2

Combinatorial Perturbations

Predicting interaction effects and synergy in multi-perturbation scenarios

This task evaluates model ability to predict responses to combinations of perturbations, including non-additive interaction effects and synergistic responses critical for drug combination therapy.

Task Overview

Task 2 focuses on the complex challenge of combinatorial perturbation prediction. Models must predict cellular responses when multiple perturbations are applied simultaneously, capturing both additive and non-additive interaction effects.

Key Challenges:

  • Interaction Effects: Non-linear responses beyond simple additive effects
  • Synergy Detection: Identifying synergistic and antagonistic combinations
  • Combinatorial Explosion: Scaling to large numbers of possible combinations
  • Buffering vs. Synergy: Distinguishing different types of non-additive effects

Interaction Types:

  • Additive: A + B = sum of individual effects
  • Buffering: Combined effect less than sum
  • Synergy: Combined effect greater than sum
Task 2 Results

Figure 3: Task 2 results showing combinatorial perturbation prediction performance and interaction effect analysis

Performance Results

Best R²

scGPT_epoch5
R² = 0.784 average

Best DE Recovery

PerturbNet
Superior DE overlap (paper)

Best Delta Accuracy

scFoundation
89.8% direction accuracy

Interactive Performance Visualizations

Delta Performance

Pearson Correlation Delta vs Delta Agreement Accuracy

Interaction Effect Analysis

Additive vs. Non-Additive

Most models struggle with non-additive effects, with performance dropping significantly compared to simple additive predictions.

Buffering Effects

Cellular buffering mechanisms create complex response patterns that challenge current model architectures.

Synergistic Responses

True synergistic effects are rare but critical for drug discovery applications. Models show variable success in detection.

Dataset Dependencies

Interaction prediction quality varies dramatically across different experimental conditions and cell types.

Key Insights

Complexity Challenge

Combinatorial effects are inherently harder to predict than single perturbations

Model Architecture

Graph-based approaches show advantages in capturing interaction patterns

Training Data

Quality and diversity of combination training data critically affects performance

Clinical Relevance

Improved combination prediction could revolutionize drug discovery

Previous: Task 1
Next: Task 3 - Cell Transfer