Batch-Effect Invariant Graph Neural Networks for Predicting Chemotherapy Response in Triple-Negative Breast Cancer Patients

Abstract

Triple-negative breast cancer (TNBC) is a particularly aggressive subtype of breast cancer that is usually treated with chemotherapy. However, the effectiveness of the treatment can vary widely. Accurate prediction of the response to chemotherapy is crucial in preparing effective personalized treatment. This paper introduces a machine learning framework that uses imaging mass cytometry (IMC) data from clinical trials to train graph neural networks (GNNs) to predict whether a patient will respond to chemotherapy. Our approach combines single-cell protein expression and spatial cell-cell contact information extracted from IMC images. To account for staining variability known as batch effects, we introduce a surrogate loss function that enables learning of a representation space predictive of response, yet invariant to batch artefacts. We investigate different graph construction methods (k-nearest neighbors, k-atmost neighbors, Delaunay triangulation) to capture cell-cell contact delineating tumor microenvironment. Our framework demonstrates improved predictive performance through batch effect correction and effective integration of protein expression with spatial cellular relationships.