Cross-Species Transfer Learning of Immune Cell States in Cancer Immunotherapy

Resources

• Manuscript: https://docs.google.com/document/d/1_vWyhbPIflFh9kcLZwrNSObb8NSu-Xq_wsdEYDqbkXA/edit?usp=sharing
• Presentation: https://docs.google.com/presentation/d/1PW6mh7ONZ2x5APxo0zb5LW4L9QDGnU5hzC0jEEIiToo/edit?usp=sharing
• Proposal: https://docs.google.com/document/d/1SIH9UfPUh1StrsREJMyBc70y5e2Ntlf_Oq63nIAOa3Y/edit?usp=sharing

Final Research Project for JHU BME EN.580.488 (Foundations of Computational Biology and Bioinformatics), mentored by Dr. Chris Bradburne and Dr. Jessica Resnick.

Overview

This project investigates the evolutionary conservation of tumor microenvironment (TME) transcriptional programs between mouse models and human cancer patients undergoing immune checkpoint blockade (ICB) therapy. Using Coordinated Gene Activity in Pattern Sets (CoGAPS) and transfer learning via projectR, we identify conserved gene expression patterns and assess their ability to predict immunotherapy response.

Key Finding: Myeloid (macrophage) transcriptional programs show 219-fold stronger cross-species conservation than lymphoid (T cell) programs, with conserved macrophage patterns predicting treatment non-response (AUC = 0.917).

Repository Structure

FCBB_Final/
├── data/
├── notebooks/
│   ├── Helper_CoGAPS_to_projectR_Export.ipynb
│   ├── Helper_Figure_Heatmap_TopPatterns.ipynb
│   ├── Helper_Figure_UMAP_CellTypes.ipynb
│   ├── Phase01_Data_Preprocessing_and_QC.ipynb
│   ├── Phase02_PCA_NMF_BatchCorrection.ipynb
│   ├── Phase03_Leiden_Clustering_Annotation.ipynb
│   ├── Phase04_CoGAPS_Pattern_Discovery_Mouse.ipynb
│   ├── Phase05_projectR_Cross_Species_Projection.ipynb
│   ├── Phase06_Response_Prediction_Mixed_Timepoints.ipynb
│   ├── Phase07_Response_Prediction_Timepoint_Stratified.ipynb
│   └── Phase08_Conservation_CrossCellType_Treatment.ipynb
└── README.md

Note: Data files (h5ad, csv, rds) are not included due to size constraints. See Data Availability section below.

Analysis Pipeline

Mouse scRNA-seq → CoGAPS → Ortholog Mapping → projectR → Response Prediction

Phase	Notebook	Description
01	Phase01_Data_Preprocessing_and_QC	Load data, QC, filtering
02	Phase02_PCA_NMF_BatchCorrection	Dimensionality reduction, batch correction
03	Phase03_Leiden_Clustering_Annotation	Cluster cells, annotate cell types
04	Phase04_CoGAPS_Pattern_Discovery_Mouse	Discover latent patterns in mouse TME
05	Phase05_projectR_Cross_Species_Projection	Project mouse patterns onto human cells
06	Phase06_Response_Prediction_Mixed_Timepoints	Test patterns as response predictors
07	Phase07_Response_Prediction_Timepoint_Stratified	Pre/post treatment stratified analysis
08	Phase08_Conservation_CrossCellType_Treatment	Compare conservation across cell types

Helper Notebooks

Notebook	Description
Helper_CoGAPS_to_projectR_Export	Extract and format CoGAPS matrices
Helper_Figure_Heatmap_TopPatterns	Generate heatmap of top predictive patterns
Helper_Figure_UMAP_CellTypes	Generate UMAP visualizations

Data Sources

Dataset	Species	Description	Reference
Gubin et al.	Mouse	Tumor-infiltrating immune cells	GEO: GSE119352
Sade-Feldman et al.	Human	Melanoma patients pre/post ICB	GEO: GSE120575

Cell Types Analyzed

Cell Type	Human Cells	Patterns Discovered
T cell	9,257	2
NK	2,847	4
Macrophage	1,274	10
Mki67hi	876	8

Key Results

1. Conservation Varies by Cell Type

Cell Type	Mean Conservation	Lineage
Macrophage	0.687	Myeloid
Mki67hi	0.450	Proliferating
NK	0.084	Lymphoid
T cell	0.003	Lymphoid

Myeloid vs Lymphoid: 40-fold difference (p < 2.2e-16, Cohen's d = 3.15)

2. Top Predictive Patterns

Pattern	Cell Type	AUC	Higher In
Pattern8	Macrophage	0.917	Non-responder
Pattern8	Mki67hi	0.833	Non-responder
Pattern3	NK	0.732	Responder

3. Treatment-Specific Conservation

Cell Type	anti-CTLA4	Combination	anti-PD1
Macrophage	0.615	0.830	0.590
NK	0.112	0.081	0.084
T cell	0.003	0.003	0.003

Dependencies

R Packages

install.packages(c("tidyr", "dplyr", "ggplot2", "pROC"))
BiocManager::install(c("CoGAPS", "projectR", "biomaRt", "org.Hs.eg.db", "anndata"))

Python Packages

pip install scanpy anndata matplotlib

Data Availability

Due to file size limitations, the following data files are not included:

File Type	Description	How to Obtain
mouse_annotated.h5ad	Mouse scRNA-seq data	GEO: GSE119352
human_annotated.h5ad	Human scRNA-seq data	GEO: GSE120575