Overview
TME-VI started as a question: can we predict which melanoma patients will respond to anti-PD-1 immunotherapy, using only single-cell RNA-seq data from their tumor microenvironment?
The Core Idea
The tumor microenvironment is an ecosystem. Immune cells don’t act in isolation — they form neighborhoods, communicate through ligand-receptor interactions, and collectively determine whether a tumor is “hot” or “cold” to immunotherapy.
A graph neural network is the natural tool here. Cells become nodes, spatial and transcriptomic proximity becomes edges, and the GAT learns which neighborhoods are predictive of response.
What Worked
scVI as the universal representation was the right call. The variational autoencoder compresses 17,000+ gene dimensions into a 10-dimensional latent space that captures biological variation while correcting batch effects across patients.
Monte Carlo Dropout for uncertainty turned out to be more than a technical nicety. The model’s uncertainty correlates with biological ambiguity — high-entropy predictions cluster around transitional cell states and mixed phenotypes that are genuinely hard to classify.
What Didn’t Work
Joint cell-patient supervision failed. Trying to optimize cell-level and patient-level loss simultaneously caused the model to collapse toward the majority class. Single-task cell-level supervision with post-hoc patient aggregation was the right approach.
The Deployment Stack
FastAPI backend + Streamlit frontend in a single Docker container on HuggingFace Spaces. The LLM report layer uses Claude API streaming to translate GNN outputs into clinical language.
Next Steps
32 patients is a hard ceiling. The next step is pan-cancer extension — zero-shot transfer to NSCLC using CELLxGENE Census.