Multiscale Flux Balance Analysis: Bridging Multi-Omics and Single-Cell Data for Therapy Response Prediction in Hormone Receptor-Positive Breast Cancer

This repository contains code and experiments for predicting treatment response in HR+ breast cancer using single-cell RNA sequencing (scRNA-seq) data, fluxomics features inferred with scFBApy, and a multimodal integration of both. The project evaluates predictive performance when models are trained on transcriptomics, fluxomics, or on their combined multimodal features, in order to assess whether integration provides improved accuracy over single-modality approaches.

---

About

This project provides a reproducible pipeline for prediction of breast cancer treatment response at the single-cell level. It includes data acquisition, preprocessing, feature extraction, machine learning modeling, and explainability analyses.

Key goals:

• Evaluate predictive performance of transcriptomic and fluxomic features individually.
• Investigate the added value of multimodal integration.
• Provide a fully documented workflow for reproducibility and future extension.

---

Repository Structure

project/
│── README.md
│── requirements.txt
│── .gitignore
│
│── notebooks/ 
│ ├── 0_data_acquistion.ipynb
│ ├── 1_preprocessing_genes.ipynb
│ ├── 2_generate_flux.ipynb
│ ├── 3_experiment_1.ipynb
│ ├── 3_experiment_2.ipynb
│ └── 3_experiment_3.ipynb
│
|
│── models/model.xml (scFBA XML model) 
|
│── scripts (scFBA python scripts)
│ ├── pipeline_scFBApy.py
│ └── utils_scFBApy.py
│
└── dataset/
│ ├── csv
│ └── h5ad

Notes:

1. notebooks/ contains Jupyter notebooks with descriptive, numbered filenames that reflect the logical order of the analysis workflow.
2. dataset/ contains all datasets used or generated throughout experiments for full reproducibility.
3. models/ contains cobra model of scFBApy.
4. scripts/ contains scFBApy scripts.
5. requirements.txt lists all Python dependencies for reproducibility.

---

Data Source

The primary dataset comes from the NCBI Gene Expression Omnibus (GEO):

🔗 GSE300475 – Single-cell RNA-seq of HR+ breast cancer

• Includes single-cell transcriptomes and TCR repertoire data across multiple patients.
• Patient response status (Responder vs Non-responder) was mapped using the associated publication:
  🔗 npj | Breast Cancer – Dynamic single-cell systemic immune responses in immunotherapy-treated early-stage HR+ breast cancer patients.
• Processed into feature matrices for downstream analysis in this project.
• Preprocessed data is available at: https://figshare.com/articles/dataset/Multiscale_Flux_Balance_Analysis_Bridging_Multi-Omics_and_Single-Cell_Data_for_Therapy_Response_Prediction_in_Hormone_Receptor-Positive_Breast_Cancer/30385879

---

How to Use

1. Clone the repository:

   git clone https://github.com/Occhipinti-Lab/HRplus-BC-Multimodal.git
   cd HRplus-BC-Multimodal

---

2. Install Dependencies

   pip install -r requirements.txt

---

3. Run notebooks in order: Execute the notebooks in the following order to reproduce the full pipeline:

• 00_data_acquisition.ipynb → Download and organize the raw datasets.
• 01_preprocessing_genes.ipynb → Perform QC, outlier removal, normalization, and differential expression (DE) analysis on transcriptomic data.
• 02_generate_flux.ipynb → Infer single-cell metabolic fluxes using scFBApy.
• 3_experiment_1.ipynb → Train, tune (Optuna), and evaluate machine-learning models using transcriptomic (gene expression) features only, including SHAP contribution plots for top genes.
• 3_experiment_2.ipynb → Train, tune (Optuna), and evaluate models using fluxomic (metabolic reaction) features only, with feature importance and SHAP interpretability visualizations.
• 3_experiment_3.ipynb → Train, tune (Optuna), and evaluate multimodal models integrating transcriptomic + fluxomic features, and visualize SHAP contribution plots comparing top features across both modalities.

---