3

Cell State Transfer

Transferring learned perturbation effects across different cell types and conditions

This task evaluates model ability to transfer knowledge about perturbation effects from one cellular context to another, critical for clinical translation and broader applicability.

Task Overview

Task 3 addresses the critical challenge of cross-context generalization. Models trained on perturbation data from one cell type or experimental condition must predict responses in completely different cellular contexts.

Key Challenges:

  • Cell Type Specificity: Different cell types show distinct perturbation responses
  • Context Adaptation: Experimental conditions and protocols vary across datasets
  • Domain Transfer: Bridging gaps between different experimental platforms
  • Biological Variability: Accounting for inherent biological differences

Transfer Scenarios:

  • Inter-cell type: From one cell line to another
  • Inter-dataset: Across different experimental studies
  • Inter-condition: Different culture or treatment conditions
Task 3 Results

Figure 4: Task 3 results showing cell state transfer performance across different datasets and contexts

Performance Results

Best Overall (R²)

scGEN
R² = 0.772 across cell types

Best Correlation

scGEN
Pearson = 0.938 average

Best Delta Accuracy

scGEN
88.1% direction accuracy

Interactive Performance Visualizations

Delta Performance

Pearson Correlation Delta vs Delta Agreement Accuracy

Transfer Learning Analysis

Cell Type Distance

Transfer success correlates with biological similarity between source and target cell types.

Effect Size Conservation

Large perturbation effects transfer better than subtle cellular changes across contexts.

Dataset Compatibility

Technical differences between datasets significantly impact transfer performance.

Model Architecture

Foundation models show promise for cross-context generalization but with limitations.

Cross-Dataset Performance

Best Source Datasets

  • Kang: High-quality reference for transfer
  • Hagai: Diverse perturbation coverage
  • Srivatsan: Well-characterized responses

Transfer Challenges

  • Batch Effects: Technical confounders affect transfer
  • Protocol Differences: Experimental variations impact results
  • Gene Coverage: Different gene sets measured across studies

Key Insights

No DE Recovery

All models failed at differential expression prediction across cell types

E-Distance Matters

Smaller transcriptomic distance between cell types improves transfer

scGEN Leads

scGEN achieved best performance across most metrics

Limited Models

Only 7 models evaluated: scGEN, trVAE, scPRAM, scVIDR, scPreGAN

Previous: Task 2
View All Results