Research map / Label-efficient and robust learning / Semi-supervised learning
Multilevel Attention Network with Semi-supervised Domain Adaptation for Drug-Target 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 MlanDTI, a drug-target interaction predictor combining two pre-trained BERT encoders, a multilevel cross-attention transformer, and a pseudo-labeling domain adaptation method with an auxiliary classifier.
Problem
Drug-target interaction prediction suffers from limited labeled data, hidden bias/shortcut learning toward drug features, and poor generalization to out-of-domain drug-target pairs.
Approach
Uses frozen-or-pretrained ChemBERTa-2 and ProtTrans embeddings feeding a transformer whose decoder cross-attends to fused multi-layer protein features (multilevel attention), plus a top-M pseudo-labeling scheme with an auxiliary BERT-input classifier and conflict-penalty loss for cross-domain adaptation.
Key results
- On the cross-domain BindingDB split, the method with pseudo-labeling reaches AUC 0.687 versus 0.616 for DrugBAN with CDAN (DrugBANCDAN), reported as an 11.52% gain.
- On the cross-domain BioSNAP split, the method with pseudo-labeling reaches AUC 0.749 versus DrugBANCDAN's 0.673 (reported 11.29% gain).
- On the human dataset random split, the method achieves AUROC 0.988 and AUPRC 0.990, the best among compared methods; on BioSNAP random split it reports AUROC 0.909.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add pseudo-labeling | MlanDTI · BindingDB cross-domain | AUC: 0.657 -> 0.687 | Helped |
| Add pseudo-labeling | MlanDTI · BioSNAP cross-domain | AUC: 0.728 -> 0.749 | Helped |
| Add BERT embeddings instead of Word2Vec and GCN embeddings (ablation) | MlanDTI · BindingDB cross-domain | AUC: 0.573 -> 0.687 | Helped |
| Add BERT embeddings instead of Word2Vec and GCN embeddings (ablation) | MlanDTI · BioSNAP cross-domain | AUC: 0.648 -> 0.749 | Helped |
| Use multilevel attention instead of standard Transformer multi-head attention (ablation) | MlanDTI · BindingDB cross-domain | AUC: 0.628 -> 0.687 | Helped |
| Use multilevel attention instead of standard Transformer multi-head attention (ablation) | MlanDTI · BioSNAP cross-domain | AUC: 0.731 -> 0.749 | Helped |
| Add pseudo-labeling (ablation) | MlanDTI · BindingDB cross-domain | AUC: 0.657 -> 0.687 | Helped |
| Add pseudo-labeling (ablation) | MlanDTI · BioSNAP cross-domain | AUC: 0.728 -> 0.749 | Helped |
| Add the auxiliary classifier to pseudo-label training (ablation) | MlanDTI · BindingDB cross-domain | AUC: 0.626 -> 0.687 | Helped |
| Add the auxiliary classifier to pseudo-label training (ablation) | MlanDTI · BioSNAP cross-domain | AUC: 0.739 -> 0.749 | Helped |
Surprising findings
- On the BindingDB random split, the proposed model is not competitive, which the authors attribute to hidden bias from highly similar drugs in the dataset.
- Some baseline models perform no better than random guessing (AUC about 0.5) under the cross-domain BindingDB split.
Limitations
- Performance on BindingDB random split is weaker than some baselines due to hidden bias in the dataset.
- Pseudo-labeling accuracy depends on the quality of BERT representations; ablation shows drops without them.
Experimental setup
Datasets: human C.elegans BindingDB BioSNAP
Benchmarks: human C.elegans BindingDB BioSNAP
Metrics: AUROC AUPRC F1-score
Baselines: SVM Random Forest GraphDTA DeepConvDTI MolTrans TransformerCPI HyperAttDTI DrugBAN
Builds on: TransformerCPI DrugBAN HyperAttDTI Lmser network pseudo-labeling
Closest papers by meaning
- Blend the Separated: Mixture of Synergistic Experts for Data-Scarcity Drug-Target Interaction Prediction (AAAI 2025 · similarity 0.86)
- R-DTI: Drug Target Interaction Prediction Based on Second-Order Relevance Exploration (AAAI 2025 · similarity 0.85)
- Customized Subgraph Selection and Encoding for Drug-drug Interaction Prediction (NeurIPS 2024 · similarity 0.84)
- MOTIVE: A Drug-Target Interaction Graph For Inductive Link Prediction (NeurIPS 2024 · similarity 0.84)
- MKG-FENN: A Multimodal Knowledge Graph Fused End-to-End Neural Network for Accurate Drug–Drug Interaction Prediction (AAAI 2024 · similarity 0.83)
- Dual-Channel Learning Framework for Drug-Drug Interaction Prediction via Relation-Aware Heterogeneous Graph Transformer (AAAI 2024 · similarity 0.83)
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.