acceptodds
Under review as a conference paper at ICLR 2027

MolRIME: Reciprocal Innovation-driven Multimodal Exchange for Molecular Representation Learning

Abstract

Learning transferable molecular representations benefits from integrating complementary information from 2D chemical topology and 3D molecular geometry. Existing molecular pretraining methods exploit this complementarity through cross-modal alignment, knowledge transfer, or representation fusion. However, these approaches typically operate on modality representations as a whole, without distinguishing between information already captured by the other modality and information that has not yet been captured. This highlights a central challenge in cross-modal interaction: determining what information should actually be exchanged between modalities. To address this challenge, we propose Reciprocal Innovation-Driven Multimodal Exchange (RIME), which makes cross-modal communication conditional on what the receiving modality can already explain. RIME uses reciprocal cross-modal prediction to identify prediction residuals as cross-modal innovations, which are dynamically re-estimated and exchanged as the modality representations evolve. Building on RIME, we develop MolRIME, a molecular pretraining framework that realizes innovation-driven 2D–3D exchange through symmetry-preserving updates and combines it with geometric denoising and topology reconstruction. We evaluate MolRIME across molecular dynamics, quantum-chemical, physicochemical, and biological tasks. Controlled frozen-encoder comparisons further demonstrate the benefit of exchanging cross-modal innovations over complete modality representations, with MolRIME reducing test MAE by 6.1-8.5% on four of six evaluated QM9 properties using only 10% of the training labels.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
molecular representation learning, molecular pretraining, multimodal learning, cross-modal information exchange

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_rWU9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes RIME, a cross-modal exchange mechanism where, instead of communicating complete source representations, two modalities reciprocally predict each other's representations and exchange only the prediction residuals ("innovations") — i.e., the information the receiver cannot yet explain. Innovations are re-estimated at every interaction block. Building on RIME, MolRIME performs innovation-driven exchange between 2D topology (GINE) and 3D geometry (TorchMD-ET) with symmetry-preserving updates (invariant coefficients modulating equivariant vector bases), combined with geometric denoising, topology reconstruction, reciprocal-prediction, and multi-conformer objectives. Pretrained on PCQM4Mv2, it is evaluated on MD22 (forces), QM9 (12 properties), MoleculeNet (8 datasets), plus a controlled frozen-encoder ablation against a "Direct Exchange" variant and an innovation-dynamics analysis.

Critical strengths

1,Shifting the question from how modalities interact to what should be exchanged is a genuinely useful reframing; the residual-as-innovation formulation is simple, intuitive, and connects to ideas from information decomposition and predictive coding. 2,The equivariant vector update (invariant coefficients × equivariant bases, centered positions, bias-free projection) is carefully constructed, with a complete O(3)-equivariance verification in Appendix A.3. 3,The Direct Exchange baseline differs only in the exchanged signal (same encoders, projections, gates, protocol, same labeled subsets, frozen encoders, 3 seeds) — this is the right way to isolate the contribution, and the reported gains (6.1–8.5% MAE reduction on 4/6 QM9 properties at 10% labels) are credible evidence for the mechanism.

Critical weaknesses

1,Best on only 3/7 MD22 systems (with large regressions elsewhere, e.g., Buckyball 0.214 vs LiTEN 0.114); best/tied on ~3/12 QM9 properties with small margins (2.4%/2.2%); best on only 3/8 MoleculeNet datasets, trailing M2UMol on 4. The core mechanism does not clearly translate into SOTA performance. 2,The "6.1–8.5% MAE reduction" in the abstract is measured against the authors' own ablation variant, not against external baselines; a reader could easily misread it as improvement over SOTA. 3,The predictor is trained with a cosine (direction-only) objective, but the residual Δ = Y − Ŷ contains magnitude mismatch, which is then exchanged as if it were "unexplained information". Magnitude error is not semantic novelty. 4,Residual magnitude partially increases at Block 3 while the story predicts innovations should shrink as information becomes predictable. The explanation offered is plausible but hand-wavy. 5,Residual exchange is close in spirit to modality factorization (Barezi & Fung, 2019, cited) and information-theoretic decomposition (Yang et al., 2026, cited); the delta over these works should be sharpened.

Questions for authors

see Weaknesses.

2 comments
AuthorsOpenReview ↗

Response part 1

We sincerely thank the reviewer for their careful evaluation of our work, recognition of the relevance of our research direction, and constructive feedback. The reviewer's main concerns focus on the consistency of performance across benchmarks, the interpretation and effectiveness of prediction-residual exchange, and the distinction from existing multimodal approaches. We have carefully addressed these concerns by updating benchmark comparisons, clarifying the interpretation of prediction residuals, refining the presentation of our experimental findings, and strengthening the discussion of our methodological contributions. We provide detailed point-by-point responses below.

  1. We appreciate the reviewer's concern regarding performance consistency across benchmarks. To provide a clearer and more rigorous comparison, we have updated the MD22 and MoleculeNet results by retraining representative baselines using exactly the same training, validation, and test samples as MolRIME.

     

    On MD22, we retrained MACE and LiTEN and updated the MolRIME results. Under this evaluation setting, MolRIME achieves the lowest force MAE on 3 of the 7 MD22 systems.

     

    Updated MD22 results. Force MAE (kcal mol⁻¹ Å⁻¹; lower is better). MACE and LiTEN were retrained using the same training, validation, and test trajectory frames as MolRIME. ViSNet results are taken from the original publication and are included as cross-study references rather than matched-split comparisons.

    MethodAc-Ala3-NHMe↓DHA↓AT-AT↓Stachyose↓AT-AT-CG-CG↓Buckyball↓Double-walled↓
    MACE0.0860.0680.0940.0670.1150.0970.290
    ViSNet0.0830.0600.0810.0880.1480.1840.362
    LiTEN0.0800.0550.0850.0610.1220.1280.236
    MolRIME0.1120.0780.0950.0770.1110.0940.197

     

    On MoleculeNet, we retrained MoleculeSDE (VE), MoleculeSTM, and M2UMol using exactly the same training, validation, and test samples as MolRIME. Under this controlled comparison, MolRIME achieves the highest mean ROC–AUC on 5 of the 8 datasets.

     

    Updated MoleculeNet results. Test ROC–AUC (%, mean ± standard deviation over three seeds; higher is better). All listed baselines were retrained using exactly the same training, validation, and test scaffold splits as MolRIME.

    MethodBBBP↑Tox21↑MUV↑BACE↑ToxCast↑SIDER↑ClinTox↑HIV↑
    MoleculeSDE(VE)65.9±3.275.5±0.374.5±3.380.1±1.363.3±0.460.2±0.769.8±0.475.9±0.9
    MoleculeSTM65.9±3.276.6±0.576.6±4.381.5±0.564.9±0.360.7±1.788.4±2.976.8±0.9
    M2UMol67.8±1.675.9±0.977.4±0.977.2±1.965.7±0.361.8±0.284.7±0.772.7±1.1
    MolRIME69.3±0.577.0±0.374.9±1.584.7±1.866.5±0.761.2±1.483.3±3.978.1±0.8

     

    The variation in performance may also reflect the task-dependent benefits of multimodal exchange. For MD22, molecular connectivity remains fixed within each system, while atomic forces depend strongly on the changing 3D configurations. Consequently, 2D topology may provide limited additional information for distinguishing trajectory frames, potentially reducing the benefit of cross-modal exchange on certain systems. For QM9, the prediction targets span spatial, electronic, vibrational, and thermodynamic properties, for which the relative contributions of topology and geometry may differ. These observations offer plausible explanations for the mixed results, although they do not establish a causal relationship.

     

    More broadly, MD22, QM9, and MoleculeNet cover substantially different molecular prediction settings, including force prediction, quantum property estimation, and molecular property classification, where the relative importance of topology and geometry can vary. MolRIME provides a unified framework for integrating 2D molecular topology and 3D geometry, enabling its application across these diverse prediction tasks. We believe this flexibility highlights the broader applicability and potential of the proposed multimodal exchange mechanism.

     

  2. To avoid potential misunderstanding, we have revised the final sentence of the abstract as follows: "Controlled ablations show that exchanging dynamically updated prediction residuals can improve downstream performance over full-representation exchange or fixed residuals, supporting the effectiveness of the proposed core exchange mechanism."

AuthorsOpenReview ↗

Response part 2

  1. We agree that the cosine-based objective encourages directional alignment without explicitly constraining feature magnitudes, so prediction residuals may contain magnitude discrepancies that do not necessarily represent novel semantic information. Our motivation for using this objective is to learn cross-modal representational correspondence without requiring exact reconstruction of source features, whose scales may differ across modality-specific encoders. Accordingly, RIME treats prediction residuals as receiver-conditioned discrepancy signals for cross-modal exchange, rather than as a strict measure of unexplained information.

     

    We retain residuals at their original scale because normalizing each residual to a fixed norm could remove potentially useful magnitude variations while amplifying small residuals. Importantly, the residuals are processed through learnable projections and adaptive gates rather than directly added to the receiving representations, allowing the model to regulate their contributions during cross-modal interaction.

     

  2. As clarified in our response to Comment 3, residual magnitude does not directly quantify the amount of complementary or semantically novel information. Instead, it reflects the discrepancy between a modality representation and its cross-modal prediction, which can also be influenced by differences in feature magnitudes. Therefore, the partial increase observed at Block 3 should not be interpreted as direct evidence of increased semantic novelty or reduced effectiveness of cross-modal exchange.

     

    We would also like to clarify that RIME recomputes prediction residuals at each interaction block as both modality representations evolve. Consequently, the residuals at different blocks are calculated from different representations and predictions, rather than reflecting the progressive reduction of a fixed reconstruction error. In this setting, a monotonic decrease in residual magnitude is not necessarily expected. Moreover, the high cosine similarity at Block 3 can coexist with a relatively large residual because cosine similarity measures directional alignment without constraining feature magnitudes.

     

  3. We agree that RIME shares the broader motivation of leveraging complementary information across modalities with the cited works. Building on this important research direction, RIME introduces a reciprocal prediction-residual exchange mechanism that dynamically determines cross-modal messages based on the current representations of both modalities. This provides a distinct approach to cross-modal interaction, complementing existing factorization- and decomposition-based methods.

     

    Specifically, Barezi and Fung (2019) investigate modality-based factorization for multimodal fusion, whereas RIME constructs cross-modal exchange messages through reciprocal prediction residuals. Yang et al. (2026) explicitly decompose multimodal interactions into redundant, unique, and synergistic components using an information-theoretic framework. In comparison, RIME does not require an explicit decomposition of these information components. Instead, each modality predicts the representation of the other, and the resulting prediction residual is used to construct a receiver-conditioned exchange message. Moreover, the predictions and residuals are dynamically recomputed at each interaction block, allowing the exchanged messages to adapt as the modality representations evolve.

     

    Beyond the general exchange mechanism, MolRIME further contributes a symmetry-preserving realization for molecular 2D–3D representation learning. Specifically, prediction residuals are constructed in invariant feature spaces, while invariant coefficients are used to update equivariant geometric features, maintaining geometric equivariance during cross-modal interaction. We have expanded the related-work discussion and clarified these methodological distinctions in the revised manuscript.

     

We sincerely thank the reviewer once again for the thoughtful and constructive feedback. The reviewer’s comments have been particularly valuable in helping us clarify the positioning of our work, deepen our explanation of the prediction-residual exchange mechanism, strengthen the experimental comparisons, and better articulate the distinctions between our approach and existing multimodal methods. We believe these revisions have substantially improved the clarity and rigor of our manuscript.

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

This main idea of this paper is how to exchange information between 2D topology and 3D geometry in molecular pretraining. The authors argue that the two modalities share a lot of information, so exchanging the full representation is redundant. The authors propose Reciprocal Innovation-Driven Multimodal Exchange (RIME), a framework that adaptively guides cross-modal information exchange based on what the receiving modality can already capture. The model emphasizes complementary information that remains unexplained.

Critical strengths

The question is interesting and the method is clean

Critical weaknesses
  1. I suspect the residual provides really “new information”? The paper assumes that the part the predictor cannot predict is information missing from the receiver. But a residual can also come from a weak predictor or poor training. The predictors are trained with a cosine loss. This loss only makes the prediction point in the right direction. It does not ask the prediction to have the right norm. For example, for 2D→3D in Block 3, the cosine similarity in Fig. 2(a) is about 0.997. If the two vectors had the same norm, this would give a normalized residual of only about 0.08 (higher cosine similarity lead to lower residual). But Fig. 2(b) shows about 0.62, which is quite big. Therefore, a large part of the innovation may not be information that the receiver is missing. Could the authors explain this?

  2. The Direct Exchange comparison may be unfair. The method uses the 2D representation H to predict the 3D representation G. It then sends only the residual, G − p(H), to the 2D side. To show that this is useful, the authors compare it with Direct Exchange, which sends the full G instead. However, this comparison is not fair. The message in Direct Exchange only uses G. The message in MolRIME uses both G and H. So MolRIME may win simply because its message sees more information, not because the residual design is better.

  3. I think the main idea of this paper is quite similar to some existing work in computer vision. In multimodal learning, there are related ideas such as TokenFusion (CVPR 2022) and shared/private disentanglement. Although these methods are not exactly the same as MolRIME, they address a very similar problem: how to reduce redundant information and better use complementary information across different modalities. The paper should discuss these and make clear whether its contribution is the general idea or its symmetry-preserving use in 2D–3D molecular learning.

Questions for authors

The same as weaknesses.

4 comments
AuthorsOpenReview ↗

Response part 1

We sincerely thank the reviewer for their interest in our work, recognition of the importance of the research topic, and thoughtful feedback. The reviewer's main concerns focus on whether prediction residuals provide useful information beyond the receiving modality, the fairness of the Direct Exchange comparison, and the distinction between RIME and existing multimodal fusion approaches. We have carefully addressed these concerns through additional probing experiments, clarification of the experimental comparisons, and a more explicit discussion of our methodological contributions and related work. We provide point-by-point responses below.

  1. We agree that prediction residuals may contain not only complementary information but also discrepancies arising from imperfect cross-modal prediction. In RIME, we use the term "innovation" to describe the difference between a modality representation and its cross-modal prediction, without assuming that this difference constitutes a strict decomposition of shared and modality-specific information.

     

    Regarding the high cosine similarity in Figure 2(a) and the relatively large normalized residual in Figure 2(b), we appreciate the reviewer's insightful observation. As the reviewer points out, the cosine-based objective encourages directional alignment but does not constrain the norms of the predicted representations. Therefore, even when cosine similarity is close to 1, differences in vector magnitudes can still lead to substantial Euclidean residuals. The equal-norm assumption used in the reviewer's calculation provides a useful reference, while our predictors do not explicitly enforce equal norms.

     

    We would also like to clarify that the residual magnitude is not intended to directly quantify the amount of genuinely new information. Instead, RIME uses prediction residuals to construct cross-modal messages whose contributions are regulated through learnable projections and adaptive gating mechanisms. We have refined the discussion in the revised manuscript to better distinguish prediction discrepancies from complementary information and to clarify the interpretation of Figure 2.

     

    To further investigate whether prediction residuals provide useful information beyond the receiving modality, we conducted additional probing experiments using frozen representations from 2,000 molecules sampled from the pretraining validation partition. These molecules were divided into molecule-disjoint training, validation, and test sets (60%/20%/20%). We trained ridge regression probes on standardized features, comparing the receiving-modality representations alone (H or G) with their concatenation with the corresponding prediction residuals ([H, ΔG] or [G, ΔH]). Specifically, we evaluated two geometric targets (radial distance and nonbonded mean distance) and two topological targets (bond type and ring membership) across three interaction blocks. The results are summarized below, with experimental details provided in Appendix B.5 of the revised manuscript.

     

    Table: Probing analysis across interaction blocks. Geometry targets use MAE (Å; lower is better), while bond type and ring membership use balanced accuracy and ROC–AUC, respectively (higher is better). Relative improvement is calculated against the receiving-representation-only baseline using the displayed rounded values. Here, ΔG and ΔH denote the geometry and topology prediction residuals, respectively, and brackets indicate feature concatenation.

    TargetProbe inputBlock 1Block 2Block 3
    Radial distanceH only0.6630.6240.591
    [H,ΔG]0.5710.5050.500
    Relative improvement13.9%19.1%15.4%
    Nonbonded mean distanceH only0.6440.6200.594
    [H,ΔG]0.5830.5280.526
    Relative improvement9.5%14.8%11.4%
    Bond typeG only0.9420.9300.925
    [G,ΔH]0.9530.9420.933
    Relative improvement1.2%1.3%0.9%
    Ring membershipG only0.9900.9770.972
    [G,ΔH]0.9980.9960.992
    Relative improvement0.8%1.9%2.1%

     

Across all evaluated blocks and probe targets, incorporating prediction residuals consistently improves probe performance. For geometric targets, the relative MAE reductions range from 9.5% to 19.1%, while the improvements for topological targets range from 0.8% to 2.1%. These results suggest that prediction residuals contain information useful for the evaluated probe tasks beyond what is captured by the receiving-modality representations alone.

AuthorsOpenReview ↗

Response part 2

  1. We would like to clarify that Direct Exchange also incorporates information from both modalities. Specifically, Direct Exchange passes the full 3D representation (G) to the 2D branch, where it is integrated with the existing 2D representation (H) through adaptive gating. In contrast, MolRIME uses H to predict G and constructs the exchange message from the prediction residual (G − p(H)).

     

    As described in Appendix B.2, both variants retain the same modality encoders, message projections, and adaptive gating mechanisms, and are evaluated under matched experimental settings. Thus, the comparison is intended to evaluate two cross-modal communication strategies under a shared interaction architecture: exchanging full source-modality representations versus exchanging receiver-conditioned prediction residuals. We have clarified this distinction in the revised manuscript.

     

  2. We thank the reviewer for highlighting the connections to TokenFusion and shared/private disentanglement methods. We agree that these approaches share the broader motivation of leveraging complementary information across modalities.

     

    While sharing this broader motivation, RIME adopts a different approach to constructing and exchanging cross-modal information. Specifically, TokenFusion identifies less informative modality-specific tokens and replaces them with projected features from other modalities, while shared/private disentanglement methods explicitly separate shared and modality-specific representations. In comparison, RIME uses reciprocal cross-modal prediction to construct receiver-conditioned residual messages, based on the discrepancy between the source representation and its prediction from the receiving modality. These residuals are dynamically recomputed at each interaction block as the modality representations evolve, without requiring an explicit shared/private decomposition.

     

    To clarify the scope of our contribution, we distinguish two complementary aspects. At the methodological level, we propose a reciprocal prediction-residual exchange mechanism, rather than claiming novelty in the general idea of leveraging complementary multimodal information. At the molecular level, MolRIME realizes this mechanism for 2D–3D molecular representation learning by constructing residuals in an invariant representation space and using invariant coefficients to update equivariant geometric features, thereby preserving the required geometric symmetry. Following the reviewer's suggestion, we have expanded the related-work discussion and clarified these contributions in the revised manuscript.

We sincerely thank the reviewer once again for the thoughtful and constructive feedback. The reviewer's comments have been particularly valuable in helping us clarify the interpretation of prediction residuals, strengthen the experimental evidence for their effectiveness, provide a more precise explanation of the Direct Exchange comparison, and better articulate the distinctions between RIME and existing multimodal fusion approaches. We believe these suggestions have significantly improved the clarity, rigor, and presentation of our work.

I thank for the reviewers' additional experiments and clarifications. Overall, I think the proposed approach has its nature limitations and the clarifications and experiments do not address my concerns, especially for the question 1. I will keep my current score.

AuthorsOpenReview ↗

Response: Clarification on cosine supervision and cross-modal innovations

We sincerely thank the reviewer for their careful consideration and continued feedback.

The reviewer’s central concern is whether cross-modal innovations genuinely capture useful information beyond the receiving modality, given that their magnitudes may also reflect prediction imperfections, particularly norm mismatches under cosine supervision. We would like to further clarify the motivation and interpretation of our design.

An important consideration is the one-to-many relationship between molecular 2D topology and 3D geometry. Although the same molecular graph may correspond to multiple conformations, its connectivity and local chemical environments still constrain the possible geometric structures. This provides a meaningful basis for learning directional cross-modal correspondence in the latent feature space, even when the complete geometric representation of a specific conformation cannot be uniquely inferred from topology.

Enforcing exact magnitude matching can introduce conflicting prediction targets under this one-to-many relationship. Specifically, different conformations of the same molecular graph may produce geometric representations with different feature norms, while a deterministic predictor receiving the same topology representation can only produce one prediction. It therefore cannot simultaneously match the distinct target magnitudes, even if their feature directions are aligned. RIME instead employs cosine supervision to encourage directional correspondence without requiring exact reconstruction of potentially conformation-dependent feature magnitudes. This establishes a predictive reference for cross-modal exchange while avoiding the requirement to match potentially incompatible magnitude targets.

Through predictive re-centering, an innovation expresses the observed source representation relative to the receiver-conditioned predictive reference. Importantly, innovation magnitude measures feature discrepancy, not the amount of useful or complementary information. Such discrepancies may contain task-relevant variations as well as prediction imperfections, including norm mismatches. RIME therefore treats innovations as candidate exchange signals, whose contributions are learned through message projections and contextual gating, rather than assuming they exclusively represent missing information.

We sincerely appreciate the reviewer’s thoughtful feedback and hope this clarification better conveys the motivation and interpretation of our proposed cross-modal innovations.

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

This paper proposes a molecular representation learning framework that integrates 2D molecular topology and 3D geometry through reciprocal innovation-driven multimodal exchange. Instead of directly exchanging complete modality representations, the method uses cross-modal prediction residuals to identify information that is not explained by the other modality. These residuals are exchanged through gated updates while preserving the geometric transformation properties of the representations. The framework is pretrained using geometric denoising, topology reconstruction, and cross-modal objectives. Experiments are conducted on MD22, QM9, and eight MoleculeNet datasets, along with a frozen-encoder study under limited training labels.

Critical strengths
  • The paper addresses an interesting problem in multimodal molecular representation learning. Distinguishing shared information from complementary information between molecular topology and geometry is a reasonable motivation.

  • The proposed framework considers the different transformation properties of 2D and 3D representations, and the symmetry-preserving design is a useful aspect of the method.

Critical weaknesses
  • The interpretation of prediction residuals as cross-modal innovations is not sufficiently justified. The main idea of RIME is to use the residual between a modality representation and its prediction from the other modality as the information to be exchanged. However, a prediction residual does not necessarily represent meaningful modality-specific information. It may also contain prediction errors, noise, or information that the cross-modal predictor simply fails to capture. This distinction is important because the entire framework is motivated by exchanging information that is not already available to the receiving modality.

  • The paper uses a cosine-based objective to train the reciprocal predictors, but it is not clear why the resulting residuals should correspond to complementary information rather than imperfect prediction. I would like to see more direct evidence about what these residuals contain and whether they actually capture information missing from the other modality.

  • The experimental improvements look quite inconsistent across different tasks. While MolRIME achieves some good results, its overall performance is not particularly convincing. On MD22, it obtains the best force MAE on only three of the seven molecular systems. For several other systems, the performance is substantially worse than existing methods. For example, on Buckyball and Double-walled, the force MAEs are 0.214 and 0.463, compared with the best competing results of 0.085 and 0.214, respectively. These are fairly large differences. The QM9 results are also mixed. MolRIME achieves the lowest error on two properties and ties for the best on LUMO, but remains behind other methods on most of the remaining properties. A similar pattern appears on MoleculeNet, where it performs best on three of the eight datasets. Given these results, the proposed exchange mechanism does not seem to provide a consistent advantage across molecular tasks. The paper would be stronger with a more detailed analysis of why the method works well for certain properties but performs considerably worse for others.

  • The ablation studies do not sufficiently isolate the main contributions. The comparison with Direct Exchange is useful, but it is limited to frozen-encoder transfer on six selected QM9 properties. It does not establish whether the proposed residual exchange mechanism is responsible for the results on MD22 or MoleculeNet, where the full models are evaluated. I am also not convinced that the dynamic re-estimation of innovations, which is repeatedly emphasized as a key contribution, has been adequately validated. What happens if the cross-modal residuals are computed only once instead of being updated at every interaction block? Comparisons with alternative shared/private representation decomposition or gated cross-modal fusion strategies would also help clarify whether the benefits come specifically from prediction-residual exchange. At present, the individual contributions of these design choices remain somewhat unclear.

  • Some experimental comparisons are not sufficiently controlled. Several important baseline results are taken directly from previous publications rather than reproduced under the same experimental conditions. This is quite concerning for the MoleculeNet experiments, where the authors acknowledge that the comparison is not based on identical scaffold splits. Differences in data splits, preprocessing, and fine-tuning settings can have a noticeable impact on these benchmarks. The comparisons on MD22 and QM9 also rely on published results from different methods. Although this is common for broad benchmark evaluations, it makes it harder to attribute the observed differences specifically to the proposed exchange mechanism. More controlled comparisons with representative baselines, especially under matched training and evaluation settings, would make the experimental conclusions more reliable.

Questions for authors

See weaknesses.

4 comments
AuthorsOpenReview ↗

Response part 1

We thank the reviewer for the careful assessment. The reviewer's main concerns are whether cross-modal prediction residuals contain useful information beyond prediction error, and whether the current ablations adequately isolate residual exchange and its dynamic re-estimation. The reviewer also raises questions about task-dependent performance and the comparability of results obtained under different experimental settings. We address these points below with additional analyses, expanded ablations, matched-split evaluations, and a more precise discussion of the scope of our claims.

  1. We agree that prediction residuals do not necessarily represent meaningful modality-specific information, as they may also contain prediction errors or noise. In RIME, we use the term "innovation" to refer to the difference between a modality representation and what can be predicted from the other modality. Rather than assuming that every part of this difference is useful, RIME uses it as a candidate signal for cross-modal exchange. The prediction objective encourages the model to capture information that can be inferred across modalities, while the learnable projections and adaptive gates regulate how the remaining differences contribute to the receiving modality. Therefore, our method does not rely on the assumption that prediction residuals perfectly separate useful and redundant information. Instead, we investigate whether exchanging these residuals provides a more effective way to integrate the two modalities than directly exchanging complete representations.

     

  2. We agree that the cosine-based prediction objective does not guarantee that the resulting residuals contain exclusively complementary information. In RIME, the cosine-based objective encourages the reciprocal predictors to align their predictions with the corresponding modality representations, while the residuals measure the discrepancies between these predictions and the source representations. We therefore use the residuals as receiver-conditioned signals for cross-modal exchange, rather than assuming that they exclusively contain complementary information.

     

    We have added a probing analysis (Figure 3 and Appendix B.5) to examine whether prediction residuals provide information beyond the receiving-modality representations. Using frozen MolRIME features from 2,000 molecules divided into molecule-disjoint training, validation, and test sets, we compare linear probes using the receiving representation alone against those additionally incorporating the corresponding cross-modal residual at each interaction block.

     

    Table: Probing analysis across interaction blocks. Geometry targets use MAE (Å; lower is better), while bond type and ring membership use balanced accuracy and ROC–AUC, respectively (higher is better). Relative improvement is calculated against the receiving-representation-only baseline using the displayed rounded values. Here, ΔG and ΔH denote the geometry and topology prediction residuals, respectively, and brackets indicate feature concatenation.

     

    TargetProbe inputBlock 1Block 2Block 3
    Radial distanceH only0.6630.6240.591
    [H,ΔG]0.5710.5050.500
    Relative improvement13.9%19.1%15.4%
    Nonbonded mean distanceH only0.6440.6200.594
    [H,ΔG]0.5830.5280.526
    Relative improvement9.5%14.8%11.4%
    Bond typeG only0.9420.9300.925
    [G,ΔH]0.9530.9420.933
    Relative improvement1.2%1.3%0.9%
    Ring membershipG only0.9900.9770.972
    [G,ΔH]0.9980.9960.992
    Relative improvement0.8%1.9%2.1%

    For 3D-to-2D exchange, adding the geometry residual reduces radial-distance MAE by 13.9–19.1% and nonbonded-mean-distance MAE by 9.5–14.8%. For 2D-to-3D exchange, despite the already high performance using geometry representations alone, adding the topology residual consistently improves bond-type balanced accuracy and ring-membership ROC–AUC across all three blocks. These results indicate that the residuals contain additional task-relevant information that is not as readily accessible from the receiving representations through linear probing.

AuthorsOpenReview ↗

Response part 2

  1. As detailed in our response to Comment 5, we conducted additional experiments to strengthen the benchmark comparisons, including retraining representative baselines under matched data splits. The updated results are presented below.

     

    Updated MD22 results. Force MAE (kcal mol⁻¹ Å⁻¹; lower is better). We retrained MACE and LiTEN using exactly the same training, validation, and test trajectory frames as MolRIME. ViSNet results are taken from the original publication and are included as cross-study references.

    MethodAc-Ala3-NHMe↓DHA↓AT-AT↓Stachyose↓AT-AT-CG-CG↓Buckyball↓Double-walled↓
    MACE0.0860.0680.0940.0670.1150.0970.290
    ViSNet0.0830.0600.0810.0880.1480.1840.362
    LiTEN0.0800.0550.0850.0610.1220.1280.236
    MolRIME0.1120.0780.0950.0770.1110.0940.197

     

    Updated MoleculeNet results. Test ROC–AUC (%, mean ± SD over 3 seeds; higher is better). All baselines were retrained on the same train/validation/test scaffold splits as MolRIME.

    MethodBBBP↑Tox21↑MUV↑BACE↑ToxCast↑SIDER↑ClinTox↑HIV↑
    MoleculeSDE(VE)65.9±3.275.5±0.374.5±3.380.1±1.363.3±0.460.2±0.769.8±0.475.9±0.9
    MoleculeSTM65.9±3.276.6±0.576.6±4.381.5±0.564.9±0.360.7±1.788.4±2.976.8±0.9
    M2UMol67.8±1.675.9±0.977.4±0.977.2±1.965.7±0.361.8±0.284.7±0.772.7±1.1
    MolRIME69.3±0.577.0±0.374.9±1.584.7±1.866.5±0.761.2±1.483.3±3.978.1±0.8

     

    Under the updated comparisons, MolRIME achieves the lowest listed force MAE on three of seven MD22 systems and the highest mean ROC–AUC on five of eight MoleculeNet datasets. Nevertheless, we acknowledge that MolRIME does not consistently outperform the strongest competing methods across all systems and properties.

     

    The variation also highlights the task-dependent nature of multimodal exchange. In MD22, molecular connectivity remains fixed within each system, so 2D topology may provide limited additional information for distinguishing trajectory frames, whereas atomic forces depend strongly on the changing 3D configurations. Consequently, the additional benefit of incorporating topology information may be limited. QM9 presents a different challenge: its targets span spatial, electronic, vibrational, and thermodynamic properties, so the benefits of exchanging topology and geometry information may differ across properties. These are plausible factors behind the mixed results, rather than causal explanations established by benchmark rankings alone.

     

    We therefore distinguish overall benchmark performance from evidence for the proposed mechanism. While MolRIME does not consistently outperform every task-specific method, its core exchange mechanism is further evaluated through controlled ablation studies, as discussed in our Response 4. We have revised the manuscript to clarify this distinction.

AuthorsOpenReview ↗

Response part 3

  1. To address these concerns, we have expanded the experiments to evaluate (1) prediction-residual exchange, (2) dynamic residual updates, (3) the effectiveness of residual exchange beyond QM9, and (4) scope and interpretation of the ablations.

     

    (1) Extending the QM9 ablations to all 12 properties. We extended the frozen-encoder comparisons to all 12 QM9 properties and three label budgets. All variants were separately pretrained and evaluated using identical labeled subsets and prediction-head settings. Direct Exchange replaces prediction residuals with full source representations while retaining the same message projections and gated updates. MolRIME achieves lower mean test MAE in 34/36 comparisons, supporting residual exchange over gated full-representation exchange. Results for both ablations are shown below.

     

    QM9 frozen-encoder ablation results. Relative reduction in test MAE of MolRIME over Direct Exchange and Static Innovation, computed from the corresponding three-seed mean MAEs. Positive values favor MolRIME, while negative values favor the ablation variant.

    PropertyDirect Exchange (1%)Direct Exchange (5%)Direct Exchange (10%)Static Innovation (1%)Static Innovation (5%)Static Innovation (10%)
    R²8.0%7.2%8.5%9.9%4.3%4.4%
    ZPVE3.2%4.5%6.1%5.8%4.3%5.8%
    HOMO1.7%0.2%-0.3%1.4%0.5%0.9%
    α0.2%3.2%4.9%5.0%3.4%5.2%
    μ2.5%6.9%7.9%7.3%12.8%12.9%
    G₂₉₈9.9%8.8%6.4%-2.3%1.4%0.8%
    Cᵥ3.8%3.5%5.1%5.1%1.9%4.1%
    Gap0.6%-0.1%0.4%2.3%0.9%0.2%
    H₂₉₈9.8%8.8%6.2%0.4%1.0%1.9%
    LUMO1.1%0.4%0.7%1.5%-0.2%0.5%
    U₀9.8%9.0%6.3%0.3%1.0%1.6%
    U₂₉₈9.8%8.8%6.3%-1.7%-0.4%-2.3%

     

    (2) Evaluating dynamic residual updates with Static Innovation. Static Innovation computes residuals in the first interaction block and reuses them in subsequent blocks, whereas MolRIME recomputes them from updated representations at each block. Under the same frozen-encoder settings, MolRIME achieves lower mean test MAE in 31/36 QM9 comparisons, supporting the benefit of dynamic residual updates.

     

    (3) Extending Direct Exchange comparisons to MD22 and MoleculeNet. To examine whether the benefit of residual exchange extends beyond frozen-encoder QM9 evaluation, we additionally compare MolRIME with Direct Exchange under downstream fine-tuning.

     

    We further evaluate Direct Exchange on three selected MD22 systems, including two systems where MolRIME achieves the best benchmark results (AT-AT-CG-CG and Double-walled) and one where it does not (AT-AT). MolRIME achieves lower force MAE than Direct Exchange on all three evaluated systems. The results are summarized below.

     

    MD22 Direct Exchange ablation. Test force MAE (kcal mol⁻¹ Å⁻¹; lower is better). Both methods are evaluated under matched settings.

    MethodAT-AT↓AT-AT-CG-CG↓Double-walled↓
    Direct Exchange0.1000.1230.206
    MolRIME0.0950.1110.197

     

    On MoleculeNet, MolRIME achieves higher mean ROC–AUC on all eight datasets, averaged over three random seeds. The complete results are presented below.

     

    MoleculeNet Direct Exchange ablation. Test ROC–AUC (%, mean ± standard deviation over three seeds; higher is better). Both methods are evaluated using matched scaffold splits. The best mean for each dataset is shown in bold.

    MethodBBBPTox21MUVBACEToxCastSIDERClinToxHIV
    Direct Exchange66.0±0.973.1±0.871.5±2.576.4±4.462.7±0.858.1±1.979.6±3.275.5±1.7
    MolRIME69.3±0.577.0±0.374.9±1.584.7±1.866.5±0.761.2±1.483.3±3.978.1±0.8

     

    These results demonstrate that the performance advantage over Direct Exchange is also observed when the full model is fine-tuned for molecular force prediction and property classification.

     

    (4) Scope and interpretation of the ablations. MolRIME and Direct Exchange share the same message projections and adaptive gating mechanism. Thus, Direct Exchange serves as a gated full-representation exchange baseline, allowing us to evaluate prediction-residual exchange within the same interaction architecture. The improvements across QM9, MD22, and MoleculeNet provide empirical support for the proposed mechanism.

     

    We agree that shared/private representation decomposition is another relevant approach, but introducing such modules would involve additional architectural changes. Our ablations therefore focus on the use of prediction residuals instead of full representations and their dynamic updates across interaction blocks. We have clarified this scope in the revised manuscript.

AuthorsOpenReview ↗

Response part 4

  1. We agree that results reported in different publications should not be treated as strictly matched comparisons, although cross-study comparisons are widely adopted in molecular representation learning benchmarks. To address this concern, we have retrained representative baselines on MD22 and MoleculeNet using the same data splits as MolRIME. The updated results and detailed comparisons are presented in Response 3. Specifically, MACE and LiTEN on MD22, as well as all listed MoleculeNet baselines, were retrained using our data splits. ViSNet results on MD22 are retained from the original publication and explicitly identified as literature-reported results.

     

    For QM9, we retain the published baseline results, following the common practice of cross-study comparisons in molecular property prediction benchmarks. Given the limited rebuttal period, we therefore prioritized retraining representative baselines on MD22 and MoleculeNet under matched data splits, while expanding our controlled QM9 ablation experiments to provide more direct evidence for the effectiveness of the proposed exchange mechanism, as detailed in our response to Comment 4. We have also clarified in the revised manuscript that the published QM9 results serve as benchmark references rather than strictly matched comparisons.

We sincerely thank the reviewer once again for the careful evaluation and thoughtful, constructive feedback. The reviewers comments have been particularly valuable in helping us clarify the interpretation of prediction residuals, strengthen the experimental validation of the proposed exchange mechanism, improve the fairness and comprehensiveness of our benchmark comparisons, and more carefully articulate the scope of our findings. We believe these suggestions have significantly improved the clarity, rigor, and overall quality of our work.

Comments on the paper

1 comment

Dear reviewers,

The rebuttal period is ending soon. Please read the authors rebuttal and check if they addressed your concerns. Please acknowledge the rebuttal and adjust your scores accordingly.

Thanks.

-AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.