Citation details will be provided upon publication of the associated manuscript. Interim DOI to cite the repo is
This repository contains an interactive Streamlit application that accompanies a manuscript investigating obesity-driven remodeling of the tumor microenvironment in pancreatic ductal adenocarcinoma (PDAC). The viewer allows readers to explore cell-type-resolved molecular signatures, BMI-associated effects, and survival-relevant patterns from the published analysis.
The statistical modeling code and upstream pipeline are maintained in a separate repository: obese-model
Figure 1. Abstract illustration
- Source: CPTAC Pancreatic Adenocarcinoma (PAAD) cohort
- Samples: 140 tumor samples with clinical annotation
- Cell types: Immune and non-immune populations inferred via BayesPrism deconvolution
- Signatures: 30+ metabolic and functional gene signatures per cell type
 1. Interactome — Chord Diagram |
 2. Interactome — Interaction Network |
 3. Signature Explorer |
 4. Ask the Model |
The application provides six analysis interfaces accessible from the sidebar.
Signature Explorer — Browse the gene signature database. View signature definitions and composition across cell types.
Categorical Analysis — Compare BMI groups (Normal < 25, Overweight 25-30, Obese >= 30) across cell types and signatures. Displays posterior effect sizes, 95% Highest Density Intervals (HDI), heatmaps, and ridge plots from the Bayesian hierarchical model.
Continuous Analysis — Treats BMI as a continuous variable. Displays the estimated slope (effect per 1 SD increase in BMI) from the dose-response model.
Signature Survival Analysis — Cox proportional hazards regression linking signature expression and BMI to clinical outcomes. Displays hazard ratios and confidence intervals.
Interactome Analysis — Cell-cell interaction network visualization based on ligand-receptor pair enrichment. Supports chord diagram rendering and comparison between BMI groups.
Individual Interaction Explorer — Gene-level drill-down for specific cell-cell pairs with detailed enrichment statistics.
In addition to the analysis interfaces, the application includes a conversational interface powered by a remote RAG (retrieval-augmented generation) system. It queries a pre-indexed document store of 73,000+ records derived from all analysis outputs. Queries are routed to the relevant analysis type and answered using a Groq-hosted LLM (Llama 3.3-70B).
BayesPrism — Bayesian cell-type deconvolution framework applied to bulk RNA-seq data to estimate cell-type proportions and cell-type-specific expression profiles per sample.
STABL — Stability-driven feature selection using repeated subsampling and bootstrapping to identify robust BMI-associated molecular features.
Bayesian Hierarchical Modeling — MCMC-based three-group model estimating cell-type-specific obesity effects on signature scores while accounting for between-sample variability.
Convergence Diagnostics — Model validity assessed via R-hat (target < 1.01), Effective Sample Size (> 400), and Hamiltonian Monte Carlo energy diagnostics.
Cox Proportional Hazards — Survival analysis linking BMI group and signature expression to patient outcomes.
cell-analysis-viewer/
├── streamlit_app_with_explorer.py # Main application entry point
├── streamlit_remote_rag.py # RAG UI components
├── config.py # Semantic config and LLM settings
├── requirements.txt
├── src/
│ ├── data_backend.py # DuckDB and Parquet interface
│ ├── remote_rag.py # Remote RAG client
│ ├── true_rag.py # Local RAG with ChromaDB
│ └── vocabulary.py # Cell type vocabulary
├── data/ # Raw analysis outputs
└── data2/ # Parquet and DuckDB views for the app
└── agent.db # DuckDB database
Citation details will be provided upon publication of the associated manuscript.