CFM-GP: Unified Conditional Flow Matching to Learn Gene Perturbation Across Cell Types

Overview

Understanding gene perturbation effects across diverse cellular contexts is a central challenge in functional genomics, with significant implications for therapeutic discovery and precision medicine. While single-cell technologies enable high-resolution measurement of transcriptional responses, collecting such data remains expensive and time-intensive, especially when repeated for each cell type. Existing computational methods attempt to predict these responses but typically require separate models per cell type, limiting scalability and generalization.

CFM-GP (Conditional Flow Matching for Gene Perturbation) is a deep learning framework that models perturbation as a continuous transformation between control and perturbed gene expression distributions, conditioned on cell type. A single model generalizes across all cell types, eliminating the need for cell type–specific training.

---

Key Features

• Cell Type–Agnostic Prediction: Single model across all cell types
• Continuous Trajectory Modeling: Learns time-dependent perturbation dynamics
• Generalization Across Contexts: Works across datasets and species
• Biological Fidelity: Recovers pathway-level signals

---

CFM-GP Framework

---

Installation

git clone https://github.com/abrarrahmanabir/CFM-GP.git
cd CFM-GP
pip install -r requirements.txt

---

Dataset Access

Download processed datasets:

👉 https://drive.google.com/file/d/1sJxHM4te1CNShBLUrLVEGPrkEbOjM7mk/view?usp=sharing

Place them in:

./data/

---

Data Processing Pipeline

Data Sources

We use five public single-cell RNA-seq datasets:

• COVID-19 (GSE145926)
• PBMC IFN-β (GSE96583)
• Glioblastoma drug response (GSE148842)
• Lupus IFN-β (GSE96583)
• Statefate cytokine stimulation (GSE140802)

All datasets contain paired control and perturbed expression profiles.

---

Preprocessing

Upstream preprocessing includes:

• Normalization
• Log-transformation
• Highly variable gene (HVG) selection
• Batch harmonization (when applicable)

---

Paired Sample Construction

For each cell type:

• Extract control cells
• Extract perturbed cells
• Match by minimum size to ensure strict pairing

min_n = min(adata_ctrl.shape[0], adata_pert.shape[0])
X_ctrl = adata_ctrl.X[:min_n]
X_pert = adata_pert.X[:min_n]

This ensures 1:1 control–perturbation pairing per cell type.

---

Cell Type Filtering

Only cell types with both control and perturbed samples are retained.

---

Feature Representation

• x_ctrl ∈ ℝ^{N × G} → control expression
• x_pert ∈ ℝ^{N × G} → perturbed expression

Gene ordering is consistent across both.

---

Cell Type Encoding

Cell types are encoded as integers:

from sklearn.preprocessing import LabelEncoder
cell_type_encoded = LabelEncoder().fit_transform(cell_types)

---

Final Dataset Format

Each dataset is saved as .pt containing:

• x_ctrl
• x_pert
• cell_type
• cell_type_mapping
• gene_names

---

Train / Validation / Test Splits

• Stratified by cell type
• Maintain paired structure
• Provided splits per dataset
• Donor-level separation applied when available

---

Implementation Notes

• Uses scanpy for .h5ad loading
• Converts sparse matrices to dense
• Uses PyTorch for tensor storage

---

Quick Start (Reproduce Results)

bash test_script.sh

Outputs:

• <dataset>_results/r2.csv
• <dataset>_results/spearman.csv
• <dataset>_results/mmd.csv

---

Running Inference

Pretrained models are available in:

./model/

Run:

bash test_script.sh

---

Training

Train from scratch:

bash train_script.sh

---

Hardware Requirements

• GPU recommended (V100 / A100)
• RAM ≥ 16GB

---

Evaluation Metrics

• R² → prediction accuracy
• Spearman correlation → rank consistency
• MMD → distribution similarity

---

Citation

@article{abir2025cfm,
  title={CFM-GP: Unified Conditional Flow Matching to Learn Gene Perturbation Across Cell Types},
  author={Abir, Abrar Rahman and Dip, Sajib Acharjee and Zhang, Liqing},
  journal={arXiv preprint arXiv:2508.08312},
  year={2025}
}

---

Authors

• Abrar Rahman Abir (abrarrahmanabir156@gmail.com)
• Sajib Acharjee Dip (sajibacharjeedip@vt.edu)
• Liqing Zhang (lqzhang@cs.vt.edu)

---