AGIPool: Attention-Guided Influence Pooling

• The paper presents AGIPool as a weighted global pooling operator that uses incoming attention to identify chemically influential BRICS-derived substructures for DDI prediction.
• AGIPool operates on a hierarchical interaction graph by integrating GCN and GAT outputs to create interpretable, pair-level molecular embeddings.
• Empirical results show that AGIPool enhances the model's predictive performance and mechanistic interpretation by matching influential fragments with known chemical motifs.

Attention-Guided Influence Pooling (AGIPool) is the molecular graph readout module introduced in MolecBioNet for drug-drug interaction (DDI) prediction. It converts a drug-pair hierarchical interaction graph into a single pair embedding by assigning larger weight to chemically influential substructures, where influence is defined by the amount of final-layer graph attention a node receives from other nodes. In MolecBioNet, AGIPool serves the micro-level molecular branch, while Context-Aware Subgraph Pooling (CASPool) summarizes the macro-level biomedical knowledge branch; the two pooled views are then fused for multi-class DDI classification and mechanistic interpretation (Chen et al., 12 Jul 2025).

1. Architectural role and scope

AGIPool is not a standalone graph classifier. It is one component inside MolecBioNet, a pair-centric framework that models a drug pair as a unified entity at two scales. The macro-level branch constructs a task-specific biomedical knowledge graph, obtains contextual node embeddings with GraphSAGE, extracts a $k$-hop enclosing subgraph around the drug pair, augments that subgraph with positional and categorical encodings, applies a Graph Transformer, and then uses CASPool to produce the pair-level biomedical embedding $\mathbf{h}_{u,v}$. The micro-level branch converts each drug’s SMILES into an atom-level molecular graph, decomposes it into BRICS-derived substructures, builds a hierarchical interaction graph over the two drugs’ substructures, learns node embeddings with GCN and GAT, and then applies AGIPool to produce the pair-level molecular embedding $\mathbf{z}_{u,v}$ (Chen et al., 12 Jul 2025).

The downstream representation combines these pair-level summaries with two individual drug embeddings. The paper gives

$$\mathbf{h}_u = \mathrm{MLP}\left(\left[\mathbf{h}_u^{(L_2)} \Vert \mathbf{h}_u^M \right]\right),$$

and analogously for $\mathbf{h}_v$. The final pair representation is

$$\mathbf{f}_{u,v} = \left[\mathbf{h}_{u,v} \Vert \mathbf{z}_{u,v} \Vert \mathbf{h}_u \Vert \mathbf{h}_v\right],$$

with prediction

$$\mathbf{\hat{y}}_{u,v} = \mathbf{W}_{c} \mathbf{f}_{u,v}.$$

Within this design, AGIPool is the mechanism that determines which molecular fragments dominate the pair-level molecular summary (Chen et al., 12 Jul 2025).

This division of labor is central to the meaning of AGIPool. CASPool emphasizes biomedically relevant entities such as proteins, diseases, transporters, and pathways in the local enclosing subgraph, whereas AGIPool emphasizes influential molecular substructures inside the pairwise molecular interaction graph. The two modules therefore answer different explanatory questions: contextual biomedical relevance versus fragment-level chemical influence (Chen et al., 12 Jul 2025).

2. Hierarchical interaction graph and message passing substrate

AGIPool operates on the hierarchical interaction graph $\mathcal{G}_{u,v}^H$, not on each drug independently. Each node in this graph is a BRICS-derived chemically meaningful substructure, such as a functional group or ring system. Initial node features are Morgan fingerprints computed from substructure SMILES. The graph is constructed by taking both drugs’ substructure graphs and connecting every substructure of drug $u$ to every substructure of drug $v$, so the representation explicitly contains both intra-drug and inter-drug structure (Chen et al., 12 Jul 2025).

Node embeddings are learned through two coupled channels. The intra-drug channel uses a GCN: $\mathbf{h}_{u,v}$0 where $\mathbf{h}_{u,v}$1 denotes intra-drug neighbors. The inter-drug channel uses a GAT: $\mathbf{h}_{u,v}$2 where $\mathbf{h}_{u,v}$3 denotes cross-drug neighbors and

$\mathbf{h}_{u,v}$4

The two channels are then merged by

$\mathbf{h}_{u,v}$5

AGIPool takes the final node embeddings $\mathbf{h}_{u,v}$6 and the final-layer attention coefficients as its inputs (Chen et al., 12 Jul 2025).

This construction matters because AGIPool does not score raw fragments in isolation. Its weights are derived after pairwise interaction modeling has already propagated information across both drugs. A fragment can therefore become influential not merely because of its own descriptor, but because other substructures attend to it strongly during the learned inter-drug interaction process (Chen et al., 12 Jul 2025).

3. Definition of influence and pooling operator

The defining idea of AGIPool is that influence is measured by incoming attention rather than outgoing attention. The paper states that “a node that receives more attention from its neighbors is considered more influential, as it contributes more significantly to their representation updates.” Formally, the influence score of node $\mathbf{h}_{u,v}$7 is

$\mathbf{h}_{u,v}$8

where $\mathbf{h}_{u,v}$9 is the attention weight assigned by node $\mathbf{z}_{u,v}$0 to node $\mathbf{z}_{u,v}$1 in the final layer (Chen et al., 12 Jul 2025).

The final pooled representation is then

$\mathbf{z}_{u,v}$2

This is the complete AGIPool readout. The operator is therefore a weighted global pooling rule over the nodes of the hierarchical interaction graph, with node weights inherited from the final GAT layer rather than produced by a separate scoring network (Chen et al., 12 Jul 2025).

Several negative characterizations are equally important. AGIPool does not use top-$\mathbf{z}_{u,v}$3 node selection, thresholding, pooling masks, assignment matrices, graph coarsening, structure-preserving reduced graphs, or hierarchical cluster pooling. It is a weighted global pooling operator, not a graph reduction method. The paper also states that, for simplicity, the final GAT layer whose attentions feed AGIPool uses a single attention head, even though inter-drug message passing is described more generally as a multi-head GAT (Chen et al., 12 Jul 2025).

The normalization semantics are inherited from the GAT. The coefficients $\mathbf{z}_{u,v}$4 are softmax-normalized upstream over candidate neighbors, but AGIPool itself introduces no additional softmax over $\mathbf{z}_{u,v}$5. Consequently, the final readout is not written as a normalized convex combination $\mathbf{z}_{u,v}$6; it is literally the mean of the weighted embeddings. The paper explicitly notes that a faithful implementation should preserve this distinction unless one intentionally alters the formulation (Chen et al., 12 Jul 2025).

AGIPool is trained end-to-end without fragment-level supervision. The main DDI prediction loss is

$\mathbf{z}_{u,v}$7

and the full objective is

$\mathbf{z}_{u,v}$8

where $\mathbf{z}_{u,v}$9 is a center loss on drug embeddings and $$\mathbf{h}_u = \mathrm{MLP}\left(\left[\mathbf{h}_u^{(L_2)} \Vert \mathbf{h}_u^M \right]\right),$$0 is a mutual-information minimization term encouraging complementarity between $$\mathbf{h}_u = \mathrm{MLP}\left(\left[\mathbf{h}_u^{(L_2)} \Vert \mathbf{h}_u^M \right]\right),$$1 and $$\mathbf{h}_u = \mathrm{MLP}\left(\left[\mathbf{h}_u^{(L_2)} \Vert \mathbf{h}_u^M \right]\right),$$2. AGIPool is therefore learned only through DDI labels and the global representation regularizers, not through direct supervision on influence scores (Chen et al., 12 Jul 2025).

4. Interpretability and empirical evidence

AGIPool is designed to be both predictive and explanatory. Because each node corresponds to a chemically meaningful BRICS fragment and each node receives a scalar influence score $$\mathbf{h}_u = \mathrm{MLP}\left(\left[\mathbf{h}_u^{(L_2)} \Vert \mathbf{h}_u^M \right]\right),$$3, the model can rank substructures by their contribution to the pair-level molecular embedding. The paper reports literature-supported motifs including 1,3-Benzdioxole, Acetylene, Imidazole, Hydrazine, Pyridine, Thiophene, and Triazole as influential for metabolism-related DDIs (Chen et al., 12 Jul 2025).

The main interpretability study uses a curated set of 8,339 drug pairs involving 19 drugs, each known to participate in metabolism-related DDIs through specific chemical substructures. Using AGIPool, the model identifies the top two most influential substructures per interaction. The reported outcomes are that the top-1 predicted substructure matches literature-reported functional groups in 2,131 DDIs, and that at least one of the top-2 predictions aligns in 5,662 DDIs (Chen et al., 12 Jul 2025).

A concrete case study is Itraconazole–Promazine. MolecBioNet identifies three key substructures in Itraconazole, with Triazole ranked highest. The paper states that this agrees with known evidence that Triazole binds to CYP3A4 and inhibits its enzymatic activity, thereby affecting Promazine metabolism. This is the clearest illustration of AGIPool’s intended explanatory role: the prediction is not only that the pair interacts, but that a specific moiety is the dominant molecular driver (Chen et al., 12 Jul 2025).

Evidence for AGIPool’s contribution to predictive performance is less isolated. The ablation labeled “w/o. Designed Pooling” replaces both CASPool and AGIPool with MeanPooling and leads to performance degradation on both datasets, which the authors interpret as evidence that context-aware pooling in biomedical subgraphs and substructure-guided prioritization in hierarchical molecular graphs are both important. However, the paper does not provide a dedicated “w/o AGIPool only” ablation, so the exact numerical contribution of AGIPool by itself is not identified. The full MolecBioNet system achieves $$\mathbf{h}_u = \mathrm{MLP}\left(\left[\mathbf{h}_u^{(L_2)} \Vert \mathbf{h}_u^M \right]\right),$$4 ACC, $$\mathbf{h}_u = \mathrm{MLP}\left(\left[\mathbf{h}_u^{(L_2)} \Vert \mathbf{h}_u^M \right]\right),$$5 F1, $$\mathbf{h}_u = \mathrm{MLP}\left(\left[\mathbf{h}_u^{(L_2)} \Vert \mathbf{h}_u^M \right]\right),$$6 PR-AUC, and $$\mathbf{h}_u = \mathrm{MLP}\left(\left[\mathbf{h}_u^{(L_2)} \Vert \mathbf{h}_u^M \right]\right),$$7 Cohen’s $$\mathbf{h}_u = \mathrm{MLP}\left(\left[\mathbf{h}_u^{(L_2)} \Vert \mathbf{h}_u^M \right]\right),$$8 on Ryu’s dataset, and $$\mathbf{h}_u = \mathrm{MLP}\left(\left[\mathbf{h}_u^{(L_2)} \Vert \mathbf{h}_u^M \right]\right),$$9 ACC, $\mathbf{h}_v$0 F1, $\mathbf{h}_v$1 PR-AUC, and $\mathbf{h}_v$2 Cohen’s $\mathbf{h}_v$3 on DrugBank, but these are full-model results rather than AGIPool-only measurements (Chen et al., 12 Jul 2025).

5. Position within learned pooling research

AGIPool belongs to the broader family of learned importance-weighted pooling, but its mechanism is domain-specific. PiNet uses two separate message-passing networks on the same graph, one to learn node features and one to learn attention coefficients, and then forms a graph representation as a dense, differentiable weighted sum over all nodes (Meltzer et al., 2020). “Attentional Pooling for Action Recognition” replaces global average pooling with attention-weighted spatial aggregation and derives the resulting operator as a rank-1 low-rank approximation of bilinear pooling (Girdhar et al., 2017). AFA in visuomotor policy learning uses a single trainable query token and cross-attends over local frozen visual features to produce a weighted readout (Tsagkas et al., 13 Nov 2025). DAMIP in MSSDMPA-Net preserves all values inside each $\mathbf{h}_v$4 window through index pooling and then amplifies them with supervised probability maps during downsampling (Bose et al., 2023).

Method	Weight source	Pooling behavior
AGIPool (Chen et al., 12 Jul 2025)	Sum of incoming final-layer GAT attention weights	Weighted global mean over BRICS-substructure embeddings
PiNet (Meltzer et al., 2020)	Separate GNN attention branch	Weighted global node readout with all nodes retained
AFA (Tsagkas et al., 13 Nov 2025)	Single trainable query over local tokens	Cross-attentive pooling over frozen local visual features
DAMIP (Bose et al., 2023)	Supervised probability maps after index pooling	Guidance-weighted downsampling that preserves all local window values

What distinguishes AGIPool in this landscape is not the generic idea of weighting before aggregation, but the precise definition of influence and the chemistry-driven graph construction. It does not learn a separate pooling score from an auxiliary projection head, as many graph pooling methods do. It does not sparsify the graph through node dropping, nor does it assign nodes to clusters. Instead, it reuses the interaction attentions already produced by the final inter-drug GAT and interprets incoming attention mass as fragment influence. This places AGIPool closer to attention-based global readouts than to hierarchical graph coarsening, while still differing from general-purpose attention pooling because its nodes are chemically meaningful substructures and its weights arise from cross-drug interaction modeling rather than generic node saliency (Chen et al., 12 Jul 2025).

The contrast with CASPool is equally instructive. CASPool is a contextual relevance operator over enclosing biomedical subgraphs, whereas AGIPool is an influence readout over pairwise molecular substructure graphs. MolecBioNet’s contribution is partly the pairing of these two notions: contextual biomedical relevance on one side and fragment-level molecular influence on the other (Chen et al., 12 Jul 2025).

6. Limitations, ambiguities, and open questions

The AGIPool formulation is explicit at the level of core equations, but several implementation details remain underspecified. The paper does not provide the hidden dimension $\mathbf{h}_v$5, the number of hierarchical interaction layers $\mathbf{h}_v$6, the head counts in earlier GAT layers, dropout rates, optimizer, learning rate, batch size, weight decay, or whether self-loops are used in the inter-drug GAT. This does not obscure the conceptual operator, but it constrains exact reproducibility of the full training setup (Chen et al., 12 Jul 2025).

There is also a slight notational inconsistency. Cross-drug neighbors in the GAT are denoted $\mathbf{h}_v$7, whereas the AGIPool influence score uses $\mathbf{h}_v$8. The paper’s explanation makes the intended semantics clear—sum incoming attention assigned to node $\mathbf{h}_v$9—but a reimplementation must decide precisely whether the summation should run over only cross-drug neighbors or over all neighbors in the hierarchical interaction graph (Chen et al., 12 Jul 2025).

A deeper limitation is epistemic rather than notational. AGIPool assumes that incoming final-layer GAT attention mass is a suitable proxy for substructure importance. The paper presents this as plausible and chemically interpretable, and its curated evaluation supports that claim, but it does not test alternative influence definitions such as gradients, perturbation scores, or integrated attention across layers. The article on AGIPool therefore occupies the broader debate over whether attention weights should be treated as faithful explanations, without resolving that debate experimentally (Chen et al., 12 Jul 2025).

Two further caveats are structural. First, the “w/o. Designed Pooling” ablation removes CASPool and AGIPool together, so AGIPool’s isolated performance effect remains uncertain. Second, the hierarchical interaction graph connects all substructures in drug $$\mathbf{f}_{u,v} = \left[\mathbf{h}_{u,v} \Vert \mathbf{z}_{u,v} \Vert \mathbf{h}_u \Vert \mathbf{h}_v\right],$$0 to all substructures in drug $$\mathbf{f}_{u,v} = \left[\mathbf{h}_{u,v} \Vert \mathbf{z}_{u,v} \Vert \mathbf{h}_u \Vert \mathbf{h}_v\right],$$1, which may introduce noisy or biologically implausible interactions and implies an upstream cost that can grow quadratically with the number of cross-drug substructure pairs. AGIPool itself is computationally lightweight once attentions exist—it only sums incoming attention and computes a weighted mean—but its efficacy depends on this dense pairwise interaction substrate (Chen et al., 12 Jul 2025).

Taken together, these properties define AGIPool as a chemically motivated, attention-derived global pooling mechanism rather than a general graph compression scheme. Its distinctive contribution is to turn inter-drug attention flow into a fragment ranking and a pair-level molecular embedding, thereby linking DDI prediction to interpretable substructure influence in a way that is tightly aligned with BRICS-based molecular decomposition and pair-centric graph modeling (Chen et al., 12 Jul 2025).