acceptodds
Under review as a conference paper at ICLR 2027

Many-to-many Relationship Modeling via Bi-order Entailment Cones

Abstract

Hyperbolic entailment cones are effective for modeling hierarchical and multimodal relationships and have been applied to diverse fields. However, existing multimodal formulations typically assign one modality to be universally more general than the other. This one-sided inductive bias is suitable for one-to-many relationships, but becomes suboptimal in many-to-many settings such as recommendation, protein–ligand binding, and protein function annotation, where neither modality is inherently more general. To alleviate this limitation, we propose Bi-Order Entailment (BiOE), a product-manifold framework that decomposes many-to-many relationships into directional latent orders modeled by specialized hyperbolic entailment factors. To handle noisy and incomplete data, BiOE further introduces confidence-weighted pseudo-order supervision from neighborhood and semantic-cluster coverage, together with a lightweight residual compatibility factor for non-hierarchical interactions. We theoretically and empirically show that single-entailment models cannot faithfully capture both compatibility and induced order structure, and that they introduce direction-dependent retrieval biases. Across recommender systems, protein–ligand binding, and protein function prediction, BiOE significantly improves over hyperbolic baselines and single-entailment variants. Our experiments also introduce new hyperbolic downstream pipelines, including the first hyperbolic Collaborative Filtering–LLM alignment method and one of the first hyperbolic gene ontology classification models

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Hyperbolic Geometry, Multi-Modal, Recommendation Systems, Ai4Sci

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_ohmLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes BiOE for predicting many-to-many relations. It uses separate hyperbolic spaces to model the two entailment directions. The full model also adds order information estimated from the data and a residual component. Experiments cover recommendation, protein-ligand retrieval, and protein function prediction.

Critical strengths
  1. Using separate spaces for the two directions is a reasonable design. It avoids forcing one type of entity to always be above the other in the hierarchy.
  2. The method is tested on three different tasks and improves several metrics. The recommendation results provide useful evidence that modeling both directions can help.
Critical weaknesses
  1. The reason for using hyperbolic space needs to be clearer. Many-to-many relations do not necessarily form a hierarchy. The authors should explain why the relations in these datasets are suited to hierarchical modeling in hyperbolic space, and provide evidence from the data to support this choice.
  2. The comparisons could better show the benefit over existing graph models. LightGCN is already included. The authors should also consider comparisons with heterogeneous GNNs, which can model different node types and relations. Using the same input information and similar model sizes would help clarify what the proposed method adds.
Questions for authors
  1. The reason for using hyperbolic space needs to be clearer. Many-to-many relations do not necessarily form a hierarchy. The authors should explain why the relations in these datasets are suited to hierarchical modeling in hyperbolic space, and provide evidence from the data to support this choice.
  2. The comparisons could better show the benefit over existing graph models. LightGCN is already included. The authors should also consider comparisons with heterogeneous GNNs, which can model different node types and relations. Using the same input information and similar model sizes would help clarify what the proposed method adds.
3 comments
AuthorsOpenReview ↗

Response to Reviewer ohmL (Part 1 of 2)

We thank the reviewer for validating the design and breadth of our method and experiments. We provide clarifications and additional results below.

(Q1, W1) The reason for using hyperbolic space needs to be clearer

We’d like to first highlight that hierarchy and many-to-many structure are two separate properties that motivate two separate design choices. The hierarchical structure present in these datasets (discussed below) is what motivates hyperbolic models in the first place. The many-to-many nature of the relations is what motivates BiOE as prior hyperbolic baselines face limitations: as neither side is uniformly more general than the other, a single entailment direction imposes the wrong inductive bias, which BiOE corrects by modeling both directions. The motivation for using hyperbolic geometry in these tasks come from the structures of the underlying dataset and domain.

In recommender systems, interactions are long-tailed, where there are a few popular items that are consumed by most users with a long tail of niche items, and a few highly active users interact broadly while most users are much less active. This gives a tree-like structure, with few general nodes and many specific ones, on both sides. Below we show δ\delta-hyperbolicity values for all three recommendation datasets used in the paper. δ\delta-hyperbolicity measures how similar a dataset is to a perfect tree, with lower values meaning more tree like.

DataAmazonYelpSteamPerfect TreePoincare Space
δ\delta-hyperbolicity (↓\downarrow)0.3160.3190.2080.00.10±0.080.10\pm0.08

As shown, the 3 datasets have low hyperbolicity values, indicating the relationships are tree like. Furthermore, to quantify the long-tail distribution of items and users, we employ the Gini coefficient, which is 0 when every node has the same number of interactions, and approaches 1 as interactions concentrate on a few nodes. Additionally, we measure the nestedness index (NODF) [a] to test that this long tail induces hierarchical structure. NODF measures how often the interactions of tail nodes are contained in those of head nodes, with high nestedness means the relations form a general-to-specific hierarchy. We compare each dataset with a random graph with the same numbers of users, items and interactions.

DatasetGini, users / items (random) (↑\uparrow)NODF ratio to random, users / items (↑\uparrow)
Amazon-Book0.45 / 0.37 (0.17 / 0.16)3.5× / 4.0×
Yelp0.43 / 0.39 (0.14 / 0.14)3.5× / 3.9×
Steam0.40 / 0.67 (0.15 / 0.07)7.9× / 5.4×

As shown above, the interactions in the 3 datasets form non-trivial long-tail distributions as shown by the Gini coefficients and the long-tail distributions also induce a general-specific hierarchical structure as shown by the NODF ratios.

In protein function annotation, the GO labels are themselves organized as a curated hierarchy. It is a directed acyclic graph of general to specific functions, and a protein annotated with a term is also annotated with all of its ancestors. As a result, the proteins of a specific function are contained within those of its more general ones, while multifunctional proteins carry annotations that cover those of more specialized proteins, so the hierarchy extends to both sides of the relation. Below shows the δ\delta-hyperbolicity, Gini coefficients, and NODF values for the protein annotation dataset, demonstrating that the protein-GO dataset is tree-like, exhibits long-tail distributions, and the long-tail distribution induces a general-specific hierarchy

MeasureProtein–GO
δ-hyperbolicity0.248
Gini, proteins (random)0.43 (0.24)
Gini, GO terms (random)0.64 (0.04)
NODF ratio to random, proteins4.6×
NODF ratio to random, GO terms3.2×

Protein ligand binding also exhibits hierarchical structure, which prior work has used to motivate hyperbolic models of binding (Wang et al., 2026, Hypseek). Proteins are organized into families and superfamilies with conserved binding sites, so a ligand that binds a conserved site can bind many members of a family, while a selective ligand binds only a few of them. Likewise, a target can bind a broad range of ligands that covers the binders of more specialized targets. We have added these measurements and this discussion to Appendix C of the revision (Table 9), and refer to them at the beginning of Section 5.

[a] Mário Almeida-Neto, Werner Ulrich. A straightforward computational approach for measuring nestedness using quantitative matrices. Environmental Modelling & Software, Volume 26, Issue 2, 2011.

AuthorsOpenReview ↗

Response to Reviewer ohmL (Part 2 of 2)

(Q2, W2) Comparison with heterogeneous GNNs

We thank the reviewer for this suggestion. BiOE models are based on HICF, which is a hyperbolic LightGCN variant, and BiOE loss terms operate only on the observed user item interactions, so it uses a single type of relation and requires no additional node types, edge types, attributes or knowledge graph relations. We therefore compare BiOE to two additional heterogeneous GNNs, Simple-HGN[b] and R-GCN[c], on exactly the same interaction graph and data splits, with the two interaction directions (user to item and item to user) as their only relation types. Both heterogeneous GNNs use embedding dimension 256, the same as HICF and BiOE. The results are average over 3 runs, where STD is omitted for presentation clarity. The results are shown below, where BiOE outperforms both heterogeneous GNNs in nearly all cases, demonstrating that BiOE can outperform heterogeneous GNNs even with a simpler backbone. These results are added to Table 1 and Section 5.1 of the revision.

ModelAmazon-bookYelpSteam
R@5R@20N@5N@20R@5R@20N@5N@20R@5R@20N@5N@20
R-GCN0.04440.11180.04440.06680.03020.08520.03520.05330.04610.12290.05140.0772
+ LLM Alignment0.05020.11390.04620.06690.02980.08880.03570.05600.04630.12620.05230.0804
Simple-HGN0.08710.17670.08830.11040.04480.12130.05190.07720.05480.13960.06010.0889
+ LLM Alignment0.09150.18010.09060.11290.04490.12490.05110.08030.05600.14610.06470.0920
BiOE$_{Bi}$0.08100.17610.08250.11350.04670.13170.05480.08240.06030.14840.06360.0996
+ LLM Alignment0.08330.18190.08520.11700.04910.13470.05700.08500.06510.15360.06960.1026
BiOE$_{PO}$0.08410.17880.08280.11450.04910.13300.05590.08370.06310.14950.06510.1010
+ LLM Alignment0.08620.18490.08590.11830.05030.13660.05810.08700.06880.15510.07220.1047

To ensure a fair comparison, we tune both heterogeneous GNN baselines extensively with grid search, with the search ranges listed in Appendix B.4.1 of the revision

[b] Qingsong Lv, Ming Ding, Qiang Liu, Yuxiang Chen, Wenzheng Feng, Siming He, Chang Zhou, Jianguo Jiang, Yuxiao Dong, Jie Tang. Are we really making much progress? Revisiting, benchmarking, and refining heterogeneous graph neural networks. KDD 2021.

[c] Michael Schlichtkrull, Thomas N. Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov, Max Welling. Modeling Relational Data with Graph Convolutional Networks

AuthorsOpenReview ↗

Follow Up and Kind Reminder

Dear Reviewer ohmL,

Thank you again for your detailed review and insightful suggestions. As the discussion period is coming to an end, we would like to kindly follow up on our response. We have added additional heterogenous GNN baselines and demonstrated BiOE advantages in comparison, and added several numerical and intuitive analysis demonstrating the presence of hierarchical structures in all datasets used in the paper to justify the use of hyperbolic geometry. We hope they have fully addressed your concerns; if so, we would greatly appreciate it if you would consider updating your assessment. We truly appreciate your time and effort in reviewing our paper.

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

The authors focus on improving existing multimodal hyperbolic entailment cone models that assume one modality is more general than others, which is poor for many-to-many settings such as recommendation, protein-ligand binding, and protein function annotation. They propose a product-manifold framework, so-called Bi-Order Entailment (BiOE), in which many-to-many relationships are split into directional latent orders, each encoded with a hyperbolic entailment factor. BiOE is enriched with confidence-weighted pseudo-order supervision from neighborhood and semantic-cluster coverage, as well as a lightweight residual compatibility factor for non-hierarchical interactions. Experiments focus on tasks such as recommendation (HICF backbone), protein-ligand binding (hyperbolic LigUnity backbone), and protein function prediction (hyperbolic StarGO backbone). The proposed method outcompetes baselines and single-entailment variants, while introducing new hyperbolic pipelines.

Critical strengths

The study focuses on the well-motivated and clear problem of fixed cone direction inside entailment-cone models, showcasing its limitations for many-to-many data. The proposed solutions are also kept simple by design and are widely applicable to three different domains and backbones, offering some performance gains. The experimental setup (no entailment/single direction/BiOE) also clearly isolates the effect of the entailment direction.

Critical weaknesses
  • No statistical tests are reported for the performance improvements, which are in several cases incremental.
  • Additionally, most baselines form hyperbolic adaptations by the authors (e.g., HyperGO), not standard SOTA methods per application.
  • No complexity/scalability aspects are analyzed, while building neighborhood-level pseudo-orders and cluster coverage can be expensive in practice.
Questions for authors
  1. Consider comparing BiOE with strong SOTA recommenders on the same benchmark datasets.
  2. Please show the statistical significance of the obtained results for BiOE and the baselines.
  3. How do the methods’ main components scale to large graphs?
  4. Is a perfect hyperbolic model (radius monotone in generality) a strong requirement for the derived theoretical guarantees concerning bias analysis? To what extent do the learned embeddings satisfy this monotonicity?
3 comments
AuthorsOpenReview ↗

Response to Reviewer AGDx (Part 1 of 2)

We thank the reviewer for recognizing that our paper solves a well defined and motivated problem, and that our method is widely applicable. We provide clarifications and additional results below.

(W1, Q2) Statistical Significance

For protein function annotation, Table 3 already reports the mean and standard deviation over 3 runs. The gains of BiOE_PO over the strongest hyperbolic baseline (HyCoGO) and the Euclidean SOTA (StarGO) are at least 3.6x the larger standard deviation in Micro AUPR, Macro AUPR, and Fmax, and a one-sided Welch t-test over the hyperbolic baselines gives p < 0.01 for each of these metrics. BiOE also improves over the strongest baseline by at least 4.7% in all metrics except AUC, and over the single entailment variants with the same number of parameters by up to 36.4%, showing that its gains are not marginal.

For recommendation, Table 1 reports the average of 3 runs, with standard deviations omitted for clarity and consistency with prior Euclidean work, as noted in its caption. Table 7 in the appendix of the revision now reports standard deviations on all three datasets, and the improvements of BiOE_PO over the strongest baseline are larger than the standard deviations in 11 of the 12 columns. The protein-ligand binding experiments are computationally intensive, and we follow prior work in reporting a single training run (Feng et al., LigUnity; Wang et al., HypSeek). Nevertheless, BiOE outperforms the single entailment baselines by up to 28.4%, demonstrating that the improvements are not marginal. Overall, our experiments over three distinct tasks demonstrate that BiOE consistently improves over the hyperbolic baselines with statistical significance.

(W3, Q3) Scaling to Large Graphs

Below we show how the pseudo-order preprocessing time grows with graph size. We build copies of Amazon and Steam at 1 to 32 times their original size by replicating every user and item, which preserves the degree structure of each graph. We chose these two datasets because they are similar in size but differ in shape (Amazon has roughly as many users as items, while Steam has about 4.5 times more users than items), and their largest copies reach 3.9M and 10.1M training edges. Across all scales, preprocessing takes at most as long as 4.3 LightGCN training epochs, measured on NVIDIA A6000 GPUs with batch size 1024. This is a small fraction of a full training run, and the fraction generally shrinks as the graph grows, falling to 0.53 and 0.29 epochs on the largest Amazon and Steam graphs.

1x2x4x8x16x32x
Amazon-Books# Edges0.12 M0.24 M0.48 M0.96 M1.93 M3.85 M
PO Processing time42 s1.5 min2.8 min6.8 min18.7 min38.0 min
Equivalent LightGCN epochs4.33.92.51.72.30.53
Steam# Edges0.32 M0.63 M1.26 M2.53 M5.06 M10.1 M
PO Processing time58 s2.2 min5.1 min13.8 min34.7 min71.6 min
Equivalent LightGCN epochs2.01.62.60.770.480.29

We have added this analysis to Section 5.1 of the revision, with the full results in Table 6 of Appendix B.4.1.

(Q4) Is radius monotonicity a requirement and do learned embeddings satisfy this monotonicity

The perfect hyperbolic model formalizes the property that hyperbolic entailment models are trained to have. Since the cone aperture shrinks as the radius grows, the entailment loss pushes general entities toward the origin and specific ones outward. For this reason, prior work uses radius as the measure of generality (Desai et al., 2023). This also matches the classical result that trees embed in hyperbolic space with arbitrarily low distortion when the root is placed at the origin (Sarkar, 2011). Analyzing this ideal outcome isolates the bias that each entailment direction induces by design, and the proof uses the assumption only through its consequence that head targets lie at smaller radii than tail targets. In practice, noisy and incomplete relations and competing objectives can leave some head targets farther from the origin than some tail targets. Proposition 4.2 could be relaxed by letting ϵ\epsilon be the probability that a random head target has a larger radius than a random tail target. Because the bias averages over head and tail pairs, correctly ordered pairs still push it in the predicted direction, while violated pairs can push it the other way. Since hyperbolic distance changes at most at unit rate along a radial geodesic, the effect of the violations is bounded and linear in ϵ\epsilon.

[response continues]

AuthorsOpenReview ↗

Response to Reviewer AGDx (Part 2 of 2)

(Q4) Radius monotonicity Continued

Empirically, the learned embeddings largely satisfy monotonicity. In Figure 3, which shows the radius distributions of BiOE_PO on protein annotation, the cone-owning modality in each factor concentrates near the origin, and within each modality, head proteins and GO terms generally lie at smaller radii than their tail counterparts. To test whether radius reflects generality beyond popularity, we use the GO hierarchy, in which every ancestor term is more general than its descendants. Among ancestor-descendant pairs whose annotation counts are within 5% of each other, the ancestor lies closer to the origin in 79.3% of cases in the GO to protein factor. Thus, the learned radii follow the GO hierarchy even when popularity cannot distinguish the two terms, suggesting that the monotonicity assumed in Proposition 4.2 largely holds in practice. This analysis is also added to Section 5.4 of the revision.

(W2, Q1) Additional Baselines, Hyperbolic Baselines are adaptations by us

Our experiments are designed to show that bi-order entailment improves over single entailment for many-to-many relationships. BiOE is a general framework instead of a model designed for any single application, and we do not focus on proposing a new SOTA model for each task. Our experiments therefore compare BiOE against single entailment directions on the same backbone and with the same number of parameters, so that any difference in performance comes from the entailment structure itself. Nevertheless, the backbones we build on are strong recent models in each application, many of which are the current Euclidean SOTA:

  • For protein-ligand binding, we compare with LigUnity (Patterns 6(10), 2025), the Euclidean SOTA model, and both the hyperbolic baseline and the BiOE variants outperform it in most metrics. We also report HypSeek (AAAI 2026) for reference, the hyperbolic SOTA model, which uses an additional modality in training and seven loss terms. Even so, BiOE_PO outperforms HypSeek by 27.3% in EF(1%) on PCBA with fewer modalities and a simpler model.
  • For protein function annotation, we compare with StarGO (Bioinformatics 42(4), 2026), a recent Euclidean SOTA model, and BiOE outperforms it in all metrics except the saturated AUC. Our hyperbolic baselines are also not weaker substitutes for the Euclidean SOTA. We developed HyperGO and HyCoGO by adapting StarGO to hyperbolic space, and to our knowledge they are the first hyperbolic models for protein function annotation. Both outperform StarGO, with HyCoGO improving by 2.9% in Micro AUPR, 6.4% in Macro AUPR, and 5.1% in Fmax. We therefore consider these models a contribution of the paper, and BiOE improves on both of them further.
  • For recommendation, we build on FACE (NeurIPS 2025), a recent SOTA framework for aligning collaborative filtering embeddings with LLM semantics. We extend it to hyperbolic space, which to our knowledge is the first method for aligning hyperbolic collaborative filtering models with LLMs, and show that hyperbolic models improve performance and BiOE improves it further. Below, we further compare BiOE with two heterogeneous GNNs, R-GCN and Simple-HGN, trained on the same interaction graph with and without LLM alignment. BiOE_PO outperforms both in all metrics on Yelp and Steam, and in R@20 and N@20 on Amazon-book, demonstrating that BiOE can outperform heterogeneous GNNs even with a simpler backbone. These results are added to Table 1 and Section 5.1 of the revision.
ModelAmazon-bookYelpSteam
R@5R@20N@5N@20R@5R@20N@5N@20R@5R@20N@5N@20
R-GCN0.04440.11180.04440.06680.03020.08520.03520.05330.04610.12290.05140.0772
+ LLM Alignment0.05020.11390.04620.06690.02980.08880.03570.05600.04630.12620.05230.0804
Simple-HGN0.08710.17670.08830.11040.04480.12130.05190.07720.05480.13960.06010.0889
+ LLM Alignment0.09150.18010.09060.11290.04490.12490.05110.08030.05600.14610.06470.0920
BiOE$_{Bi}$0.08100.17610.08250.11350.04670.13170.05480.08240.06030.14840.06360.0996
+ LLM Alignment0.08330.18190.08520.11700.04910.13470.05700.08500.06510.15360.06960.1026
BiOE$_{PO}$0.08410.17880.08280.11450.04910.13300.05590.08370.06310.14950.06510.1010
+ LLM Alignment0.08620.18490.08590.11830.05030.13660.05810.08700.06880.15510.07220.1047

To ensure a fair comparison, we tune both heterogeneous GNN baselines extensively with grid search, with the search ranges listed in Appendix B.4.1 of the revision.

AuthorsOpenReview ↗

Follow Up and Kind Reminder

Dear Reviewer AGDx,

Thank you again for your detailed review and insightful suggestions. As the discussion period is coming to an end, we would like to kindly follow up on our response. We have added additional statical details for our experiments that demonstrate that our improvements are not marginal, added analysis for how pseudo-order scales with graph size that shows the computation time is a tiny fraction of the total training time, and clarified our comparisons with SOTA models and added additional heterogenous GNN baselines. We hope they have fully addressed your concerns; if so, we would greatly appreciate it if you would consider updating your assessment. We truly appreciate your time and effort in reviewing our paper.

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

This paper is about many-to-many relationships embedded in hyperboic space, including proteins and GO terms. Existing models encode hierarchy with entailment cones and require items to fall inside a cone owned by a more general one. This paper is about what happens when neither side is more general and what are the consequences. The authors propose in thisss case to distribute the embedding across sevral hypoerbolic factors and show that it outperforms base models.

Critical strengths

I am an out of field reviewer and not familiar with the state of the art in hyperbolic representation learning, so this is an assessment of the paper’s internal logic rather than its novelty or its standing relative to prior work.

  1. The problem is well defined. The assumption being challenged, that one side of a multimodal relation is universally the more general, is stated precisely, and the settings where it fails are concrete and recognizable, including protein function annotation.

  2. The proposed solution follows from the diagnosis rather than being added on top of it. Separating the two entailment directions into different hyperbolic factors is motivated by the fact that two points cannot entail each other within a single space.

  3. Each claim is paired with matching evidence. The predicted direction dependent bias is tested directly through a head and tail item split (Figure 2) rather than inferred from aggregate performance, and the embedding visualization (Figure 3) shows the two factors specializing in opposite directions as intended.

  4. The evaluation covers three unrelated application domains with ablations on each component, and the authors report that single direction entailment sometimes performs worse than no entailment at all, which is what their own argument predicts.

To the extent I can judge, the approach appears sound, but I cannot verify the proofs in Appendix A or assess whether the chosen baselines are the appropriate comparisons.

Critical weaknesses

I am out of field and cannot assess novelty, the proofs, or whether the baselines are current. On that basis I have not identified a weakness that would support rejection. I have one concern from my own area, which I have raised as a question rather than a weakness, since I am not positioned to judge how consequential it is.

Questions for authors

The pseudo-order is built from how much of one entity’s observed neighborhood is covered by another. In GO annotation data, the number of annotations an entity carries largely reflects curation effort and how well studied a protein is, rather than biological generality. Have you checked whether the learned order correlates with annotation count, and whether the head and tail radius pattern in Figure 3 survives controlling for it?

2 comments
AuthorsOpenReview ↗

Response to Reviewer siAY

We thank the reviewer for recognizing that our paper solves a well defined problem with solution that follow from evidence. We provide clarifications and additional results below.

(Q1) Does the learned order correlate with annotation count?

We thank the reviewer for this question. First we’d like to clarify that in our setting, generality is defined by the observed relation and BiOE learns the order implied by the data it is trained on. In recommender systems, this represents how active a user is (e.g. how many items has this user interacted with) and how popular/desirable an item is. However, as noted by the reviewer, the observed data is biased by curation effort in certain fields, e.g. annotation counts in GO also reflect how well studied a protein is on top of how many function a chain has. This bias could be partially represented in Figure 3, as generality here is defined through annotation count. As a result, below we perform further analysis of our model to test whether the learned order represents more than just curation effort.

We split each protein's annotation count into breadth, the number of its most specific annotated GO terms (annotated terms with no annotated descendant), and specificity, the mean depth of those terms in the GO hierarchy. Specificity is a proxy for curation effort, since deep GO terms require more experiments, so better studied proteins in general have deeper annotations [a]. In the protein to GO factor, we measure each protein's closeness to the origin (inverse of its radius), where closer means more general. We find that the Spearman correlation of closeness to origin v.s. breadth is significantly higher than closeness to origin v.s. specificity (0.57 v.s. 0.12 on proteins in the training set and 0.50 v.s. 0.14 for held-out proteins), suggesting that the order BiOE learns is driven mainly by how many distinct functions a protein has, instead of how extensively it has been studied. Furthermore, for GO terms, we examine the relationship between the learned order and ancestor-descendant pairs in the GO hierarchy. For a given GO term, its ancestors are the more general terms above it in the GO hierarchy, and its descendants are the more specific terms below it in the hierarchy. Among ancestor-descendant pairs where the annotation count for the ancestor and descendant are nearly equal (within 5% of each other), we find that in 79.3% of the cases the ancestor lies closer to the origin than its descendant in the GO to protein factors, suggesting that BiOE's order over GO terms captures the generality encoded in the GO hierarchy itself instead of annotation counts. For example, "RNA polymerase activity" and its descendant "5'-3' RNA polymerase activity" both annotate exactly 227 proteins, yet BiOE places the former closer to the origin. We have added this analysis to Section 5.4 of the revision.

Nevertheless, GO annotations of proteins are curated, so some curation bias is inherent to any model trained on them and cannot be fully isolated from annotation data alone. However, our analyses still show that BiOE's learned order goes beyond just annotation frequency and captures the inherent generality of proteins and GO terms.

[a] Alexandra M. Schnoes, David C. Ream, Alexander W. Thorman, Patricia C. Babbitt, Iddo Friedberg. "Biases in the experimental annotations of protein function and their effect on our understanding of protein function space", PLoS Computational Biology 9(5), 2013

AuthorsOpenReview ↗

Follow Up and Kind Reminder

Dear Reviewer siAY,

Thank you again for your review and insightful questions. As the discussion period is coming to an end, we would like to kindly follow up on our response. We have conducted analysis to further separate the bias induced by curation effort and annotation count from the learned order by BiOE on the protein annotation experiment. We hope they have fully addressed your concerns; if so, we would greatly appreciate it if you would consider updating your assessment. We truly appreciate your time and effort in reviewing our paper.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Our Rebuttal and Manuscript Revisions

We thank the reviewers for their careful reviews and the AC for facilitating the discussion. Our work shows that hyperbolic entailment cones, which assume one modality is more general than the other, are not suited for many to many relationships in hierarchical data, where they cannot capture both induced orders and introduce direction dependent retrieval biases. We introduce BiOE, which models the two directional orders in separate hyperbolic factors and adds pseudo order supervision and a residual compatibility factor. We support the design with theoretical results and evaluate it on recommendation, protein ligand binding and protein function annotation, where it consistently improves over strong hyperbolic and Euclidean baselines. The reviewers note the well defined and motivated problem (siAY, AGDx), the broad applicability of the method (siAY, AGDx, ohmL), and experiments that clearly isolate the effect of the entailment direction (siAY, AGDx).

The reviewers' main concerns, and our responses, were as follows.

  • Comparison with heterogeneous GNNs (ohmL, AGDx). We trained heterogeneous GNN baselines on the same interaction graph with extensive tuning, and showed that BiOE_PO consistently outperform these baselines, with and without LLM alignment (Table 1, Section 5.1).
  • Motivation for hyperbolic space (ohmL). We clarified that hierarchy motivates hyperbolic geometry while the many to many structure motivates BiOE, and showed that all datasets are hierarchical, with low δ hyperbolicity and a long tail distribution that induced nested hierarchies (Appendix C, Table 9).
  • Statistical significance (AGDx). We added standard deviations for all recommendation datasets, where the gains of BiOE_PO over the strongest baseline exceed the standard deviation in 11 of 12 columns (Table 7), and on protein function annotation a one sided Welch t test gives p < 0.01 against the hyperbolic baselines.
  • Scalability (AGDx). We scaled Amazon Book and Steam up to 32 times their size, reaching 3.9M and 10.1M training edges, and pseudo order preprocessing never exceeds a fraction of the training cost even in large graphs (Section 5.1, Appendix B.4.1).
  • Learned order and generality (siAY, AGDx). We showed that the order BiOE learns reflects true generality rather than annotation counts or curation effort. Proteins with more distinct functions lie closer to the origin, while better studied proteins do not, and GO terms lie closer to the origin than their more specific descendants even when both have nearly the same number of annotations (Section 5.4).

All new results are in the revised PDF, which we will add to the camera ready version. We thank the reviewers again for their constructive feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.