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

Area of Science:

• Computational biology
• Pathology
• Biostatistics

Background:

• Single-cell multiplex imaging offers insights into disease subtypes and prognoses.
• Quantitative models for cell-cell interactions at single-cell resolution are needed for population-scale survival prediction.

Purpose of the Study:

• To develop and validate a quantitative model for predicting patient survival using single-cell resolution cell-cell interaction features.
• To identify novel patient survival subtypes and atypical subpopulations within breast cancer.

Main Methods:

• Quantified hundreds of single-cell resolution cell-cell interaction features via neighborhood calculation and cellular phenotypes.
• Applied features to a neural-network-based Cox-nnet survival model and used non-negative matrix factorization (NMF) for subtype identification.
• Validated identified subpopulations using label transferring with the UNION-COM method.

Main Results:

• The Cox-nnet model achieved high predictive accuracy for patient survival (Concordance Index > 0.8).
• Identified seven distinct survival subtypes based on epithelial, immune, and fibroblast cell interactions.
• Revealed atypical subpopulations in triple-negative breast cancer (TNBC) and Luminal A breast cancer with distinct prognoses, validated in external datasets.

Conclusions:

• This work presents a novel approach to integrate single-cell level data for population-level survival prediction.
• The findings enable more precise patient stratification and prognostic assessment in cancer research.