Research map / Label-efficient and robust learning / Recognition and long-tail

Association Pattern-aware Fusion for Biological Entity Relationship Prediction

NeurIPS 2024 · Label-efficient and robust learning › Recognition and long-tail · approach: Dense features and foundation priors · 1 citations (OpenAlex)

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 Pattern-BERP, which samples hop-distance-based association patterns around each entity in a hypergraph and fuses them via transformer attention plus a bind-relation module for hard negative sampling, to predict triple-wise biological associations.

Problem

Predicting triple-wise biological entity associations (e.g., drug-microbe-disease, drug-target-adverse reaction) with existing methods that only aggregate entity-centric features and ignore association patterns among entities.

Approach

Samples association patterns (hyperedges) at multiple hop distances for each entity node, applies transformer self-attention over these patterns to learn shared commonality (with a pattern commonality coefficient for interpretability), and adds a bind-relation module that reconstructs low-order pairwise features to generate hard negative triplets.

Key results

• On DPA, hit@1 rises from 33.24 (HyperSAGNN, best baseline) to 43.52, about a 23.6% relative improvement.
• On DMD, hit@1 reaches 93.94 versus 90.04 for the best baseline MCHNN, and on DDC it reaches 48.01 versus 41.91 for MCHNN.
• Ablations on DDC show removing the API block drops hit@1 from 48.01 to 40.29, and removing hard negative sampling drops it to 42.81.

Measured effects reported in the paper

Surprising findings

• On the sparse DPA dataset, removing the hypergraph convolution block improves performance (hit@1 46.16 vs 43.52 for full model), which the authors attribute to conflicting aggregation modes on sparse structures.
• Graph-based methods outperform non-graph methods on dense DMD and DDC but not on sparse DPA, where non-graph methods are competitive with hypergraph methods.

Limitations

• Restricted to fixed-length association patterns; variable-length pathways are not captured.
• Evaluated only on three biological datasets; generalization to other domains such as knowledge graph completion is untested.
• Complexity grows with the number of entity nodes and sampled patterns N, requiring graph sampling for larger datasets.

Experimental setup

Datasets: DMD DDC DPA ADReCS-Target

Benchmarks: DMD DDC DPA

Metrics: hits@1 hits@3 hits@5 ndcg@1 ndcg@3 ndcg@5

Baselines: RF MLP CP Tucker CoSTCo GCN GraphSAGE GAT GIN DHNE HyperSAGNN HGSynergy

Builds on: hypergraph convolution Transformer BGNN MCHNN

Closest papers by meaning

• Association Pattern-enhanced Molecular Representation Learning (AAAI 2025 · similarity 0.86)
• Knowledge Enhanced Representation Learning for Drug Discovery (AAAI 2024 · similarity 0.83)
• Biomedical Knowledge Graph Embedding with Householder Projection (Student Abstract) (AAAI 2024 · similarity 0.83)
• Dual-Channel Learning Framework for Drug-Drug Interaction Prediction via Relation-Aware Heterogeneous Graph Transformer (AAAI 2024 · similarity 0.81)
• Learning Molecular Representation in a Cell (ICLR 2025 · similarity 0.81)
• Hypergraph Representation Learning with Hyperlink Random Effects (NeurIPS 2026 · similarity 0.81)