Research map / Label-efficient and robust learning / Recognition and long-tail
R-DTI: Drug Target Interaction Prediction Based on Second-Order Relevance Exploration
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 R-DTI, a drug-target interaction predictor that builds a structure-aware protein feature extractor from atomic covariance matrices and a Riemannian (SPD) classifier over joint protein-drug features, reporting gains in AUC-ROC on BindingDB and DrugBank.
Problem
Drug-target interaction prediction suffers from protein extractors that cannot capture complex 3D protein structure from sequences, and from the protein-drug feature gap that makes the classifier fragile and biased toward protein features.
Approach
Represent amino acids by covariance matrices of their atomic property vectors (SPD matrices) processed with SPDNet, pre-train a sequence-based extractor to mimic structural features, and use a Riemannian classifier that computes second-order covariance relevance over a unified protein-drug joint feature space, with an auxiliary classifier to ease gradient flow.
Key results
- On BindingDB, R-DTI reaches 0.983 AUC-ROC versus 0.969 for the second-best PSC-CPI (about 1.4% gain).
- On DrugBank, R-DTI reaches 0.914 AUC-ROC, 1.9% above PSC-CPI.
- On KIBA interaction, AUC-ROC is 0.935, slightly below PSC-CPI's 0.936, while KIBA affinity CI and MSE are best (0.906, 0.127).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the protein structural feature extractor (ablation) | R-DTI · BindingDB | AUC-ROC: 0.952±0.003 -> 0.960±0.002 | Helped |
| Add the feature joint block (ablation) | R-DTI with protein structural features · BindingDB | AUC-ROC: 0.960±0.002 -> 0.965±0.001 | Helped |
| Add the Riemannian classifier (ablation) | R-DTI with protein structural features · BindingDB | AUC-ROC: 0.960±0.002 -> 0.963±0.001 | Helped |
| Add the auxiliary classifier (ablation) | R-DTI · BindingDB | AUC-ROC: 0.977±0.001 -> 0.983±0.001 | Helped |
| Add the protein structural feature extractor (ablation) | R-DTI with feature joint block and Riemannian classifier · BindingDB | AUC-ROC: 0.962±0.002 -> 0.977±0.001 | Helped |
| Use R-DTI instead of PSC-CPI | BindingDB · BindingDB | AUC-ROC: 0.969±0.001 -> 0.983±0.001 | Helped |
| Use R-DTI instead of PSC-CPI | DrugBank · DrugBank | AUC-ROC: 0.895±0.001 -> 0.914±0.001 | Helped |
| Use R-DTI instead of PSC-CPI | KIBA interaction prediction · KIBA | AUC-ROC: 0.936±0.001 -> 0.935±0.001 | Hurt |
| Use R-DTI instead of PSC-CPI | KIBA interaction prediction · KIBA | AUC-PR: 0.824±0.001 -> 0.830±0.002 | Helped |
Surprising findings
- Riemannian classifier alone performed worse than with an auxiliary classifier, since second-order relevance computation complicates gradients to feature extractors.
- Most SOTA methods performed worse on the Unseen target subset than the Unseen drug subset, suggesting protein-biased parameters.
Limitations
- Riemannian second-order relevance complicates gradient computation and training of upstream extractors, requiring an auxiliary classifier.
- Protein Riemannian representation and DTI model efficiency are acknowledged as areas for improvement.
- Highlighted high-response regions are not verified to correspond to actual binding sites or functional groups.
Experimental setup
Datasets: PDB (3D structures) Human C.elegans BindingDB DrugBank Davis KIBA
Benchmarks: Human C.elegans BindingDB DrugBank Davis KIBA
Metrics: AUC-ROC AUC-PR CI MSE
Baselines: TransformerCPI HADTI CPInformer TriMulDTI MFR-DTA PSC-CPI
Builds on: SPDNet ProtBert DenseNet TransformerCPI CPInformer
Closest papers by meaning
- Blend the Separated: Mixture of Synergistic Experts for Data-Scarcity Drug-Target Interaction Prediction (AAAI 2025 · similarity 0.87)
- MOTIVE: A Drug-Target Interaction Graph For Inductive Link Prediction (NeurIPS 2024 · similarity 0.86)
- M²N: A Progressive Macro-to-Micro 3D Modeling Scheme for Unveiling Drug-Target Affinity (AAAI 2025 · similarity 0.86)
- Multilevel Attention Network with Semi-supervised Domain Adaptation for Drug-Target Prediction (AAAI 2024 · similarity 0.85)
- Knowledge Enhanced Representation Learning for Drug Discovery (AAAI 2024 · similarity 0.84)
- PPDiff: Diffusing in Hybrid Sequence-Structure Space for Protein-Protein Complex Design (ICML 2025 · similarity 0.84)
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.