Explainable Machine Learning Reveals the Role of the Breast Tumor Microenvironment in Neoadjuvant Chemotherapy Outcome

Abstract

Recent advancements in single-cell RNA sequencing (scRNA-seq) have enabled the identification of phenotypic diversity within breast tumor tissues. However, the contribution of these cell phenotypes to tumor biology and treatment response has remained less understood. This is primarily due to the limited number of available samples and the inherent heterogeneity of breast tumors. To address this limitation, we leverage a state-of-the-art scRNA-seq atlas and employ CIBER-SORTx to estimate cell phenotype fractions by de-convolving bulk expression profiles in more than 2000 samples from patients who have undergone Neoad-juvant Chemotherapy (NAC). We introduce a pipeline based on explainable Machine Learning (XML) to robustly explore the associations between different cell phenotype fractions and the response to NAC in the general population as well as different subtypes of breast tumors. By comparing tumor subtypes, we observe that multiple cell types exhibit a distinct association with pCR within each subtype. Specifically, Dendritic cells (DCs) exhibit a negative association with pathological Complete Response (pCR) in Estrogen Receptor positive, ER+, (Luminal A/B) tumors, while showing a positive association with pCR in ER-(Basal-like/HER2-enriched) tumors. Analysis of new spatial cyclic immunoflu-orescence data and publicly available imaging mass cytometry data showed significant differences in the spatial distribution of DCs between ER subtypes. These variations underscore disparities in the engagement of DCs within the tumor microenvironment (TME), potentially driving their divergent associations with pCR across tumor subtypes. Overall, our findings on 28 different cell types provide a comprehensive understanding of the role played by cellular compo-nents of the TME in NAC outcomes. They also highlight directions for further experimental investigations at a mechanistic level.

Loss of chromosome cytoband 13q14.2 orchestrates breast cancer pathogenesis and drug response

“…We estimated cell fractions using either "Imput Cell Fractions" module of CIBER-SORTx webtool with default settings or the Docker version of CIBERSORTx with a similar setting. Following the pipeline we developed for more reliable and reproducible deconvolution (23), we estimate cell fractions using previously created 10 SMs (created using high resolution single-cell RNA sequencing (scRNA-seq) from ( 24)), and then averaged over all 10 estimates. We also perform an analysis on association of loss of 13q14.2 and fraction of cell types enumerated from spatial omics data.…”

“…To explore the potential impact of 13q14.2 loss on the cellular composition in the tumor microenvironment, we estimated the fractions of different cell types within the breast tumor microenvironment. Employing our established pipeline (23), we utilized high-resolution scRNA-seq data (24) as a reference for the breast TME and employed CIBERSORTx as a deconvolution method (22). The scRNA-seq data encompassed a diverse array of cell types, including cancer cells divided into seven recurrent gene modules (GenMod1-7).…”

Feature Importance in Predicting Clinical Outcome: Statistics vs. Explainable Artificial Intelligence

“…This concept is particularly relevant when dealing with cancer patients data where one has to handle complex and high-dimensional biological data collected at different scales 2 . Identifying the key features within cancer patients data can reveal crucial biomarkers and pathways associated with disease mechanisms, aiding in the development of targeted therapies and personalized medicine approaches 3,4 . Thus, feature importance not only streamlines the analytical process by focusing on the most impactful variables but also significantly contributes to advancing clinical research and improving patient outcomes 5 .…”