acceptodds
Withdrawn from ICLR 2027

Fused Gromov-Wasserstein Contrastive Learning for Effective Enzyme-Reaction Screening

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Enzymes are crucial catalysts that enable a wide range of biochemical reactions. Efficiently identifying specific enzymes from vast protein libraries is essential for advancing biocatalysis. Traditional computational methods for enzyme screening and retrieval are time-consuming and resource-intensive. Recently, deep learning approaches have shown promise. However, these methods focus solely on the interaction between enzymes and reactions, overlooking the inherent relational structure within each domain. To address these limitations, we introduce FGW-CLIP, a novel contrastive learning framework based on optimizing the fused Gromov-Wasserstein distance. FGW-CLIP incorporates multiple alignments, including inter-domain alignment between reactions and enzymes and intra-domain alignment within enzymes and reactions. By introducing a tailored regularization term, our method minimizes the Gromov-Wasserstein distance between enzyme and reaction spaces, which enhances information integration across these domains. Extensive evaluations demonstrate the superiority of FGW-CLIP in challenging enzyme-reaction tasks. On the widely-used EnzymeMap benchmark, FGW-CLIP achieves state-of-the-art performance in enzyme virtual screening, as measured by BEDROC and EF metrics. Moreover, FGW-CLIP consistently outperforms across all three splits of ReactZyme, the largest enzyme-reaction benchmark, demonstrating robust generalization to novel enzymes and reactions. These results position FGW-CLIP as a promising framework for enzyme discovery in complex biochemical settings, with strong adaptability across diverse screening scenarios.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
contrastive learning, enzyme-reaction screening, Gromov-Wasserstein distance

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_YEPdOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes to combine cross-domain enzyme-reaction contrastive learning, EC-based intra-domain objectives, a GW-inspired regularizer, and an auxiliary EC prediction loss for the task of effective enzyme-reaction screening. The reported gains on ReactZyme can outperform existing methods to some extent.

Critical strengths

The paper proposes to combine cross-domain enzyme-reaction contrastive learning, EC-based intra-domain objectives, a GW-inspired regularizer, and an auxiliary EC prediction loss for the task of effective enzyme-reaction screening. The reported gains on ReactZyme can outperform existing methods to some extent.

Critical weaknesses

The methodological distinction from combining established objectives is insufficiently established, and the EnzymeMap screening evidence is too narrow to support the broad SOTA claim. The detailed comments are as follows:

  1. The novelty of the proposed method seems limited. The method mainly combines cross-domain contrastive learning, EC-based intra-domain losses, and a GW-inspired regularizer, which are existing techniques more of less and makes the novelty of this paper limited. The authors need to clarify what is new beyond this combination and how the actual objective relates to standard FGW.
  2. Table 4 (page 9) does not separately isolate the effects of the two intra-domain losses and EC prediction. More detailed ablation studies are suggested.
  3. The proposed method needs to compare with more methods. Table 1 compares it against CLIPZyme and two variants of that same method. Table 2 includes only CLIPZyme. This is a narrow basis for a general SOTA claim on enzyme virtual screening. Table 2 also shows a lower EF₀.₁ for FGW-CLIP than CLIPZyme (7.61 versus 7.81), which should be discussed rather than described only as comparable performance.
  4. Preliminaries appropriately introduces OT/GW/FGW, but the link to the implemented objective needs to be clearer. Figure 1 should better explain the GW regularizer, rather than merely display loss names.
  5. At Table 6(a) (page 19), in terms of Top5, the score of the proposed FGW-CLIP is 0.9746, which is smaller than the score of the comparison method CLIPZyme (UniMol-3D + ESM), namely 0.9478. However, the result of FGW-CLIP (0.9746) is bolded, which is not consistent with the claim that the best results are bolded.
Questions for authors
  1. The novelty of the proposed method seems limited. The method mainly combines cross-domain contrastive learning, EC-based intra-domain losses, and a GW-inspired regularizer, which are existing techniques more of less and makes the novelty of this paper limited. The authors need to clarify what is new beyond this combination and how the actual objective relates to standard FGW.
  2. Table 4 (page 9) does not separately isolate the effects of the two intra-domain losses and EC prediction. More detailed ablation studies are suggested.
  3. The proposed method needs to compare with more methods. Table 1 compares it against CLIPZyme and two variants of that same method. Table 2 includes only CLIPZyme. This is a narrow basis for a general SOTA claim on enzyme virtual screening. Table 2 also shows a lower EF₀.₁ for FGW-CLIP than CLIPZyme (7.61 versus 7.81), which should be discussed rather than described only as comparable performance.
  4. Preliminaries appropriately introduces OT/GW/FGW, but the link to the implemented objective needs to be clearer. Figure 1 should better explain the GW regularizer, rather than merely display loss names.
  5. At Table 6(a) (page 19), in terms of Top5, the score of the proposed FGW-CLIP is 0.9746, which is smaller than the score of the comparison method CLIPZyme (UniMol-3D + ESM), namely 0.9478. However, the result of FGW-CLIP (0.9746) is bolded, which is not consistent with the claim that the best results are bolded.
Reviewer_iQbUOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This work introduces FGW‑CLIP, a contrastive‑learning framework for enzyme‑reaction virtual screening and bidirectional retrieval, optimized using the Fused Gromov‑Wasserstein (FGW) distance. Existing enzyme‑reaction contrastive models only model cross‑domain pairing between enzymes and reactions while ignoring intrinsic relational structures inside the enzyme domain and reaction domain. FGW‑CLIP jointly optimizes inter‑domain alignment (enzyme‑reaction contrastive loss), intra‑domain contrastive alignment guided by EC numbers, a GW‑based structural regularization loss, and multi‑level EC number classification auxiliary loss. Reaction representations are produced by frozen Uni‑Mol 3D molecular encoder; enzyme embeddings come from frozen ESM2 protein language model with a lightweight projection head. The method is validated on two benchmarks: EnzymeMap for enzyme virtual screening and ReactZyme (time‑split, enzyme‑similarity split, reaction‑similarity split) for bidirectional enzyme‑reaction retrieval.

Critical strengths
  1. Prior contrastive enzyme‑reaction models (CLIPZyme, CARE) only perform paired cross‑modal alignment. This work brings Fused Gromov‑Wasserstein into biochemical representation learning, explicitly enforcing consistency of within‑enzyme and within‑reaction relational structures when learning cross‑domain soft couplings.

  2. The model achieves SOTA performance on EnzymeMap and all three challenging splits of ReactZyme, covering real‑world generalization scenarios: temporally unseen data, sequence‑dissimilar unseen enzymes, and fully unseen reactions. Ablations isolate the gain from GW regularization itself. Even with frozen backbones (ESM2, Uni‑Mol), FGW‑CLIP consistently outperforms matched‑encoder baselines, confirming improvements stem from the proposed learning objective rather than encoder capacity.

Critical weaknesses
  1. Lack of real‑world wet‑lab validation; evaluation remains fully confined to retrospective benchmark datasets.

  2. The best performance is obtained at a specific α=0.1 for GW weight; larger or smaller α degrades performance. The paper does not provide guidance or analysis on how to select α for new/unseen datasets outside EnzymeMap / ReactZyme.

Questions for authors
  1. Your intra‑domain contrastive loss and GW structural regularizer rely heavily on EC annotations. How robust is FGW‑CLIP when a substantial fraction of EC labels are missing or incorrect? Could you provide quantitative results simulating EC label corruption?
Reviewer_UFA9OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes FGW-CLIP, a contrastive learning framework for enzyme-reaction screening. The method combines cross-domain contrastive learning, intra-domain alignment based on EC labels, and a Gromov-Wasserstein regularization term to capture structural relationships between enzymes and reactions. The authors also provide a theoretical interpretation from the inverse optimal transport perspective. Experiments on EnzymeMap and ReactZyme show improvements over several existing baselines, with additional ablations supporting the proposed GW regularization.

Critical strengths
  1. The paper addresses an important problem in enzyme screening. Modeling both enzyme-reaction correspondence and intra-domain relationships is a reasonable direction that goes beyond standard cross-modal contrastive learning.
  2. The method achieves strong results on EnzymeMap and three ReactZyme splits. The authors also include comparisons using the same encoder backbone and ablation studies to examine the contribution of GW regularization.
Critical weaknesses
  1. The overall framework mainly combines existing components, including Uni-Mol, ESM2, contrastive learning, EC classification, and GW regularization. While the combination is reasonable, the paper does not clearly establish what is fundamentally new beyond applying these techniques to enzyme-reaction retrieval. The theoretical connection to FGW also seems largely an interpretation of the proposed objective rather than a new optimization algorithm.
  2. The method introduces several additional objectives at once, but Table 4 only compares CLIP-only, the model without GW, and the full model. It is still unclear how much each intra-domain contrastive loss and the EC prediction loss contribute individually. More importantly, the paper does not compare GW against simpler structural regularization methods, making it difficult to tell whether the gains specifically come from GW.
  3. The method uses EC numbers to construct intra-domain positive pairs and adds an auxiliary EC classification objective. However, EC numbers already encode enzyme functional information closely related to the retrieval targets. This raises concerns about whether the improvements come from better enzyme-reaction alignment or simply stronger supervision from functional labels. The paper does not sufficiently investigate performance under incomplete, noisy, or unavailable EC annotations.
  4. Although the reported improvements are promising, the paper lacks repeated runs, standard deviations, or statistical significance tests. Some improvements are relatively small, such as MRR on the ReactZyme enzyme-similarity split (0.8581 vs. 0.8546). The paper also provides little biological interpretation or case-level analysis to show whether the retrieved enzymes are genuinely meaningful beyond benchmark metrics.
  5. The paper presents its objective as optimizing an FGW-type distance, but the actual GW regularizer uses detached similarity matrices and a softmax-based alignment rather than explicitly solving the standard GW transport problem. The derivation also treats certain marginal-dependent terms as constants. These assumptions need clearer justification, especially since the proposed coupling is not obviously constrained to satisfy both transport marginals.
Questions for authors
  1. What is the main technical novelty of FGW-CLIP compared with directly adding intra-domain contrastive losses and structural regularization to CLIPZyme? Could the authors compare against simpler alternatives such as pairwise similarity matching?
  2. Can the authors provide separate ablations for the reaction-level contrastive loss, enzyme-level contrastive loss, and EC classification loss? It would also be useful to report the results without EC supervision.
  3. How does the method handle missing or incorrect EC annotations? Since EC labels are closely related to enzyme function, could this additional supervision explain much of the improvement?
  4. In Proposition 4.1, how is the softmax-based alignment guaranteed to satisfy the transport constraints required by GW? Could the authors clarify the assumptions used to simplify the GW objective?
  5. Can the authors report results over multiple random seeds and provide statistical significance tests, particularly for the relatively small improvements on the enzyme-similarity split?
  6. The paper cites EnzymeCAGE, a recent enzyme retrieval method, but does not include it in the experimental comparisons. Could the authors explain this omission and provide a comparison under a matched evaluation protocol?

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

Related papers

Discussion (0)

Sign in to comment.