Research map / Label-efficient and robust learning / Domain adaptation and generalization
Learning Cross-Domain Representations for Transferable Drug Perturbations on Single-Cell Transcriptional Responses
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 XTransferCDR, a domain-separation autoencoder that disentangles basal-state and perturbation embeddings from single-cell expression and cross-transfers perturbation embeddings to predict unseen drug and genetic perturbation responses.
Problem
Predicting single-cell transcriptional responses to novel drugs, genetic perturbations, and cellular contexts requires perturbation representations that transfer across cell contexts, which prior methods do not learn well.
Approach
Encodes perturbed profiles into a basal-state encoder and a perturbation encoder with orthogonality and similarity constraints, then swaps perturbation embeddings across paired profiles under a linear additivity assumption and reconstructs them with a shared decoder (cross-transfer loss).
Key results
- On sci-Plex3, XTransferCDR reached mean R^2 0.81 (all genes) and 0.62 (DEGs), versus 0.72 and 0.55 for cycleCDR.
- On K562 single-gene perturbations, XTransferCDR achieved R^2 0.90 (all genes) and 0.68 (DEGs), above cycleCDR's 0.88 and 0.37.
- On A549 combinatorial perturbations, DEG R^2 rose from 0.166 (baseline) to 0.837 and PCC on DEGs from 0.711 to 0.877.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use XTransferCDR instead of cycleCDR | single-cell drug response prediction · sci-Plex3 | mean R² (All genes): 0.72 -> 0.81 | Helped |
| Use XTransferCDR instead of cycleCDR | single-cell drug response prediction · sci-Plex4 | R² (DEGs): 0.629 -> 0.806 | Helped |
| Use XTransferCDR instead of cycleCDR | single-gene perturbation prediction · K562 | R² (DEGs): 0.37 -> 0.68 | Helped |
| Use XTransferCDR instead of cycleCDR | single-gene perturbation prediction · RPE-1 | R² (DEGs): 0.45 -> 0.78 | Helped |
| Use XTransferCDR instead of baseline | combinatorial genetic perturbation prediction · single-cell combinatorial genetic perturbations | R² (DEGs): 0.166 -> 0.837 | Helped |
| Add cross-transfer loss (ablation) | XTransferCDR · sci-Plex3 | R² (All genes): 0.660 -> 0.817 | Helped |
| Add cross-transfer loss (ablation) | XTransferCDR · sci-Plex3 | R² (DEGs): 0.183 -> 0.626 | Helped |
| Add orthogonality constraint (ablation) | XTransferCDR · sci-Plex3 | R² (All genes): 0.724 -> 0.817 | Helped |
| Add orthogonality constraint (ablation) | XTransferCDR · sci-Plex3 | R² (DEGs): 0.521 -> 0.626 | Helped |
| Add basal-state similarity loss (ablation) | XTransferCDR · sci-Plex3 | R² (All genes): 0.752 -> 0.817 | Helped |
| Add basal-state similarity loss (ablation) | XTransferCDR · sci-Plex3 | R² (DEGs): 0.521 -> 0.626 | Helped |
Surprising findings
- Removing the similarity loss had a smaller effect than removing the cross-transfer or orthogonality losses, per the ablation.
- The baseline that uses unperturbed profiles already reaches a mean R^2 of 0.50 on sci-Plex3, showing how much of the all-gene signal is basal.
Limitations
- Evaluation uses only a few cell lines and held-out drugs; generalization to other contexts is untested.
- Performance on DEG metrics shows large variance, and the SARS gene in RPE-1 was an exception where the model did not beat the baseline.
Experimental setup
Datasets: sci-Plex3 sci-Plex4 K562 single-gene CRISPR perturbation RPE-1 single-gene CRISPR perturbation A549 combinatorial CRISPR perturbation
Benchmarks: sci-Plex3 sci-Plex4 K562 genetic perturbation RPE-1 genetic perturbation A549 combinatorial perturbation
Metrics: R^2 explained variance Pearson correlation mean R^2 median R^2
Baselines: chemCPA chemCPA+pretraining cycleCDR GEARS Baseline (unperturbed profile)
Builds on: domain separation networks CPA cycleCDR style transfer
Closest papers by meaning
- Learning Identifiable Factorized Causal Representations of Cellular Responses (NeurIPS 2024 · similarity 0.87)
- A benchmark for prediction of transcriptomic responses to chemical perturbations across cell types (NeurIPS 2024 · similarity 0.85)
- PerturBench: Benchmarking Machine Learning Models for Cellular Perturbation Analysis (NeurIPS 2025 · similarity 0.85)
- Knowledge-Guided Domain Adaptation Model for Transferring Drug Response Prediction from Cell Lines to Patients (AAAI 2025 · similarity 0.85)
- Chem-PerturBridge: a harmonized compendium of small molecule perturbation transcriptomic effects (NeurIPS 2026 · similarity 0.85)
- Cradle-VAE: Enhancing Single-Cell Gene Perturbation Modeling with Counterfactual Reasoning-based Artifact Disentanglement (AAAI 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.