CASPool: Context-Aware Subgraph Pooling

• The paper introduces CASPool, a pair-conditioned attention pooling method that emphasizes biologically relevant nodes in a drug pair's local biomedical subgraph.
• It employs Graph Transformer layers to integrate structural, positional, and entity-type information with pair-level attention, resulting in a unified embedding for drug interactions.
• CASPool generates node-level influence scores for interpretability, enabling actionable insights by highlighting key biomedical entities driving DDI predictions.

Searching arXiv for the cited papers and closely related pooling work. Context-Aware Subgraph Pooling (CASPool) is the pooling module on the biomedical knowledge graph branch of MolecBioNet, a graph-based framework for drug-drug interaction (DDI) prediction that models a drug pair $(u,v)$ as a unified entity rather than encoding the two drugs separately and concatenating them later. In this setting, CASPool summarizes an extracted local $k$-hop enclosing subgraph $\mathcal{G}_{u,v}^k$ from the task-specific biomedical knowledge graph (tsBKG) into a single drug-pair-level embedding $\mathbf{h}_{u,v}$, with the explicit goals of improving prediction by emphasizing biologically relevant entities and supporting interpretability through node-level influence scores (Chen et al., 12 Jul 2025).

1. Definition and design motivation

CASPool was introduced within MolecBioNet to address a specific limitation of graph-based DDI models: not every node in the local biomedical neighborhood contributes equally to a given interaction. In the tsBKG surrounding a drug pair, some proteins, diseases, pathways, side effects, or other biomedical entities are more relevant to the pair than others. CASPool therefore assigns higher weight to nodes that are more relevant to the central drug pair $(u,v)$, so that the resulting embedding reflects the surrounding biomedical context rather than an indiscriminate average over all nearby entities (Chen et al., 12 Jul 2025).

The motivating contrast is with naive pooling such as mean pooling, which would treat all nodes equally and therefore dilute the signal from biologically meaningful entities. Within MolecBioNet, this design is tied directly to the framework’s broader premise that DDIs are context dependent and should be modeled at the level of the pair as a unified object. CASPool is the mechanism that operationalizes this principle on the biomedical graph side.

A plausible implication is that CASPool should be understood not as a generic graph readout, but as a pair-conditioned pooling operator whose semantics depend on the specific drug pair under consideration. This distinguishes it from pooling methods that are agnostic to the identity of the central query.

2. Placement within the MolecBioNet pipeline

On the tsBKG branch, CASPool is the final aggregation step after contextual node encoding. The pipeline begins with node embedding generation on the task-specific biomedical knowledge graph $\mathcal{G}_{BKG}$ using a GraphSAGE-style GNN, producing initial contextual node embeddings $\mathbf{x}_v^{(L_1)}$. For a drug pair $(u,v)$, the model then extracts the induced $k$-hop enclosing subgraph $\mathcal{G}_{u,v}^k$ and constructs node features by concatenating three components for each node $k$0: the pretrained embedding $k$1, a positional encoding $k$2 based on distances to $k$3 and $k$4, and an entity-type encoding $k$5, yielding

$k$6

These subgraph node embeddings are then updated by a Graph Transformer for $k$7 layers, producing final node representations $k$8. CASPool uses the final embeddings of the two central drugs together with all subgraph node embeddings to compute attention scores and generate the subgraph-level embedding $k$9 (Chen et al., 12 Jul 2025).

This position in the pipeline is significant because CASPool operates after both structural and feature enrichment have already occurred. The subgraph nodes reaching CASPool already encode graph topology, positional information relative to the two drugs, and entity type information. As a result, the pooling step acts on representations that are already contextualized rather than on raw tsBKG features.

3. Mathematical formulation

CASPool first constructs a base vector from the two central drugs: $\mathcal{G}_{u,v}^k$0 For each node $\mathcal{G}_{u,v}^k$1 in the subgraph $\mathcal{G}_{u,v}^k$2, it computes an attention score

$\mathcal{G}_{u,v}^k$3

where $\mathcal{G}_{u,v}^k$4 and $\mathcal{G}_{u,v}^k$5 project node embeddings and the pair base vector into a shared latent space. The softmax is taken over all nodes in the subgraph, so the scores satisfy

$\mathcal{G}_{u,v}^k$6

The final pooled representation is then defined as

$\mathcal{G}_{u,v}^k$7

This $\mathcal{G}_{u,v}^k$8 is the biomedical-context embedding for the drug pair and is later fused with molecular information in MolecBioNet (Chen et al., 12 Jul 2025).

The formal structure makes explicit that CASPool is pair conditioned. A node is weighted not merely by its prominence in the subgraph, but by the compatibility between its Graph Transformer representation and the base vector formed from the two central drugs. This suggests that relevance is defined relationally: the same biomedical node could receive different attention scores in different drug-pair subgraphs even if its local graph position were similar.

4. Meaning of “context” and pair-conditioned aggregation

In CASPool, context means the local biomedical neighborhood of the drug pair in the tsBKG, specifically the proteins, diseases, pathways, side effects, and other biomedical entities connected within $\mathcal{G}_{u,v}^k$9 hops of either drug. This context is represented in two complementary ways. First, there is the structural context given by the extracted $\mathbf{h}_{u,v}$0-hop subgraph $\mathbf{h}_{u,v}$1. Second, there is the node context encoded in each node’s learned representation after Graph Transformer message passing, which incorporates graph topology, positional information relative to $\mathbf{h}_{u,v}$2 and $\mathbf{h}_{u,v}$3, and entity type information (Chen et al., 12 Jul 2025).

CASPool compares each node embedding $\mathbf{h}_{u,v}$4 against the pair-level base vector $\mathbf{h}_{u,v}$5, making the pooling step explicitly pair conditioned. The paper states that CASPool “adaptively aggregates node-level information within the subgraph, assigning greater importance to nodes more relevant to the central drug pair $\mathbf{h}_{u,v}$6.” This formulation is central to the method’s interpretation: relevance is not defined as a generic property of a biomedical entity, but as relevance to a specific candidate interaction.

This pair-conditioned design distinguishes CASPool from other graph pooling paradigms in related literature. The knowledge-graph method called “Context Pooling” constructs a query-specific graph for link prediction, but the acronym CASPool does not appear there; the method is consistently called Context Pooling and operates by node-drop pooling over logically relevant neighbors for a query relation (Su et al., 10 Jul 2025). In a different domain, subgraph pooling for logical formulae uses DAG LSTM pooling and attention between premise and conjecture, again reflecting context-aware aggregation but under a different graph semantics and objective (Crouse et al., 2019). These comparisons indicate a broader family resemblance—context-sensitive graph reduction—while preserving the specific meaning of CASPool in MolecBioNet.

5. Interpretability and explanatory role

CASPool is one of MolecBioNet’s key interpretability mechanisms because it assigns a normalized attention score $\mathbf{h}_{u,v}$7 to each biomedical node in the local subgraph. These scores can be used to rank nodes by influence and highlight the top nodes as an explanation of the predicted DDI. In practice, CASPool is used to visualize and extract biologically meaningful local subgraphs around the drug pair (Chen et al., 12 Jul 2025).

The reported case study is Itraconazole–Promazine. For this pair, the local tsBKG subgraph is extracted and the top 10% most influential nodes are highlighted. The visualization suggests that Itraconazole inhibits the ABCG2 transporter, providing a mechanistically plausible explanation. Interpretability is also evaluated using Fidelity metrics and Sparsity metrics. The paper reports that as sparsity increases, $\mathbf{h}_{u,v}$8 decreases gradually, while $\mathbf{h}_{u,v}$9 remains low, suggesting that small subsets of nodes are sufficient to preserve predictive performance. This is presented as evidence supporting the faithfulness of CASPool-derived explanations (Chen et al., 12 Jul 2025).

A common misconception would be to treat these attention scores as a complete mechanistic proof of causality. The paper does not make that claim. Instead, CASPool provides a ranked, biologically grounded explanation of which entities in the local biomedical context most strongly influence the model’s prediction. The explanatory role is therefore model-internal and interpretive, not equivalent to experimental validation.

6. Performance contribution, ablations, and relation to AGIPool

The ablation study shows that removing the designed pooling mechanisms and replacing them with MeanPooling (“w/o. Designed Pooling”) degrades performance. The paper attributes this drop to the loss of context-aware relational analysis in biomedical subgraphs and substructure-guided prioritization at the molecular level. This ablation does not isolate CASPool alone from AGIPool, but it shows that the custom pooling design, including CASPool, is important for performance (Chen et al., 12 Jul 2025).

Within MolecBioNet, CASPool and AGIPool are analogous in intent but distinct in domain. CASPool operates on the biomedical knowledge graph branch, pooling nodes in a $(u,v)$0-hop local biomedical subgraph and using the drug pair itself as context through the base vector $(u,v)$1. AGIPool operates on the hierarchical interaction graph (HIG) built from substructure-level molecular graphs, where the influence score is

$(u,v)$2

and the final pair embedding is

$(u,v)$3

CASPool emphasizes biologically relevant entities such as proteins and diseases, whereas AGIPool highlights influential chemical substructures (Chen et al., 12 Jul 2025).

This division clarifies why CASPool is needed. AGIPool alone would explain molecular substructures but could not explain which surrounding biomedical entities in the knowledge graph drive the interaction. CASPool is therefore the component that makes the knowledge-graph branch interpretable and biologically grounded. Together, the two pooling mechanisms provide what the paper describes as multi-scale explanations: CASPool yields macro-level biomedical explanation, and AGIPool yields micro-level molecular or substructure explanation.

7. Hyperparameters, sensitivity, and broader context

Several implementation details are directly relevant to CASPool. The local biomedical context is extracted as a $(u,v)$4-hop enclosing subgraph, and sensitivity analysis shows that performance improves from $(u,v)$5 to $(u,v)$6, then slightly drops at $(u,v)$7; the default is

$(u,v)$8

The Graph Transformer produces the node embeddings used by CASPool after $(u,v)$9 layers, and CASPool itself includes learnable matrices $\mathcal{G}_{BKG}$0 and $\mathcal{G}_{BKG}$1. The overall model also uses a mutual-information minimization term and center loss in training, although these are not specific to CASPool. For interpretability evaluation, the top 10% most influential nodes from the CASPool scores are highlighted in the case study (Chen et al., 12 Jul 2025).

In the broader graph-learning literature, subgraph pooling has been used for different purposes, including transfer robustness under structural mismatch, where local subgraph embeddings can be more stable than node embeddings across semantically similar graphs (Wang et al., 2024). That work is not a DDI method and does not define CASPool in MolecBioNet’s sense, but it reinforces a general pattern: moving from node-level representations to local subgraph-level summaries can improve robustness and relevance when graph neighborhoods are noisy or only partially informative. This suggests a broader methodological context for CASPool’s use in biomedical knowledge graphs.

Taken together, CASPool is best characterized as a pair-conditioned attention pooling operator for tsBKG subgraphs in DDI prediction. Its function is simultaneously representational and explanatory: it compresses a local biomedical subgraph into a single embedding for downstream fusion while retaining a node-level importance distribution that can be inspected as an explanation of the predicted interaction (Chen et al., 12 Jul 2025).