A5 - Interpreting Breast Cancer Liver Metastatic Characteristics predicted from Paired Primary Tumors using Shapley Value-explanative Machine Learning

Introduction: : BCLM are hypovascularized lesions that rely on surrounding hepatic vasculature for nourishment, making conventional vasculature-mediated chemotherapy delivery ineffective. Consequently, understanding the composition of patient-specific tumor microenvironments (TME), with varying concentrations of macrophages, T-cells, as well as connective tissues (e.g., extracellular matrix [ECM]) is critical to improving patient therapy and prognosis. Acquisition of BCLM tissue is rare since biopsy is rarely performed and resection is generally discouraged. This hinders patient-specific understanding of the metastatic TME. In contrast, biopsies of primary breast cancer lesions are routinely performed, providing a glimpse into the TME characteristics that could potentially be inferred to the BCLM TME.

In our recent work, multiple machine learning models were trained on patient-paired primary TME cell cluster intensities to classify 20 different median-stratified cell clusters of BCLM TME with AUROC ≥0.75 during cross validation [1]. These models predicted a variety of BCLM TME characteristics, including immune cells, tumor markers, hypoxia, vascularity, and ECM. However, feature rankings in [1] were model-dependent and opaquely link dataset features to model classifications. Thus, it is unknown how “black box” machine learning approaches would translate primary intensities to metastatic TME predictions.

To better understand how primary TME cluster densities are associated with the BCLM TME, Extreme Gradient Boosting (XGB) models were trained using Shapley feature selection to predict metastatic TME cluster densities from primary TME cluster densities. Shapley values provide a more transparent understanding of feature rankings that could translate into actionable steps for improving patient outcomes.

Materials and

Methods: : The paired primary and metastatic imaging mass cytometry (IMC) data from our previous study were used [1]. Briefly, de-identified patient-matched primary breast cancer and BCLM paraffin samples (n = 15 pairs) were commercially obtained from BioCoreUSA. This sample size was deemed sufficient for this proof-of-concept analysis. IMC was performed on five and three ROIs selected at random in primary and metastatic tissues, respectively. IMC identified 23 different markers in the TME pertaining to immune cells, tumor markers, hypoxia, vascularity, and ECM. Unsupervised clustering was performed on normalized mean single-cell marker intensities to obtain 23 marker clusters associated with different cell types, 20 of which were in common between the two analytical batches. Marker clusters were then preprocessed using base 10 logarithm, normalized using the total cluster intensities, centered about mean intensity, and finally averaged across ROIs on a per-patient basis.

Like in [1], for each BCLM cluster, patients were separated into low (< median) and high (≥median) expression groups. Using primary cluster intensities, an Extreme Gradient Boosting with linear booster (xgbLinear) was trained in R ≥4.2.2 using the caret package to predict metastatic expression classification. Models were trained using 5-fold cross validation with 30 repeats and a tune length of 5 with Cohen’s Kappa as the optimization metric. Forward feature selection was performed using Shapley values computed on primary cluster densities using the SHAPforxgboost package. Visualizations were made using ggplot2. AUROC and PRAUC were computed using pROC and PRROC packages, respectively. PRAUC distribution was computed using bootstrapping with 10,000 iterations.

Results, Conclusions, and Discussions:: Twenty primary marker clusters shared across the two analytical batches were used in this analysis, comprising passive immune response markers (monocyte: CD14, macrophage: CD68, CD163, and CD206, natural killer cell: CD56), adaptive immune response (helper T-cell: CD4, cytotoxic T-cell: CD8a), immune suppression (PD1, PD-L1), and ECM (αSMA, collagen, and MMP9), hypoxia (HIF1α), vascularity (CD31), tumor tissue (Ki-67 and phosphorylated ERK), cell adhesion (E-cad). To predict metastatic characteristics from primary TME, XGB models were trained to classify metastatic TME cell cluster densities into either low (< median) or high (≥median) groups. All 20 primary lesions were ranked by Shapley value for each BCLM TME prediction. Optimal models were selected with validation subset AUROC and mean bootstrapped PRAUC ≥0.75. Out of the 20 metastatic TME cell clusters, 10 were predicted with the minimally acceptable criteria. Fewer than 5 features were used by XGB models with the highest validation subset AUROC (“optimal models”). Rankings were consistent with prior variable importances, namely, that clusters denoting macrophage influence were selected as features by the majority of optimal models. However, dependence plots between primary clusters and cluster concentrations in optimal model Shapley values indicated a subset of cluster densities changed from positive and negative associations depending on the metastatic cluster density being predicted.

This study shows that BCLM TME characteristics predicted from paired primary tumors can be interpreted using Shapley values in machine learning. Macrophage densities were consistently within the top 5 Shapley value in optimal XGB models. However, trends for higher-Shapley value features were occasionally inconsistent across different optimal models, indicating that multiple models must be trained to better link the primary and metastatic regions. Furthermore, this methodology successfully consolidates predictions from multiple machine learning and feature selection approaches into a single machine learning and feature selection approach.

While the study sample size is limited due to the difficulty of obtaining paired samples, future work will aim to prospectively increase sample size to further validate the Shapley value interpretation. Longer term, the implementation of Shapley values on additional paired patient samples could be used to validate patient-specific therapeutic options to improve patient outcome.

Acknowledgements and/or References (Optional):: Partial support by Department of Defense/US Army Medical Research grant W81XWH2110012 (B. Godin and H.B. Frieboes) is acknowledged. B. Godin acknowledges support from a Houston Methodist Cancer Center Innovation grant and a Cancer Prevention and Research Institute of Texas (CPRIT, RP200619) grant.

Reference:
1. Goodin, D.A., et al., Characterization of the Breast Cancer Liver Metastasis Microenvironment via Machine Learning Analysis of the Primary Tumor Microenvironment. Cancer Res Commun, 2024. 4(10): p. 2846-2857. DOI: https://doi.org/10.1158/2767-9764.crc-24-0263