acceptodds
Under review as a conference paper at ICLR 2027

Learning Transferable Reaction Mechanisms from Visual Chemical Knowledge

Abstract

Reaction mechanisms describe the step-by-step transformations underlying chemical reactions and are central to reaction analysis and synthesis. Learning-based models have achieved strong performance on established mechanism-prediction benchmarks, but transferring them to unseen chemistry remains challenging. Such transfer is difficult because familiar mechanisms must be applied to unfamiliar molecular structures, and some target mechanisms may be poorly covered by the training data. To address these challenges, we introduce MechaVLM, a visual framework that combines transferable chemical representations with external mechanistic knowledge. It learns reusable visual features through multiscale chemical grounding and cross-rendering contrastive learning. For open-book prediction, MechaVLM retrieves a fixed set of precedents from 70,384 literature mechanism figures and re-reads relevant visual evidence as the molecular state evolves, directly using the figures without symbolic mechanism parsing. An atom-indexed language decoder then recursively generates executable electron edits to construct the complete mechanism. We further introduce MechBench, a challenging literature-derived benchmark with 2,184 mechanisms and 9,146 elementary steps. Across cross-dataset and literature-derived benchmarks, MechaVLM establishes strong zero-shot mechanism prediction. Its closed-book model alone improves Step/Pathway Top-1 by 12.50/13.93 on FlowER to ReactMech, while external visual precedents unlock further gains on challenging OOD reactions. The learned representation also generalizes beyond mechanism prediction to atom mapping and reaction center prediction. Codes and benchmarks will be released.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Reaction Mechanism Prediction, Zero-Shot Transfer, Chemical Vision-Language Models

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_tE7xOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper proposes a relatively novel idea: learning reaction mechanisms from images rather than SMILES or molecular graphs. The authors develop MechaVLM to implement this idea and demonstrate the advantages of visual inputs for learning reaction mechanisms and generalizing the acquired knowledge to OOD data. Through detailed experiments, the paper demonstrates the effectiveness of the proposed method and its individual components.

Critical strengths
  1. The idea of learning reaction mechanisms from images is relatively novel. The training tasks for the vision encoder and the designs for precedent retrieval and utilization offer valuable insights.
  2. Extensive experiments demonstrate MechaVLM’s superior generalization capability in reaction mechanism prediction relative to the baselines. The paper also investigates and analyzes the contributions of its individual components.
  3. The experiments further demonstrate that the proposed training approach of the vision encoder benefits other related tasks.
Critical weaknesses

I have concerns regarding the setup and implementation of some experiments, which leave me uncertain about the soundness of the corresponding conclusions. Please see the questions in "Questions For Authors" for details.

Questions for authors
  1. Atom types in molecules exhibit a highly imbalanced, long-tailed distribution, with neutral carbon atoms accounting for the vast majority. How do you address this imbalance during your training of the vision encoder? Similarly, the appendix states that bonded and non-bonded atom pairs are sampled "at approximately equal frequency". However, among bonded pairs, single bonds account for the vast majority. Do you take any further measures to address this imbalance?
  2. What is the source of the molecular data used to train the vision encoder, and how exactly are different visual forms of molecular images under different rendering conditions generated? These details are not clearly explained in the method section.
  3. Considering that MechaVLM is a task-specific model and that this work places particular emphasis on OOD generalization, a fair comparison requires the baseline models to undergo corresponding task-specific fine-tuning as well. In the experiments reported in Tables 1 and 2, were all baseline models evaluated after additional task-specific fine-tuning on the corresponding source dataset (FlowER or ReactMech)? Why are results for some models, such as ChemVLR and DeepMech, reported for only one of the two source datasets?
  4. Similarly, did the VLMs undergo task-specific fine-tuning for the experiments in Tables 3 and 4?
  5. For the experiments in Table 5, how exactly were the SMILES-input and graph-input variants implemented, and did they undergo task-specific fine-tuning? These details are critical to proving whether learning from images provides stronger generalization. Furthermore, if comparable precedent retrieval were incorporated into the SMILES-input and graph-input variants, could they achieve similar improvements in generalization?
  6. Purely out of curiosity: what do you see as the value of reaction mechanism prediction for AI for chemistry? Given that current OOD performance remains far from practically usable, how could this capability be applied in concrete chemical scenarios? Alternatively, what benefits or insights could it offer for models addressing specific practical chemical tasks or problems?
2 comments

Thanks for the detailed response. Your explanations have addressed my questions about the experimental setup as well as those raised out of curiosity. I have therefore decided to raise my soundness score to 3. I appreciate your rigorous work and inspiring insights.

Thank you for your positive follow-up and kind words. We sincerely appreciate your thoughtful questions and constructive feedback. Your suggestions have helped us improve the clarity and rigor of our work, and we are grateful for your careful consideration.

Reviewer_Jk65OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a visual framework that combines transferable chemical representations with external mechanistic knowledge. Specifically, the paper considers transferable mechanism prediction in a zero-shot transfer setting. The paper grounds the chemical intermediates in multiple scales. The paper uses contrastive learning to capture rendering differences. For open-book, it retrieves mechanism figures from literature and re-reads relevant visual evidence as the molecular state evolves. The paper includes a step-wise cross-image attention for retrieval. The paper also adds atom-indexed electron-edit decoding. The paper provides a new MechBench with 2,184 pathways and 9,146 steps.

Critical strengths
  1. The framework mainly includes two contributions. (1) The multiscale grounding and cross-rendering contrastive learning provide better representations. (2) The step-wise cross-image attention addresses settings where only part of the retrieved figure is useful. Both are supported by additional experiments. Table 1 shows that the method has strong transfer for both datasets. Table 5 provides cumulative ablation for closed-book and open-book settings, and different input modalities. In Table 5, the consistent gains in OOD show that each component can help with transfer.
  2. The new benchmark can be a useful resource.
Critical weaknesses
  1. The related work on representation learning is insufficient. The MaskMol [1] includes a pixel-masking-based pretraining task with atomic knowledge, bond knowledge, and motif knowledge. It overlaps with multiscale chemical grounding. The MoleClip [2] includes structural classification and contrastive learning, which aligns with cross-rendering contrastive learning. These two can be discussed in the related work.
  2. Leakage control still has some problems. The paper excludes DOI, semantically similar images, etc. However, the same chemistry may appear in differently rendered figures or different publications. Moreover, retaining related mechanisms is intentional, so instance-level novelty does not imply mechanistic-family novelty.

[1] Cheng, Z., Xiang, H., Ma, P., Zeng, L., Jin, X., Yang, X., ... & Zeng, X. (2024). MaskMol: knowledge-guided molecular image pre-training framework for activity cliffs. arXiv preprint arXiv:2409.12926. [2] Harnik, Y., Shalit Peleg, H., Bermano, A. H., & Milo, A. (2025). Data efficient molecular image representation learning using foundation models. Chemical Science, 16(24), 10833-10841.

Questions for authors
  1. Appendix B reports leakage prevention and deduplication. What did the paper remove for each leakage-prevention method, such as DOI, semantic similarity, etc.? The paper said "We intentionally retain mechanistically related but distinct reactions". Can you provide some examples?
Reviewer_GyfeOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

MechaVLM predicts reaction mechanisms from molecular images using chemically grounded visual features and atom-indexed electron edits. Its open-book version retrieves literature mechanism figures and re-reads them at each step. The paper also introduces the 2,184-pathway MechBench benchmark.

Critical strengths
  • The closed-book transfer gains are substantial, especially from FlowER to ReactMech. Table 5 also separates the contributions of chemical grounding, cross-rendering learning, and precedent transfer.
  • The retrieval study is more informative than a simple with/without-retrieval comparison. Random precedents, alternative retrievers, fixed versus step-wise use, and bank size are all examined. MechBench adds a harder literature-derived test with expert-checked pathways.
Critical weaknesses
  • The modality comparison in Table 5 is difficult to interpret without more detail on the SMILES and graph variants. The image model also benefits substantially from auxiliary chemical grounding. I couldn't tell whether the other modalities receive equivalent auxiliary supervision and a comparable training budget, which matters for attributing the transfer gains to visual modeling.
  • Appendix B describes DOI exclusion and VLM-assisted reaction-overlap screening, which are useful safeguards. However, it does not quantify the screening results or audit unflagged figures. With a 70,384-figure memory, some evidence about missed overlaps would strengthen the claim that open-book gains reflect transfer to unseen reaction instances.

Overall: Weak acceptance. The closed-book gains and the retrieval ablations make a useful case for the approach. Complete-pathway accuracy on MechBench remains low even in the best setting (2.52% Top-1), so I view this as progress on transfer rather than a reliable mechanism predictor.

Questions for authors
  • How are the SMILES and graph variants in Table 5 implemented, and what auxiliary supervision and training budget do they receive? Does the visual advantage remain with matched chemical-grounding supervision?
  • How many figures were removed by each overlap check, and was an unflagged sample independently audited for remaining reaction-instance overlap?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.