Characterization of the Breast Cancer Liver Metastasis Microenvironment via Machine Learning Analysis of the Primary Tumor Microenvironment

Abstract

Breast cancer liver metastases (BCLM) are hypovascular lesions that resist intravenously administered therapies and have grim prognosis. Immunotherapeutic strategies targeting BCLM critically depend on the tumor microenvironment (TME), including tumor-associated macrophages. However, a priori characterization of the BCLM TME to optimize therapy is challenging because BCLM tissue is rarely collected. In contrast to primary breast tumors for which tissue is usually obtained and histologic analysis performed, biopsies or resections of BCLM are generally discouraged due to potential complications. This study tested the novel hypothesis that BCLM TME characteristics could be inferred from the primary tumor tissue. Matched primary and metastatic human breast cancer samples were analyzed by imaging mass cytometry, identifying 20 shared marker clusters denoting macrophages (CD68, CD163, and CD206), monocytes (CD14), immune response (CD56, CD4, and CD8a), programmed cell death protein 1, PD-L1, tumor tissue (Ki-67 and phosphorylated ERK), cell adhesion (E-cadherin), hypoxia (hypoxia-inducible factor-1α), vascularity (CD31), and extracellular matrix (alpha smooth muscle actin, collagen, and matrix metalloproteinase 9). A machine learning workflow was implemented and trained on primary tumor clusters to classify each metastatic cluster density as being either above or below median values. The proposed approach achieved robust classification of BCLM marker data from matched primary tumor samples (AUROC ≥ 0.75, 95% confidence interval ≥ 0.7, on the validation subsets). Top clusters for prediction included CD68+, E-cad+, CD8a+PD1+, CD206+, and CD163+MMP9+. We conclude that the proposed workflow using primary breast tumor marker data offers the potential to predict BCLM TME characteristics, with the longer term goal to inform personalized immunotherapeutic strategies targeting BCLM.

Significance: BCLM tissue characterization to optimize immunotherapy is difficult because biopsies or resections are rarely performed. This study shows that a machine learning approach offers the potential to infer BCLM characteristics from the primary tumor tissue.

Figure 1

Workflow of study design. A, Study profile. Primary breast cancer samples were taken from 15 patients. After subsequent diagnosis of BCLM, a core needle biopsy was also obtained. B, Summary of analysis. ROIs from primary tumor and BCLM were identified using H&E staining for tumor tissue and TME. Multiple marker clusters were quantified by IMC, producing 20 common clusters between two analytical batches. IMC ROI missing multiple tumor markers (Ki-67, E-cad⁺, or αSMA) were excluded. Multiple ML models were trained to classify BCLM cluster expression into low (< median) or high (≥ median) groups using primary tumor cluster data. Forward feature selection was performed on preprocessed data using varImp to identify primary TME markers associated with BCLM classification. C, Diagram of model training and validation. Primary tumor data were randomly sorted and split into k folds (subsets; here, k = 5). Each model was trained with k-1 folds and validated with the kth fold. This process was repeated until all folds were used once as the validation set. Twenty permutations were performed in total, repeating the validation process for each fold within each permutation. Final results of each model are the averages of the validations across all folds and all iterations (n = 100). H&E, hematoxylin and eosin.

Figure 2

Representative H&E-stained slices of normal and tumor tissue, and corresponding IMC images from (A) breast and (B) liver tissue. H&E, hematoxylin and eosin.

Figure 3

IMC raw data. A, UMAP representation of 20 identified clusters from breast primary and metastatic liver tumors. B, Heatmap of corresponding cluster and marker intensities from the IMC assay. UMAP, Uniform Manifold Approximation and Projection for Dimension Reduction.

Figure 4

IMC cluster correlations. A, Correlations between primary tumor (breast) and BCLM (liver) marker clusters. B, Pairs of strongly correlated clusters (|ρ| ≥ 0.75).

Figure 5

AUROC curves of ML models using primary IMC tumor cluster densities as input features to predict BCLM IMC cluster densities as being either above or below median values. One ML model was created per BCLM cluster.

Figure 6

Primary tumor IMC clusters sorted by average relative rank across the ML models (lowest number denotes the highest rank) based on their importance for classification of the BCLM IMC clusters. Each box denotes the relative rank of a cluster in a particular model (with lowest rank as most important), whereas a checked box highlights the relative rank of a cluster when included by a model for prediction of BCLM cluster densities.