Graph-based analysis of multiplex immunofluorescence

Graph Neural Networks over multi-channel images, capturing spatial cell-cell relationships in multiplex IF for downstream phenotype classification.

AstraZeneca · Drug Discovery ML

Problem

Multiplex immunofluorescence (mIF) images carry rich information about cell type, marker expression, and — crucially — spatial relationships between cells. Conventional analysis pipelines flatten this to per-cell feature tables, discarding the spatial structure that often correlates most directly with biological mechanism and clinical outcome.

Approach

Treat mIF images as graphs. Each cell becomes a node with marker-expression features; spatial proximity defines edges. Graph Neural Networks then learn representations that explicitly model cell-neighborhood interactions for downstream phenotype classification.

Result

The approach is described in a preprint on bioRxiv and was presented at SITC 2021 and at the AACR Tumor Immunology and Immunotherapy 2021 meeting:

Innocenti C, Zhang Z, Selvaraj B, Gaffney I, Frangos M, Cohen-Setton J, et al. “An unsupervised graph embeddings approach to multiplex immunofluorescence image exploration.” bioRxiv 2021.06.09.447654 (preprint).

bioRxiv preprint · SITC 2021 abstract · AACR Clinical Cancer Research abstract

Stack

PyTorch Geometric, scikit-learn, custom mIF preprocessing pipeline.