← Projects Cellular Scale

TME-VI

Tumor Microenvironment Variational Inference

An uncertainty-aware decoder virtual instrument for immunotherapy response prediction in melanoma, grounded in the AIVC framework.

0.885 Cell-level AUROC
0.97 Confidence-filtered AUROC
0.78 Patient-level AUROC
16,290 CD45+ immune cells
32 Melanoma patients
8 Cell types

The Problem

Anti-PD-1 immunotherapy produces durable responses in some melanoma patients but fails in others. The tumor microenvironment — the ecosystem of immune cells surrounding a tumor — holds the key to predicting who will respond. But extracting actionable signal from tens of thousands of single cells per patient, while quantifying the model's uncertainty, remains a hard problem.

The Approach

TME-VI implements a Decoder Virtual Instrument in the AIVC framework. The pipeline has three stages:

01
scVI Encoder

Variational autoencoder learns a 10-dimensional cellular universal representation from raw scRNA-seq counts, correcting for batch effects across patients.

02
GATDecoderVI

Graph Attention Network operates on a k-NN cell graph built from latent space proximity. MC Dropout applied at inference time for uncertainty quantification.

03
Patient Aggregation

Cell-level predictions aggregated per patient using confidence-weighted scoring. Top 20% most confident cells yield AUROC 0.97.

Dataset

GSE120575 (Sade-Feldman et al. 2018) — 16,290 CD45+ immune cells from 32 melanoma patients profiled before and during anti-PD-1 treatment. 8 annotated cell types including Cytotoxic CD8 T, CD4 T, Naive/Memory CD8 T, B cell, Macrophage, γδ T, Plasma cell, and pDC.

AIVC Connection

This project directly instantiates the AIVC vision from Bunne et al. (Cell, 2024): scVI embeddings serve as the cellular universal representation, the GAT decoder is a decoder virtual instrument, and MC Dropout implements the uncertainty quantification explicitly called for in the paper. The LLM report layer translates numerical GNN outputs into human-readable biological language.

Limitations & Next Steps

The 32-patient cohort is a hard ceiling on patient-level generalization. The immediate next step is pan-cancer extension — zero-shot transfer to NSCLC using GSE176021 (Caushi et al. 2021) via CELLxGENE Census, targeting a universal representation that generalizes across cancer types.

Stack

Model
scVI-tools PyTorch Geometric GAT MC Dropout
Data
Scanpy / AnnData GSE120575 CELLxGENE Census
Deployment
FastAPI Streamlit HuggingFace Spaces Docker Claude API