Deep-learning and transfer learning identify new breast cancer survival subtypes from single-cell imaging data

Area of Science:

• Computational pathology
• Cancer systems biology
• Translational oncology

Background:

• Quantitative models for predicting patient survival at a population scale using single-cell resolution cell-cell interaction features are lacking.
• Breast cancer prognosis prediction can be enhanced by integrating detailed cellular interaction data.

Approach:

• Extracted hundreds of features describing single-cell interactions and phenotypes from cyto-images of breast cancer patients.
• Applied these features to a neural-network based Cox-nnet survival model for survival prediction.
• Identified seven survival subtypes and validated atypical subpopulations in independent datasets (TCGA-BRCA, METABRIC).

Key Points:

• Achieved high accuracy in predicting patient survival (Concordance Index > 0.8) using single-cell features.
• Discovered seven distinct survival subtypes characterized by epithelial, immune, fibroblast cell interactions.
• Identified atypical TNBC and Luminal A subpopulations with distinct prognostic profiles and molecular markers.

Conclusions:

• Single-cell imaging mass cytometry (IMC) data offers clinical utility as a novel patient prognosis marker.
• The developed model bridges single-cell information towards population-level survival prediction.
• This work provides a framework for integrating multi-scale data for improved cancer patient stratification and treatment strategies.