MolecBioNet: Graph-based DDI Predictor

• MolecBioNet is a graph-based framework for multi-class DDI prediction that models drug pairs as unified entities to capture pair-specific interactions.
• It fuses macro-level biomedical context from a task-specific knowledge graph with micro-level molecular graphs using context-aware and attention-guided pooling operators.
• Empirical evaluations demonstrate high accuracy and enhanced interpretability through multi-scale representation learning and mutual information minimization.

MolecBioNet is a graph-based framework for multi-class drug-drug interaction (DDI) prediction that integrates macro-level biomedical context from a task-specific biomedical knowledge graph (tsBKG) with micro-level molecular structure from drug molecular graphs. Its defining architectural choice is to model a drug pair $(u,v)$ as a unified entity rather than encode each drug independently and concatenate the results, so that pair-specific interaction mechanisms can be learned at both the biomedical-network and molecular-substructure scales. The framework further couples this pair-centric representation with two domain-specific pooling operators—context-aware subgraph pooling (CASPool) and attention-guided influence pooling (AGIPool)—and with mutual information minimization regularization intended to preserve complementarity between network-level and molecular-level embeddings (Chen et al., 12 Jul 2025).

1. Problem setting and conceptual basis

MolecBioNet formulates DDI prediction as a multi-class classification problem over relation types. With $\mathcal{D}$ denoting the set of drugs, $\mathcal{R}$ the set of DDI relation types, and

$$\mathcal{S}=\{(u,r,v)\mid u,v\in\mathcal{D},\ r\in\mathcal{R}\},$$

the goal is to learn

$$f:\mathcal{D}\times\mathcal{D}\rightarrow\mathcal{R}.$$

This framing is explicit in treating the task as prediction of which interaction type occurs, not merely whether two drugs interact. A common misconception is therefore avoided at the outset: MolecBioNet is not a binary interaction detector, and it is not built around independently learned drug embeddings that are only fused at the classifier stage. Instead, it introduces a pair-level contextual embedding $\mathbf{h}_{u,v}$ from the tsBKG branch and a pair-level molecular interaction embedding $\mathbf{z}_{u,v}$ from the hierarchical interaction graph, both defined directly for the ordered pair under analysis (Chen et al., 12 Jul 2025).

The pair-as-entity formulation is central to the framework’s interpretive claims. The model argues that DDI mechanisms are pair-specific and context-dependent, so the effect does not emerge as an independent property of each drug alone. This suggests a shift from drug-centric encoding toward relation-centric representation learning. In practical terms, MolecBioNet seeks to preserve two distinct sources of evidence: biomedical context surrounding the pair in a knowledge graph, and interaction patterns among substructures derived from the two molecular graphs.

2. Biomedical-context branch and task-specific knowledge graph

The biomedical branch operates on a tsBKG built by merging a DDI graph $\mathcal{G}_{DDI}$ with an external biomedical knowledge graph $\mathcal{G}_{KG}$. The merged graph is

$$\mathcal{G}_{BKG}=(\mathcal{V}_{BKG},\mathcal{E}_{BKG},\mathcal{R}_{BKG}),$$

with

$\mathcal{D}$0

Global node embeddings are first learned on the whole tsBKG with GraphSAGE. For each pair $\mathcal{D}$1, the model then extracts a $\mathcal{D}$2-hop enclosing local subgraph $\mathcal{D}$3, induced by nodes $\mathcal{D}$4 satisfying

$\mathcal{D}$5

where $\mathcal{D}$6 is shortest-path distance in $\mathcal{D}$7. This local subgraph is the pair-specific biomedical neighborhood from which contextual evidence is aggregated (Chen et al., 12 Jul 2025).

Each node in the extracted subgraph receives three kinds of features: a global embedding $\mathcal{D}$8 from GraphSAGE, a positional encoding based on distances to $\mathcal{D}$9 and $\mathcal{R}$0, and a type or category encoding. The initial node representation is

$\mathcal{R}$1

When a node is disconnected from one of the central drugs, the model replaces infinite distance with a pseudo-distance so that positional encodings remain well-defined. After initialization, node states are updated with a Graph Transformer, and the resulting subgraph is pooled with CASPool.

CASPool is explicitly context-conditioned on the central drug pair. Let

$\mathcal{R}$2

For each node $\mathcal{R}$3, the attention score is

$\mathcal{R}$4

and the final biomedical pair embedding is

$\mathcal{R}$5

In the framework’s own interpretation, CASPool emphasizes biomedical entities most relevant to the pair and provides a subgraph-level explanation by identifying influential proteins, diseases, pathways, and related entities. The best reported subgraph radius is $\mathcal{R}$6, which is described as balancing relevant biomedical context and computational efficiency.

3. Molecular branch and hierarchical interaction graph

The molecular branch begins with each drug’s SMILES string, which is converted into a molecular graph

$\mathcal{R}$7

where nodes are atoms and edges are chemical bonds. The atom-level graph is then fragmented by BRICS into chemically meaningful substructures, yielding

$\mathcal{R}$8

whose nodes are substructure fragments such as functional groups or ring systems. For a pair $\mathcal{R}$9, MolecBioNet constructs a hierarchical interaction graph

$$\mathcal{S}=\{(u,r,v)\mid u,v\in\mathcal{D},\ r\in\mathcal{R}\},$$0

by connecting all nodes in $$\mathcal{S}=\{(u,r,v)\mid u,v\in\mathcal{D},\ r\in\mathcal{R}\},$$1 with all nodes in $$\mathcal{S}=\{(u,r,v)\mid u,v\in\mathcal{D},\ r\in\mathcal{R}\},$$2. The resulting graph explicitly contains both intra-drug substructure relations and inter-drug substructure relations (Chen et al., 12 Jul 2025).

This branch uses a GCN to encode substructures, a GCN to model intra-drug relations, and a GAT to model inter-drug relations. The model’s molecular representation is therefore not a flat fingerprint but a hierarchical interaction graph in which cross-drug substructure coupling is represented directly. This is the molecular analogue of the pair-centered biomedical subgraph: in both branches, the central object is the pair rather than the isolated drug.

AGIPool aggregates the final HIG node embeddings into the pair-level molecular interaction embedding $$\mathcal{S}=\{(u,r,v)\mid u,v\in\mathcal{D},\ r\in\mathcal{R}\},$$3. Using the final GAT attention coefficients $$\mathcal{S}=\{(u,r,v)\mid u,v\in\mathcal{D},\ r\in\mathcal{R}\},$$4, the influence score for node $$\mathcal{S}=\{(u,r,v)\mid u,v\in\mathcal{D},\ r\in\mathcal{R}\},$$5 is

$$\mathcal{S}=\{(u,r,v)\mid u,v\in\mathcal{D},\ r\in\mathcal{R}\},$$6

The pooled molecular representation is then

$$\mathcal{S}=\{(u,r,v)\mid u,v\in\mathcal{D},\ r\in\mathcal{R}\},$$7

According to the reported design rationale, AGIPool prioritizes influential molecular substructures and supports substructure-level mechanistic interpretation. In this sense, MolecBioNet’s molecular explainability is not external to the predictor; it is embedded in the pooling operator itself.

4. Fusion, regularization, and training objective

MolecBioNet combines pair-level and drug-level representations. In addition to $$\mathcal{S}=\{(u,r,v)\mid u,v\in\mathcal{D},\ r\in\mathcal{R}\},$$8 and $$\mathcal{S}=\{(u,r,v)\mid u,v\in\mathcal{D},\ r\in\mathcal{R}\},$$9, it computes per-drug embeddings: a biological embedding $$f:\mathcal{D}\times\mathcal{D}\rightarrow\mathcal{R}.$$0 from the tsBKG branch and a molecular embedding $$f:\mathcal{D}\times\mathcal{D}\rightarrow\mathcal{R}.$$1 from the molecular graph. These are fused by an MLP to obtain

$$f:\mathcal{D}\times\mathcal{D}\rightarrow\mathcal{R}.$$2

The final pair representation is

$$f:\mathcal{D}\times\mathcal{D}\rightarrow\mathcal{R}.$$3

and the class logits are predicted by

$$f:\mathcal{D}\times\mathcal{D}\rightarrow\mathcal{R}.$$4

This fusion makes explicit that MolecBioNet does not replace drug-specific information with a pair-only representation; rather, it couples pair-level embeddings with single-drug embeddings in the classifier head (Chen et al., 12 Jul 2025).

A further architectural feature is mutual information minimization between the biomedical embedding and the molecular embedding. The model defines

$$f:\mathcal{D}\times\mathcal{D}\rightarrow\mathcal{R}.$$5

with the stated objective of reducing redundancy and encouraging complementary information during embedding fusion. The training loss combines three terms. The prediction loss is cross-entropy over DDI types,

$$f:\mathcal{D}\times\mathcal{D}\rightarrow\mathcal{R}.$$6

and a center loss is added to make embeddings of the same drug consistent across many pairs,

$$f:\mathcal{D}\times\mathcal{D}\rightarrow\mathcal{R}.$$7

The total objective is

$$f:\mathcal{D}\times\mathcal{D}\rightarrow\mathcal{R}.$$8

The reported best value for the center-loss weight is $$f:\mathcal{D}\times\mathcal{D}\rightarrow\mathcal{R}.$$9.

5. Empirical performance, ablation evidence, and interpretability

MolecBioNet is evaluated on two benchmark multi-class DDI datasets: Ryu’s dataset, derived from DeepDDI / Ryu et al. (2018), and a DrugBank dataset updated to version 5.1.12 on March 14, 2024. Performance is reported with 5-fold cross-validation using ACC, macro-F1, PR-AUC, and Cohen’s $\mathbf{h}_{u,v}$0. Compared methods include DeepDDI, SSI-DDI, GGI-DDI, KGNN, SumGNN, LaGAT, KnowDDI, and TIGER. On Ryu’s dataset, MolecBioNet reports ACC $\mathbf{h}_{u,v}$1, F1 $\mathbf{h}_{u,v}$2, PR-AUC $\mathbf{h}_{u,v}$3, and Cohen’s $\mathbf{h}_{u,v}$4 $\mathbf{h}_{u,v}$5. On DrugBank, it reports ACC $\mathbf{h}_{u,v}$6, F1 $\mathbf{h}_{u,v}$7, PR-AUC $\mathbf{h}_{u,v}$8, and Cohen’s $\mathbf{h}_{u,v}$9 $\mathbf{z}_{u,v}$0 (Chen et al., 12 Jul 2025).

The framework is also evaluated in cold-start settings on Ryu’s dataset. In the Novel Drug–Existing Drug setting, it reports ACC $\mathbf{z}_{u,v}$1, F1 $\mathbf{z}_{u,v}$2, PR-AUC $\mathbf{z}_{u,v}$3, and $\mathbf{z}_{u,v}$4 $\mathbf{z}_{u,v}$5. In the Novel Drug–Novel Drug setting, it reports ACC $\mathbf{z}_{u,v}$6, F1 $\mathbf{z}_{u,v}$7, PR-AUC $\mathbf{z}_{u,v}$8, and $\mathbf{z}_{u,v}$9 $\mathcal{G}_{DDI}$0. These results are described as the best among the compared models, but they also show that generalization to previously unseen drugs remains markedly harder than performance on standard splits.

Ablation studies attribute performance gains to all major components. Removing the tsBKG branch reduces performance, indicating the contribution of macro-level biological context. Removing the hierarchical interaction graph causes an even larger drop, indicating the importance of substructure-level molecular interactions. Replacing CASPool and AGIPool with mean pooling degrades results, and removing mutual information minimization reduces both performance and generalization. The appendix values reported for InfoMin are especially explicit: on Ryu’s dataset, MolecBioNet test F1 is $\mathcal{G}_{DDI}$1 versus $\mathcal{G}_{DDI}$2 without InfoMin; on DrugBank, test F1 is $\mathcal{G}_{DDI}$3 versus $\mathcal{G}_{DDI}$4 without InfoMin.

Interpretability is evaluated at two levels. At the network level, the paper uses Fidelity$\mathcal{G}_{DDI}$5, Fidelity$\mathcal{G}_{DDI}$6, and Sparsity, reporting that across increasing sparsity, Fidelity$\mathcal{G}_{DDI}$7 decreases smoothly while Fidelity$\mathcal{G}_{DDI}$8 stays low. At the chemical substructure level, AGIPool is tested on a curated set of 8,339 drug pairs involving 19 drugs with known metabolism-related DDIs. The top-1 influential substructure matches literature in 2,131 DDIs, and at least one of the top-2 influential substructures matches in 5,662 DDIs. Reported influential motifs include 1,3-Benzdioxole, Acetylene, Imidazole, Hydrazine, Pyridine, Thiophene, and Triazole. In the case study of Itraconazole–Promazine, the model highlights Triazole as the most influential substructure and identifies ABCG2 in the tsBKG explanation; the biological interpretation given is that Itraconazole inhibits the metabolism of Promazine and that Triazole binds CYP3A4 and inhibits its activity.

6. Position within multi-scale biomolecular network modeling

MolecBioNet belongs to a broader methodological movement in which biological systems are treated as multi-scale, multi-relational, and often time-varying networks rather than as single aggregated graphs. Multilayer network modeling has argued that aggregation causes loss of information and that explicit coupling across genetic, physical, metabolic, signaling, anatomical, functional, synaptic, electrical, modulatory, temporal, and condition-specific layers is often necessary for biologically faithful analysis (Domenico, 2018). In that context, MolecBioNet’s combination of a tsBKG branch and a hierarchical molecular branch can be read as a task-specific instance of coupled network representation.

Its logic also aligns with network-constrained inference frameworks that treat interaction structure as a prior and seek context-specific activity within connected regions. NetResponse, for example, models connected subnetworks with multiple latent transcriptional states across subsets of conditions, emphasizing that network function is condition-dependent rather than globally uniform (Lahti et al., 2012). This suggests a conceptual affinity with MolecBioNet’s insistence that DDI mechanisms are pair-specific and context-dependent, even though the underlying tasks differ.

At a finer level of biochemical representation, rule-based spatial frameworks such as MCell-R address combinatorial complexity by coupling particle-based spatial simulation with network-free rule evaluation, showing that biomolecular networks may require explicit representation of structured molecules, internal states, and spatial heterogeneity rather than only enumerated reaction networks (Tapia et al., 2018). MolecBioNet does not simulate biochemical dynamics in that sense, but its hierarchical interaction graph similarly rejects flat molecular encodings in favor of structured, relational representation.

More broadly, graph-theoretic accounts of biomolecular network origin and evolution frame biological interaction maps as historically structured objects characterized by hubs, modularity, clustering, preferential attachment, and information asymmetry (Janwa et al., 2018). A plausible implication is that MolecBioNet is best situated not as an isolated DDI predictor, but as one member of a larger class of biomolecular network methods that attempt to preserve relational structure across levels of biological organization. Within that class, its distinguishing contribution is the explicit unification of pair-level biomedical context, pair-level molecular substructure interaction, and built-in explanatory pooling in a single DDI prediction architecture.