acceptodds
Under review as a conference paper at ICLR 2027

MolTokSyn: Molecule-Text Retrieval With Text-Conditioned Tokens and Structure-Grounded Synthetic Captions

Abstract

Natural language provides a flexible way to search for molecules and describe the chemical properties that matter for a given task. However, existing molecule-text datasets rely largely on database or literature descriptions that vary widely in content and do not consistently emphasize properties that can be derived from molecular structure. We address this limitation with MolSynText, a synthetic molecule-text dataset whose descriptions are generated from structured molecular information and screened for consistency with their source. We further introduce MolTokSyn, a graph-text retrieval model that represents each molecule with multiple learned molecular tokens and uses the text query to determine how these tokens are combined for scoring, allowing different aspects of the same molecule to be emphasized for different descriptions. Across bidirectional molecule-text retrieval, MolTokSyn consistently outperforms strong existing methods on MolSynText and remains competitive on matched PubChem-derived text, while its learned molecular representations transfer strongly to downstream property-prediction tasks. Because retrieval alone does not reveal which chemical concepts are captured by the learned representations, we additionally introduce a controlled Semantic Benchmark covering structural, physicochemical, and relational properties. On this benchmark, MolSynText supervision improves MolTokSyn on most semantic categories, with particularly strong gains for descriptor-grounded properties. Together, these results show that structure-grounded synthetic descriptions can provide informative chemical supervision, especially when combined with text-conditioned molecular representations.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
molecule retrieval, contrastive learning, synthetic datasets, representation learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_FYb8OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents two complementary contributions for molecule–text retrieval. 1. MolSynText: a synthetic molecule–text dataset in which textual descriptions are generated by Qwen3.8-27B from RDKit-derived “molecular cards” and subsequently filtered using the same model as a judge. Each molecule in MolSynText is also paired with its corresponding PubChem-derived text from STM. 2. MolTokSyn: a retrieval model that augments a Chemprop D-MPNN with K one-way virtual readout nodes, following the design of Llip. Experimentally, the authors report substantial improvements on MolSynText. The proposed approach also yields improvements in linear probing and on the Semantic Benchmark.

Critical strengths
  1. The experimental setup is generally well controlled, e.g., using the same molecules, splits, and architectures across the two text sources allows a relatively direct comparison of textual supervision. Retrieval is evaluated over 12K candidates and three scaffold splits.

  2. The paper reports the data construction pipeline in useful detail, including prompts, judge criteria, acceptance thresholds, and rejected examples.

  3. The architecture is simple and easy to understand, preserves the pretrained message-passing backbone, and allows molecular tokens to be cached.

Critical weaknesses
  1. The ablations do not show that text conditioning is responsible for the main gains. Several results make it difficult to determine whether the improvements come from text conditioning, the molecular backbone, pretraining, or training configuration.
  • With K=1, the softmax in Eq. (4) is identically 1, so the model effectively reduces to a dual encoder. Yet K=1 is best on all eight STM metrics in Table 12, and K=1 and 8 are nearly identical on MolSynText in Table 13. The only evidence in favor of query conditioning is Table 11, but that comparison uses InfoNCE and a smaller batch, does not specify how the “Global” variant pools, and does not report K=1 under the same setting.
  • The linear probing study compares against a randomly initialized model rather than frozen CheMeleon. Without a frozen CheMeleon control, it is difficult to separate the effect of molecule–text training from the pretrained backbone.
  1. MolSynText needs stronger validation
  • The same model is used to generate and judge the captions, which may produce correlated errors. The paper does not report the rejection rate, retry count, or a human-audited error rate. A blind expert audit of a random sample would strengthen the dataset contribution.
  • Some captions contain exact descriptor values or IUPAC names. With ~12K retrieval candidates, combinations of exact numerical values may strongly narrow down molecule identity.
  • Higher absolute retrieval performance on MolSynText than on STM does not by itself show that MolSynText provides more informative supervision; it may simply indicate that MolSynText is an easier retrieval task.
  1. The Semantic Benchmark contains only 150 items, including 20 Relational questions. The benchmark was written by the authors using phrasing close to the MolSynText templates, and some labels follow the same RDKit conventions. The improvement may therefore partly reflect alignment between the training and evaluation text rather than stronger chemical understanding.

  2. The architectural novelty is limited, as the model largely adapts Llip to molecular graphs, while its multi-token readout is related to set-based graph pooling methods such as PMA in Set Transformer and Graph Multiset Transformer. These connections could be discussed more clearly.

Questions for authors
  1. What pooling is used in the “Global” variant in Table 11, and how does it differ from MolTokSyn with (K=1)?
  2. Do the baselines use their original molecular encoders (e.g., GIN or Uni-Mol) or the same CheMeleon-initialized D-MPNN? If the former, how large is MolTokSyn's advantage on MolSynText when a strong baseline uses the same CheMeleon encoder?
  3. What are the linear-probe results for frozen CheMeleon and for RDKit descriptors?
  4. What are the judge rejection rate and the human-audited error rate of the final captions?
  5. What fraction of MolSynText test captions contain exact descriptor values or IUPAC names, and how does retrieval change after removing them?
  6. For the Semantic Benchmark, the paper states that the test cases were manually defined. How were the correct answers verified, e.g., programmatically with RDKit for computable properties, or by independent annotators?
  7. is CheMeleon pretrained on classical molecular descriptors, many of which overlap with MolSynText content, including logP, TPSA, HBD/HBA, and ring counts? This may contribute to the gains on MolSynText.
8 comments
AuthorsOpenReview ↗

Response to Reviewer FYb8 (1)

We thank the Reviewer for the valuable feedback. All changes are marked in blue in the revised paper. We address each point below.

W1: Sources of the gains. We agree that part of the previous model ablations was run under varying setup. To address this, we reran the key experiments using the final training recipe with SigLIP, physical batch size 512, and matched optimization settings. Here is what we completed so far: the full K sweep on MolSynText (K=1, 4, 8, 16), the STM multi-token sweep (K=1, 4, 8, 16), the MolSynText-15x sweep (K=1, 8, 32), and the matched K=8 Query vs Uniform aggregation experiment.

We also added a matched comparison with conventional Global pooling (Table 15), using the final training recipe with SigLIP and physical batch size 512. Replacing Global pooling with a single learned virtual token improves R@1 by 17.61 and 17.89 points in the two retrieval directions. Increasing capacity from K=1 to query-conditioned K=8 provides further gains of 4.32 and 3.53 points, respectively.

In addition, we added a matched control that keeps the same K=8 molecular tokens, keys, values, and scoring architecture, but replaces text-dependent attention with uniform aggregation, which can be seen in section Appendix D.2. As shown in Table 16, query conditioning improves R@1 by 6.96 percentage points for molecule-to-text and 6.61 points for text-to-molecule retrieval, with corresponding MRR gains of 5.49 and 5.05 points. This shows that the gain is not explained simply by using multiple molecular tokens, but also conditioning it with text. To address it further we added a token analysis showing that the learned routing actually changes with the query, which can be seen in Appendix D.2. With the molecular graph and K=8 tokens fixed, 84.75% of molecule-attention-head combinations change their dominant slot across natural descriptions. Under eight controlled semantic-focus queries, the dominant slot differs in 48.1% of pairwise molecule-head comparisons on average (Figure 6). Summing it up, these results support query-dependent reweighting of the molecular representation.

Additionally, K=1 is indeed a special case of our model. With only one molecular token, the attention weight in Eq. (4) is always 1, so the molecular representation no longer depends on the text query. Therefore, at the readout level, it behaves as a standard molecular embedding, which makes the K=1 comparison very useful, because the gains for K>1 show the benefit of having multiple tokens that can be weighted differently for different queries.

Additionally, on MolSynText, increasing K from 1 to 8 improves R@1 from 74.41% to 78.76% and from 75.08% to 79.02% in the two retrieval directions, while K=16 gives the best central value on 7 of 8 metrics. On STM, increasing K from 1 to 4 improves R@1 by 3.61 and 3.57 percentage points in the two retrieval directions. Performance then remains largely stable across K=4, 8, and 16, with differences lying within the reported standard deviations. This supports our claim that extra token capacity is useful when the textual supervision is rich enough to exploit it in both MolSynText and STM setup (what was a main goal of this architecture).

CheMeleon pretraining. To assess the contribution of CheMeleon pretraining, we added an ablation in Appendix D.6, “Initialization of the Molecular Encoder”, comparing pretrained and random initialization of MolTokSyn’s molecular encoder. Across both pretraining datasets, the randomly initialized variant achieves higher mean scores on every retrieval metric and lower mean errors on all regression tasks, while the relative performance on classification tasks varies by task. We provide the full ablation results and further discussion in our general response A on molecular encoder initialization, downstream transfer and our response to Question 3 (What are the linear-probe results for frozen CheMeleon and for RDKit descriptors?).

AuthorsOpenReview ↗

Response to Reviewer FYb8 (2)

W2: Caption validation and identity cues. Please see our general response B on caption validation, which reports the independent judges, the regeneration rates and a blinded expert review (Appendix B.4). In addition, we want to address hypothesis that the larger number of stored captions could translate into more training examples. For that we created two datasets based on STM and MolSynText with exactly one molecule-text pair per molecule (so each of them has 125k captions in total), which can be seen in Appendix D.1, Table 14. Even in this setting, MolSynText reaches 78.11% and 78.37% R@1, compared with 51.94% and 51.38% for STM. This shows that the improvement cannot be explained by the larger number of stored MolSynText captions. Moreover, to address higher descriptor density hypothesis we created new dataset (one caption per molecule), with removed explicit molecular names (IUPAC), SMILES, and numerical sequences. MolSynText still achieves 73.10% and 73.67% R@1. Summing it up, these controls show that the observed gains are not simply a consequence of greater text diversity or an easier retrieval problem driven by simple distinguishing cues. Moreover, MolSynText supervision improves downstream linear probing and the Controlled Semantic Benchmark across multiple architectures, showing that the richer information is useful beyond simply matching easier retrieval captions (what was also a goal of this paper). These results are discussed in Appendix D.1 and Sections 5.2-5.3.

W3. Train/evaluation text alignment and benchmark size. We partly agree, and the revised paper now states this explicitly (Section 5.3, Table 12, and Appendix E in the revised paper). MolSynText verbalizes an RDKit card (atom and ring counts, H-bond donors/acceptors, rotatable bonds, TPSA, logP, charge, functional groups). Categories that test these concepts are intentionally close to the training signal and share its RDKit conventions: they check whether a model acquired what the text teaches, and this is where MolSynText helps most (Global family: +21.4 points for our model, Table 6).

The other categories are out of distribution. Regiochemistry, ring topology, tautomer and protonation forms, analog edits and bioisosteric replacement are absent from the card and are phrased as a medicinal chemist would query the model. Several are nonetheless solved well: the Relational family reaches 61.7% for our model against 25% chance, bioisosteric replacement is among the highest-scoring categories for most models, and tautomer/protonation is well above chance (Figure 12). Ring topology and regiochemistry remain near chance, which we report. These five categories are not computed in the cards, so performance there is less tied to the cards’ RDKit conventions.

Two further points limit what alignment can explain. First, MolSynText has no fixed sentence templates: the templates are instructions telling the generator which aspect of a molecule to emphasize, and the descriptions themselves are free-form text of 2-4 sentences. Benchmark items are short single-claim statements, and several deliberately avoid descriptor vocabulary (e.g. "greasiest" or "most water-averse" instead of logP wording). Second, all six models are trained on the same text, so alignment is identical across them and does not by itself explain the between-model ranking.

We agree the benchmark is small. Per-category results (10 items) are diagnostic only; we draw conclusions from the overall score and the Global and Local families, and the revised paper states this limitation in Appendix E.

AuthorsOpenReview ↗

Response to Reviewer FYb8 (3)

W4: Relation to Llip and set-based pooling. We thank the Reviewer for pointing out these connections. We agree that these relations should be discussed more clearly.

Llip. Our model is inspired by the mixture-token idea of Llip (Lavoie et al., 2024), which we credit in Sections 1 and 4. Our adaptation is on the molecular side and is designed for message-passing networks: we add K learned one-way virtual readout nodes to the D-MPNN. These nodes collect information from the molecular graph but send no messages back, so message passing between the real atoms and bonds is unchanged (Section 4.1). Each virtual node has its own learned features and projection, producing several molecular representations without predefined chemical meaning. To our knowledge, this one-way virtual-node readout has not previously been used for molecule-text retrieval.

The second difference is how these representations are used. The molecular tokens are computed once per molecule and can be cached, and each text query combines them differently (Section 4.2). The query-conditioned representation is therefore the retrieval score itself, rather than a re-ranking step after first-stage retrieval as in MolCA and MolFM, and the text combines the tokens with learned attention weights rather than a hard maximum over them.

Set-based pooling. Producing several pooled vectors per graph from learned slot parameters is related to attention-based pooling with learned seed vectors, as in PMA of the Set Transformer (Lee et al., 2019) and the Graph Multiset Transformer (Baek et al., 2021), which computes its keys and values with a GNN. MolTokSyn differs in two ways. First, our slots do not attend over atoms: each virtual node aggregates atom messages through one-way edges inside the D-MPNN. Second, PMA and GMT produce a query-independent graph representation, whereas our slot outputs are mixed by the text query at scoring time, following Llip. The query-independent controls in Appendix D.2, a single virtual token (K=1K=1) and uniform weights over the same K=8K=8 tokens, measure what this query-dependent mixing adds on top of a multi-slot readout.

Ablations. Finally, we also provide ablations separating these design choices rather than treating the architecture as one block. The K sweep tests the value of multiple molecular tokens, the Query vs Uniform experiment isolates text-dependent aggregation, and the attention analysis verifies that routing actually changes with the query. Additionally, we study the effect of the contrastive objective and physical batch size separately.

Contribution. More broadly, the main contribution is the combination of a structure-grounded textual supervision pipeline with a molecular architecture designed to exploit that richer signal. The aim is not only better retrieval but molecular representations that capture chemical concepts from text. This is why we also evaluate downstream property prediction (Section 5.2) and the Controlled Semantic Benchmark (Section 5.3), which tests specific chemical concepts directly.

Q1: Global pooling versus K=1. We thank the Reviewer for this question and we clarified this distinction in the revised Appendix D.2. The Global variant uses the standard Chemprop graph readout: the final atom-level representations are aggregated using MeanAggregation into a single CheMeleon graph fingerprint, which is then projected to the shared 128-dimensional space, L2-normalized, and compared with the independently encoded SciBERT text representation. It therefore does not use virtual molecular nodes, molecular-token keys or values, or query-dependent attention.

MolTokSyn with K=1 is different. It replaces the standard global graph pooling with one learned one-way virtual readout node and uses the MolTokSyn token/value/output projection pathway. Because there is only one molecular token, the softmax weight in Eq. (4) is always 1, so the resulting molecular representation is query-independent. Therefore, K=1 is not equivalent to the Global baseline. Global uses conventional mean-pooled graph readout, while K=1 tests our learned virtual-token readout in its query-independent special case. Table 15 compares both variants under the same training setup. Global pooling obtains 58.78% and 59.40% R@1, whereas the learned K=1 virtual readout reaches 76.39% and 77.29%. The full query-conditioned K=8 model further improves these scores to 80.71% and 80.82%, respectively.

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

This paper proposes MolTokSyn, a molecule–text retrieval framework pairing a structure-grounded synthetic text dataset with text-conditioned multi-token molecular representations. Evaluated on molecule–text retrieval, property prediction, and a new Controlled Semantic Benchmark, MolTokSyn shows strong gains across multiple architectures.

Critical strengths
  • Chemically grounded synthetic supervision
  • Broad empirical gains
Critical weaknesses

W1. Unvalidated synthetic text quality

Relies entirely on Qwen3.8-27B for both caption generation and quality control using identical prompts, risking undetected systematic LLM errors without expert human or deterministic validation.

W2. Confounded text supervision baselines

Gains over the STM dataset may stem from MolSynText's larger volume (216K vs 154K), longer length, and higher descriptor density rather than the generation method itself.

W3. Limited real-world query generalization

Largest gains occur when testing on MolSynText's own synthetic distribution; evaluation lacks cross-dataset transfer to natural, human-written queries by chemists.

W4. Unconvincing multi-token architecture gains

Ablations use sub-optimal setups (InfoNCE, lower batch size), show single-token setups winning on STM, and lack token-attention analysis proving query-specific selection.

Questions for authors

Provided above

4 comments
AuthorsOpenReview ↗

Response to Reviewer RWpG (1/2)

We thank the Reviewer for the valuable feedback. All changes are marked in blue in the revised paper. We address each point below.

W1: Validation of synthetic text quality. Please see our general response B on caption validation, which reports the independent judges, the regeneration rates and a blinded expert review (Appendix B.4).

W2: Caption count, length and descriptor density. We want to thank the Reviewer for pointing out the possible differences between STM and MolSynText. First we want to address hypothesis that the larger number of stored captions could translate into more training examples. For that we created two datasets based on STM and MolSynText with exactly one molecule-text pair per molecule (so each of them has 125k captions in total), which can be seen in Appendix D.1, Table 14. Even in this setting, MolSynText reaches 78.11% and 78.37% R@1, compared with 51.94% and 51.38% for STM. This shows that the improvement cannot be explained by the larger number of stored MolSynText captions. Moreover, MolSynText descriptions are indeed slightly longer, but this is closely connected to the purpose of our generation procedure rather than an independent source of additional training examples. Our goal is specifically to produce descriptions that expose more structure-derived chemical information than the often sparse PubChem text. On the contrary, to address higher descriptor density hypothesis we created new dataset (one caption per molecule), with removed explicit molecular names (IUPAC), SMILES, and numerical sequences. MolSynText still achieves 73.10% and 73.67% R@1. Summing it up, these controls show that the observed gains are not simply a consequence of greater text diversity or an easier retrieval problem driven by simple distinguishing cues. Moreover, MolSynText supervision improves downstream linear probing and the Controlled Semantic Benchmark across multiple architectures, showing that the richer information is useful beyond simply matching easier retrieval captions (what was also a goal of this paper). These results are discussed in Appendix D.1 and Sections 5.2-5.3.

W3: Transfer to human-written queries. We have added a “Cross-View Knowledge Transfer” subsection D.7 in the Appendix, evaluating models trained on STM or MolSynText with queries from both corpus and human-written ChEBI-20 descriptions, without further training. ChEBI-20 provides descriptions for 1,179-1,207 held-out test molecules per split, approximately 10% of each test set. Each query is ranked against the full test-molecule pool.

Table R1: Zero-shot cross-view text-to-molecule retrieval (%). mean ±\pm std, over three splits, against the full test-molecule pool. Higher is better.

Training corpusQuery sourceR@1R@10
MolSynTextMolSynText79.02 ±\pm 1.6398.08 ±\pm 0.16
MolSynTextSTM22.84 ±\pm 1.0939.51 ±\pm 0.19
MolSynTextChEBI-2021.75 ±\pm 0.5047.95 ±\pm 2.55
STMChEBI-2050.79 ±\pm 2.9783.33 ±\pm 2.42
RandomAny query source0.00800.0800

The results confirm that retrieval remains sensitive to the query distribution. The MolSynText-trained model achieves 79.02% R@1 on MolSynText queries but 22.84% on STM queries, evaluated over the same test molecules. On human-written ChEBI-20 descriptions, it achieves 21.75% R@1 and 47.95% R@10. These scores substantially exceed the expected random-ranking performance of 0.0080% R@1 and 0.0800% R@10 against the full test-molecule pool. The STM-trained model achieves 50.79% R@1 on ChEBI-20.

We explicitly acknowledge that the PubChem descriptions used in STM include ChEBI annotations: although the query molecules are held out, this shared text source limits ChEBI-20’s independence as a transfer benchmark. These additional experiments provide evidence of partial transfer to curated human-written descriptions while highlighting the remaining challenge of generalization across query distributions.

AuthorsOpenReview ↗

Response to Reviewer RWpG (2/2)

W4: Ablations under the final setup and query-specific routing. We agree that part of the previous model ablations was run under varying setup. To address this, we reran the key experiments using the final training recipe with SigLIP, physical batch size 512, and matched optimization settings. Here is what we completed so far: the full K sweep on MolSynText (K=1, 4, 8, 16), the STM multi-token sweep (K=1, 4, 8, 16), the MolSynText-15x sweep (K=1, 8, 32), and the matched K=8 Query vs Uniform aggregation experiment. The revised paper now also reports a matched Global versus K=1 versus query-conditioned K=8 comparison under the same final training setup (Table 15). The results separate the contribution of the learned virtual readout from the additional benefit of multiple query-conditioned tokens.

In addition, we added a matched control that keeps the same K=8 molecular tokens, keys, values, and scoring architecture, but replaces text-dependent attention with uniform aggregation, which can be seen in section Appendix D.2. As shown in Table 16, query conditioning improves R@1 by 6.96 percentage points for molecule-to-text and 6.61 points for text-to-molecule retrieval, with corresponding MRR gains of 5.49 and 5.05 points. This shows that the gain is not explained simply by using multiple molecular tokens, but also conditioning it with text. To address it further we added a token analysis showing that the learned routing actually changes with the query, which can be seen in Appendix D.2. With the molecular graph and K=8 tokens fixed, 84.75% of molecule-attention-head combinations change their dominant slot across natural descriptions. Under eight controlled semantic-focus queries, the dominant slot differs in 48.1% of pairwise molecule-head comparisons on average (Figure 6). Summing it up, these results support query-dependent reweighting of the molecular representation.

Additionally, on MolSynText, increasing K from 1 to 8 improves R@1 from 74.41% to 78.76% and from 75.08% to 79.02% in the two retrieval directions, while K=16 gives the best central value on 7 of 8 metrics. On STM, increasing K from 1 to 4 improves R@1 by 3.61 and 3.57 percentage points in the two retrieval directions. Performance then remains largely stable across K=4, 8, and 16, with differences lying within the reported standard deviations. This supports our claim that extra token capacity is useful when the textual supervision is rich enough to exploit it in both MolSynText and STM setup (what was a main goal of this architecture).


We hope our responses and revisions address the Reviewer’s concerns. We would be happy to clarify any remaining questions during the discussion period.

Thank you for detailed explanations on my concerns. In overall, I believe authors addressed my concerns.

I will increase my score to 3.

We thank the Reviewer for the positive response and for raising the score. The Reviewer’s suggestions, in particular validating the captions with chemists and testing transfer to human-written queries, helped us strengthen our work. We thank the Reviewer again for the thoughtful review and for considering our responses.

Reviewer_reV6OpenReview ↗
Rating
3
Confidence
4(+1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies molecule-text retrieval using structure-grounded synthetic descriptions and text-conditioned molecular representations. It introduces MolSynText, whose descriptions are generated from RDKit molecular cards and screened by an LLM judge. A matched PubChem-derived text source covers the same molecules, allowing comparison of the two forms of supervision. The proposed MolTokSyn model produces multiple molecular tokens and uses the text query to combine them for retrieval scoring. Experiments evaluate bidirectional retrieval, transfer to property prediction through linear probing, and recognition of chemical concepts on a semantic benchmark.

Critical strengths
  1. I appreciate that the synthetic descriptions are evaluated across several molecule-text architectures using the same underlying molecules and splits. Their benefits on downstream property prediction and semantic evaluation suggest that the corpus offers useful supervision beyond the proposed model. This gives the dataset a clear potential value for other researchers working on molecular representations and retrieval.

  2. The semantic benchmark provides a useful complement to aggregate retrieval scores. Its categories distinguish global physicochemical properties, local structural features, and relational edits, making it easier to see which kinds of chemical information a model captures. Reporting these differences, including categories where MolTokSyn does not lead, helps clarify the scope of the learned representations.

Critical weaknesses
  1. The property-prediction evaluation does not include a direct comparison with the pretrained CheMeleon encoder before molecule-text training. Since MolTokSyn is initialized from this encoder, the current results leave some uncertainty about the additional transfer benefit provided by the proposed training. Evaluating the original encoder under the same linear-probing protocol would help clarify this point.

  2. The specific benefit of query-dependent aggregation could be clearer. In Tables 12–13, the single-token variant performs similarly to the default eight-token variant, although its aggregation does not depend on the text query under Eq. (4). A comparison with query-independent aggregation of the same molecular tokens would help clarify when text conditioning provides an additional benefit.

Questions for authors
  1. Could the authors provide, or at least discuss, the linear-probing performance of the pretrained CheMeleon encoder before molecule-text training? This would help assess how much the proposed training improves property prediction over its initialization.

  2. The motivation for adapting molecular representations to semantic context also appears in MolEmb [1], while MV-CLAM [2] explores complementary molecular views and finer-grained molecule-text alignment. Could the authors briefly discuss how MolTokSyn relates to these approaches? This would help place the contribution alongside other efforts to represent multiple aspects of a molecule, while making the role of query-conditioned retrieval scoring clearer.

[1] MolEmb: Multimodal Large Language Models Can Be Strong Molecular Embedding Models. ICML 2026 Workshop on FM4LS.

[2] MV-CLAM: Multi-View Molecular Interpretation with Cross-Modal Projection via Language Model. Findings of EMNLP 2025.

4 comments
AuthorsOpenReview ↗

Response to Reviewer reV6 (1/2)

We thank the Reviewer for the valuable feedback. All changes are marked in blue in the revised paper. We address each point below.

W1 and Q1: Comparison with pretrained CheMeleon. We thank the Reviewer for this suggestion. We have added an encoder initialization ablation in Appendix D.6, “Initialization of the Molecular Encoder”, comparing MolTokSyn initialized with pretrained CheMeleon weights against MolTokSyn trained from scratch. Training from scratch yields higher mean scores on all retrieval metrics and lower mean errors on all regression tasks across both pretraining datasets, while classification results are task dependent. Please see our general response A on molecular encoder initialization and downstream transfer for further discussion and the full ablation results.

W2: Query-dependent aggregation. We want to thank the Reviewer for highlighting that the previous comparison did not address enough specific contributions of query-dependent aggregation. That is why, we added a matched control that keeps the same K=8 molecular tokens, keys, values, and scoring architecture, but replaces text-dependent attention with uniform aggregation, which can be seen in section Appendix D.2. As shown in Table 16, query conditioning improves R@1 by 6.96 percentage points for molecule-to-text and 6.61 points for text-to-molecule retrieval, with corresponding MRR gains of 5.49 and 5.05 points. This shows that the gain is not explained simply by using multiple molecular tokens, but also conditioning it with text. To address it further we added a token analysis showing that the learned routing actually changes with the query, which can be seen in Appendix D.2. With the molecular graph and K=8 tokens fixed, 84.75% of molecule-attention-head combinations change their dominant slot across natural descriptions. Under eight controlled semantic-focus queries, the dominant slot differs in 48.1% of pairwise molecule-head comparisons on average (Figure 6). Summing it up, these results support query-dependent reweighting of the molecular representation.

We also added a matched comparison with conventional Global pooling (Table 15), using the final training recipe with SigLIP and physical batch size 512. Replacing Global pooling with a single learned virtual token improves R@1 by 17.61 and 17.89 points in the two retrieval directions. Increasing capacity from K=1 to query-conditioned K=8 provides further gains of 4.32 and 3.53 points, respectively.

Additionally, K=1 is indeed a special case of our model. With only one molecular token, the attention weight in Eq. (4) is always 1, so the molecular representation no longer depends on the text query. Therefore, at the readout level, it behaves as a standard molecular embedding, which makes the K=1 comparison very useful, because the gains for K>1 show the benefit of having multiple tokens that can be weighted differently for different queries.

Additionally, on MolSynText, increasing K from 1 to 8 improves R@1 from 74.41% to 78.76% and from 75.08% to 79.02% in the two retrieval directions, while K=16 gives the best central value on 7 of 8 metrics. On STM, increasing K from 1 to 4 improves R@1 by 3.61 and 3.57 percentage points in the two retrieval directions. Performance then remains largely stable across K=4, 8, and 16, with differences lying within the reported standard deviations.This supports our claim that extra token capacity is useful when the textual supervision is rich enough to exploit it in both MolSynText and STM setup (what was a main goal of this architecture).

AuthorsOpenReview ↗

Response to Reviewer reV6 (2/2)

Q2: Relation to MolEmb and MV-CLAM. We thank the Reviewer for pointing us to these two papers. Both are now discussed in the Related Work section of the revised paper.

MolEmb shares our motivation that one molecule may need different representations in different contexts, but implements it differently. A task instruction (for example, "Encode this molecule for BBB penetration") is placed in the input of a multimodal LLM together with a 2D depiction and the SMILES string, so each new context requires a new LLM forward pass for each molecule. In MolTokSyn, the context is the retrieval query itself: the graph encoder runs once per molecule, its KK tokens are cached, and each text query only reweights them (Section 4.2). MolEmb is concurrent work (arXiv, August 2026) and shares no baselines with ours, so we compare the two conceptually. Its conclusion that context-aware embedding is "primarily a data property of the supervision" is consistent with our controlled finding that the text source matters (Tables 5 and 6).

MV-CLAM aligns 2D-graph and 3D-conformer views of a molecule with text through 12 learnable query tokens per view and a token-level contrastive loss, and passes these tokens to LLaMA2 for captioning and question answering; it also reports retrieval. Our tokens are readouts of a single 2D graph, and the text is not matched to them token by token: it weights them into one pair-specific molecule vector, and the retrieval model itself uses no LLM.

Summary. All three works move beyond a single fixed molecule vector. MolTokSyn makes the retrieval score itself query-conditioned while keeping the molecule tokens cacheable. The matched control in Appendix D.2 (Table 16) isolates this step: with the same K=8K=8 tokens, replacing the query-dependent weights with uniform weights lowers R@1 on MolSynText by 6.96 points (molecule to text) and 6.61 points (text to molecule).


We hope our responses and revisions address the Reviewer’s concerns. We would be happy to clarify any remaining questions during the discussion period.

Thank you for the author's reply. I believe I have no further questions and will maintain my positive assessment and increase my confidence level.

We thank the Reviewer for the positive assessment and constructive feedback. The Reviewer’s suggestions, evaluating the pretrained CheMeleon encoder, comparing with query-independent aggregation of the same tokens, and relating our work to MolEmb and MV-CLAM, have improved the paper. We are grateful for the thoughtful review and for considering our responses.

Comments on the paper

3 comments
AuthorsOpenReview ↗

General Response (1/2)

Dear Area Chair and Reviewers,

We thank the Reviewers for their careful assessment of our work and their thoughtful, constructive suggestions. Their feedback has guided revisions that strengthen the manuscript and improve its clarity. Our detailed responses cover all comments raised by the Reviewers. We discuss two recurring concerns below and address reviewer-specific questions in the individual replies.

A. Response to the Reviewers on molecular encoder initialization

We thank the Reviewers for highlighting the need to distinguish the contribution of molecule-text training from the representations inherited from CheMeleon pretraining. We have added an ablation in Appendix D.6, “Initialization of the Molecular Encoder”, comparing MolTokSyn initialized with pretrained CheMeleon weights against MolTokSyn with a randomly initialized molecular encoder, while keeping the remaining configuration unchanged. We evaluate both variants on retrieval and downstream property prediction.

Retrieval. Random initialization yields higher mean scores than CheMeleon initialization on all eight retrieval metrics across both pretraining datasets. The gains are most pronounced in R@1 and MRR.

Table G1: Cross-modal retrieval results on STM and MolSynText. Retrieval results for MolTokSyn with CheMeleon and random initialization. Higher is better; bold indicates the higher mean within each retrieval direction and corpus.

Retrieval directionCorpusEncoder initializationR@1R@5R@10MRR
Molecule -> textSTMCheMeleon53.36 ±\pm 1.2968.44 ±\pm 1.2772.66 ±\pm 1.2660.27 ±\pm 1.25
Molecule -> textSTMFrom scratch55.11 ±\pm 1.1468.81 ±\pm 1.4772.85 ±\pm 1.3061.54 ±\pm 1.23
Molecule -> textMolSynTextCheMeleon78.76 ±\pm 1.6196.45 ±\pm 0.6398.56 ±\pm 0.2286.57 ±\pm 1.20
Molecule -> textMolSynTextFrom scratch82.57 ±\pm 1.9397.23 ±\pm 0.8198.75 ±\pm 0.3989.02 ±\pm 1.43
Text -> moleculeSTMCheMeleon52.74 ±\pm 1.1067.94 ±\pm 1.3872.31 ±\pm 1.1259.75 ±\pm 1.17
Text -> moleculeSTMFrom scratch54.78 ±\pm 1.1268.76 ±\pm 1.3972.45 ±\pm 1.2761.28 ±\pm 1.21
Text -> moleculeMolSynTextCheMeleon79.02 ±\pm 1.6396.01 ±\pm 0.5998.08 ±\pm 0.1686.58 ±\pm 1.20
Text -> moleculeMolSynTextFrom scratch82.79 ±\pm 2.0496.73 ±\pm 0.7898.38 ±\pm 0.3288.97 ±\pm 1.47

Linear probing. The effect of initialization is task dependent. CheMeleon initialization yields higher mean AUROC on four of seven classification tasks for each corpus, while random initialization performs better on the remaining three. For regression, random initialization yields lower mean RMSE on all three tasks across both pretraining datasets.

Table G2: Linear probing performance of frozen MolTokSyn representations on molecular classification tasks, comparing CheMeleon and random encoder initialization. Higher scores are better; bold indicates the better mean within each task and corpus.

CorpusEncoder initializationHIVToxCastTox21BBBPBACEClinToxSIDER
STMCheMeleon74.03 ±\pm 0.1062.98 ±\pm 0.8372.91 ±\pm 0.9064.11 ±\pm 0.4873.11 ±\pm 1.2882.29 ±\pm 2.4361.70 ±\pm 0.54
STMFrom scratch71.36 ±\pm 2.5963.82 ±\pm 0.3673.85 ±\pm 0.3964.06 ±\pm 1.7977.88 ±\pm 2.9778.59 ±\pm 2.1461.25 ±\pm 0.82
MolSynTextCheMeleon76.78 ±\pm 0.7764.81 ±\pm 0.5274.51 ±\pm 0.0666.63 ±\pm 1.2577.78 ±\pm 1.2586.59 ±\pm 2.2261.22 ±\pm 0.18
MolSynTextFrom scratch71.78 ±\pm 2.9265.42 ±\pm 0.0574.44 ±\pm 0.5466.21 ±\pm 1.8777.93 ±\pm 1.8883.95 ±\pm 1.8661.32 ±\pm 0.34

Table G3: Linear probing performance of frozen MolTokSyn representations on molecular regression tasks, comparing CheMeleon and random encoder initialization. Lower RMSE is better; bold indicates the better mean within each task and corpus.

CorpusEncoder initializationESOLFreeSolvLipophilicity
STMCheMeleon1.201 ±\pm 0.0833.118 ±\pm 0.1180.920 ±\pm 0.028
STMFrom scratch1.042 ±\pm 0.0562.661 ±\pm 0.1990.896 ±\pm 0.016
MolSynTextCheMeleon0.940 ±\pm 0.0382.611 ±\pm 0.1050.879 ±\pm 0.013
MolSynTextFrom scratch0.931 ±\pm 0.0182.330 ±\pm 0.1150.848 ±\pm 0.011

These findings show that MolTokSyn’s strong performance does not require CheMeleon pretraining: training the molecular encoder from scratch improves all reported mean retrieval scores and reduces mean errors on every regression task tested.

AuthorsOpenReview ↗

General Response (2/2)

B. Response to the Reviewers on caption validation: independent judges and expert review

We thank the Reviewers for highlighting the risk of relying on the same model for generation and quality control. We extended the validation pipeline with two independent model families, DeepSeek-R1-14B and Gemma3-12B, both with medium reasoning, which evaluate each caption after the initial Qwen check. A caption is kept only if all judges accept it; otherwise it is regenerated and evaluated again. Appendix B now reports the rejection and regeneration process in detail. The Qwen judge rejects 5.68% of first-pass candidates; 0.96% of the dataset needs a third attempt and 0.23% further attempts. Of the captions that pass Qwen, DeepSeek rejects 5.28% and Gemma 1.12% (6.18% rejected by at least one); 2.15% of the dataset proceeds to another attempt, 0.61% to a third, and 0.45% requires later attempts.

To measure caption quality directly, four chemists with a master’s degree in chemistry rated captions sampled at random from three groups, blind to the judges’ decisions (207 ratings of 147 descriptions; Appendix B.4, Figure 5). Among captions accepted by all three judges, which are the ones kept in the dataset, 18% (95% CI 11-27%; n = 69) were rated incorrect, i.e. containing at least one clear error, and 66% fully correct. Most flagged errors concern misnamed functional groups or substitution patterns rather than descriptor values, which are copied from the RDKit card.

Table G4: Expert ratings of MolSynText captions that passed the Qwen judge, grouped by the DeepSeek and Gemma verdicts.

Independent judges' verdictnRated incorrectCorrect or minor (lenient)Correct (strict)
Accepted by both (kept)6918%82% (73–89%)66% (55–76%)
Rejected by DeepSeek3417%83% (71–94%)61% (46–76%)
Rejected by Gemma4320%80% (67–90%)69% (56–82%)

We note two limitations. Agreement between raters was low (Fleiss’ κ = 0.14), so the human reference is itself noisy. And within captions that had already passed the Qwen judge, the DeepSeek and Gemma judges did not separate accurate from inaccurate descriptions: rejected captions were rated as accurate as accepted ones. We report this in the revised paper rather than claiming the additional judges improve caption quality.

Experiments in progress

We also note that several Reviewers indicated issues with inconsistent setup in ablations. We are rerunning them on all three splits now. We hope the results will arrive before the rebuttal period ends. For now, we provide new results on one split (seed 0) in the revision.

AuthorsOpenReview ↗

Final Response and Summary of Changes

Dear Area Chair and Reviewers,

We thank the Area Chair for handling our submission and the Reviewers for their time and constructive feedback. We have just uploaded the final revised version, in which all results are complete. The changes are marked in blue. The main revisions are:

  • Caption validation: two independent judges (DeepSeek, Gemma), regeneration statistics, and a blinded review by four chemists (Appendix B).
  • Text-supervision controls: one caption per molecule and a version with names, SMILES and numbers removed (Appendix D.1); transfer to human-written ChEBI-20 descriptions (Appendix D.7).
  • Architecture and training ablations: rerun under the final training setup across all three scaffold splits (Appendix D.2–D.5; Tables 15–21): Global vs K=1 vs K=8, a uniform-weight control, number of tokens, physical batch size at fixed effective batch size 512, and SigLIP vs InfoNCE, alongside additional analysis of query-dependent routing.
  • Encoder initialization: random vs CheMeleon initialization, MolCA with the same CheMeleon encoder, and frozen CheMeleon as a linear-probe reference (Appendix D.6).
  • Semantic Benchmark: an audit and re-run on the corrected items (Appendix C.4), with the limitations stated in Appendix E.
  • Related work: extended the discussion of text-conditioned molecular representations and added set-based graph pooling.

The ablations that our earlier replies reported on one split are now averaged over three. The conclusions are unchanged.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.