acceptodds
Under review as a conference paper at ICLR 2027

GRASP: Graph Representation Learning with Assay Supervision for Molecular Properties

Abstract

Large unlabeled molecular datasets provide an opportunity to learn reusable representations before expensive property measurements are available. Yet molecular pretraining is commonly followed directly by property-specific fine-tuning, leaving open how large but sparsely labelled bioactivity collections can provide an intermediate source of supervision. We introduce GRASP, a 93.5M-parameter molecular graph Transformer trained through a two-stage representation-learning framework: large-scale self-supervised pretraining followed by sparse multi-assay adaptation. We use replaced-token detection (RTD), where a compact generator produces contextual atom-token replacements and a graph-relative discriminator learns from every valid input position, before the encoder is adapted using sparsely observed labels across 642 ChEMBL assays. Across 23 cluster-held-out OpenADMET endpoints, full fine-tuning achieves a mean MAE of 0.3742, compared with 0.4075 for ECFP4-LightGBM and 0.4395 for Chemprop D-MPNN. Full fine-tuning also improves over frozen features, with LoRA falling between the two. In an OpenADMET comparison of replaced-token and masked-token pretraining, the mean difference is small with frozen features and nearly zero after fine-tuning. Together, these results support a staged approach to molecular representation learning: large-scale self-supervision, followed by sparse assay supervision and downstream fine-tuning. This allows abundant unlabeled molecules and limited experimental measurements to contribute at different stages of training.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Protein Ligand, Drug Discovery, Binding Affinity

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_q2P2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GRASP, a 93.5M-parameter graph Transformer for molecular property prediction. It combines large-scale replaced-token detection (RTD) pretraining on ZINC20 with subsequent multi-assay bioactivity adaptation on ChEMBL. GRASP achieves competitive performance on 23 OpenADMET endpoints and 22 TDC ADMET tasks. The paper also evaluates RTD against MLM and examines the contribution of ChEMBL adaptation.

Critical strengths

1. Careful and Transparent Evaluation

The paper evaluates GRASP across diverse ADMET endpoints using strong baselines and consistent data splits. The authors also provide statistical analyses, structural-overlap checks, and a candid discussion of limitations.

2. Consistent Benefit of ChEMBL Adaptation

The ChEMBL adaptation stage reduces mean MAE from 0.407 to 0.374 and improves performance on 21 of 23 OpenADMET endpoints and 19 of 22 TDC tasks, demonstrating the practical value of the staged training approach.

Critical weaknesses

1. Limited Methodological Novelty

The main components are drawn from prior work, including graph-relative attention, ELECTRA-style replaced-token detection, and multi-task bioactivity supervision. Although their integration is useful, the methodological novelty remains limited.

2. Modest Gains over the Strongest Baseline

GRASP improves mean MAE over CheMeleon by only 2.4%, with better results on 12 of 23 OpenADMET endpoints. CheMeleon is also absent from the paired statistical comparisons in Table A3. Given GRASP's substantially larger model size, the practical significance of this improvement remains unclear.

3. Confounded Training Comparisons

The ChEMBL adaptation stage introduces assay labels, additional training, and exposure to new molecules simultaneously, making it difficult to isolate the benefit of assay supervision. The RTD–MLM comparison also changes both the training objective and prediction targets, and its improvement on OpenADMET is not statistically significant after correction (p = 0.097).

Questions for authors

Please address the concerns raised in the Critical Weaknesses section, particularly regarding the improvement over CheMeleon and the independent contributions of ChEMBL adaptation and RTD pretraining.

Additionally, since radius-0 atom tokens may contain information that can be inferred from graph topology, could the authors clarify how much of the RTD task can be solved using structural cues alone, and whether topology-constrained token replacement would affect the results?

1 comment

Thank you for recognizing the evaluation's transparency and the value of ChEMBL adaptation. We have completed the requested continuation control, added the paired CheMeleon comparison and completed the closer radius-0 objective comparison.

1. Methodological contribution

Revised GRASP changes the molecular representation and structural task jointly: seven explicit atom attributes, separate generator heads and whole-bundle replacement, together with explicit bond/stereochemical encoder inputs and removal of absolute atom-index embeddings. The equations and cumulative ablations are given in response to reviewers KZ5q and pwYW . We credit the established attention, RTD and two-stage training foundations; our contribution is their specific molecular formulation and controlled evaluation.

The reference/A1/A2/revised-GRASP means are 0.3702/0.3686/0.3662/0.3589. Separate reference-architecture GDES/ES/NES means are 0.3702/0.3638/0.3655.

2. Accuracy relative to CheMeleon and practical significance

The requested paired comparison is complete and the retained predictions have been audited. Revised GRASP reaches 0.3589 versus 0.3834 mean MAE, a 6.39% reduction(a significant improvement in this context over the SOTA ), and achieving lower error on 87% of endpoints (20/23) across three distinct organizations (ASAP, Biogen, and ExpansionRx) proves the gain is systemic rather than driven by one or two outlier datasets.

The paired difference is−0.0245 (pointwise 95% CI [−0.0379,−0.0119]); exact Wilcoxon raw p=0.000552 and expanded Holm-25 p=0.00883. The sign-test sensitivity also remains significant (p=0.01025).

This strengthens the accuracy evidence beyond the submitted comparison. Model size remains relevant: the encoder has approximately 93M parameters, and we distinguish accuracy from parameter or compute efficiency. We also completed revised-model FT, LoRA and frozen-mixing evaluations, with mean MAEs 0.3589, 0.3746 and 0.3821, respectively. LoRA uses 986,374 trainable parameters; the FT–LoRA difference is not statistically resolved after correction.

For a comparison between frozen systems, revised GRASP achieves 0.3821 versus 0.4123 for Mol-JEPA, with lower error on 20/23 endpoints (Holm-4 Wilcoxon p=0.000511). GRASP uses learned layer mixing/attention pooling while Mol-JEPA uses CLS features, so this result compares the evaluated frozen systems rather than identical readouts.

3. Contributions of ChEMBL adaptation and RTD

The continuation control matches the structural starting checkpoint,511,897 ChEMBL structures, 50k updates, effective batch 512 and learning-rate schedule. Label-free continuation loads no assay labels. Mean MAEs are 0.3915 before adaptation, 0.3871 after continuation and 0.3589 after supervised adaptation. Supervised adaptation improves over continuation by 7.30% on 19/23 endpoints, with Δ−0.0283, 95% CI [−0.0532,−0.0122] and Holm-3 p=0.0000987. This directly answers whether additional label-free exposure accounts for the supervised-stage improvement at the matched update budget.

The completed original-architecture radius-0 RTD/MLM control gives FT means 0.3658/0.3663, with no detected paired difference (Holm-2 p=0.3146). Frozen MLM is better (0.3797 versus 0.3989;18/23 endpoints; Holm-2 p=0.0271, expanded Holm-25 p=0.1353). We therefore separate the evidence for the complete revised design from claims about objective superiority. Nominal exposure and adaptation budgets are matched; FLOPs and distributed numerical trajectories are not.

4. Structural cues and topology-constrained replacement

We completed a diagnostic on 10,000 validation-partition molecules and 217,489 atoms, using the trained revised structural checkpoint. A deterministic detector checks whether the realized atom description's degree or ring membership contradicts bond adjacency, without seeing the original attributes or corruption mask. These rules identify 14.75% of realized replacements, with a molecule-bootstrap 95% CI of 12.99–16.63%.

After excluding atoms with those contradictions, the discriminator retains AUROC 0.9364. Under replacements constrained to degree/ring consistency, its AUROC is 0.9444, versus 0.9331 for a positive-rate-matched unconstrained diagnostic. These observations show that discrimination persists beyond the two tested inconsistencies; they quantify shortcut availability rather than establish the model's causal reliance or exclude other structural cues.

We respectfully ask that the stronger baseline evidence, completed attribution controls and direct response to the structural-cue question be considered for upgrading the rating of revised manuscript.

Reviewer_KZ5qOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This submission introduces GRASP, a graph Transformer for molecular property prediction. GRASP is trained in two stages: replaced-token detection on ZINC20 molecular presentations, followed by supervised adaptation on 642 sparsely observed ChEMBL assays. The resulting encoder is evaluated using full fine-tuning, LoRA, and frozen layer mixing.

Critical strengths
  • The motivation is clear and practically relevant.
  • The evaluation includes competitive classical and neural baselines under common data splits.
  • ChEMBL adaptation improves most OpenADMET endpoints.
  • The structural-overlap analysis addresses potential molecule-level leakage.
Critical weaknesses
  • The novelty relative to MolE[1] is unclear. MolE already uses structural pretraining followed by supervised biological adaptation. The main difference appears to be RTD, but no protocol-matched comparison with MolE is provided.

  • RTD detects replaced radius-0 tokens, while MLM predicts radius-2 atom environments. Both the objective and prediction target are changed, so the experiment does not isolate the effect of RTD. The OpenADMET difference is also not significant after correction.

  • GRASP-S1 and GRASP differ in additional training, exposure to ChEMBL molecules, and assay labels. A control using continued self-supervised training on the same ChEMBL molecules is needed.

[1] MolE: a foundation model for molecular graphs using disentangled attention.

Questions for authors
  1. What is the main methodological contribution beyond MolE?
  2. Does ChEMBL adaptation remain beneficial when assay labels are removed?
3 comments

1. Main methodological contribution beyond MolE

Revised GRASP changes the atom representation, generator prediction space, replacement operator and chemical attention inputs.

Each atom is represented by seven attributes—element, degree, hydrogen count, formal charge, aromaticity, ring membership and hybridization—with embedding

$

ei=17∑k=17E(k)[xi(k)]+Estereo[si].\mathbf e_i=\frac 1{\sqrt 7}\sum_{k=1}^{7}E^{(k)}[x_i^{(k)}]+E^{\mathrm{stereo}}[s_i].

$

The second term encodes atom stereochemistry. No learned absolute atom-index embedding is added. For attention head h, graph-relative attention additionally receives learned bond-type, aromaticity and bond-stereochemistry biases:

$

aij(h)=(qi(h))⊤kj(h)+(qi(h))⊤kr(i,j)rel,(h)2dh+bhtype(tij)+bharom(uij)+bhstereo(vij).a_{ij}^{(h)}= \frac{(\mathbf q_i^{(h)})^\top\mathbf k_j^{(h)}+(\mathbf q_i^{(h)})^\top\mathbf k_{r(i,j)}^{\mathrm{rel},(h)}}{\sqrt{2d_h}} +b_h^{\mathrm{type}}(t_{ij})+b_h^{\mathrm{arom}}(u_{ij})+b_h^{\mathrm{stereo}}(v_{ij}).

$

Here r(i,j) encodes graph distance. We evaluate these chemical-input changes as part of the configuration rather than attribute an accuracy gain to the number of attention terms alone.

Whereas MolE's principal structural objective predicts radius-2 environments, our generator predicts the seven attributes through separate heads sharing molecular context C:

$

pG(xi∣C)=∏k=17pG(k)(xi(k)∣C),LG=−1∣M∣∑i∈M∑k=17log⁡pG(k)(xi(k)∣C).p_G(\mathbf x_i\mid C)=\prod_{k=1}^{7}p_G^{(k)}(x_i^{(k)}\mid C),\qquad \mathcal L_G=-\frac 1{|M|}\sum_{i\in M}\sum_{k=1}^{7}\log p_G^{(k)}(x_i^{(k)}\mid C).

$

The loss shown assumes supported targets; unsupported field targets are omitted from that field's mean. All seven attributes at selected atoms are masked together. Head-wise argmax predictions form a complete replacement bundle, with one atom-level label:

$

yi=1[∃k:x~i(k)≠xi(k)],L=LG+50LD.y_i=\mathbb 1[\exists k:\widetilde x_i^{(k)}\ne x_i^{(k)}],\qquad \mathcal L=\mathcal L_G+50\mathcal L_D.

$

The discriminator thus learns to detect changes to an explicit chemical description using graph context. Graph connectivity and chemical side inputs remain fixed during corruption. This formulation does not guarantee chemical validity of generated bundles; it defines a representation-learning task.

We also evaluated the cumulative architecture changes under matched approximately 1.538B structural presentations:

ConfigurationMean MAE ↓
Retrained reference architecture with GDES0.3702
A1: remove absolute atom-index embeddings0.3686
A2: add bond/stereochemical inputs to A10.3662
Revised GRASP: explicit attributes and bundle generation/replacement0.3589

Separate reference-architecture GDES/ES/NES controls give 0.3702/0.3638/0.3655. The complete revised configuration has the lowest observed mean.

2. Protocol-matched comparison with MolE

The full pretrained MolE model was not available to us for protocol-matched evaluation. MolE has public code, but the pretrained-model download section of its official repository remains marked “TODO.”

Our completed released-checkpoint comparison uses CheMeleon under identical OpenADMET partitions and full fine-tuning. Revised GRASP achieves 0.3589 versus 0.3834 mean MAE, with lower error on 20/23 endpoints. The 6.39% improvement survives the expanded Holm-25 sensitivity (Wilcoxon p=0.00883). We will release our revised and control checkpoints and evaluation code to enable further protocol-matched comparisons.

3. Does ChEMBL adaptation remain beneficial when assay labels are removed?

Mean MAE is 0.3915 without adaptation, 0.3871 after label-free continuation and 0.3589 after supervised adaptation. Continuation gives a small 1.13% improvement and lower error on 18/23 endpoints (Holm-3 Wilcoxon p=0.0384); its pointwise mean-difference interval spans zero. Supervised adaptation improves over continuation by 7.30% on 19/23 endpoints. Thus, continued ChEMBL exposure provides a small gain.

4. RTD–MLM objective and target-radius confounding

Your feedback directly motivated the completed radius-0 comparison. FT MAEs are 0.3658 for RTD and 0.3663 for MLM (Δ−0.00047;95% CI [−0.01405, 0.00945]; Holm-2 p=0.3146). Frozen MLM is better (0.3797 versus 0.3989;18/23 endpoints; Holm-2 p=0.0271, broader Holm-25 p=0.1353).

We now state the conclusion at the level supported by this control: no FT winner is statistically established, and the frozen comparison favors MLM within the dedicated two-contrast family. This does not establish equivalence. The new result strengthens the attribution analysis by holding the prediction radius fixed; differing distributed execution and one pretrained checkpoint per arm remain part of its scope. We will release the MLM checkpoint and evaluation configuration as a reusable outcome of this reviewer-motivated experiment.

We respectfully ask that the completed controls and revised accuracy evidence be considered for upgrading rating for the article.

Thank you for your response and the supplementary experiments. However, I still have concerns. The MolE paper has already proposed a similar two-stage framework, and the superior performance reported in the current submission relies heavily on a substantial number of new experiments added after the initial submission. While these supplementary experiments are informative, they also highlight questions regarding the novelty of the proposed method. Consequently, I consider this paper to be on the borderline, if other reviewers and the AC are inclined to accept it, I will not object.

We thank the reviewer for this thoughtful assessment. We agree that MolE established an important precedent for combining structural molecular pretraining with subsequent bioactivity supervision, and we do not intend to claim the general two-stage training paradigm as an independent novelty of our work.

Our contribution lies instead in the specific representation-learning strategy and training objectives employed in GRASP. In particular, GRASP uses replaced-token detection (RTD), rather than masked-token reconstruction, to learn molecular representations, alongside graph-relative attention and gradient-disentangled embedding sharing. The distinction is therefore not simply the sequencing of two training stages, but how molecular structural representations are learned and subsequently adapted using sparse, multi-assay experimental supervision.

Regarding the supplementary experiments, we acknowledge that the revised evaluation provides substantially stronger evidence for the method's effectiveness than the original submission. These experiments were introduced to address the reviewers' concerns, isolate the contributions of the training stages, and strengthen the statistical reliability of our comparisons. We believe they reinforce the central methodological contribution rather than introduce a fundamentally different method.

We appreciate the reviewer's recognition of these additional analyses and their willingness to defer to the overall assessment of the other reviewers and the Area Chair.

Reviewer_pwYWOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The authors present GRASP, a graph transformer pre-trained with replaced token detection (RTD) on ZINC20 and then further pre-trained on ChEMBL dataset. Evaluation covers OpenADMET and TDC tasks. Ablation studies compared the RTD with masked-token prediction and the effect of ChEMBL adaptation. GRASP shows modest improvements over Chameleon, with lower MAE on 12 of 23 OpenADMET tasks and better reported scores on 13 of 22 TDC tasks.

Critical strengths

1. Broad empirical evaluation

The paper evaluates GRASP on OpenADMET and TDC benchmarks, compares full fine-tuning, LoRA, and frozen representations, and includes ablations examining RTD pre-training and ChEMBL adaptation.

Critical weaknesses

1. Limited methodological novelty

MolE [1] already combines large-scale structural pre-training with supervised bioactivity pre-training on ChEMBL. GRASP follows a similar framework, with RTD replacing masked-token prediction as the main methodological difference. While Table 3 shows comparisons between GRASP-RTD and GRASP-MLM, the gains are marginal. GRASP also achieves performance similar to Chameleon on OpenADMET, reducing mean MAE by 2.4% and outperforming on 12 of 23 endpoints. Taken together, the methodological changes and reported results provide limited evidence of a substantial contribution beyond existing approaches.

[1] MolE: a foundation model for molecular graphs using disentangled attention. Nature Communications. (2024)

Questions for authors

RTD improves performance on LogD and solubility tasks, while MLM outperforms on several clearance and binding tasks. How do the authors interpret this task-dependent pattern, and what does it suggest about the molecular information learned by the two different objectives?

3 comments

Thank you for recognizing the broad empirical evaluation. Your comments motivated both a clearer molecular formulation and the completed objective control.

1. “The main methodological difference from MolE appears to be replacing MLM with RTD.”

We credit MolE for structural pretraining followed by bioactivity adaptation. Revised GRASP changes the atom representation, generator prediction space, replacement operator and chemical attention inputs.

Each atom is represented by seven attributes—element, degree, hydrogen count, formal charge, aromaticity, ring membership and hybridization—with embedding

$

ei=17∑k=17E(k)[xi(k)]+Estereo[si].\mathbf e_i=\frac 1{\sqrt 7}\sum_{k=1}^{7}E^{(k)}[x_i^{(k)}]+E^{\mathrm{stereo}}[s_i].

$

The second term encodes atom stereochemistry. No learned absolute atom-index embedding is added. For attention head h, graph-relative attention additionally receives learned bond-type, aromaticity and bond-stereochemistry biases:

$

aij(h)=(qi(h))⊤kj(h)+(qi(h))⊤kr(i,j)rel,(h)2dh+bhtype(tij)+bharom(uij)+bhstereo(vij).a_{ij}^{(h)}= \frac{(\mathbf q_i^{(h)})^\top\mathbf k_j^{(h)}+(\mathbf q_i^{(h)})^\top\mathbf k_{r(i,j)}^{\mathrm{rel},(h)}}{\sqrt{2d_h}} +b_h^{\mathrm{type}}(t_{ij})+b_h^{\mathrm{arom}}(u_{ij})+b_h^{\mathrm{stereo}}(v_{ij}).

$

Here r(i,j) encodes graph distance. We evaluate these chemical-input changes as part of the configuration rather than attribute an accuracy gain to the number of attention terms alone.

Whereas MolE's principal structural objective predicts radius-2 environments, our generator predicts the seven attributes through separate heads sharing molecular context C:

$

pG(xi∣C)=∏k=17pG(k)(xi(k)∣C),LG=−1∣M∣∑i∈M∑k=17log⁡pG(k)(xi(k)∣C).p_G(\mathbf x_i\mid C)=\prod_{k=1}^{7}p_G^{(k)}(x_i^{(k)}\mid C),\qquad \mathcal L_G=-\frac 1{|M|}\sum_{i\in M}\sum_{k=1}^{7}\log p_G^{(k)}(x_i^{(k)}\mid C).

$

The loss shown assumes supported targets; unsupported field targets are omitted from that field's mean. All seven attributes at selected atoms are masked together. Head-wise argmax predictions form a complete replacement bundle, with one atom-level label:

$

yi=1[∃k:x~i(k)≠xi(k)],L=LG+50LD.y_i=\mathbb 1[\exists k:\widetilde x_i^{(k)}\ne x_i^{(k)}],\qquad \mathcal L=\mathcal L_G+50\mathcal L_D.

$

The discriminator thus learns to detect changes to an explicit chemical description using graph context. Graph connectivity and chemical side inputs remain fixed during corruption. This formulation does not guarantee chemical validity of generated bundles; it defines a representation-learning task.

We also evaluated the cumulative architecture changes under matched approximately 1.538B structural presentations:

ConfigurationMean MAE ↓
Retrained reference architecture with GDES0.3702
A1: remove absolute atom-index embeddings0.3686
A2: add bond/stereochemical inputs to A10.3662
Revised GRASP: explicit attributes and bundle generation/replacement0.3589

Separate reference-architecture GDES/ES/NES controls give 0.3702/0.3638/0.3655. The complete revised configuration has the lowest observed mean.

2. “The improvement over CheMeleon is modest.”

The review accurately describes the submitted 2.4%/12-of-23 result. Revised GRASP achieves 6.39% lower mean MAE (a significant improvement in this context over the SOTA ) and achieving lower error on 87% of endpoints (20/23) across three distinct organizations (ASAP, Biogen, and ExpansionRx) proves the gain is systemic rather than driven by one or two outlier datasets. The paired advantage remains significant under the expanded Holm-25 sensitivity (Wilcoxon p=0.00883).

3. “The RTD–MLM gains are marginal.”

The completed radius-0 control gives FT MAEs 0.3658 for RTD and 0.3663 for MLM: a 0.13% aggregate difference, with RTD lower on 10/23 endpoints and no detected paired difference (Holm-2 p=0.3146). Frozen MLM performs better, as reported above. We therefore base the contribution on the evaluated molecular design and adaptation recipe, rather than a general RTD-over-MLM claim.

4. Why does RTD help LogD/solubility while MLM helps some clearance/binding endpoints?

Our completed matched radius-0 comparison shows that the pattern is not consistent across datasets. Under full fine-tuning, RTD improves ASAP LogD (MAE 0.4781 versus 0.6001), whereas MLM improves ExpansionRx LogD (0.2549 versus 0.2710). Solubility is similarly mixed: RTD performs slightly better on Biogen, while MLM performs better on ASAP and ExpansionRx. Clearance and binding endpoints also show both directions.

We therefore interpret these as endpoint-dependent performance differences rather than evidence that the objectives specialize in particular property classes. Across all 23 endpoints, FT means are 0.3658 for RTD and 0.3663 for MLM, with no detected paired difference (Holm-adjusted Wilcoxon p=0.3146).

We hope the completed controls and stronger baseline evidence address the concerns about the scope and significance of the contribution, and respectfully ask that they increase the rating for this paper.

Thank you for the response. However my concerns still remain unresolved.

The improved results come from substantial changes to the representation, attention inputs, and replacement procedure compared to the previously submitted version of model. My concern was about the novelty of the submitted paper's method, not a request for redesign.

More importantly, the matched RTD-MLM comparison shows almost no difference. The submitted paper claims the contribution of RTD at first, and it has the section titled "4.2. RTD PROVIDES GAINS OVER MLM FOR PRETRAINING ". Reframing the contribution around the overall training recipe does not resolve this concern because structural pre-training followed by ChEMBL adaptation was already used in MolE.

Based on the submitted manuscript, originally claimed novelties, and rebuttal, I maintain my original score.

Thank you for the further engagement. We address the two remaining concerns directly.

1. Novelty of the submitted method

We credit MolE for the two-stage framework of structural pretraining followed by bioactivity adaptation. The submitted GRASP differs from MolE in its objective and formulation. MolE's principal structural objective predicts radius-2 environments. GRASP instead trains a generator–discriminator on molecular graphs, using atom-token replacement and gradient-disentangled embedding sharing (GDES). This is the formulation we submitted and evaluated, and it is the first contribution.

The redesign presented in the rebuttal is a second, separate contribution, and we have kept it apart from the submitted configuration throughout. It consists of the following:

  • explicit seven-attribute atom representations;
  • a factorized generator;
  • whole-atom-bundle replacement;
  • bond-type, aromaticity and stereochemistry attention inputs.

We introduced it because the review questioned whether the work provides "substantial contribution beyond existing approaches", and it is direct evidence on that question. The revised model lowers mean MAE relative to CheMeleon by 6.39%, with gains on 20/23 endpoints across ASAP, Biogen and ExpansionRx.

  • Downstream utility: Releasing the encoder and control checkpoints will let others build property-prediction models on these representations and integrate them with protein or other modalities—for example, in binding-affinity models. Our verified ADMET results support their use for property prediction; multimodal applications are a prospective extension.

2. RTD versus MLM

We agree with the reviewer on this evidence. We also note that the per-endpoint pattern is mixed. Section 4.2 will be retitled and rewritten. We remove the general RTD-superiority claim and unsupported task-specialization interpretations.

This correction does not change what the paper contributes. The contribution is the evaluated molecular-graph design and adaptation recipe, and the gains are not attributed to RTD alone.

3. "Structural pretraining followed by ChEMBL adaptation was already used in MolE"

The two-stage ordering is shared with MolE. What GRASP adds is the first controlled attribution of where the adaptation benefit comes from. Within the revised configuration, structural pretraining, label-free ChEMBL continuation and supervised adaptation reach mean MAEs of 0.3915, 0.3871 and 0.3589. Supervised adaptation outperforms continuation by 7.30% on 19/23 endpoints. The benefit therefore comes from the supervised signal itself, not from additional label-free exposure.

4. Value to the community

The MolE checkpoint is not publicly available, so the community has no open pretrained model of this kind to build on or compare against. We will release all GRASP models, including the ablation checkpoints (RTD and MLM, the GDES, ES and NES controls, and the ChEMBL continuation and adaptation branches), together with code. Every comparison in the paper can then be independently verified and extended.

And our evaluation is more robust and one more challenging splits. We also deliberately used the hardest splits so the results reflect real generalization.

Requested assessment

We have corrected the one claim the evidence does not support and completed the controls the review asked for. We ask that the final assessment weigh these as distinct contributions: the submitted formulation, the corrected objective claim, the attribution controls, the revision gains and the open release. The redesign should not be treated as part of the original submission, and its gains should not be attributed to RTD alone. On that basis, we respectfully ask the reviewer to reconsider the rating.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Shared response to all reviewers/AC/SAC

We thank the reviewers for recognizing the practical motivation and breadth of our evaluation. Their feedback motivated additional controls, a revised molecular pretraining configuration, and prediction-level statistical audits. We address three central questions: the methodological contribution, improvements over CheMeleon, and the contributions of individual training stages.

Revised GRASP configuration. Revised GRASP represents each atom using seven explicit chemical attributes, predicts them through a factorized generator, and replaces complete atom-attribute bundles. Its encoder incorporates bond and stereochemical information while removing learned absolute atom-index embeddings. This formulation builds on MolE's two-stage framework and established RTD and graph-attention methods.

Accuracy and statistical evidence. Under identical reconstructed cluster-held-out partitions and full fine-tuning, revised GRASP achieves a mean MAE of 0.3589 versus 0.3834 for CheMeleon, a 6.39% reduction, with lower error on 20/23 endpoints. The paired difference is −0.0245 (pointwise 95% endpoint-bootstrap CI: [−0.0379, −0.0119]; exact Wilcoxon p=0.000552; exploratory Holm-25 adjusted p=0.00883).

GRASP also outperforms the rerun classical and neural baselines:

ComparatorMean MAEReductionEndpoints won
CheMeleon FT0.38346.39%20/23
ECFP4–LightGBM0.407511.95%21/23
RDKit–LightGBM0.411912.87%21/23
RDKit random forest0.430616.67%22/23
Chemprop D-MPNN0.439518.36%23/23
ECFP4 random forest0.452520.70%22/23

Contribution of supervised adaptation. We completed the requested matched ChEMBL continuation control. Both branches start from the same structural checkpoint and use 511,897 ChEMBL structures, 50,000 updates, effective batch 512, and identical learning-rate schedules. The label-free branch does not access assay labels.

Training stageMean MAE ↓
Structural pretraining0.3915
Label-free continuation0.3871
Supervised adaptation0.3589

Supervised adaptation improves MAE by 7.30% over label-free continuation, winning on 19/23 endpoints (ΔMAE −0.0283; 95% CI [−0.0532, −0.0122]; Holm-3 Wilcoxon p=0.0000987). This isolates the benefit of supervised adaptation beyond additional label-free exposure.

Reproducibility and scope. We will release revised GRASP and radius-0 MLM checkpoints, configurations, and evaluation code with full public access after the review process is complete(at this time it is private due to anonymity requirement of conference) .

Hence we request Reviewers/ACs/SACs for to consider these points while making their decision to upgrade their final ratings.

Area_Chair_JpnqOpenReview ↗

Please check the authors' rebuttal ASAP.

Dear reviewers,

As the discussion period will end very soon, please check the authors' rebuttal and see if your concerns have been addressed. Thank you.

-AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.