Research map / Probabilistic and scientific modeling / Molecules and proteins
MOTIVE: A Drug-Target Interaction Graph For Inductive Link 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.
Introduces MOTIVE, a graph dataset linking ~11.5K genes and ~3.6K compounds via seven databases, with JUMP Cell Painting node features and cold-source/cold-target splits, and benchmarks GNNs against feature-based and topology baselines for inductive DTI link prediction.
Problem
Drug-target interaction prediction lacks graph benchmarks with rich empirical node features and realistic cold-start (inductive) data splits; existing compound-gene graphs have no morphological profile features.
Approach
Use JUMP Cell Painting morphological profiles as node features for compounds and genes in a compound-gene-compound knowledge graph, with graph-based random and cold-source/cold-target splits and a custom negative sampler that prevents leakage of held-out nodes.
Key results
- GIN_CP achieved the best random-split results with F1 0.5238, Hits@500 0.4552 and Precision@500 0.9920, versus 0.3836 F1 for GraphSAGE_CP and 0.1703 for Bilinear.
- In cold-source split, GIN_CP reached F1 0.2827 and Hits@500 0.1078, outperforming the feature-based MLP (F1 0.1876).
- In cold-target split, GIN_CP reached F1 0.3916 and Precision@500 0.9624, versus 0.3408 F1 for MLP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GIN with Cell Painting features and graph structure instead of the strongest listed non-GIN baseline | MOTIVE link prediction · MOTIVE random split | F1: 0.3836 ± 0.029 -> 0.5238 ± 0.040 | Helped |
| Use GIN with Cell Painting features and graph structure instead of the strongest listed non-GIN baseline | MOTIVE link prediction · MOTIVE cold-source split | Hits@500: 0.0439 ± 0.013 -> 0.1078 ± 0.008 | Helped |
| Use GIN with Cell Painting features and graph structure instead of the strongest listed non-GIN baseline | MOTIVE link prediction · MOTIVE cold-target split | F1: 0.3494 ± 0.012 -> 0.3916 ± 0.016 | Helped |
| Add Cell Painting node features to graph structure | GraphSAGE link prediction · MOTIVE random split | Precision@500: 0.8476 ± 0.029 -> 0.9056 ± 0.012 | Helped |
| Add graph structure to Cell Painting features | DTI link prediction · MOTIVE random split | F1: 0.3829 ± 0.008 -> 0.5238 ± 0.040 | Helped |
| Add graph structure to Cell Painting features | DTI link prediction · MOTIVE cold-source split | F1: 0.1876 ± 0.006 -> 0.2827 ± 0.018 | Helped |
| Add graph structure to Cell Painting features | DTI link prediction · MOTIVE cold-target split | F1: 0.3408 ± 0.014 -> 0.3916 ± 0.016 | Mixed |
| Use ORF gene profiles instead of CRISPR knockout profiles (ablation) | MOTIVE · Downstream DTI prediction | performance | Helped |
| Add source-source and target-target edges to the bipartite graph (ablation) | GraphSAGE_CP and GraphSAGE_embs · MOTIVE random split | test metrics | Mixed |
| Use Cell Painting features instead of random input embeddings (ablation) | GraphSAGE link prediction · MOTIVE random split | Precision@500: 0.8476 ± 0.029 -> 0.9056 ± 0.012 | Helped |
Surprising findings
- The shortest path topology heuristic performed near-chance (Hits@500 0.0025), and GraphSAGE_embs (structure only) reached Precision@500 0.8476 in random split.
- GATv2_CP had a lower F1 than the MLP in the cold-target split (0.2805 vs 0.3408) despite other GNNs outperforming feature-only models.
Limitations
- Cell Painting features may be less effective when the graph is densely connected.
- Some perturbations induce no significant morphological change, leading to uninformative node embeddings and false positives.
- Image-based profiles are unavailable for many compounds and genes, limiting graph expansion.
Experimental setup
Datasets: MOTIVE JUMP Cell Painting BioKG DGIdb DRHub Hetionet OpenBioLink PharMeBINet PrimeKG
Benchmarks: MOTIVE random split MOTIVE cold-source split MOTIVE cold-target split
Metrics: F1 Hits@500 Precision@500 average precision
Baselines: GraphSAGE_embs Shortest Path MLP Bilinear GATv2_CP GraphSAGE_CP
Builds on: GraphSAGE GIN GATv2 JUMP Cell Painting PyG negative sampling
Closest papers by meaning
- Blend the Separated: Mixture of Synergistic Experts for Data-Scarcity Drug-Target Interaction Prediction (AAAI 2025 · similarity 0.88)
- Customized Subgraph Selection and Encoding for Drug-drug Interaction Prediction (NeurIPS 2024 · similarity 0.88)
- Dual-Channel Learning Framework for Drug-Drug Interaction Prediction via Relation-Aware Heterogeneous Graph Transformer (AAAI 2024 · similarity 0.87)
- M²N: A Progressive Macro-to-Micro 3D Modeling Scheme for Unveiling Drug-Target Affinity (AAAI 2025 · similarity 0.86)
- R-DTI: Drug Target Interaction Prediction Based on Second-Order Relevance Exploration (AAAI 2025 · similarity 0.86)
- Motif-Oriented Representation Learning with Topology Refinement for Drug-Drug 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.