Research map / Label-efficient and robust learning / Domain adaptation and generalization
Knowledge-Guided Domain Adaptation Model for Transferring Drug Response Prediction from Cell Lines to Patients
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 TransDRP, a multi-label GNN pre-trained on GDSC cell lines and adapted to TCGA patients via global-local (category-aware prototypical contrastive) domain adversarial training for drug response prediction.
Problem
Drug response prediction models trained on label-rich cell lines degrade on label-scarce patients due to distribution shift; existing UDA methods train one model per drug and align data globally, ignoring cancer-category confusion.
Approach
Combines a multi-label GNN decoder over a drug label graph (ECFP-based molecular similarity, GAT) with a global-local domain adversarial strategy in which prototypical contrastive learning aligns same-cancer-category samples across domains and separates different categories.
Key results
- Average over 9 drugs: TransDRP AUC 0.6654 and AUPR 0.7029, with ACC 0.6418 and F1 0.6287.
- Removing the global adversarial loss drops average AUC from 0.6654 to 0.5944.
- TransDRP outperforms baselines on 6 of 9 drugs in AUC and AUPR, and is competitive on the other 3.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the autoencoder for shared representation extraction (ablation) | TransDRP · TCGA, average of 9 drugs | AUC: 0.6473 -> 0.6654 | Helped |
| add the autoencoder for shared representation extraction (ablation) | TransDRP · TCGA, average of 9 drugs | AUPR: 0.6908 -> 0.7029 | Helped |
| use the GNN decoder and label graph instead of an MLP (ablation) | TransDRP · TCGA, average of 9 drugs | AUC: 0.6455 -> 0.6654 | Helped |
| use the GNN decoder and label graph instead of an MLP (ablation) | TransDRP · TCGA, average of 9 drugs | AUPR: 0.7001 -> 0.7029 | Helped |
| use multi-label prediction instead of separate single-drug tasks (ablation) | TransDRP · TCGA, average of 9 drugs | AUC: 0.6602 -> 0.6654 | Helped |
| use multi-label prediction instead of separate single-drug tasks (ablation) | TransDRP · TCGA, average of 9 drugs | AUPR: 0.6813 -> 0.7029 | Helped |
| add the local category-contrastive loss (ablation) | TransDRP · TCGA, average of 9 drugs | AUC: 0.6396 -> 0.6654 | Helped |
| add the local category-contrastive loss (ablation) | TransDRP · TCGA, average of 9 drugs | AUPR: 0.6871 -> 0.7029 | Helped |
| add the global adversarial loss (ablation) | TransDRP · TCGA, average of 9 drugs | AUC: 0.5944 -> 0.6654 | Helped |
| add the global adversarial loss (ablation) | TransDRP · TCGA, average of 9 drugs | AUPR: 0.6330 -> 0.7029 | Helped |
Surprising findings
- Removing the multi-label design (per-drug separate training) lowered ACC and F1 but AUC was nearly unchanged (0.6602 vs 0.6654).
Limitations
- Performance differences vary strongly by drug, attributed to class imbalance.
- Average AUC around 0.67 indicates modest absolute accuracy.
- Novel drug screening case study is qualitative with limited validation.
Experimental setup
Datasets: CCLE TCGA GDSC
Benchmarks: TCGA patient drug response (9 drugs)
Metrics: AUC AUPR ACC F1
Baselines: Celligner Velodrome scDEAL AITL CODE-AE WISER tCNNs DeepCDR GraphCDR BiGNN
Builds on: DANN Deep Domain Confusion-style shared/private autoencoder (Bousmalis et al. 2016) CODE-AE WISER InfoNCE
Closest papers by meaning
- WISER: Weak Supervision and Supervised Representation Learning to Improve Drug Response Prediction in Cancer (ICML 2024 · similarity 0.87)
- Learning Cross-Domain Representations for Transferable Drug Perturbations on Single-Cell Transcriptional Responses (AAAI 2025 · similarity 0.85)
- Do Multiple Instance Learning Models Transfer? (ICML 2025 · similarity 0.83)
- Predicting Dose-Response Curves with Deep Neural Networks (ICML 2024 · similarity 0.83)
- Knowledge Transfer Scaling Laws for 3D Medical Imaging (NeurIPS 2026 · similarity 0.82)
- Learning Identifiable Factorized Causal Representations of Cellular Responses (NeurIPS 2024 · similarity 0.82)
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.