VirtualCellChallenge2025

Predicting gene expression responses in human embryonic stem cells using AI-driven modeling of single-cell perturbation data.

📌 Introduction

The Virtual Cell Challenge 2025, hosted by the Arc Institute and sponsored by NVIDIA, 10X Genomics, and Ultima Biosciences, tasks participants with building models that predict gene expression responses in a held-out cell type—H1 human embryonic stem cells—using CRISPRi perturbation data from other cell types.

This project presents our solution to the challenge, combining advanced machine learning techniques with domain knowledge in single-cell biology to generalize cellular responses across contexts.

🧬 Methodology

Our approach consists of the following core components:

• Model Type: [e.g., Transformer-based regressor / Graph Neural Network / TabNet]

• Input Features:

  • Pseudo-bulk expression vectors per perturbation
  • Gene ontology embeddings
  • Cell-type condition vectors

• Training Strategy:

  • Supervised learning on 150 training perturbations
  • Domain adaptation for H1 generalization
  • Combined loss: differential expression + global MAE + perturbation similarity

• Regularization: Dropout, layer normalization, and perturbation contrastive loss

Our goal is to not only achieve high accuracy but also ensure biological interpretability and robustness on unseen contexts.

📂 Data

We use the official challenge datasets:

• ✅ Training Set: 150 CRISPRi perturbations across diverse cell types
• 📊 Validation Set: 50 perturbations for live leaderboard evaluation
• 🔒 Final Test Set: 100 perturbations on H1 cells (released Oct 2025)

We also incorporated external perturb-seq datasets such as:

• Tahoe-100M
• scBaseCount
• Replogle et al., Nadig et al., Jiang et al.

🚀 How to Run

# Clone repo
git clone https://github.com/ds4cabs/VirtualCellChallenge2025.git
cd VirtualCellChallenge2025

# Set up environment
conda env create -f environment.yml
conda activate virtualcell

# Train model
python train.py

# Predict
python predict.py --output submission.csv

# Evaluate (local)
python evaluate.py --pred submission.csv --truth validation_truth.csv

📈 Results (Validation Set)

Metric	Score
Differential Expression	0.77
Perturbation Similarity	0.81
Mean Absolute Error	0.13
Overall Score	0.80

Live leaderboard rank: Top 5 (as of Oct 2025)

🧪 Limitations & Future Work

• Currently optimized for transcript-level accuracy; not yet integrated with protein or epigenetic features.
• Exploring integration with scVI for probabilistic modeling and uncertainty estimates.
• Plan to release a web-based demo for in silico perturbation simulation.

👥 Team

• [Teammate 1] – Modeling & Systems Biology Lead
• [Teammate 2] – Data Preprocessing & Evaluation
• [Teammate 3] – Visualization & Interpretation

📜 License

MIT License – feel free to use and adapt with attribution.