Research map / Label-efficient and robust learning / Recognition and long-tail
WISER: Weak Supervision and Supervised Representation Learning to Improve Drug Response Prediction in Cancer
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 WISER, which learns drug-specific discrete representations with a supervised triplet loss on cell-line data, then pseudo-labels patient profiles via weak supervision (majority vote) and selects a low-noise subset via cut statistics for drug response prediction in patients.
Problem
Predicting patient drug response from cell-line data is hampered by distribution shift between cell lines and patients, heterogeneous drug responses, and scarce labeled patient data, while abundant unlabeled patient genomic profiles go unused in the classifier.
Approach
Represents each genomic profile as a weighted combination of per-drug discrete embeddings, trained with a supervised triplet loss on cell-line response labels plus domain-adversarial (WGAN-GP) alignment; then applies majority-vote weak supervision to unlabeled patient data and selects a least-noisy subset via cut-statistics-based scoring on nearest-neighbor graphs before training the final classif
Key results
- WISER reports AUROC of 0.851 on Cisplatin versus 0.652 for CODE-AE and the best AUROC on four of five drugs.
- Ablation shows weak supervision and subset selection improve AUROC by 4.58% and AUPRC by 3.4% on average over WISER without WS.
- Subset selection with budget below 100% beats the full non-abstained pseudo-labeled set for several drugs (e.g., +10.2% AUROC for 5-Fluorouracil at b=40%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use WISER instead of the strongest baseline | patient drug response prediction · Cisplatin | AUROC: 0.652/0.071 -> 0.851/0.007 | Helped |
| Use WISER instead of the strongest baseline | patient drug response prediction · Temozolomide | AUROC: 0.751/0.017 -> 0.760/0.006 | Helped |
| Use WISER instead of the strongest baseline | patient drug response prediction · Gemcitabine | AUROC: 0.594/0.016 -> 0.649/0.037 | Helped |
| Use WISER instead of the strongest baseline | patient drug response prediction · Sorafenib | AUROC: 0.631/0.020 -> 0.727/0.007 | Helped |
| Use WISER instead of the strongest baseline | patient drug response prediction · 5-Fluorouracil | AUROC: 0.868/0.030 -> 0.715/0.036 | Hurt |
| Add weak supervision and subset selection (ablation) | WISER · five evaluated drugs, average | AUROC: 4.58% | Helped |
| Add weak supervision and subset selection (ablation) | WISER · five evaluated drugs, average | AUPRC: 3.4% | Helped |
| Add supervised triplet loss (ablation) | WISER(Z) discrete representation · 5 evaluated drugs | AUROC | Helped |
| Add supervised triplet loss (ablation) | WISER(Z) discrete representation · 5-Fluorouracil | AUPRC: 0.745/0.047 -> 0.688/0.047 | Hurt |
| Add embedding loss along with triplet loss (ablation) | WISER(Z) discrete representation · 5 evaluated drugs | AUROC and AUPRC | Helped |
| Use a subset instead of the complete non-abstained patient dataset (ablation) | WISER subset selection · 5-Fluorouracil, Cisplatin, Gemcitabine | AUROC: 10.2%, 2.4% and 7.7% | Helped |
| Use a subset instead of the complete non-abstained patient dataset (ablation) | WISER subset selection · Cisplatin, Gemcitabine, Sorafenib, 5-Fluorouracil | AUPRC: 0.8%, 28.2%, 1.8% and 6.9% | Helped |
Surprising findings
- On 5-Fluorouracil, CODE-AE (0.868 AUROC) outperforms WISER (0.715), so WISER is not best on all drugs.
- Naively using all pseudo-labeled patient samples did not improve performance over the baseline.
Limitations
- Evaluation limited to five drugs with at least 20 documented patient responses, and only 179 labeled patient samples.
- Performance limited by the available labeled cell-line dataset and the set of drugs used for discrete representation learning.
- Hyperparameters (epochs, temperature, thresholds, budget) selected via grid search that appears to be tuned on the reported evaluation results.
Experimental setup
Datasets: DepMap cell line data TCGA patient gene expression
Benchmarks: TCGA patient drug response (5-Fluorouracil, Temozolomide, Sorafenib, Gemcitabine, Cisplatin)
Metrics: AUROC AUPRC
Baselines: CODE-AE VAEN DAE CORAL VELODROME DSN (MMD/DANN) ADAE COXEN TCRP MLP ElasticNet Celligner
Builds on: CODE-AE Wasserstein GAN with gradient penalty cut statistics
Closest papers by meaning
- Knowledge-Guided Domain Adaptation Model for Transferring Drug Response Prediction from Cell Lines to Patients (AAAI 2025 · similarity 0.87)
- Learning Identifiable Factorized Causal Representations of Cellular Responses (NeurIPS 2024 · similarity 0.84)
- ScreenShot: A Foundation Model for Few-Shot Combination Drug Screening (NeurIPS 2026 · similarity 0.83)
- Predicting Dose-Response Curves with Deep Neural Networks (ICML 2024 · similarity 0.82)
- Contrastive Balancing Representation Learning for Heterogeneous Dose-Response Curves Estimation (AAAI 2024 · similarity 0.82)
- Learning Cross-Domain Representations for Transferable Drug Perturbations on Single-Cell Transcriptional Responses (AAAI 2025 · 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.