acceptodds
Withdrawn from ICLR 2027

SoCoBio: Socialized Coevolution Based Hierarchical Biomolecular Function Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Biomolecular function prediction assigns sequences to functional taxonomies whose data-rich ancestors branch into a long tail of data-scarce leaf functions. Previous methods do not treat vertical generalization, which carries evidence from ancestors to tail classes, and horizontal discrimination, which distinguishes a class from its most similar siblings, as distinct interactions. To tackle this, we introduce SoCoBio, a hierarchical multi-label classifier for biomolecules that treats each class prototype as a learner in the sense of socialized learning. Whereas static hierarchical models weight relatives by a fixed rule, SoCoBio learns two interactions among prototypes, namely a Vertical Controller that passes gated information from each parent to its children and a Horizontal Controller that then lets each class select messages from a few similar siblings. A class thus receives shared context before it weighs messages from its closest siblings. Viewing SoCoBio as a within-predictor special case of socialized learning, we derive a sufficient condition for a learned update to lower the multi-label loss and a per-class direction bound whose required positive support grows only logarithmically with the number of classes. As a complete predictor on peptide and enzyme benchmarks and an independent enzyme test set, SoCoBio attains the highest leaf-level and mean-level F1 among the compared methods, and matched retraining with a short budget on the largest benchmark shows that its leaf-F1 gain from interaction, concentrated on tail classes, requires both controllers.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
multi-label, socialized coevolution, biomolecular, hierarchical classification

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_Y6mcOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SoCoBio for hierarchical multi-label classification in biomolecular function prediction. It treats each class prototype as a learner and treats parent and sibling classes as collaborators. In this way, “whom to learn from and how much to learn” becomes class prototype collaboration inside a single predictor. The method has two parts: a vertical controller and a horizontal controller. The vertical controller uses gated parent-child calibration to pass ancestor information. The horizontal controller uses Gumbel-Softmax to sparsely select same-level peers and then performs gated aggregation. Theoretically, it gives a sufficient condition for the joint update to reduce BCE loss, and a direction gain bound determined by the positive sample support of each class. Experiments are on therapeutic peptides, CLEAN-50, CLEAN-100, and an independent enzyme test set. The authors report better leaf-level and mean-level F1 than baselines, especially on long-tail classes and strict path consistency. Ablations show that the vertical controller, horizontal controller, and learned sibling selection are all necessary.

Critical strengths
  1. Clear problem decomposition. The paper separates parent-class generalization from sibling-level discrimination via vertical and horizontal controllers. This directly targets the core difficulty of hierarchical multi-label prediction and makes the method design well motivated.

  2. Strong mechanism evidence. The 0/V/H/J ablation and peer-selection study show gains mainly come from learnable sibling selection, not just gating. Error analysis further links gains to fewer cross-parent and multi-leaf errors, supporting the claimed mechanism beyond raw F1.

Critical weaknesses
  1. Limited conceptual novelty. The "socialized learning / agents" framing seems stronger than the actual methodological novelty. The class prototypes share one encoder, one prediction head, and one loss, rather than operating as independent agents. Stripped of terminology, the method is closer to gated parent-child message passing plus sparse peer prototype interaction. Its distinction from existing hierarchy-aware message-passing methods remains unclear, and the claimed conceptual novelty is weaker than the framing suggests.

  2. Incomplete empirical validation. The paper lacks direct comparisons with hierarchy-aware enzyme predictors such as GloEC, EnzymeHunter, and HIT-EC, limiting the strength of its performance claims. Some gains are small, and the independent halogenase subset performs worse than FGVC. The paper also does not establish homology-separated evaluation, so the evidence for long-tail robustness and generalization is not yet sufficient to support the stronger performance claims.

Questions for authors
  1. Beyond the terminology, what concrete design decision or testable prediction does the "socialized learning" view provide that distinguishes SoCoBio from a standard hierarchical GNN or gated message-passing model?

  2. Could you clarify why SoCoBio was not compared with GloEC or another strong hierarchy-aware enzyme predictor under the same ESM-2 features and data splits? If such a comparison is not feasible, what is the main obstacle?

  3. Does the horizontal controller's peer selection depend on the input sequence? The peer scores in Eq. (4) appear to depend only on prototypes, so what exactly is "dynamic" in the dynamic interaction?

  4. Given that Theorem 3.2 requires sufficient positive sample support, what practical guidance does it offer for tail classes with very few positive samples? Does the empirical analysis provide evidence that the theorem is relevant to these classes?

Reviewer_saa2OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper introduces SoCoBio, a new method to predict protein function. The main novelty of the method is treating protein function labels as a hierarchy graph, making a prediction based on the hierarchy rather than simple protein -> function mapping. The proposed method is evaluated on two datasets: Peptide and CLEAN-50/CLEAN-100, and the theorems are provided to justify the performance.

While the results indicate SoCoBio can outperform other methods in some cases, the results are not very convincing and their significance is not clear. The contributions of the paper are not clear as the paper is hard to read and understand. Key proposed baselines seem to be missing. The experimental setup is not clear. Therefore, the significance of the experimental results is hard to judge. In the current state the paper is not ready.

Critical strengths
  • Protein function prediction is an important and largely unsolved problem, any well-supported progress can be impactful

  • The proposed method is compared against a rich set of baselines

  • Appendix provides many details

Critical weaknesses
  1. The main contributions, Vertical Controller and Horizontal Controller, seem to be at least partially reinventing graph neural networks (GNN)/message-passing neural networks. A comparison to some well-established GNN architectures such as GCN (https://arxiv.org/abs/1609.02907) or GAT (https://arxiv.org/abs/1710.10903) is missing.

  2. Statistical significance of the results is not measured. It is important since improvements on the Peptide dataset are not very substantial for example in F1. There is an attempt for this in section C.2 but it is still does not explicitly show that results are statistically significant.

  3. Text is difficult to understand. For example: [line 026, line 106] It would be good to mention the specific F1 and how much higher it is compared to other methods. [line 039] Typo “:” -> “.” [Figure 1A] The blue and dark blue colors seem to be swapped by mistake (as text annotations are opposite to color legend)? [Figure 1B] It is not clear what the x axis shows [Line 160] What is v? [Equation 7] What is L_{Proto}. It is not defined [Line 269] What is u? [Section 3.6] The intuitive explanation of what theorems demonstrate is missing. Section 3.6 asks “WHEN DOES PROTOTYPE COLLABORATION HELP?” but the question does not seem to be answered in the section’s text. Table 1 and Figure 4 have many shortcuts/acronyms. It would be good to explain them, for example, in the caption.

The clarity is a major issue, which is not easy to address.

Questions for authors

What exactly are the CLEAN-50 and CLEAN-100 datasets? The official CLEAN publication (https://www.science.org/doi/10.1126/science.adf2465) has different dataset naming and the performance numbers do not seem to match. The evaluation should ideally be exactly consistent with the official CLEAN publication, or the discrepancy should at least be clearly explained in the main text.

Reviewer_jv1pOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper aims to exploit the natural hierarchy that characterizes biological labels, which exhibit a horizontal dependency derived from evolution and a vertical dependency derived from ancestral information. Given this structure, the paper introduces a social coevolutionary training framework based on message-passing networks that learns hierarchically related embeddings, encoding both the label hierarchy and integrated biological features.

The proposed framework, SoCoBio, combines ESM-2 sequence representations with explicit biochemical features and maintains a learnable prototype for each function class. The loss utilizes a vertical controller that transfers gated parental information and a horizontal controller that selects peer messages to update these prototypes. Predictions combine sequence–prototype similarity with a linear classification head.

Linear classification heads trained on the extracted features, together with sequence–prototype similarity, are evaluated on two distinct tasks: hierarchical multi-label prediction of therapeutic peptide activities and enzyme functions under the EC taxonomy.

SoCoBio achieves the highest leaf-level and mean-level macro-F1 among compared methods across the three main benchmarks and the independent enzyme test set. The improvements are generally modest and vary across datasets, with larger benefits observed in low-resource settings and no uniform advantage across external subsets.

The authors also provide component ablations, controlled comparisons of peer-selection mechanisms, and evaluations focusing on rare classes, hierarchical consistency, and external generalization. These analyses support the individual contributions of both controllers and learned peer selection.

Critical strengths

The coevolutionary prior evaluated in this work is currently underexploited in representation learning. Moreover, the addressed problems are of significant practical relevance to the community.

The paper is well-written and motivated, establishing a clear bridge between the nature of biological data and the AI methodology employed. Additionally, separating parent–child transfer from learned same-level interactions provides a strong architectural motivation.

The transparent experimental setup, clear description of baselines, and detailed dataset compositions make the evaluation easy to follow. The ablation section, controlled comparisons, and theoretical analysis provide valuable insight into the individual components and the role these embeddings play in classification.

Critical weaknesses

The biological interpretation of learned interactions remains uncertain. Same-level similarity and message exchange do not necessarily establish true evolutionary or functional relationships, and the reported analysis does not demonstrate a clear connection between these learned relations and improved predictive performance.

While the proposed architecture directly adopts a parent–child-inspired graph network, it lacks a comparison against simpler or more naive graph baselines.

Finally, the connection between the theoretical analysis and the primary evaluation metrics remains limited. The theoretical guarantees address binary cross-entropy and prototype-update directions, whereas the main empirical claims focus on macro-F1. Consequently, the theoretical analysis does not directly explain when or why these guarantees translate in those gains.

Questions for authors

a. Could a simpler hierarchical graph baseline or gated prototype aggregation achieve comparable performance when using the same encoder, biological features, and training objectives?

b. How does the method handle missing or partially specified EC labels? Are unannotated functions treated as implicit negatives, and could this introduce false-negative supervision or distort learned class relationships?

c. Can you quantify which specific performance gains arise from parent–child transfer versus same-level interactions, or their combination? Does their relative importance vary according to hierarchy depth or class frequency?

d. How does the model handle promiscuous enzymes that carry multiple EC annotations, particularly when those functions belong to distant branches of the taxonomy?

e. can you give more details on the Therapeutic Peptide dataset you constructed.

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

Related papers

Discussion (0)

Sign in to comment.