaneously determining the new fragment and its attachment mode. Appendix A.4 also uses in-batch positive and negative samples for contrastive training. The manuscript does not discuss this important precedent. The authors need to clearly explain what capability the wildcard-anchored readout in this paper adds relative to existing attachment point query and joint selection mechanisms. The existing readout ablation provides some evidence, but lacks direct comparisons such as attachment point node representations, global pooling and concatenation with attachment point representations, and stepwise fragment/attachment point prediction. The t-SNE in Figure 3 mainly shows that representations vary with attachment points; it cannot sufficiently prove that attachment point selection is more accurate, nor that there is a consistent geometric transformation in the original representation space. In addition, the distinction between "continuous retrieval" and "classification" needs to be defined more precisely. Applying softmax scoring to dynamic candidate structure representations is not the same mechanism as using a classification head with fixed class parameters. MiCaM already belongs to the former candidate matching approach. Its vocabulary transfer experiment involves parameter fine-tuning, so it cannot yet be used to deny the value of this paper's "library expansion without retraining"; however, this paper should verify this added capability under the same frozen library expansion protocol. W2. Coverage of related work on retrieval and vocabulary expansion is insufficient, and the closest methods are not adequately discussed. DeepFrag has already adopted the approach of predicting fragment fingerprints and then matching fragments through a candidate library, and allows the candidate set to be independent of the training set. RetMol guides controllable generation by retrieving example molecules and dynamically updating the database. f-RAG directly studies fragment-level retrieval and expanding the fragment vocabulary during generation. These are respectively relevant to this paper's claims regarding representation retrieval, controllable generation, and candidate library expansion. These methods are not identical to Fraglingo, but the current manuscript does not fully explain the differences or establish added value through representative quantitative comparisons. In particular, f-RAG should be regarded as an important reference for fragment retrieval and vocabulary expansion. GEAM's research on dynamic target-related fragment vocabularies should also be included in the contribution positioning. The main issue here is not the number of baselines itself, but that the existing comparisons do not adequately cover the closest mechanisms and capabilities. W3. The strongest performance is confounded with the fragment partitioning scheme, and existing ablations cannot fully attribute it to continuous retrieval. In Table 2, the standard BRICS version has NJD 1.9808 and Joint@1× 40.88%, lower than MolGPT and MolRWKV on these property control metrics; its advantage is mainly reflected in the 100% validity rate. Therefore, the results reflect trade-offs across different dimensions and cannot yet support the claim that the standard version is generally superior in property control. The BFE version raises Joint@1× to 69.65%. In comparison, removing the attachment point readout reduces the BRICS version from 40.88% to 34.66%. Although these differences cannot be directly interpreted as the proportional contribution of each component, they indicate that fragment granularity has a significant impact on the results. The authors need to clarify whether different partitioning schemes use exactly the same molecule sets and test targets, and compare generation mechanisms under matched partitioning schemes. Appendix A replaces InfoNCE with MSE, but both retain continuous vector prediction and nearest-neighbor retrieval. Therefore, this experiment verifies the role of the training objective and cannot independently prove that continuous retrieval is superior to discrete output mechanisms. The most direct missing control is: fixing the encoder, data, fragment partitioning, and condition input, and changing only stepwise prediction, compositional classification, and continuous retrieval. In addition, filtering training molecules according to different vocabulary coverage conditions causes vocabulary changes to be accompanied by data distribution changes. Therefore, some scale experiments cannot fully isolate the effect of vocabulary size. W4. The vocabulary expansion experiment mainly demonstrates that new fragments are selected, but does not sufficiently prove that expansion brings effective generalization. "Open vocabulary" as a capability to add new candidates after training is reasonable. However, in Table 4, after the 500-fragment model is expanded to 2000 fragments, NJD rises from 1.4928 to 1.8628, Partial drops from 6.56 to 6.16, and Diversity only rises from 0.8218 to 0.8382. These results show a trade-off between coverage and property control. A higher OOV fragment selection ratio alone is not sufficient to show that new fragments are correctly utilized, because it may also be affected by the number of candidates and the embedding distribution. It is recommended to report independent performance for molecules containing OOV fragments, Joint@1× after expansion, and the structural distance between new fragments and training fragments. A few cases showing that new fragments help satisfy objectives that are difficult to achieve with the original candidate library would also be more convincing than reporting only the OOV selection ratio. W5. Property and task evaluation is not yet sufficient to fully test attachment mode modeling and broader molecular design capability. The seven properties are mainly two-dimensional physicochemical descriptors and overlap in definition. QED itself integrates information such as MW, logP, polar surface area, hydrogen bond donors and acceptors, and rotatable bonds, so the seven metrics cannot be regarded as validation of seven independent property categories. The low computational cost of properties does not mean joint control is necessarily easy. However, molecules produced by different attachment positions can have the same MW, HBD, and HBA, as well as similar other descriptors. Therefore, the current evaluation may not sufficiently distinguish the quality of attachment modes. It is recommended to add attachment point selection evaluation involving symmetric equivalent positions, and to expand the target types using representative PMO activity prediction or complex MPO tasks, such as DRD2, GSK3β, JNK3, and Sitagliptin or Ranolazine MPO. The original PMO does not include docking; its activity targets use predictive models; evaluation should also unify the oracle budget and report sample efficiency. Molecular optimization also lacks a baseline that "directly returns the original molecule." Because training and test pairs have high structural similarity and limited variation, the original molecule may already satisfy some looser objectives. Reporting only final NJD and similarity is insufficient to judge the actual degree of improvement. The change in target distance before and after optimization and the improvement rate should be reported. In addition, scaffold generation and decoration mainly rely on self-comparison. MoLeR and PromptSMILES already have scaffold-constrained generation or decoration capabilities, and interface differences are not sufficient to completely replace representative comparisons. For overall optimization capability, it is valuable to appropriately add search-based methods, but MCTS should serve as a supplementary reference rather than a necessary substitute for validating attachment point representations. W6. There are specific contradictions and missing key details in the experimental reporting. The following issues need to be explicitly checked:

  • Contradiction in threshold success rates in Table 5: The MSE row has Joint@1× 26.85%, while Joint@2× is 24.94%. Under the same samples and statistical method, relaxing the tolerance should not reduce the success rate.
  • Inconsistent data split description: The main text states an 8:1:1 scaffold split, while Appendix F describes a 90/10 training/validation split. If there is a nested relationship, it needs to be explained.
  • Generation randomness is unclear: The appendix describes selecting the highest-similarity candidate at each step, but does not fully explain where the random differences among 20 samples under the same property condition come from.
  • Incomplete statistical reporting: The main text claims to report means and standard deviations over three random seeds, but the strongest BFE configuration and several key experiments do not provide standard deviations.
  • Unclear optimization baseline condition interface: Appendix F.2 does not fully explain how HierVAE uses the seven-dimensional target properties, nor does it provide the complete JT-VAE experimental setup.
  • Numerical values in qualitative examples need verification: In the left side of Figure 5, the MW target is 500, the generated value is 344, yet the reported error is −28.5, which is inconsistent with the displayed value. These issues are not all of equal severity, but together they limit reproducibility and interpretation of the results.
Questions for authors
  1. Relative to MiCaM's existing attachment point query, joint fragment selection, and contrastive training, what is the substantive new contribution of this paper? Can you provide mechanism comparisons with matched backbone and data?
  2. Do BFE and BRICS use the same molecule sets and test targets? What are the repeated-experiment statistics for the strongest BFE configuration?
  3. When expanding the candidate library with frozen parameters, what advantage does this paper have over existing candidate representation matching methods? Can molecules containing OOV fragments still reliably satisfy the objectives?
  4. Can the current seven-property evaluation effectively distinguish different attachment modes? Can attachment-point-specific evaluation and a small number of representative activity/MPO tasks be used to strengthen the evidence?
  5. How much does molecular optimization improve over directly returning the input molecule? Do all baselines receive the same target information?
  6. How should the threshold contradiction in Table 5, the data split description, and the generation randomness mechanism be explained? Overall Assessment I acknowledge the potential value of attachment-point-conditioned representations and inference-time candidate library expansion. However, the current version has not yet sufficiently answered three decisive questions: what is added relative to the closest existing mechanisms; whether the strongest performance comes from the generation mechanism or the fragment partitioning; and whether expanding the candidate library and applying it across tasks brings verifiable benefits. My inclination to reject is mainly based on these evidence gaps and the unresolved experimental reporting issues. If the authors can clearly define the differences from MiCaM and retrieval-based work, provide targeted matched comparisons, and correct key experimental contradictions, I would be willing to reassess the recommendation.
Reviewer_w7jEOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces Fraglingo, a deep learning framework for generating molecules with desired physicochemical properties through autoregressive fragment assembly. It introduces attachment-aware latent retrieval to jointly select fragments and their attachment sites, enabling inference-time vocabulary expansion without retraining. Rather than treating next-fragment prediction as classification over a fixed vocabulary, the model predicts a continuous embedding conditioned on the growing molecule, its active attachment site, and target properties, and retrieves a candidate from an attachment-aware fragment library. Experiments evaluate generation quality, property control, and vocabulary expansion. The results demonstrate that the model can incorporate previously unseen fragments during inference, while showing varying levels of property control depending on the fragmentation scheme and retrieval vocabulary size. These findings highlight a distinction between extending the accessible chemical space and effectively using that flexibility for property-directed molecular design.

Critical strengths
  • Fraglingo provides a coherent framework for jointly selecting fragments and their attachment configurations. The wildcard-anchored readout addresses the problem of same fragment connected to different sites which is a substantive ambiguity in fragment assembly . Representing each candidate relative to its own attachment site lets one retrieval decision specify both the fragment and its connection to the growing molecule's active site. Figure 3 provides a qualitative case study that the readout does respond to the local attachment environment. The BRICS readout ablation in Table 2 also worsens NJD from 1.9808 to 2.1256, supporting the contribution of site-specific contextualization. The same assembly interface is reused across the four design tasks.

  • The experiments demonstrate inference-time retrieval of fragments unseen during training. The expansion experiment tests whether newly encoded candidates are actually selected during generation. In Table 4, expanding a 500-fragment BFE vocabulary to 2,000 candidates yields 48.76% OOV fragment usage while retaining 100% validity; expanding a 1,000-fragment vocabulary to 2,000 yields 28.32% OOV usage. Because the encoder is frozen and the training-only classification head is discarded, this provides direct evidence that generation can exploit an updated fragment library without retraining. Such extensibility is a useful capability for incorporating new fragment collections, even though its benefit for satisfying design objectives remains a separate empirical question.

Critical weaknesses
  • Autoregressive fragment generation, attachment-aware representations, and embedding-based fragment selection have substantial precedents including BBAR [1] and M-FRAG [2]. BBAR already supports new building blocks without retraining, although it selects the building block and connecting atom separately. M-FRAG aligns molecular and fragment representations in a shared, property-driven space for reinforcement-learning-based generation. Although Fraglingo introduces a different formulation, the manuscript does not clearly establish its methodological advantage over these approaches or stronger recent baselines. Moreover, comparisons with more recent conditional generation methods, such as STGG+ [3] and Graph DiT [4], are missing. These comparisons should use matched training data, target properties, and sampling budgets. Fraglingo-BRICS underperforms MolGPT [5] and MolRWKV [6] on joint property-control metrics, while stronger results depend heavily on BFE fragmentation. Table 2 reports NJD of 1.9808 for Fraglingo-BRICS, compared with 1.7339 for MolGPT and 1.6090 for MolRWKV, despite Fraglingo's higher validity. The best NJD in Table 2, 1.4928, is obtained with BFE, which coarsens BRICS fragments into larger units. Appendix A and Table 11 usefully examine components and fragmentation within Fraglingo, but do not establish a cross-method advantage under matched fragmentation where the assembly rules permit this comparison.Jointly updating fragment embeddings in a shared space can introduce conflicting optimization pressures, potentially disrupting previously learned relationships elsewhere in the latent space. The manuscript does not investigate embedding stability or whether its contrastive objective learns a meaningful similarity structure for unseen fragments.

  • Jointly updating fragment embeddings in a shared space can introduce conflicting optimization pressures, potentially disrupting previously learned relationships elsewhere in the latent space. The manuscript does not investigate embedding stability or whether its contrastive objective learns a meaningful similarity structure for unseen fragments. Appendix E specifies a shared encoder for partial assemblies and candidate fragments, with InfoNCE, vocabulary-level classification, and RVQ commitment jointly shaping the retrieval space. Multiple chemically acceptable continuations may also appear among the in-batch negatives, creating a potential source of conflicting supervision. Figure 3 and Table 5 provide evidence of attachment sensitivity and the value of the training objectives, but they do not evaluate retrieval for held-out fragment identities or whether fitting some relationships degrades others during training. Inference-time expansion uses a frozen encoder; the concern is therefore the geometry learned during training and its generalization to newly encoded candidates. The reported experiments do not establish whether harmful interference actually occurs.

  • Vocabulary expansion is arguably the most important contribution of the proposed framework. Table 4 shows that expanding the fragment vocabulary increases OOV fragment usage while worsening property control. Expanding the 500-fragment vocabulary to 2,000 worsens NJD from 1.4928 to 1.8628 and lowers the mean number of satisfied property targets from 6.56/7 to 6.16/7. The 1,000-to-2,000 expansion follows the same pattern: NJD rises from 1.2318 to 1.5018 and partial satisfaction falls from 6.83/7 to 6.54/7. Validity, uniqueness, and novelty remain essentially saturated, and fingerprint diversity increases; the authors acknowledge this tradeoff. Therefore, although the model can retrieve unseen fragments, the experiments do not show whether this capability improves molecular design performance. An evaluation of diversity or chemical coverage among target-satisfying molecules, or an objective for which the original library is insufficient, would clarify when expansion provides a measurable advantage.

  • references [1] Seo, S., Lim, J., and Kim, W. Y. (2023). Molecular Generative Model via Retrosynthetically Prepared Chemical Building Block Assembly. Advanced Science, 10(8), 2206674. https://doi.org/10.1002/advs.202206674 [2] Kim, Y., Piao, Y., Lee, S., and Kim, S. (2025). Aligning Molecules and Fragments in a Shared Embedding Space for RL-Based Molecule Generation. ICLR 2025 Workshop MLGenX. https://openreview.net/forum?id=lwK6AaIAJB [3] Jolicoeur-Martineau, A., Baratin, A., Kwon, K., Knyazev, B., and Zhang, Y. (2025). Any-Property-Conditional Molecule Generation with Self-Criticism using Spanning Trees. Transactions on Machine Learning Research. https://openreview.net/forum?id=QGZd5Bfb1L [4] Liu, G., Xu, J., Luo, T., and Jiang, M. (2024). Graph Diffusion Transformers for Multi-Conditional Molecular Generation. Advances in Neural Information Processing Systems, 37, 8065-8092. https://doi.org/10.52202/079017-0260 [5] Bagal, V., Aggarwal, R., Vinod, P. K., and Priyakumar, U. D. (2022). MolGPT: Molecular Generation Using a Transformer-Decoder Model. Journal of Chemical Information and Modeling, 62(9), 2064-2076. https://doi.org/10.1021/acs.jcim.1c00600 [6] Li, X., Gong, K., Jiang, Y., Yang, Y., and Li, T. (2025). MolRWKV: Conditional Molecular Generation Model Using Local Enhancement and Graph Enhancement. Journal of Computational Chemistry, 46(10), e70100. https://doi.org/10.1002/jcc.70100

Questions for authors
  • How well does the learned embedding space generalize to fragments outside the original fragment library and training distribution? It is recommended to evaluate retrieval performance after incorporating fragments from an independent library or a structurally out-of-distribution set, rather than simply expanding the vocabulary with additional fragments from the same corpus.
  • How does InfoNCE handle cases where multiple fragments or attachment configurations are equally valid continuations? Could treating these alternatives as negative, while being false negative distort the learned embedding space?
  • Since fragment embeddings are jointly learned and share encoder parameters, it is recommended to investigate if optimizing some fragment relationships does not disrupt previously learned relationships elsewhere in the embedding space?
  • Fraglingo-BRICS underperforms several baselines, while the strongest results depend heavily on BFE fragmentation. This makes it difficult to attribute the reported performance gains to the proposed methodology. It is recommended to discuss why Fraglingo performs substantially better with BFE than with BRICS, and what characteristics of the fragmentation schemes might explain this performance gap.
Reviewer_M8cPOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The authors introduce an approach for fragment-based molecular generation that combines attachment point-aware representations with continuous fragment retrieval. This is motivated by the need to make the approach flexible to the choice of fragment libraries, which can be swapped during inference. The method is compared against baseline across several setting, in a task of generating molecules with properties similar to the provided vector of values.

Critical strengths

S1: The paper is clearly written and easy to comprehend.

Critical weaknesses

W1: Novelty and contributions: the authors claim that their main contribution is introduction of 1) attachment point-aware representations and 2) continuous fragment retrieval. Similar concepts were already explored in the literature, e.g. [1-3] for representations, [4-6] for retrieval. The motivation, a flexible fragment library that can be swapped during inference, also was demonstrated in earlier work, e.g. [7]. These references are very much not exhaustive, but neither them nor similar ones were properly discussed in the paper (with Related Works section being a single short paragraph), overstating novelty. Even if these papers did not exist, my personal feeling is that introducing a way to include attachment points in the representation, combined with reusing existing loss functions, would be a very modest methodological contribution.

W2: Experimental setup is based on very dated papers. Molecular optimization experiment, for example, uses methods from 2018 and 2020. Experimental setup is also fairly non-standard, from my experience, with property control used as the main optimization criterion. While it’s an interesting experimental choice, molecular generation has a huge body of work, with fairly established experimental setups, and probably the most straightforward one that is missing would be maximizing a desired (non-trivial) property. As is, it’s very difficult to establish how well does the method perform against SotA in typical tasks.

[1] Hit and Lead Discovery with Explorative RL and Fragment-based Molecule Generation (NeurIPS 2021) [2] Hierarchical Generation of Molecular Graphs using Structural Motifs (ICML 2020) [3] Learning to Extend Molecular Scaffolds with Structural Motifs (ICLR 2022) [4] MolPLA: a molecular pretraining framework for learning cores, R-groups and their linker joints (Bioinformatics 2024) [5] Learning to Navigate The Synthetically Accessible Chemical Space Using Reinforcement Learning (ICML 2020) [6] DeepFrag: a deep convolutional neural network for fragment-based lead optimization (Chemical Science 2021) [7] Generative Flows on Synthetic Pathway for Drug Design (ICLR 2025)

Questions for authors

S1: The authors list the ability to “exploit larger or updated fragment libraries at inference time” one of their advantages. I feel like it would advantages to discuss in more detail why this is important, from the practical point of view, since while it’s an interesting benefit, I fail to see why it would be particularly critical.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.