acceptodds
Withdrawn from ICLR 2027

From Protein Representations to Functional Hierarchies: Reshaping Representation Geometry for Enzyme Classification

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Enzyme Commission (EC) classification assigns protein sequences to hierarchically organized functional classes and is important for biotechnology. Fine-grained EC classification remains challenging, and errors can compromise downstream functional annotation and enzyme discovery. Despite increasingly powerful protein representations, existing methods still insufficiently model the representation geometry of fine-grained EC classes: at the inter-class level, sibling classes can remain poorly separated; at the intra-class level, proteins within the same class can form heterogeneous distributions poorly captured by a single center. To address these challenges, we propose HiPro-EC, a hierarchy-guided multi-prototype framework that reshapes this geometry over frozen protein language model (PLM) representations. It reshapes inter-class structure by increasing sibling-class separation using the EC hierarchy, and models intra-class structure with multiple local prototypes to capture heterogeneous class distributions. Extensive experiments on the CARE benchmark show that HiPro-EC achieves the best performance at the finest EC level across all four official test sets, with consistent gains across multiple PLM backbones. Further analyses show clearer sibling-class separation and larger gains for heterogeneous EC classes, highlighting the value of modeling representation geometry for fine-grained EC classification.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Enzyme function prediction, EC classification, protein language models, class representation learning, representation geometry, prototype learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_qsgQOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces HiPro-EC, a framework for fine-grained enzyme classification using frozen protein language model representations. It initializes a linear projection using EC hierarchy statistics and employs multiple learnable prototypes to capture within-class heterogeneity. The projection and prototypes are jointly trained for similarity-based classification. Experiments across four CARE test sets and multiple protein encoders demonstrate consistent improvements in fine-grained prediction.

Critical strengths
  1. The method addresses two well-motivated challenges in fine-grained enzyme classification: confusion between sibling classes and heterogeneous distributions within individual classes.

  2. The framework is lightweight, training only a linear projection and class prototypes over frozen protein representations.

  3. Experiments across four CARE test sets and multiple protein language model backbones demonstrate consistent improvements in fine-grained classification.

Critical weaknesses
  1. Section 3 would benefit from clearer explanations of the statistical quantities and their roles in projection initialization. Although the variables are formally defined, the rationale for constructing the hierarchy-guided scatter from parent–child offsets, applying trace normalization, and equally aggregating hierarchy levels is insufficiently explained.

  2. The methodological novelty appears limited. The approach combines a hierarchy-aware Fisher-style projection initialization with multi-prototype classification over frozen protein embeddings, both closely related to established representation-learning techniques. The authors should clarify what methodological advances this combination offers and why the challenges of enzyme classification specifically motivate the proposed design.

Questions for authors

none

Reviewer_Y7vcOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces HiPro-EC, a hierarchy-guided multi-prototype framework for fine-grained Enzyme Commission (EC) classification using pretrained protein language model representations. The authors identify two limitations of existing representation-based classification methods: insufficient separation between closely related sibling EC classes and heterogeneous feature distributions within individual classes. To address these issues, HiPro-EC combines Hierarchy-Guided Geometry (HG), which initializes a linear projection using hierarchy-aware discriminant analysis, with Multi-Prototype (MP) learning, which represents each EC4 class through multiple trainable local prototypes. The pretrained protein language model remains frozen, allowing the method to focus on reorganizing the downstream representation space. Experiments on the CARE benchmark evaluate the framework across four test sets and multiple protein language model backbones.

Critical strengths

A major strength of this paper is its clear identification of two complementary geometric challenges in fine-grained enzyme classification, together with a relatively simple and interpretable solution. The hierarchy-guided projection provides a principled way to incorporate EC taxonomy information into representation learning, while multiple prototypes offer a natural mechanism for representing functionally equivalent but evolutionarily diverse proteins.

Critical weaknesses

The most significant limitation concerns the novelty and independent contribution of the hierarchy-guided component. The HG initialization is essentially a hierarchy-conditioned extension of classical discriminant analysis, while the multi-prototype classifier builds on established prototype-based learning techniques. Furthermore, HG is used only to initialize the projection matrix, which is subsequently optimized without an explicit hierarchy-preserving constraint. Consequently, it is unclear whether the final representation consistently maintains the intended hierarchical geometry. The ablations also indicate that MP accounts for a substantially larger improvement than HG, raising questions about how essential the hierarchy-specific design actually is. Another concern is the absence of direct experimental comparisons with several hierarchy-aware enzyme classification approaches discussed in the related work, such as GloEC and HIT-EC.

Questions for authors
  1. Contribution of hierarchy-aware initialization: Since HG only initializes the projection matrix, could the authors quantify how much hierarchical information is preserved after training? Would an explicit hierarchy-aware regularization objective improve performance beyond initialization alone?

  2. Comparison with related hierarchical methods: Could the authors compare HiPro-EC with existing hierarchy-aware enzyme classification methods such as GloEC, HIT-EC, or other relevant approaches under the same CARE evaluation protocol?

  3. Multi-prototype versus alternative initialization: How much of the improvement arises from having multiple learnable prototypes rather than from the particular spherical K-means initialization? Could the authors compare with randomly initialized prototypes, adaptive prototype allocation, or standard metric-learning objectives?

  4. Low-support enzyme classes: Appendix Table 9 shows that multi-prototype modeling can degrade performance for singleton classes. Could the authors investigate support-aware regularization or prototype allocation strategies that avoid overfitting rare EC classes?

  5. Realistic multi-label inference: The evaluation uses the ground-truth number of EC annotations to determine how many labels to predict. How does HiPro-EC perform when this number is unknown and must be inferred from predicted scores? Could the authors report precision–recall, F1, or other threshold-based multi-label metrics?

  6. Quantifying representation geometry: Beyond t-SNE visualizations, could the authors provide quantitative measures of sibling-class separation, intra-class compactness, hierarchical consistency, and their relationship with prediction accuracy?

Reviewer_rsJ8OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors present HiPro-EC, a new method for transforming the latent space learned by a pretrained protein language model into a space that is better suited for enzyme function prediction across levels of label granularity. Their method utilizes the hierarchical structure of Enzyme Commission (EC) enzyme function classifications to try to encourage separation between child nodes of the same parent while maintaining tight cohesion of within-class representations. To better enable classification of enzymes with a shared function (i.e. same label) but heterogeneous representation (possibly due to differences in global structure or sequence composition), the authors use a multi-prototype learning approach for the actual enzyme function classification.

Critical strengths
  1. The authors lay out a strong motivation for the work: that PLMs or similar self-supervised protein models are increasingly being used for enzyme function prediction, but these models learn evolutionary relationships that may not reflect more recently diverged functionality
  2. Empirical results on the CARE benchmark support the claim that the authors' method results in improved enzyme function classification across levels of functional granularity
  3. Analyses looking at performance within classes with different levels of heterogeneity and training support shed light on the settings in which their method results in classification improvements and when it does not
Critical weaknesses
  1. Empirical results all rely on a single benchmark (CARE), which is widely adopted, but does limit the generality of the conclusions
Questions for authors

Questions affecting assessment

  1. What's the source of discrepancy between CARE results reported here (Tables 1 and 2) and CARE results reported in their paper (Table 2)? e.g. for BLASTp on Promiscuous L4 accuracy, HiPro-EC reports 68.0% accuracy but CARE reports 93.7
    • Evaluation metric may explain some of the differences (HiPro-EC reports top-k accuracy where k is the number of true EC annotations for the query protein, whereas CARE seems to report top-1 accuracy)
  2. At which layers were representations extracted for PLM baselines? There's strong empirical evidence that final layer representations can be suboptimal for function prediction (see Fig 1D of the ESM-C/Candido et al paper). Experimenting with intermediate layer representations could result in both stronger baselines and stronger performance for HiPro-EC on top of those representations

Suggestions not affecting assessment

  • In Figure 3 (performance across PLM backbones), it would be nice to include a bar showing the performance of the PLM backbone alone, for reference
  • Figure 5 lacks any annotation of what is actually being shown (e.g. t-SNE, UMAP, etc). It seems likely that it's t-SNE based on other figures, but would be good to state explicitly
  • Additional details of the MLP head used for the PLM-based baselines (e.g. number of layers) would aid reproducibility of the work
  • Would it be possible to adaptively select the number of prototypes used per class? Tables 4 and 9 (performance on classes stratified by heterogeneity and training support) suggest that using too many prototypes introduces unnecessary flexibility for simpler cases. It seems like the training set could be used to learn a rule for mapping class statistics (e.g. heterogeneity, support) to a number of prototypes, although this could be expensive if it requires an outer loop for hyperparameter search

Summary of review

Overall, I think the empirical results here are strong and support the assertion that HiPro-EC is an advancement in using protein representation learning methods for enzyme function prediction. The methods are likely of interest both for protein researchers and researchers working in other domains with hierarchical classification schemes. My assessment of weak accept reflects the empirical results and clarity of writing, as well as the limitation of evaluation driven by relying on a single (well-curated) benchmark.

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

Related papers

Discussion (0)

Sign in to comment.