Research map / Architectures and efficiency / Neural architecture search
Customized Subgraph Selection and Encoding for Drug-drug Interaction Prediction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CSSE-DDI, a neural architecture search framework that jointly searches query-specific subgraph selection (via Gumbel-softmax relaxation and subgraph representation approximation) and message-passing encoding functions for drug-drug interaction prediction.
Problem
Subgraph-based DDI prediction relies on fixed, hand-designed subgraph selection (fixed k-hop) and encoding functions, which cannot adapt to diverse query contexts and interaction semantics in dense drug interaction networks.
Approach
Searches data-specific subgraph scopes (i-hop of u, j-hop of v) and message-passing operations with a differentiable one-shot NAS: Gumbel-softmax relaxation over subgraph scopes, approximating subgraph representations by k-subtree node encodings of the full graph, plus message-aware partitioned supernet training with SPOS.
Key results
- On DrugBank (S0), CSSE-DDI reaches F1 92.08 ± 0.22, accuracy 95.56 and Cohen's kappa 94.72, above the best baseline KnowDDI (F1 90.06).
- On TWOSIDES (S0), CSSE-DDI reports ROC-AUC 95.47 and PR-AUC 94.21, versus 95.05 and 93.75 for KnowDDI.
- In the S1 new-drug setting, CSSE-DDI beats the other non-EmerGNN baselines on TWOSIDES (ROC-AUC 88.33 vs 87.42 for EmerGNN), but EmerGNN is best on DrugBank.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| search query-specific subgraphs and data-specific encodings instead of using fixed components | CSSE-DDI · DrugBank S0 | F1 Score: 90.06\std0.27 -> 92.08\std 0.22 | Helped |
| search query-specific subgraphs and data-specific encodings instead of using fixed components | CSSE-DDI · TWOSIDES S0 | ROC-AUC: 95.05\std0.06 -> 95.47\std0.02 | Helped |
| use CSSE-DDI rather than EmerGNN | DrugBank S1 · DrugBank S1 | F1 Score: 58.13\std1.36 -> 37.24\std1.13 | Hurt |
| use CSSE-DDI rather than EmerGNN | TWOSIDES S1 · TWOSIDES S1 | ROC-AUC: 87.42\std0.39 -> 88.33\std0.52 | Helped |
| add fine-grained subgraph selection for each query (ablation) | CSSE-DDI · DrugBank S0 | reported performance: 86.31\std 0.36 -> 92.08\std 0.22 | Helped |
| add fine-grained subgraph selection for each query (ablation) | CSSE-DDI · TWOSIDES S0 | reported performance: 94.35\std0.07 -> 95.47\std0.02 | Helped |
| search the encoding function instead of using fixed 3-layer CompGCN (ablation) | CSSE-DDI · DrugBank S0 | reported performance: 80.96\std 0.65 -> 92.08\std 0.22 | Helped |
| search the encoding function instead of using fixed 3-layer CompGCN (ablation) | CSSE-DDI · TWOSIDES S0 | reported performance: 94.26\std0.08 -> 95.47\std0.02 | Helped |
| use message-aware partitioning to train multiple sub-supernets (ablation) | CSSE-DDI · DrugBank | reported performance: 90.17\std0.29 -> 92.08\std0.22 | Helped |
| use message-aware partitioning to train multiple sub-supernets (ablation) | CSSE-DDI · TWOSIDES | reported performance: 95.12\std0.04 -> 95.47\std0.02 | Helped |
| use SPOS to decouple supernet training from architecture search (ablation) | CSSE-DDI · DrugBank | reported performance: 90.97\std0.72 -> 92.08\std0.22 | Helped |
| use SPOS to decouple supernet training from architecture search (ablation) | CSSE-DDI · TWOSIDES | reported performance: 94.89\std0.13 -> 95.47\std0.02 | Helped |
Surprising findings
- In the S1 new-drug setting on DrugBank, CSSE-DDI (F1 37.24) falls well behind EmerGNN (F1 58.13), so it is not the best method there.
Limitations
- Focused on method design rather than system-level efficiency.
- Searches only data-specific components; hyperparameters are not searched jointly.
Experimental setup
Datasets: DrugBank TWOSIDES
Benchmarks: DrugBank TWOSIDES
Metrics: F1 Score Accuracy Cohen's kappa ROC-AUC PR-AUC AP@50
Baselines: Decagon GAT SkipGNN CompGCN ACDGNN TransFOL SEAL GraIL SumGNN SNRI KnowDDI LaGAT
Builds on: SumGNN AutoGEL SPOS few-shot NAS CompGCN
Closest papers by meaning
- MOTIVE: A Drug-Target Interaction Graph For Inductive Link Prediction (NeurIPS 2024 · similarity 0.88)
- Motif-Oriented Representation Learning with Topology Refinement for Drug-Drug Interaction Prediction (AAAI 2025 · similarity 0.87)
- Dual-Channel Learning Framework for Drug-Drug Interaction Prediction via Relation-Aware Heterogeneous Graph Transformer (AAAI 2024 · similarity 0.87)
- PharmacoMatch: Efficient 3D Pharmacophore Screening via Neural Subgraph Matching (ICLR 2025 · similarity 0.86)
- ExDDI: Explaining Drug-Drug Interaction Predictions with Natural Language (AAAI 2025 · similarity 0.86)
- Blend the Separated: Mixture of Synergistic Experts for Data-Scarcity Drug-Target Interaction Prediction (AAAI 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.